CONTENTS

Preface
1. Introduction
1.1 Concepts of Measurement and Error.

1.2 Types of Error.

1.3 Elementary Concepts in Probability

1.4 Reliability of Measurements

1.5 Significant Figures

2. Error Propagation and Linearization

10

13

2.1. Error Propagation

13

2.2. Linearization

20

3. The Concept Adjustment

31

3.1. Introduction

31

3.2. Simple Adjustment Methods

34

3.3. The Least Squares Method

36

3.4. Examples or Simple Least Squares Problems

41

4. Least Squares Adjustment

5.

59

4.1. Techniques or Least Squares

59

4.2. The Concept of Weight

66

4.3. Least Squares Adjustment of Indirect Observations

63

4.4. Least Squares Adjustment of Observations Only

87

Elementary Probability Theory

106

5.1.Random Events and Probability

106

5.2. Random Variables

107

5.3. Continuous Probability Distributions

113

5.4. The Normal Distribution

115

5.5. Expectation

125

5.6. Measures of Precision and Accuracy

129

5.7. Covariance and Correlation

133

5.8. Covariance, Cofactor, and Weight Matrices

138

5.9. Introduction to Sampling

144

6. Variance-Covariance-Propagation

148

6.1. Introduction

148

6.2. Derivals of the Propagation Laws

149

6.3. Examples

154

6 . 4. Stepwise Propagation

160

6.5. Propagation for Least Squares Adjustment of

Indirect Observations

167

6.6. Propagation or Lest Squares Adjustment or Observations

Only

7. Preanalysis of Survey Measurements

170

180

7.1. Preanalysis Procedure

180

7.2. Horizontal Angle Measurement with a Theodolite

184

7.3. Distance Measurement By EDM

188

7.4. Elevation Difference by Direct Leveling

190

7.5. Survey Tolerances

193

8. Introductory Statistical Analysis

199

8.1. Samples and Statistics

199

8.2. The Chi-Square Distribution

200

8.3. The t (Student) Distribution

203

8.4. Common Sample Statistics

206

8.5. Estimation for the Mean

209

8.6. Estimation of the Variance Confidence interval for t h e lean 210

8.7. Confidence Interval for the Mean

211

8.8. Confidence interval for the Variance

214

8.9. Statistical Testing

215

8.10. Test of the Mean of a Probability Distribution

216

8.11. Test of the Variance of a Probability Distribution

218

8.12. Bivariate Normal Distribution

220

8.13. Error Ellipses

224

9. General Least Squares Adjustment

236

9.1. Introduction

236

9.2. Derivation

242

9.3. Precision Estimation

248

9.4. Special Cases

254

9.5. Summary of Symbols and Equations

256

10. Applications in Plane Coordinate Surveys

266

10.1. Introduction

266

10.2. The Distance Condition and Its Linearization

266

10.3. The Azimuth Condition and its Linearization

269

10.4. The Angle Condition and its Linearization

272

10.5. Position Fixing by Distance

274

10.6. Two-Parameter Similarity Transformation

281

10.7. Four-Parameter Similarity Transformation

297

Appendix A. An Introduction to Matrix Algebra

305

A.1. Definitions

305

A.2. Types of Matrices

305

A.3. Equality of Matrices

308

A.4. Sums of Matrices

308

A.5. Scalar Multiplication of Matrices

309

A.6. Matrix Multiplication

309

A.7. The Transpose of a Matrix

312

A.8. Symmetric Matrices

314

A.9. The Inverse of a Matrix

314

A.10.

Determinants, Minors, and Cofactors

316

A.11.

Cofactor and Adjoint Matrices

318

A.12.

Matrix Inversion Using the Adjoin Matrix

319

A.13.

Linear Equations

320

A.14.

Bilinear and Quadratic Forms

321

A.15.

Differentiation of Vectors, Bilinear Forms, and

Quadratic Forms

322

Appendix B. Tables
Table I. Values of the Standard Normal Distribution Function

326

Table II. Percentiles of the Chi-Square Distribution

328

Table III. Percentiles of the 1 Distribution

330

Bibliography

333

Index

337

INTRODUCTION

1.1.

CONCEPTS OF MEASUREMENT AND ERROR

The surveyor may be called upon to perform a variety of tasks in connection

with a survey project, from initial project design to final presentation of results in
some specified form. Much of what the surveyor does, however, involves
measurements, including their adjustment and analysis in the office as well as
their acquisition in the field. If Site surveyor is to acquire, adjust, and analyze
measurements intelligently, he must understand what the process
of measurement is.
Measurement must involve observation. No measurement is made until
something is observed. Accordingly, the terms measurement and observation
are often used synonymously.
Although we may think of a measurement as a single act, a typical survey
measurement may involve several elementary operations, including such
observing operations as centering, pointing, matching, setting, and reading. Yet
at the end of all these operations a single numerical value is used to represent
then "measurement" or "observation" of the quantity that we are seeking.
Let us consider, for example, the relatively
simple task of using a 30 m steel
tape suspended above the ground to measure the distance between two survey
stations that are slightly less than 20 mapart .The following elementary
operations are performed:
1. The head and rear [apemen take up positions hex; to their respective survey
stations and hold the tape above the ground, approximately in line with the
stations.
2. The rear tapeman suspends his plumb bob from the tape, setting and
holding the string on the 20 m graduation.

3. The rear tapeman then centers his plumb bob over his survey station.
The head tapeman suspends his plumb bob from the tape, with the plumb
bob string held near the zero graduation, and applies tension to the tape.
(The rear tapeman, of course, must pull equally in the opposite direction
while keeping his plumb bob centered over his station.)
5. While maintaining tension, the head tapeman shifts his plumb bob string
along the tape until his plumb bob is centered over his survey station.
6. The head tapeman then reads the position of his string on the tape.
7. The measured distance is obtained by subtracting the tape reading in step

trom the 20 m value set in step 2.

It is obvious that 2,3,5 and 6 are observing operations of setting, centering ,
centering, and reading, respectively, all of which are necessary to obtain a
measurement of the distance between the two survey stations. This clearly
demonstrates that even a simple measurement may involve several elementary
operations, and that what is regarded as an observation of a desired quantity,
such as the distance between two survey stations, is not just a single reading of
a scale but the result of several steps in an observing process.
Even steps 1 through 7 are not the whole story in getting a reliable measurement of the distance. For some purposes the value obtained in step 7 may
be satisfactory; for other purposes, it is not. If a "better" measurement is
required, the step 7 value must be corrected for such things as length of tape,
temperature, and sag. To evaluate then appropriate corrections, more things
have to be observed, because the tape must be compared with a standard, the
temperature must be observed on a thermometer, and the pull applied to the
tape must be observed on a tension handle or similar device. For a precise
survey measurement, these corrections and others are every bit as important as
steps 1 through 7. A measurement is, indeed, the result of several operations,
each of which makes some contribution to the measurement's ultimate
usefulness.
Measurement is a process that is subject to variation. Variation can occur if
some aspect of the measurement, such as temperature, is not taken into
account. If, for example, several measurements of the distance between the
two survey stations are made with the steel tape, and a change in temperature
takes place. while the measurements are being made, there will be a corresponding change in the length of the tape and consequently in the readings on
the tape. If no correction is made for the effect of temperature, the resulting
measurements will exhibit variation due to the temperature change. Variation is
also a natural consequence of the elementary observing operations themselves.
No observation can by repeated exactly, because of limitations in the
instrumentation used and in the ability of the observer to center, point, match,
set, and read. The small variations which occur in the elementary operations
produce corresponding variations in the measurement. Since all
measurements are subject to variation, it follows that no quantity that is
measured is completely determinable. We may seek a fixed value for a
quantity that we conceive to be the true value, but what we get in reality is
nothing more than an estimate of the true value. Mathematically, the measurement or observation must be looked upon as a variable. More will be said
about this later.
It must be emphasized that the variation in the values obtained for a
measured quantity is a natural phenomenon which is to be expected even
when the conditions under which the measurements are made are essentially
constant. If we are to expect variation, then we must expect a difference
between a measured value of a quantity and its true value, whatever it may
be. This difference is known as the error in the measured value. Although
the English meaning of the term error may suggest that something wrong
has been committed, errors in measurement should not be so regarded,
unless perhaps they are what are called gross errors (see Section 1.2).

The study of observational errors and their behavior is essentially equivalent

to the study of observations themselves. In other words, what has been
classically referred to as the theory of errors. is equivalent to what is now
known as the theory of observations.
If denotes the t rue value of a quantity (a distance, angle, etc.), and x is its
observed value, then the error in x is defined as:
=x

(1-1)

Since we will never really know what the value of is, we will never know
what the exact value of is. However, if we are able to obtain by some means
a good estimate of , we can use this estimate in place of as reference for

expressing variation in the observed values. If x denotes the estimate of

then the difference between and the observed value, x, is defined as the
residual, v; specifically,

v = x x
(1-2)
The residual, v, is the quantity that is actually used to express variation in
the measurement.
1.2. T Y P E S O F E R R O R S
Errors have been traditionally classified into three types: (I) gross errors,
(2) systematic errors, and (3) random errors. Each type will be discussed
separately.
Gross Errors
Gross errors are the results of blunders or mistakes that are due to carelessness of the observer. For example, the observer may make a pointing on the
wrong survey target, or he may read a scale or dial incorrectly, or he may read
the wrong scale, or record the wrong value of a reading by transposing numbers
(e.g., recording 41.56 m as .41.65 m). There are any number of mistakes an
observer can make if he is inattentive.
If a survey is to have any usefulness at all, mistakes and blunders cannot be
tolerated. Good field procedures are designed to assist in detecting mistakes.
These procedures include:
1. Careful checking of all pointings on survey targets.
2. Taking multiple readings on scales and checking for reasonable consistency.
3. Verifying recorded data by rereading scales.
Repeating entire measurements independently and checking for consistency.
5. Using simple geometric or algebraic checks, such as comparing the sum of
three measured angles in a plant triangle with 180.
It is very important to safeguard against the occurrence of mistakes. If they do
occur, they must be detected and eliminated from the survey measurements
before such measurements can be used.

Systematic Errors
Systematic errors are so called because they occur according to some deterministic system which, when known. can be expressed by some functional
relationship. If, for example, the expansion of a steel tape is essentially linear
with respect to temperature, and the coefficient of thermal expansion is known,
a functional relationship between the temperature and the expansion of the tape
can be established. I f the length of the tape at some specified standard
temperature is taken as reference, the change in the length of the tape from this
reference caused by change in temperature from its standard value is classified
as a systematic error.
A systematic error follows a pattern which will be duplicated if the measurement
is repeated under the same conditions. For example, measuring a distance with
a steel tape that is too short will result in the same systematic error if the same
tape is used by the same tapemen to measure the same distance under the
same conditions of temperature, pull, support, and slope.
remain the same throughout the measuring process. It is counteracting if its
sign changes while its magnitude remains the same.
The system underlying a systematic error may depend on the observer, the
instrument used, the physical or environmental conditions at the time the
measurement is made, or any combination of these factors.
The personal bias of an observer leads to systematic errors that may be
constant or counteracting, depending on the observation procedure. If the
conditions of observation vary, the natural senses of vision and hearing of the
observer may vary as well and his personal error becomes variable, too.
Imperfect instrument construction or incomplete instrument adjustment can lead
to instrumental errors that are systematic. Imperfect construction includes such
things as variation in scale graduations and eccentricity in centering
components. Incomplete instrument adjustment includes such things as not
making the axis of collimation (telescope axis) of a theodolite perpendicular to
the instrument's tilting (horizontal) axis.
Since survey measurements are acquired in the field, they are affected by
many physical and environmental factors. Temperature, pull and terrain slope,
for example, affect taped distances, while humidity and atmospheric pressure
as well as temperature affect electro-optical distance measurements (EDM),
angle measurements, and leveling. AB of these effects are functionally
expressible in terms of the factors wich cause them and so are classified as
systematic errors.
All sources of systematic error so far discussed are related directly to the
observational operations, However, systematic errors can also occur through
simplification of the geometry or mathematical model chosen to represent the
survey. If, for example, a plane triangle instead of a spherical triangle is used to
connect three survey stations that are spaced several kilometers apart, the
spherical excess will emerge as a systematic error.
In the reduction of survey measurements, it is important to detect and correct
for all possible systematic errors.
Random Errors
After all blunders are detected and removed, and the measurements are
corrected for all known systematic errors, there will still remain some variation in

the measurements. This variation results from observational errors which have
no known functional relationship based upon a deterministic system. These
errors, instead, have random behavior, and must be treated accordingly,
it was stated earlier that a measurement or observation is looked upon
mathematically as a variable. More specifically, it is a random variable
because it includes error components which exhibit random behavior. Indeed,
the random errors themselves are random variables.
Whereas systematic variations ate dealt with mathematically using functional
relationships or models, random variables must use probability models. Some
elementary concepts in probability theory are introduced in the following
section. More will be discussed in Chapter 5,
1.3. ELEMENTARY CONCEPTS IN PROBABILITY
Let us assume that a distance is measured a very large number of times and
that all measurements are free of gross errors and corrected for all systematic
errors. Whatever variation remains in the measurements is caused by random
errors only. Although it is not possible to correct measurement for specific
random errors from knowledge of the measurement system, it is possible to
study their collective behavior from their frequency distribution. It is the
frequency distribution that Is used as basis for constructing the probability
model for the measurements.
EXAMPLE 1-1
A distance of about 810 m is measured 200 times. All measurements are free of
gross errors and are corrected for systematic errors. The corrected values are
ex-pressed to 0.01m. It is noted that after correcting for the systematic errors the
resulting variation in the measurements ranges from 810.11m to 810,23m,
distributed as follows:
VALUE OF
MEASUREMENT (m)

NUMBER OF
MEASUREMENTS

810.11
810.12
810.13
810.14
810.15
810.16
810.17
810.18
810.19
810.20
810.21
810.22
310.23

1
3
7
19
20
36
38
29
24
10
11
0
2

Evaluate and plot the relative frequencies of occurrence for all listed values.
Solution
The relative frequency of occurrence is obtained by dividing the number of measurements observed for a value by the total number of measurements. Since there

are 200 measurements in total, the following relative frequencies are obtained:
VALUE OF
MEASUREMENT (m)

RELATIVE
FREQUENCY

810.11
810.12
810.13
810.14
810.15
810.16
810.17
810.18
810.19
810.20
810.21
810.22
810.23

O.005
0.015
0.035
0.095
0.100
0.180
0.190
0.145
0.120
0.050
0.055
0.000
0.010

The sum of the relative frequencies must, of course, be 1.000. The relative
frequencies are plotted in Fig. 1-1. as rectangles.
Figure 1-1 is a frequency distribution in the form of a histogram. The base of
each rectangle represents a class interval; and the height represents the
corresponding relative frequency. The base of the tallest rectangle in Fig. 1-1, for
example, represents the class of all measurements between 810.165 m and
810.175 m (expressed specifically as 810.17 m), and the height of this rectangle
represents the corresponding relative frequency, 0.190.
The frequency distribution in Fig. i-1 is centered on or very near the value 810.17
m. Highest frequencies are at or near the central value.
If the number of measurements were to be increased infinitely, it would be found
that each relative frequency would approach a stable limit. The limiting value of
the relative frequency is known as the probability.
Instead of using a histogram to represent probabilities, it is often more convenient
to use a mathematical modela probability model, or probability - distribution.
An example of such a model is given in Fig. 1-2, in which probability is
represented by the area under a continuous curve that is a mathematical function
of the measurement. Specifically, the probability that the measurement falls
between the two values x1 and x2 is given by the area of

Fig 1-1

Fig 1-2

the shaded region in Fig. 1-2. Recalling that a measurement is a random

variable, the curve in Fig. 1-2 is the probability density function of the random
variable that represents the distance measurement. The central value, , in Fig.
1-2 is the mean value of the measurement. If no systematic errors are present
in the measurement, the mean value is taken as the "true" value.

A similar density function, shown in Fig. 1-3, can be used as probability model
for the random error of the measurement. In this case, the mean value is zero.
The mean value, , is referred to as the position or location parameter of the
probability distribution. Another parameter of the distribution is its standard
deviation, o, which measures the spread or dispersion of the probability
distribution, as indicated in Fig. 1-3. If measurement A has greater variation
than measurement B, measurement A will have a standard deviation that is
larger than the standard deviation of B. The square of the standard deviation is
known as the variance.
1.4. RELIABILITY OF MEASUREMENTS
Several terms are used to express the reliability of measurements. Three
common terms are precision, accuracy, and uncertainty.

Fig1-3

1. Precision is the degree of closeness or conformity of repeated

measurements of the same quantity to each other. If the repeated
measurements are closely clustered together, they are said to have high
precision; if they are widely spread apart, they have low precision. High
precision generally reflects the high degree of care and refinement in the
instrumentation and procedure used in making the measurements. Precision
is indicated by the dispersion or spread of the probability distribution. The
narrower the distribution, the higher is the precision, and vice versa. A
common measure of precision is she standard deviation . The higher the
precision, the lower is the value
and vice versa.
2. Accuracy is the degree of conform ity or closeness of a measurement to
the true value. Accuracy includes not only the effects of random errors but
also any bias due to uncorrected systematic errors. If there is no bias, the
standard deviation can also be used as a measure of accuracy.
3. Uncertainty is the range within which it is expected the error of a

measurement will fail. A specified level of probability is generally associated

with an uncertainty. The 90% uncertainty, for example, is the range of values
within which it is 90% probable (i.e., the probability is 0.90) the en or of the
measurement will fall. In genera:, if the uncertainty of a measurement is
known, it should accompany the measured value. These terms are discussed
further in Chapter 5.
1.5. SIGNIFICANT FIGURES
The number of significant figures in a numerical quantity equals time number
of digits in the quantity less all zero digits that are used to fix the position of
the decimal point. For example,
147

has 3 significant figures;

147.64
2.1
1013
1.007
17.710
0.021
1320

has 5 significant figures;

has 2 significant figures;
has 4 significant figures;
has 4 significant figures;
has 5 significant figures:
has 2 significant figures (zeros fix decimal point);
has either 3 or 4 significant figures, depending on whether or
not the zero digit is used only to fix the position of the
implied decimal point which follows it.

A numerical value should carry all certain digits plus the first digit stint is
doubtful. if, for example, the first tour digits of the value 137.824 are certain
and the last two digits are doubtful, the value should be expressed to only
five significant figures, i.e., 137.82 (1, 3, 7 and 8 are certain, 2 is doubtful).
The number of significant figures in a directly measured quantity is not
usually difficult to determine, as it essentially depends on the least co unt of
the instrument used. For example, if a distance is measured with a tape
graduated in centimeters, with estimation to millimeters, and a reading of
462,513 m is taken, the first five digits are certain, the sixth digit is estimated
(and therefore doubtful) and so the value has six significant figures.
The number of significant figures in a numerical quantity is reduced by
rounding off. The least error will be caused if rounding off is done according
to the following rules:
1. If k significant figures are required, discard all digits to the right of the
(k+1)th digit.
2. Examine the (k+ 1)th digit.
a. If it is 0 to 4, discard it ; e.g., 12.34421 is rounded off to four
significant figures as 12.34.
b. If it is 6 to 9, discard it and increase the kth digit by one; e.g.,
1.376 is rounded off to three significant figures as 1.38.
c. If it is 5 and the kth digit is even, discard it; e.g., 12.345 is
rounded off to four significant figures as 12.34.
d. If it is 5 and the kth digit is odd, discard it and increase the kth
digit by one; e.g., 12.3435 is rounded off to five significant figures as
12.344.

The number of significant figures is not so easily determined in a quantity

obtained by computation as it is in a directly measured quantity. There are,
however, some general rules that can be applied with reasonable effectiveness. The two most important rules deal with the arithmetic operations,
addition (or subtraction) and multiplication (or division).
In the process of addition, the sum should be rounded off to the number of
decimal places that is least in the quantities being added. For example, the
sum
165.21
149.7
65.495
2.2167
382.6217
should be rounded off to 382.6 because 149.7 is expressed to only one
decimal place. In multiplication, the number of significant figures in the
product should the number of significant figures in the factor with the least
number significant figures, exact numerical factors excluded. For example:
2.15X11_1234 =23.9
2(2.15 x 11.1234) = 47.8 (2 is an exact factor).
In each case, the number of significant Figures (3) in 2.15 decides the
number of significant figures in the product.
In computations with a large number of arithmetic operations, it is good
practice to carry one extra significant figure throughout, then round off after
the computations are completed.

ERROR PROPAGATION AND

LINEARIZATION
2.1. ERROR PROPAGATION
In surveying, as in many areas of science and engineering, the quantities
that are measured directly in the field are often used to compute other
quantities of interest. In such cases, the computed quantities are
expressed as mathematical functions of the heal measurements. if the field
measurements have errors. it is inevitable that the quantities computed
from them will have errors. Evaluation of the errors in the computed
quantities as functions of the errors in the measurements is called error
propagation.
Suppose x is a measured quantity and y is a new quantity to be calculated
from x according to the function
y=ax+b

(2-1)

represented by the straight line in Fig. 2-1. The coefficients a and b are
known and assumed to be errorless.
For purposes of analysis, it is helpful to use the concept of true value, as
introduced in Chapter 1, and to define the error of a measurement as the
measured value minus the true value, as it was in Eq. (1-i), Thus, if x, represents the true value of x, and dx represents the error, then
x= x 1 + dx

(2-2)

Now the true value of y1 designated y0 can be computed directly from x 1

using the Eq. (2-1) function; i.e.,
y1=ax1+dx1
Thus, from Eq. (2-1), we get the following computed value for y:

(2-3)

Fig 2-1
y ax b

a ( x1 dx) b
a x1 b adx
y

adx

dy represents the error in y, then it follows from Eq. (2-4) that

dy adx

(2-5)

Now if elementary calculus were applied to Eq. (2-I), we would see that
derivative of y with respect to xis dy/dx = a. Thus an equivalent expression
for Eq. (2-5) is
dy
(2-6)
dy
dx
dx

Eq. (2-6) is obviously the expression for the total differential of the function
Eq. (2-1). The reason the error dy obtained from the function turns out to 4
Identical to the error as a total differential from calculus is that the function
ax + b is linear in the measured quantity x. It will be shown shortly that q, (2
6) does not hold exactly for nonlinear functions.
EXAMPLE 2-1
A land parcel is trapezoidal in shape with the dimensions given in Fig. 2-2. For a
measured distance d=3,560m , the ordinate h, is required. If the error in the

measured distance is 0.016 m, compute the corresponding error in the calculated

value of h.
Solution
The slope of line CD is
a

60 20
0.5(exact)
80

If we visualize a coordinate system with origin at A and x-axis along AB, the
equation of line CD is
y 0.5 x 20

in which quantities 0.5 and 20 are errorless.

When x= d = 23.560 m, the corresponding y-coordinate is the ordinate h. Thus,
h 0.5(23.560) 20 31.780m

From Eq. (2-5), the error in h is

dh adx 0.5(0.016) 0.008

Fig 2-2

Let us now look at a case in which the function that relates the computed
quantity y to the measured quantity x is nonlinear. For example, let
2
yx
(2-7)
If, again, x, and y, represent the true values of x and y, respectively, and dx
and dy represent the corresponding errors, then applying Eq. (2-7), we get

y x
1

(2-8)

And
y ( y dy )
1

( x1 dx) 2

x 2 x dx (dx)
1

(2-9)

From wich we obtain

dy 2 x1 dx (dx) 2

(2-10)

Recognizing from Eq. (2-7) that 2x, is the derivative of y with respect to x,
evaluated at x0 we can express Eq. (2-10) as follows:
dy

dy
dx (dx) 2
dx

(2-11)

Equation (2-11) differs from Eq. (2-6) in that it includes the additional term
(ox)'. In practice, however, the error dx is so small relative to the measurement
itself that the higher-order term (dx)2 can be neglected. This means that
instead of using point P (Fig. 2-3) or. the curve to determine y point P ' on the
tangent to the curve at T is used to determine y. Thus, the propagated error
dy is represented by (y-y1) in Fig. 2-3, instead of by (y-y1), under the
assumption that the difference y -y' =(dx) 2 is negligible.
EXAMPLE 2-2
The area y of a square tree: of land, shown In Fig. 2.4, is required. The length x
of the side of the tract is measured with a 30 m long steel tape and is
observed to be 50.170 m. This measurement is then used to calculate the area
of the tract, as follows: the calculated area is y = x2 = (50.170)2 = 2517.0289 m2,
represented by square ABCD in Fig. 2-4,
If the tape is known to be too short by 0.030 m, compute the corresponding
error in the calculated area.

Fig 2-3

Fig 2-4
the tape is too short by 0.030 m, a measured 30-meter distance is
actually m, The correct length of the side of the square is therefore

x1

29.970
(50.170) 50.120m
30.000

The correct area of the. tract is

y1 x12 (50.120) 2 2512.0144m 2

,
Error in the area is thus
dy y y1 2517.0289 2512.0144 5.0145m 2

in fig. 2-4,equais the sum of the areas of the two rectangles,

AB 1B 'A' and BC and the square, B1BB2B'. This exact error is also obtainable
directly from he error in x, using Eq. (2-10); i.e.,
dx x x1 50.170 50.120 0.050 m

dy 2 x1 dx (dx) 2
2(50.120)(0.050) (0.50) 2 5.0145m 2

Now, if Eq. (2-6) is used, we must first evaluate the derivative dy/dx at x=0.120
Thus,
dy d 2

( x ) 2(50.120) 100.240m
dx dx
dy

dy
dx (100.240)(0.050) 5.0120m 2
dx

which, in Fig. 2-4, equals the sum of the areas of rectangles AB 1B'A' and
CB 2B'C'. The difference between the exact determination of the error in the
area and its determination according to Eq. (2-6) is only 0.0025m2 which is
(dx)2 the area of the all square B1BB2B'. This difference is only 0.05 % of the
error, and is therefore significant.*
In practice it makes little or no difference if the derivative dy/dx is evaluated at
x, or at x. Since the measured value is often the convenient value to
evaluation of the derivative at the measured value is common procedureSo far we have considered only the case of a single variable y computed as
a function of a single variable x. Suppose y now represents the area of a
rectangle instead of a square. In this case, two measured quantities, the
length x1 and width x2 are involved, and y is their product; i.e., y =x1 x2.
When more than one variable is involved in a function, the rules of partial
differentiation can be applied. Specifically, if the errors in x1, x2,. xn, are
represented by the differentials dx1 dx2..., dxnrespectively, then the error in y
can be represented by:
dy

y
y
y
dx1
dx 2 ...
dx n
x1
x 2
x n

(2-12)

y
y
y
,
,
are evaluated at the given
x1 x 2
x n
numerical (measured) values of x1, x2,. xn, respectively.

in which the partial derivatives

EXAMPLE 2-3
Instead of the square dealt with in Example 2-2, a tract of land is rectangular,
measuring 50.170 m by 61.090 m. If the same 30 in tape (0.030 m too short) is
used to make the measurements, evaluate the error In the calculated area of
the tract.
Solution
The calculated area, expressed to the appropriate number of significant
figures, is
y1 x1 x 2 (50.170)(61.090) 3064.9m 2
The partial derivatives are evaluated as:
y
x 2 61 .090 m
x1
y
x1 50 .170 m
x 2

*aplication of the multiplication rule for significant figures (Section 1.5) calls for expression dy to two significant Figures only;
i.e., dy = 5.0 m'. This, too, indicates that the 0.0025 m' mince is insignificant.

The errors, dx1 and dx2 are computed on the basis of the 30 m tape being too
short by 0.030 m. Thus,
dx1

0.030
(50.170) 0.050m
30

dx2

0.030
(61.090) 0.061m
30

Eq. (2-12) is used to calculate the error in the area:

and
dy

y
y
dx1
dx 2 (61 .090 )(0.050 ) (50 .170 )(0.061) 6.1m 2
x1
x 2

2.2. LINEARIZATION
We have noticed in the propagation of known errors in Section 2.1 that when
the function is nonlinear the higher-order term (dx)2 can be dropped because
it is very small. The resulting error propagation function, with (dx)' omitted, is
an approximation. Since it was shown that error propagation involving linear
functions does not entail any approximation, an alternative approach for
nonlinear functions is possible. This approach first replaces the nonlinear
function by its linearized form and then applies propagation, rather than
applying propagation to the nonlinear function first and then dropping terms.
The basis of linearization of it Function is the Taylor series expansion which
for one function of one variable y f (x), is
dy
y y0 x higher-order terms
dx x0

(2-13)*

where y f ( x0 ), and x x x0
The linearized form includes only the first two terms on the right-hand side
of Eq (2-13); all higher-order terms are neglected. For example, the
linearized form of the function y x 2 .
y x 02 2 x 0 x

in which y x 2 and (dy/dx)0=2x0 The function in question, shown in

Fig. 2-5, is represented by the curve from the origin to point O, and then by
the straight line tangent to the curve at O (i.e., along OP'). The slope of
OP' is (dy/dx)0
If y0 is represented by the constant b, and the derivative (dy/dx) o is represented by the constant a , then Eq. (2-14) becomes y a x b , which is
obviously a linear function in x. Applying error propagation [Eq. (2-5)) to this

* The symbol (dy/dx)0 represents the first derivative or y with respect to x, evaluated at x=x0

linearized form of the original function will yield a result that is identical to the
result obtained using Eq. (2-6),
EXAMPLE 2-4
Linearize the function y 2 x 3 x 3 4 x 7 at x=2 If x=2 and the error in xis
0.01, compute the error in y by:

Fig 2-5
i.
ii.

Using the expression for the total differential Eq. (2-6).

Applying error propagation, Eq. (2-5), to the linearized function.

Solution
Eq. (2-13), with higher order terms neglected, is y y 0 ( dy

dx

) x Thus, for

x0=2,
y ( 2 x 03 x 02 4 x 0 7) (6 x 02 2 x 0 4) x
19 24 x
which is the linearized form of y (2 x 3 x 2 4 x 0 7) , with a=24 and b=19. i.

According to Eq. (2-6), the error in y is

dy
dx (6 x 02 2 x 0 4)dx 24(0.01) 0.24
dx
ii.Applying error propagation., Eq. (2-5), to the linearized function, we get
dy

dy adx 24(0.01) 0.24

If the function y is in terms of two variables, x 1 and x2 i.e.,

y f ( x1 , x 2 )
then the linearized form is
y
y
x1
y y 0
x1 x0
x 2

x 2 ,
x0

(2-15)

(2-16)

In which y0 is the function evaluated at x1 x1 0 and x 2 x 2 0 ;i.e., y f ( x1 0 , x 2 0 ) .

As the number of independent variables, x1 x2 ...,increases, expression of eq.(2-6)
becomes quite lengthy. To make Eq. (2-16) as compact as possible, matrix
notation is used (see Appendix A). If in Eq. (2-16) we denote (y ) x 1 . by j1 and
x1
(y
) . by j2 then
x 2 x 2
(2-17)
y y 0 j1 x1 j 2 x 2
Defining j as the row matrix [j1, j2], and x as the column matrix
x1
x ,
2

the multiplication of j by x yields j1 x1 j 2 x 2 Thus, it can be seen that E.q. (217) can be written compactly as
y y 0 j x

(2-18)

As matter of fact. Eq. (2-18) applies no matter how many elements j and x have,
so long as they have the same number of elements. For example, y may be a
function of four variables. i.e.,

y y 0 j x y 0 j1 , j 2 , j3 , j 4

x1
x
2
x 3

x 4

y y 0 j1 x1 j 2 x 2 j3 x3 j 4 x 4

which shows how compact the matrix form in Eq. (2-18) is.

EXAMPLE 2-5
Figure 2-6 shows a tract of land composed of a semicircle with diameter AB, a
rectangle ABCE, and a triangle ECD. Express the total area y of the tract as a
function of the three dimensions x1, x2 and x3 shown in the figure. Then
linearize this function, given x1 0 50m, x 2 0 20m, x3 0 30m,

Fig 2-6

Solution
From the figure, the rea of the tract is
y

x32 x1 x3

1
x 2 x3
2

To linearize this function, first evaluate y0

y0

x320 x10 x30

1
x 20 x30 (30) 2 (50)(30) (20)(30) 2153.43m 2
2
8
2

The evaluate j:
y y y
1

j
,
,
x 30 , x 30 , x 30 , x10 , x 20 ,
2
4
2

x1 x 2 x 3

=[30, 15, 84] m.

The function in linearized form is thus

y y 0 j x
x1
=2153.43+[31 15 84 ] x 2 (m2)
x 3

That is,
y 2153 .43 30 x1 15x 2 84 x3 ( m 2 )

Equation (2-IS) shows linearization for the case of one variable y as a function
of several variables x1, x2, ....We now extend the technique to the more general
case of several variables, y1 y2 .., each a function of a set of independent
(measured) variables, x1 x2 ...; i.e.,
y1 f 1 ( x1 , x 2 ......x n )
y 2 f 2 ( x1 , x 2 ......x n )

(2-19)

y m f m ( x1 , x 2 ......x n )

in a manner similar to Eq. (2-16) e a c h of these functions is linearized as

y
y
y 1 y1 0 1 x1 1
x1 x0
x 2
y
y 2 y 2 0 2
x1

x 2 1
x0
x n

y
x1 2
x0
x 2

x n
x0

x 2 2
x0
x n

x n
x0

(2-20)*

y
y
y m y m 0 m1 x1 m
x1 x0
x 2

x 2 m
x0
x n

x n
x0

Replacing the partial derivative of each yi with respect to each variable xk by j1k
y
j 32 ), Eq. (2-20) becomes
(e.g., 3
x 2
y1 y1 0 j11 x1 j1 2 x 2 j1 n x n
y 2 y 2 0 j 2 1 x1 j 2 2 x 2 j 2 n x n

(2-21)

y m y m 0 j m 1 x1 j m 2 x 2 j m n x n

We can compact each line in Eq. (2-21) further to a form comparable to Eq,
(2-18) by using j1 =[j11, j12, j1n], j=[ j21, j22, j2n] and so on. Thus,
y1 y1 0 j1 x
y 2 y 2 0 j 2 x

y m y m 0 j m x

*At this point, it should be clear that derivatives y1/x1, y1/x2,, y1/xn are to be evaluated at

x1 x1 0 , x 2 x 2 0 , , x n x n 0 , .Accordingly, the subscripts x1 0 , x 2 0 , , x n 0 , are omitted.

(2-22)

We may further collect all y 1 and yi0 into respective column matrices; thus,
y1 0
y1
x1

y
x
y2 0
2
2

y
; y0
. As before , x

y m 0
ym
x n
Finally, if we consider j, as the first row, j 2 the second row, and so on, of a
rectangular matrix J with m rows and it commas, Eq (2-22) becomes
j1 2 j1 n x1
y 1 0 j11
y 1 y 1 0 j1

y y j
y
x 2
j
j

2
2
2
2
0
0
22
2 n
x
21

y
j
y
y

m
m
m0
j m 2 j m n x n
m 0 jm1

(2-23)

which in its most compact form is

y y 0 J x

(2-24)

The matrix J is called the jacobian matrix, which represents the partial
derivatives of all she functions in y with respect to each of the variables in x;
i.e ,
y1 y1
y1

x
x 2 x n
1

y 2
y 2 y 2

x 2 x n
(2-25)
J m1n
x1
x

y m y m
y m

x1
x 2
x n

It can be soon that J has as many rows m as there are functions yi, and as
many columns n as there are independent (or measured) quantities xk . All
partial derivatives, as well as the quantities y1 0 , y 2 0 , y m 0 , are evaluated at
the given values x1 0 , x 2 0 , x m 0 of the independent variables.
EXAMPLE 2-6
The tract shown in Fig. 2-6 is to be divided into two parts by the broken line
conne c t ing points B and E. Express the areas of the two parts, y1 and y2,
shown in the figure, as functions of the dimensions x1, x2 and x3. For x10 = 50 m,

x20=20m and x3o=30m evaluate the Jacobian matrix, J, and express y1 and y2
in linearized form.
Solution
From the figure, the two areas arc

1
x1 x3
8
2
1
1
y x1 x3 x 2 x3
2
2
y1

x32

The jacobian matrix is

y 1
x
1
J
y 2

x 1

y1
x 2
y 2
x 2

y 1 1
x
x 3 2 3 0

y 2 1
x3 0

x 3 2

0
1
x3
2 0

x1 0 )
2

1
x 2 0 )

x3 0

4
1
( x1 0
2

15 0 49

15 15 35

Now
1
2

x3 0 x1 0 x3 0

1103.43 2
y
10
8
2
y0

m
y 2 0 1 x x ( 1 x x ) 1050.00
10 3 0
1
2
2 0 0
2

The linearized form of functions y, and y, is thus

y y 0 J x

x1
1103 .43 15 0 49

x 2 m 2

1050 .00 15 15 35
x 3

that is,
y1 (1103 .43 15 x1 49 x 3 ) m 2
y 2 (1050 .00 15 x1 15 x 2 35 x 3 ) m 2

EXAMPLE 2.7
If the dimensions in Fig, 2-6 are x 1 = 50.00m, x 2 = 20.00 m, and x3 = 30.00 m,
and the errors in these dimensions are 0.02 m, -0.04m and 0.03 m,
respectively, evaluate the errors in areas y1 and y2 by applying error
propagation to the linearized functions derived in Example 2-6.
Solution
Errors dx1, dx2 and dx3 in x1, x 2 and x3 respectively, are also errors in x1, x2
and x3, respectively. Thus, applying error propagation to
y1 1103 .43 15 x1 49 x 3
y 2 1050 .00 15 x1 15 x 2 35 x 3

We get for the errors in y1 and y2

dy1=15dx1+49dx3
=15(0.02)+49(0.03)=1.77m2

PROBLEMS
2-1 Interior angles A and B of a plane triangle are known and fixed. Side b
(opposite B) is computed from the fixed values of A and B and the measured
value of side a (opposite A). If the error in a is 0.015 rn, evaluate the resulting
error in b. (a) for A= 60 and B= 60; (b) for A= 120 and B= 15; (c) for A =
15 and B= 120.
2-2 The triangular parcel of land ABC shown in Fig. 2-7 has dimensions AB=
150.00 m, BC= 80.00 m, and CA = 110.00 m. The parcel is divided into two
parts, I and II, as shown, by setting D on AB at a distance x from B. Evaluate
the resulting error in the area of I if x has an error of 0.020 m.

Fig2-7
2-3 A building wall is shown in plan view in Fig. 2-8. Each window has width a, the
spacing between windows is b, and the distance from each building corner to the
nearest window is c. The length y of the wall is determined from measurements x1,
x 2 and x3 If the errors in measurements x1, x 2 and x 3 are 5 mm, 8 mm, and 9 mm,
respectively, determine tic: error in the calculated value of y.

2-4 In taping a distance, the correction for slope is computed as

`

v2
c
2l

where l is the length of the tape and y is the difference in elevation between the
ends of the tape. If l=30 m, v =4.0 m, and the error in v is 0.10 m, evaluate the
error in the computed slope correction.
2-5 Interior angles A, B, and C of a plane triangle are known and fixed. The area
of the triangle is computed from these angles and the measured value of side a
(opposite angle A).

Fig2-8
Show that the relative error* in the computed area is twice the relative error
in the measured side.
2-6 Interior angles A and B of a plane triangle are fixed at 3000'00' and
7000'00', respectively. Side a (oppositesA) is measured and found to be
400.000 m, and the area of the triangle is computed from the given data. If
the error in the measured value of a is 0.040 m, evaluate the error in the
computed area: (a) by determining the difference between the area
computed from the measured value of a and the area computed from the
true value of a; (b) by applying Eq. (2 .6); (c) by applying the relationship
expressed in Prob. 2-5.
2-7 In stadia leveling, the difference in elevation V is computed from the rod
intercept r and vertical angle using the function V=(1/2)kr sin 2 where k
is the stadia constant. If k= 100 (assumed errorless) and the errors in r and
a are 0.005 m and 60 seconds of arc+ , respectively, evaluate V and the
error in V for: (a)r= 1.500 m, a = 0; (b) r = 1.500 m, a=15.
2 . 8 Interior angles A and B and side b of a plane triangle are measured and
found to be 5000'00', 2000'00' and 100.000 m, respectively. The errors in
A, B and b are 15'- 25', and 0.010 m, respectively. If side a is calculated
from A, B, and b, evaluate the error in at (a) by determining the difference
between a computer from the measured values of A, B, and b, and a
computed from the corresponding true values; (b) by applying Eq. (2 -12).
* If d x is the error in x, then d x/ x is the relative error in x.
+Angular error should the expressed in radians in functions that contain derivative

2-9 Use the Taylor series expansion, Eq. (2-13), to linearize each of the
following functions at the given value then. use the linearized function to
evaluate y the value x,. Compare this value of y with the value obtained
using the origin function.
(a)

y = 10/x; x0=5.0, x1=5.5.

(b)
(c)
(d)

y 100 x 2 ; x0 =6.0, x 1=6.2

y= 50 sin x; x0 =30, x 1=31.
y = 400tanx; x 0=12.0, x1 =12.5.

2-10 The correction for the sag of a tape is given by the function
w 2l
y
24 p 2

in which L is the length of the tape, w is the unit weight of the tape and p
is the applied pull. The length and unit weight of a tape are 50 m and
0.031 kg/m respectively, and the pull is variable. (a) Linearize y at p.= 10
kg using the Taylor series expansion, Eq. (2-I3). (b) Evaluate y for p= 11
kg using the line arized function and compare the resulting value with the
value obtained usin the original function.
2-11 Linearize the function y= [(I00)2 + x2 -200x cos] 1/2 at x0=80.000 m and
0 = 4000' according to Eq. (2-16). Use this linearized function to evaluate y at
x=81.000 m and =4020', and compare the resulting value with the value of
y obtained using the original function.
y1
x
2-12 Vector x = is a function of v ector y y 2 such that
x2
y 3

x1 3 y12

y2
0 .7 y 3
y12

And
x2

y2
4 y 32
2
y1

Evaluate the Jacobian matrix J = x./ y at y1o =2, y2o = 4, and y3o = 1
d
2-13 With reference to Fig. 2-9, vector y is a nonlinear function of vector
h

s
x .Evaluate the Jacobian matrix J=x./ y at s0=1000 m and 0=20 and

express y in the linearized form of Eq (2-24).

2-14 if, in fig 2-9, measured values for s and are 1000.000 m and 200000
respectively, and their errors are 0.100 m and 20, respectively, evaluate the
errors in the computed values of d and h applying error propagation to the
linearized form of y as obtained in problem 2-13.

THE CONCEPT OF ADJUSTMENT

3.1. INTRODUCTION

Surveying measurements, such as distances and angles, are usually

obtained, not only because they are of direct interest themselves, but
because they are often needed to determine other quantities. For example,
although it may be quite important to know the value of an angle for its own
sake, it may be of greater importance to determine the area of a figure of
which the angle is an element. Thus, a new quantity, in this case the area,
is computed in terms of the angle and other measured quantities. Other
quantities calculated from survey measurements include: relative positions
of points (often in the form of Cartesian coordinates), dimensions, curve
data, elevations, slopes, and volumes. Thus, before the survey
measurements are obtained, we usually have a requirement for a particular
quantity or quantities that are functions of the needed measurements. The
general relationships which relate the measurements to the other quantities
of interest constitute what is known as the model. Since almost all the
relationships encountered in surveving computations are mathematical
representations of the underlying physical and geometric conditions, the
model is frequently called the mathematical model. Although in general
surveying practice we may not be immediately aware of the model
underlying the particular problem, it nevertheless exists the moment data
are to be reduced to yield useful information. We will show in the course of
this book that using the model concept will always be helpful to the
surveyor in understanding the elements of a problem and its reduction.
As an example, let us consider the size and shape of a plane triangle. Once
we have stated the problem, we have also specified the underlying
mathematical model, in this case, that of determining the size and shape of
a plane triangle. At this point, it is quite important to determine the minimum
number (n 0) of elements (variables) necessary for uniquely determining the
model, and the type of such elements that are consistent with the model. For
instance, we realize that while the shape of the triangle can be determined
by either angles or sides (or a combination), the size on the other hand
cannot be dete rmined from angles only but requires at least one side. It
follows that one side and two internal angles, or a total of three elements, are
the minimum required to specify the model uniquely. Again, at least one of
the three elements must be a side; the three internal angles are not

sufficient for determining the triangle 's size.

Having determined n0 the surveyor now decides on what measurements to
obtain. If exactly one side and two interior angles are measured, the triangle
will be uniquely fixed. But if a mistake or blunder is committed in any or
these three measurements, there is no means of detecting it. Therefore, it
is always good practice to measure more elements of the model than the
mini-mum necessary for its unique determination.
Each observation in excess of the minimum number, n0, is called a redundant
measurement. Thus, if a total of n measurements is obtained with respect to
(and consistent with) a certain model whose minimum number of elements is
n0 the redundancy (or statistical degrees of freedom as it is also called) is
given by
r= n n0

(3-1)

So, the redundancy is the difference between the given number of

observations, n, and the minimum required uniquely determine the
underlying model, n 0 both sets being consistent with the model.
When redundant measurements exist, an interesting situation is encountered. For each subset of n 0 measurements of the n given measurements
(n> n0) a slightly different determination of the model is likely to result, Thus,
when a side and three internal angles are measured, there exist three
possible combinations, each consisting of the side and any two of the three
angles. For each of these three possibilities, a triangle with slightly different
shape and size (area) will usually result because of the unavoidable errors in
the measurements. Consequently, because of measurement errors (or the
statistical variations in the measurements, see Chapter I), then observations
no longer fit the model exactly. This apparent inconsistency with the model is
resolved through the replacement of the given observation L by another set

of so-called estimates* l , such that the new set l fits the model exactly.
For example, when the new estimates of the side and three interior angles
are calculated, any combination of the side and any two angles will give

exactly the same triangle. This means that with l , the inconsistency with the
model is resolved, and any n 0 subset of the n estimated measurements will
always yield the same unique determination of the model.

Each estimated observation, l i , can be looked upon as a corrected observation, obtained from the measured value l ; by adding a correction v; to it,
i.e.,

l lv

(3-2)

Equation (3-2) is equivalent to Eq. (1-2), in which vi is defined as the residual.

For all n observations the corresponding relation is

*The term estimate is used in the proper statistical sense (see Chaper 8). It is not to be taken as a guess, but as the result of an
appropriate computation or graphical construction using known data,

l
n,1

l v
n,1

(3-3)

n,1

Where each term is a column vector with it elements, i.e.,

l 1
l 1

l l 2 ; l l 2 ; and v





l
n
l n

v1

v 2


v n

(3-4)

The residuals (corrections v are unknown and must be determined before l

can be computed) There is essentially an infinite number of possible sets of
residuals which would wee corresponding estimated (or corrected) observations that fit the model exactly. However, it is logical to recognize that only
one set of residuals would yield an optimum solution. To obtain such a
solution, an additional criterion is necessary so that only get of residuals
may be calculated. There area variety of ceiteria possible but
the
which is most commonly used in surveying that given he the principle of least
squares (see Section 3.3),

The operation of finding the new set of estimates l according to some

criterion is called adjustment. Least squares adjustment is used extensively,
not only in surveying and the related fields of photogrammetry and geodesy,
but also in many other fields of science and engineering. But before we
discuss this method we will first briefly review two relatively simple pro cedures of adjustment.
3.2. SIMPLE ADJUSTMENT METHODS
Let is consider a fairly common problem in surveying, that of inte rsection. In
fig. 3-1, points C1, C2, C3 are three control points with known horizontal
coordinates. The location of point P is to be determined by intersection, I.e.,
by angle (or direction) measurements at the control points. If the two angles
l1 and l2, only are measured, the intersection would be at point A, fig. 3 -I. If,
however, one more (redundant) observed angle l3 is obtained, we would
have three possible intersected positions of the required point. In addition
to A, point B is obtainable from l2 and l3 and point C from l1 and l3. This
clearly demonstrates the concept that errors in measurements make them,
inconsistent with the underlying mathematical model. It also shows that a
criterion of some sort is needed before a unique solution to the problem is
possible.
The triangle ABC in Fig. 3-I is often called the error triangle. (It is obviously
exaggerated in the figure for purposes of effective illustration.) If a
simple adjustment is to be exercised, a common criterion for such

adjustment Is to select the centroid of the error triangle as the point of

intersection of the three rays. This is shown as point P i n Fig. 3-1.
Connecting point P to C1 C2 C3 gives the three corrected directions (which
may also be called adjusted directions), shown by dashed lines in Fig. 3-1.
Thus, it can be seen

Fig 3-1

that l 1 needs to be increased by v1, l 2 , increased by v2, and l 3 , decreased

by v3 to get the corresponding l 1 , l 2 , l 3 respectively. In other words, to get P

as a unique solution, the two residuals v 1, v 2 are positive while v 3 is
negative. While Fig. 3-1 helps demonstrate the concepts of redundancy,
inconsistency with the model, residuals, and the new estimates for the
observations, it should be emphasized that the solution is almost always
obtained graphically.
Other simple procedures of adjustment are concerned with traverses. One
very common procedure of traverse adjustment is the compass or Bowditch
rule, depicted in Fig, 3-2 for a traverse that begins at a control point .4 and
ends, or closes, at another point B. The departure and latitude of a particular
leg or course of the traverse (e.g., course 2-3) are given by
departure, x 23 l 23 sin 23
latitude, y 23 l 23 cos 23

(3-5)

in which is the length of the course and a,, is its azimuth from north or the
Y-axis in Fig. 3-2. Starting with point A and using the measured angles and
distances, we end up at point B' instead of the control point B. The closing
errors are shown in the figure as the departure error ex, and the latitude error

e y. The corrections in departure and latitude for one leg, such as 2-3, are
computed as
departure correction for. 2-3 =

latitude correction for 2-3 =

(l 23 )( e x )
l ij

(3-6)

(l 23 )(e y )

(3-7)

l ij

Fig 3-2

in which

is the sum of all course lengths in the traverse. For the

if

traverse in Fig. 3 - 2 ,

if

l A1 l12 l 23 l3 B

EXAMPLE 3-1
Data for the traverse in fig 3-2 are as follows

STATION
A
1
2
3
B

LINE

LENGTH
(m)

AZIMUTH

A1
12
23
3B

212,120
321,070
315,820
304,650

60 33'' 00'
130 12'' 00'
64 03'' 00''
122 45'' 00'

X (m)
5000,000

Y (m)
5000,000

5970,010

4870,280

Adjust this traverse by the compass or Bowditch rule.

Solution
The compass or Bowditch rule yields good results when distances and
angles are measured with equivalent precision. The tale assumes no
correlation and that all measurements are of equal weight The least
squares method, on the other hand, imposes no such restrictions. Its
results are optimum, and with the accelerated increase in the availability
and efficiency of electronic computational aids, it is expected to be the
procedure that is most extensively used.
First, we compute the departures and latitudes and the provisional
coordinates of point B' to determine the closing errors e and
See first table on opposing page.
In order to apply the compass rule, the following two ratios arce
needed: 1. e, divided by the total length of the traverse:

e
l

0.134
1.16 10 4
1153.66

2. e, divided by the total length of the traverse:

e
l

0.166
1.44 10 4
1153.66

Each of these values is multiplied by each course length, with sign

changed, to give the corresponding departure and latitude corrections,
respectively [Eqs. (3-6) and (3.7)]. The departures and latitudes are
corrected accordingly, and used to calculate final coordinates for all
traverse stations. The results are summarized in the second table on the
opposite page,

STATION
A

COURSE

LENGTH
(m)

AZIMUTH

DEPARTURE

A1

212,12

60.5500

184,711

12

321,07

130.2000

245,232

LATITUDE

5104,292
-207.237

5429,943
23

315,82

64.0500

283,978

4897,055
138,199

5713,921
3B'

B'
B'

304,65
1153,66

122.7500

Y(m)
5000

104.292
5184,711

STATION
A

X(m)
5000

256,223

5035,254
-164,808

5970,144
5970,01
e=0,134

4870,446
4870,28
e=0,166

CPRRECTD
DEPARTURE DEPARTURE
LATITUDE
CORRECTED
DEPARTURE CORRECTION
(m)
LATITUDE CORRECTION LATITUDE

COURSE

LENGTH
(m)

A1

212,12

184,711

-0,025

184,686

104.292

-0,031

104,261

12

321,07

245,232

-0,037

245,195

-207.237

-0,046

-207,283

23

315,82

283,978

-0,037

283,941

138,199

-0,045

138,154

3B'

304,65
1153,66

256,223

-0,035
-0,134

256,188

-164,808

-0,044
-0,166

-164,852

1
2
3
B'
B'

check

X(m)
5000

Y(m)
5000

5184,686

5104,261

5429,881

4896,978

5713,822

5035,132

5970,01
5970,01
0

4870,28
4870,28
0

3.3 THE LEAST SQUARES METHOD

As we have indicated in the preceding two sections of this chapter, when
redundant measurements exist, an adjustment is necessary so as to get
a unique solution to the problem at hand. Although approximate
methods, both graphical and computational, may be adequate for some
limited cases, a more general and systematic procedure is needed for
application to all situations The least squares adjustment method is such
a procedure, Assuming, for the time being, that all observations are
uncorrelated and of equal precision (correlation and unequal precision
will be taken up later), the cast squares adjustment method is based
upon the following criterion,
The sum of the squares of the observational residuals must be a
minimum, i.e.,
n

v12 v 22 v n2 vi2

minimum

(3-8)

i 1

Thus, in addition to the fact that the ad j usted observations must satisfy
the model exactly, the corresponding residuals must satisfy the least
squares criterion of Eq. (3-8). To demonstrate the application of this
criterion, let us consider the very simple example of determining a
distance x by direct measurement. Since the model involves only one
element (or variable), it requires only one measurement to uniquely
determine the distance, i.e., n 0 =1. Suppose that we have two
measurements, l1 = 15.12 m and l2= 15.14 m. Then n = 2, and according
to Eq. (3-1), there is one redundancy because
r=nn 0 =2 1 = 1.
The final value of the distance can be obtained from the observations as
follows:

x l1 v1 l 1

(3-9)

x l2 v2 l 2

There are obviously many possible values for v 1 and v 2 such that these
relations are satisfied. For example, we could have v 1 = 0 and

v 2 l l v =-0.02 m; or v1 = +0.01m and v 2=-0.01 m; or v1= 0.015 m

n,1 n,1
n,1
and v 2 = -0.005 m; all of which would satisfy the model as expressed by the
relations in Eq. (3-9). Of all these possibilities, the east squares solution is
the one for which (v12 v 22 ) is a minimum. For the three possibilities the
corresponding sums of the squares of the residuals are

1 0 ( 0.02 ) 2 4 10 4 m 2
2 ( 0.01) 2 ( 0.01) 2 2 10 4 m 2
3 ( 0.015 ) 2 ( 0.005 ) 2 2.5 10 4 m 2
It is clear that 2 is the smallest of the three values, but the real question is
whether it is the very minimum value when all possible combinations of
corrections are considered. To answer this question, and also demonstrate
the criterion of least squares geometrically, we refer to Fig. 3-3. The two

adjusted measurements l 1 and l 2 , are related to each other by

l1 l 2 0

(3-10)

which is easily obtained from Eq. 3-9 by subtracting the second line from
the first. Now, if we let the abscissa of a two-dimensional Cartesian

coordinate system in Fig.3-3 represent l 1 and the ordinate l 2 then Eq.(310) would be depicted by a straight line that is inclined 45 with both axes,

as shown. The. two numerical values of the observations, l 1 = 15.12 m and

l 2 =15.14 m define a point A which falls above the line because l 1 < l 2 The
line representing Eq. (3-10) is called the condition line since it represents

the condition that must exist between the two adjuste a observations l 1

and l 2 . When this condition is satisfied, the underlying model is also

satisfied. Thus, an adjustment would be carried out if point A is replaced
by another point which falls on the condition line. It is obvious that there
exist many possibilities

Fig 3-3

for such a point on the line, three of which are indicated by A1, A2, A 3 to
correspond to the three computed values 1, 2, 3, respectively. Of all
possibilities, the least squares principle selects the one point, A2 such that
the distance AA2, is the shortest possible (i.e., minimum). From simple
geometry, the line AA2, is therefore normal to the condition line as shown in
Fig. 3-3. It can be seen that. A2 also satisfies the intuitive property that the

new estimates l 1 , l 2 of the observations must deviate as little as possible

from the given observations, l 1 , l 2

In Fig. 3-3, point A1, is obtained by moving from A straight down, thus with
v1= 0 and v2= -0.02. For point A2 the direction of AA2: is perpendicular to the
condition line, thus forming a 45 triangle from which v1 and v 2 are equal in
magnitude but opposite in sign. The adjusted observations end up being

equal, or l 1 l 2 15.13m , thus satisfying the condition expressing the

model. In fact, the final estimate of the distance is x 15.13m , which

satisfies the intuitive feeling that the adjusted value should be the arithmetic
mean of the two given observations. I. e.,

x (1 / 2)( l 1 l 2 ) (1 / 2)(15.13 15.14) 15.13m

Indeed, this is the simplest tase of the very important fact that whenever
measurements of a quantity are uncorrelated and of equal precision
(weight), the least squares estimate of rite quantity is equal to the arithmetic
mean of the measurements (see the following section).
When only n 0 observations ate obtained, the model will be uniquely determined, as for example measuring a distance once. If one additional
measurement is made, there is one r redundancy (r= I) and a corresponding
equation must be formulated to alleviate he resulting inconsistency. Thus, in
the case of two measurements of a distance just discussed, Eq. (3-10) must
be enforced in order to guarantee that the two adjusted observations end up
being equal, thus satisfying the model. Such an equation is called a
condition equation, or simply a condition, since it reflects the condition that
must be satisfied with regard to the given model.
As another example, consider the shape of a plane triangle. Any two
interior angles would uniquely determine its shape, l.e., n0=2 If all three
interior angles are measured (n=3), then there is a redundancy of one (r=
nn0=32=1). For this redundancy, one condition equation needs to be
writ ten to make the adjusted observed angles consistent, such a condition
would reflect the geometric fact that the sum of the three interior angles in
a plane triangle must equal I80. Thus, if the three measured angles

are l 1 , l 2 , l 3 , then the condition is l 1 + l 2 + l 3 =180

The condition equations discussed so far contain both observations and
constants. They are equal in number to the redundancy of the problem, r.
Thus, for any adjustment problem, there exist r condition equations among

the observations. This will lead to one technique of least squares called
adjustment of observations only.
A second least squares technique which is used frequently is called adjustment of indirect observations. In this technique, the number of conditions
is equal to the total number of observations, n. Since in terms of the observations only there should be r conditions, then in this technique the
equations must contain n - r= n0 additional unknown variables. These
additional unknowns are called parameters. However, unlike an observation
which has a value at the outset, a parameter is an unknown which has to a
priori value An example of the conditions involved in the case of adjustment

of indirect observations is given by Eqs. (3-9). In these equations, x

represents the least squares estimate of the required distance, the one

unknown parameter. x is not known beforehand, but is a calculated result of

the least squares solution. Here, there are two condition e nations (n = 2),
which is the sum of the redundancy (r= 1) and the number of parameters

(n0= 1). It can be seen that if the parameter estimate x is algebraically

eliminated from Eqs. (3-9) there will remain one equation in terms of the
observations, Eq. (3-10). In general, all techniques of least squares are
equivalent in that they yield identical results for the same problem. The
reason for having different techniques is that each class of problems is
usually better handled by one technique than by another.
3.4. EXAMPLES OF SIMPLE LEAST SQUARES PROBLEMS
Although we have not yet given a general derivation of any least squares
techniques, and will not do so until Chapter 4, some experience in working
out a few simple problems by applying the least squares criterion directly will
be helpful at this point. In all of these examples, we will assume that the
observations are uncorrelated and of equal precision (weight). This
assumption is necessary in order to keep the manipulations relatively
simple. It is important to note that, in practice, measurements may we]
have unequal weights for which, of course, the more detailed procedures
given in succeeding chapters would apply,
EXAMPLE 3-2
A distance is measured 4 times (n=4) with the following results: l1 = 32.51 m,
l 2 32.48 on, l 3 =32.52 m, and l 4 =32.53 on. What is the least squares estimate
of the distance?
SOLUTION

of then Iran squares estimate of the distance be designated by x .

Since n = 4, and n=1 (only one element, the distance, is in the mode) then r=

4 I = 3. Carrying x the Unknown parameter estimate, the number of

conditions Is 3 + 1 = 4 = n The conditions are

l1 v1 x or

v1 x 32.51

l 2 v2 x or

v 2 x 32.48

l 3 v3 x or

v 3 x 32.52

l 4 v4 x or

v 4 x 32.53

The quantity to be minimized is

v12 v 22 v 32 v 42 ( x 32 .51) 2 ( x 32 .48) 2 ( x 32 .52 ) 2 ( x 32 .53) 2

Order for to be a minimum. Its partial derivative with respect to x must be

evaluated and equated to zero. Thus

2( x 32.51) 2 2( x 32.48) 2 2( x 32.52) 2 2( x 32.53) 2 0

Of

4 x (32.51 32.48 32.52 32.53) 2 130.04m

As expected, the least squares estimate turns out to be the simple mean of
the four observations. This is always true no matter how many repeated
measurements there are as long as they are uncorrelated and of equal
precision. [When the measurements are uncorrelated but unequal in
precision. the so -called weighted mean is used; see Example 4-5 and Eq.
(4-39), Chapter 4.1]
EXAMPLE 3-3
The interior angles of a plane triangle are 1= 4133', 2= 7357' and
3=5927'. Compute the adjusted angles using the method of least squares.
Solution
Since it takes two angles to fix the triangle (i.e., ,n 0=2) and since n=3, then
the redundancy is r= 3 2 = 1. The one corresponding condition is

1 2 3 180
Or
( 1 v1 ) ( 2 v 2 ) ( 3 v 3 ) 180

Or
v1 v 2 v 3 ) 180 ( 1 2 3 )
180 ( 4133'7857'59 27' )
180 17957' 3'

From which v3 = (3- v1- v2). The quantity to be minimized is

v12 v 22 v32 v12 v 22 (3'v1 v 2 ) 2

Then,

2v1 2(3'v1 v 2 )( 1) 0
v1

2v 2 2(3'v1 v 2 )( 1) 0
v 2

These two equations can be reduced to

2v1 v 2 3'
v1 2v 2 3'

Multiplying the equation by 2 and subtracting, we get

4v1 2v 2 6'
v1 2v 2 3
3v1
3'

Hence v1=1 substituting back in the first of the pair of equations, we get
v 2 3'2v1 3'2' 1'

Finally,
v3 3'v1 v 2 3'2' 1'

The adjusted angles are

1 1 v1 4134 '

2 2 v 2 78 58'

3 3 v3 59 28'
Check 18000

EXAMPLE 3-4
Distances AB, BC, CD, AC, and BD are measured, Fig 3-4. The observed
values are 100.000 m, 100.000 m, 100.080 m 200.040 m, and 200.000
m, respectively. All measurements are uncorrelated and have the same
precision. If the measured values are adjusted in accordance with the
least squares principle, what is the resulting adjusted distance between
A and D?

Solution
The geometric model is relatively simple being that of three collinear
distances, AB, BC, and CD, which we will denote by x 1 , x2 and x3
respectively. It would obviously take a minimum of three measurements to
uniquely determine this model (i.e., n 0= 3). Since we have five measured
distances (n= 5), there are then two redundant observations, or r=5-3=2.
If we carry x 1, x 2, x 3 as three unknown parameters, then we need to write
2+3 =5= n condition equations which relate the five measurements, l1, l2,

l3, l4 and l 5 , to the parameter estimates x 1 , x 2 , and x 3

l1 v1 x1

v1 x1 l1

or

v2 x 2 l2

or

v3 x 3 l3

or

v 4 x1 x 2 l 4 x1 x 2 200.040

or

v 5 x 2 x3 l 5 x1 x 2 200.000

l 2 v2 x 2

l 3 v3 x3

l 4 v4 x1 x 2

l 5 v5 x 2 x3

or

x1 100.000

x 2 100.000

x 3 100.080

In order to obtain a least squares solution we must minimize the sum of the
squares of the residuals. Thus,

v12 v 22 v32 v 42 v52

( x 1 100.000) 2 ( x 2 100.000) 2 ( x 3 100.080) 2 ( x 1 x 2 200.040) 2 ( x 2 x 3 200.000) 2

Mu st be m in im ize d.
To minimize , its partial derivative with respect to each parameter

estimate ( x 1 , x 2 , x 3 ) is evaluated and equated to zero; thus

Fig 3-4

2( x 1 100000) 2( x 1 x 2 200.040) 0

x1

x2

x3

2( x 2 100000) 2( x 1 x 2 200.040) 2( x 2 x 3 200.000) 0

2( x 3 100080) 2( x 2 x 3 200.000) 0

Cleaing and rearranging, the three aquations become

2 x1

300 .040

(a)

x 1 3 x 2 x 3 500 .040

(b )

x2

x 2 2 x 3 300 .080

(c )

These three equations in three unknowns are called the normal equations.
They indicate that after applying the least squares criterion (of minimizing ),
an overdetermined inconsistent measurement case is transformed into a
unique (consistent) case.
Dividing (a) by 2 and subtracting from (b) gives

2.5 x 2 x 3 350.020

from which we subtract one-half (c) to get

2 x 2 199.980,

or

x 2 99.990m

Substituting this value of x 2 into (a), we get

2 x 1 300.040 99.990 200.050 ,

or

x 1 100.025m

or

x 3 100.045m

and substituting the same value into (c), we get

2 x 3 300.080 99.990 200.090

Thus, the adjusted distance between A and D is

AD x 1 x 2 x 3 100.025 99.990 100.045 300.060

EXAMPLE 3-5
The equation of a straight line in a plane is y-ax-b=0, as shown in fig. 3-5. In
order to estimate the slope, a, and the y-intercept, b, the coordinates of three
points are given:

c
1
2
3

X (cm)
2
4
6

Y (cm)
3,2
4
5

The method of least squares to calculate the respective estimates, a and b of

__________ the and y-intercept, assuming the y-coordinates are the only
observations, i.e., the coordinates are error-free constants.
solution
if the y-coordinates as well as the x-coordinates were error-free, each one of
the given points would fall exactly on the line y-ax-b=0. However, due to
measurements errors in the y-coordinates, the three points do not lie on y-axb=0,
do they necessarily lie on any straight line. The objective of this

exercise, there
to find a and b such that the line y- a x- b =0 fits the three
points as closely possible according to the least squares criterion.
Obviously, it takes a minimum of two measurements (two points) to uniquely
(determine a straight line (i.e., n0= 2). Since we have three measurements
(n =3 ), are is one redundant observation (r= 3 -2 =1 I). I f we carry two
unknown parameter, a and b, we must write n =3 condition equations which
relate the three measurements y 1 , y 2 and y 3 to the two parameter estimates

a and b . Allowing for residuals y 1 , y 2 and y 3 respectively, in the y-coordinates,

the three condition equations are:

v1 y1 a x1 b 0

v2 y 2 a x2 b 0
v3 y 3 a x3 b 0

Expressing the residuals in terms of a and b andd the given coordinate values,
we have

v1 2 a b 3.2
v 2 4 a b 4 .0
v 3 6 a b 5 .0

The least squares estimate a and b are obtained when

v12 v 22 v32

v ( 2 a b 3 .2 ) 2 ( 4 a b 4 .0 ) 2 ( 6 a b 5 .0 ) 2

Is a minimum. To minimize , partial derivates of are taken with respect to a

and b and equated to zero . thus,

2 ( 2 a b 3 .2 ) 2 ( 2 ) 2 ( 4 a b 4 .0 ) 2 ( 4 ) 2 ( 6 a b 5 .0 ) 2 ( 6 ) 0

2( 2 a b 3.2) 2 (1) 2( 4 a b 4.0) 2 (1) 2(6 a b 5.0) 2 (1) 0

Clearing and combining terms, we get the normal equations as in the

preceding example

56 a 12 b 52.4
12 a 3 b 12.2v

Solving these equations in the usual way, we obtain

a =0.45, and b =2.27 cm

Once a and b are known, the residuals can be computed:

V1=2(0.45)+2.27-3.2=-0.033 cm
V2=4(0.45)+2.27-4.0=-0.067 cm
V3=6(0.45)+2.27-5.0=-0.033 cm
These residuals are also shown in fig 3-5. Their signs indicate whether the
given y-coordinates need to be increased (+) or decreased (-) to make points
fall on the line y - 0.45x - 2.27=0

EXAMPLE 3-6
Figure 3-6 depicts a simple triangulation scheme where the six angles x1 x2
x3 are measured and have the following values:
X1=48.88
x2=42.10
x3=44.52
x4=43.80
x5=46.00
x6=44.70
Using the method of least squares, calculate the adjusted values of the six
angles.
Solution I
The geometric model of this problem is concerned with two overlapping
plane triangles, as shown in Fig. 3-6. It therefore takes its n0=4 angles to
uniquely determine the model. Having n =6 measured angle, the
redundancy is r=6-4=2. Hence, there are two independent condition
equations that can be written in terms of the six observations.These are

x 1 x 2 x 3 x 4 180.00
x 3 x 4 x 5 x 6 180.00

in which x 1 , x 2 ,, x 6 represent the least squares estimates of the six

angles, i.e., the "Adjusted" angles. Each estimate x i is equal to the

observed value x i plus a residual v i The values of the residuals may be
calculated by the method of least squares as follows. First, we re write the
two conditions as

Fig 3-6

( x1 v1 ) ( x 2 v 2 ) ( x3 v3 ) ( x 4 v 4 ) 180.00

And
( x3 v3 ) ( x 4 v 4 ) ( x5 v5 ) ( x6 v6 ) 180.00

Or when substituting the values of the measurements,

v1 v 2 v3 v 4 180 .00 ( x1 x 2 x3 x 4 )
180 .00 179 .30 0.70

and

v3 v 4 v5 v 6 180 .00 ( x3 x 4 x5 x 6 )
180 .00 179 .02 0.98

The tows values 0.70 and 0.98 are called closing errors for the tow triangles
ABD and ABC respectively.
The least squares criterion calls for minimization of

v12 v 22 v32 v 42 v52 v 62

But since there are two conditions among the six residuak there are only four
independent un knowns, thus, using the first condition to solve for v1 and the
second to solve for v6, we get
v1=0.70- v2- v3- v4
v6=0.98- v5- v4- v5
substituting, becomes

(0.70 v 2 v1 v 4 ) 2 v 22 v32 v 42 v52 (0.98 v3 v 4 v5 ) 2

And for to be a minimum, we have

2(0.70 v 2 v3 v 4 )( 1) 2v 2 0
v 2

2(0.70 v 2 v 3 v 4 )( 1) 2v3 2(0.98 v 3 v 4 v5 )( 1) 0

v 3

2(0.70 v 2 v3 v 4 )( 1) 2v 4 2(0.98 v3 v 4 v 5 )( 1) 0
v 4

2v 5 2(0.98 v3 v 4 v5 )( 1) 0
v 5

(1)
(2)

 '
2v 4 2k1 2k 2 0 or
v 4

v 4 k1 k 2

'
2v 5 2k 2 0 or
v 5

v5 k 2

'
2v 6 2k 2 0 or
v 6

v6 k 2

Now we have the six unknown residual expreseed in terms of only two
unknown constants k1, k2. Thus when substituting into the two condition
equations we get two normal equations in two unknowns, i.e.,
k1 k1 ( k1 k 2 ) ( k1 k 2 ) 0.70
( k1 k 2 ) ( k1 k 2 ) k 2 k 2 0.98

Or
4k1 2k 2 0.70
2k1 4k 2 0.70

Solving these two equations for k1 and k2 we get:

k1=0.07, and k 2=0.21
thus, the residuals are
v 2 k1

0.07

v3 k1 k 2 0.28
v 4 k1 k 2 0.28
v5 k 2

0.21

v6 k 2

0.21

Which are identical to those calculate by the firs procedure. Therefore, the

estimated values for the angles x 1 , x 2 ,, x 6 are also the same. These two
procedures represent two techniques of least squares adjustment
describes in Chapter 4
Example 3-7
Figure 3-7 shows a small level net which A is a bench mark with known
elevation 281.130 m. the following differences in elevation are observed
using a direct leveling procedure:

FROM
(LOWER
POINT)
B
D
D
B
D
A

TO (HIGHER
POINT)
A
B
A
C
C
C

OBSERVED
DIFERENCE IN
ELEVATION
l1=11,973
l2=10,940
l3=22,932
l4=21,040
l5=31,891
l6=8,983

Assuming that all observations are uncorrelated and have equal precision,
use to least squares criterion to calculate values (least squares estimates)
for the elevation of points B, C, and D.

Solution
In this problem, the minimum number of observations needed for a unique
solution is n 0= 3, and the number of redundant observations is r = 6 - 3 = 3.
If there were no observational errors, going around a loop , and closing back
the starting point would yield no discrepancy. However, this is not the case,
as example, going from A to B to D and back to A yields (-11.973 - 10.940
+22.932) 0.019 m. We must therefore allow for this discrepancy, and others,
by introduction six residuals v1, v2,,vk one for each observation, For
convenience, the elevation of a point is designated by its own symbol, i.e., B
is the elevation of point B, and s o The six condition equations are
ring and rearranging we get the following set of four normal equations in the
unknown residuals v2, v3, v4, v5,:
2v2+ v3+ v4
=0.70
(a)
v2+3v3+2 v4+ v5 =1.68
(b)
v2+2 v3+3 v4+ v5=1.68
(c)
v3+ v4+2 v5=0.93.
(d)

attracting (c) from (b) gives

v3-v4=0,

or

v3=v4

v4 In place of v3 in (a), (b), and (d) yields

2v2+ v4 =0.70
v2+5v4+ v5=1.68
2v4+2 v5=0.98

(e)
(f)
(g)

multiplying (g) by 1/2 and subtracting from (f) we obtain

v2+4v4=1.68-(1/2)(0.98)=1.19 (h)
multiplying (e) by 1/2 and subtracting from. (h.), we obtain
3v4=1.19-(1/2)(0.70)=0.84
v4=0.28= v3. Using (e) we solve for v2:
v2=(1/2)[0.70-2(0.28)]=0.07
Using (g) we solve for v5:
v5=(1/2)[0.982(0.28)] =0.21
ly, using the condition equations (1) and (2), we calculated the values of v1 and
v5:
v1=0.70-0.07-0.28-0.28=0.07
v 5=0.98-0.28-0.28-0.21=0.21
these residuals the estimated values of the angles are

x 1 x1 v1 48 .88 0.07 48 .95

x 2 x 2 v 2 42 .10 0.07 42 .17

x 3 x3 v3 44 .52 0.28 44 .80

x 4 x 4 v 4 43 .80 0.28 44 .08

Check 18000

x 3 44 .80

x 4 44 .08

x 5 x5 v5 46 .00 0.21 46 .21

x 6 x 6 v 6 44 .70 0.21 44 .91

Sum= 180.00check
also showing that the two condition equations are exactly satisfied.
Solution II
This problem can also be solved another way. The two conditions
equations are rewritten as follows:
v1 v 2 v 3 v 4 0.70 0
v 3 v 4 v 5 v 6 0.98 0

Multiplying these equations by 2k 1 and 2k2 respectively, yields

2 k1 (v1 v 2 v 3 v 4 0.70 ) 0
2 k 2 (v 3 v 4 v 5 v 6 0.98) 0

The scalars k1 and k2 are unknown constants which are called Lagrange
multipliers (see also Section 4.4). The factor 2 is introduced for
convenience only, to avoid fractions after differentiation.
Now, the least squares criterion calls for minimization of the function

v12 v 22 v32 v 42 v52 v 62

I f 2k1 2k1 (v1 v 2 v3 v 4 0.70) 0 and 2k 2 (v3 v 4 v5 v 6 0.98) 0 are
subtracted from a new function is created:

v12 v 22 v 32 v 42 v 52 v 62 2k1 (v1 v 2 v 3 v 4 0.70 )

2k 2 (v 3 v 4 v 5 v 6 0.98)

If is minimized, the least squares criterion is still satisfied because both

subtracted terms in are equal to zero. Differentiating with respect to
each residual, and setting the derivatives equal to zero, we obtain
'
2v1 2k1 0 or
v1

v1 k1

'
2v 2 2k1 0 or
v 2

v 2 k1

'
2v 3 2k1 2k 2 0 or
v 3

v 3 k1 k 2

B l1 A B

D l2 B D

D l3 A D

B l4C B

D l5C D

l1 v1 281 .130 0
l2 v2 B

l 3 v 3 281 .130 0
l4 v4 C

l5 v5 C

A l 6 C 281 .130 l 6 v 6 C

Given observed values and rearranging, the condition equations become

v1=269.157-B
v2=B-D-10.940
v3=258.198-D
v4=C-B-21.040
v5=C-D-31.891
v6=C-290.113
the leas squares criterion, the quantity to be minimized is

v12 v 22 v32 v 42 v52 v 62 2k1 (v1 v 2 v3 v 4 0.70 )

( 269 .157 B ) 2 ( B D 10 .940 ) 2 ( 258 .198 D ) 2 (C B 21 .040 ) 2 (C D 31 .891) 2
(C 290 .113) 2

mplished by taking partial derivates of with respect to the unknown B, C

and D and equating these derivates to zero. Thus

2( 269.157 B ) 2( B D 10.940) 2(C B 21.040) 0

B

2C B 21.040) 2(C D 31.891) 2(C 290.113) 0

C

2( B D 10.940) 2( 258.198 D ) 2(C D 31.891) 0

D

and collecting terms, we get:

3B C D=259.057
(a)
B + 3 C D=343.044
(b)
B C + 3D=215.367
(c)
Are the normal equations, as will be explained n Chapter 4. to
solve equations, we first eliminate C. thus, from (a)
C=3B - D - 259.057,

(d)

which when substituted into (b) and (c) gives

-B+3 (3B - D - 259.057) - D=343,044
and
-B- (3B - D - 259.057)+30=215.367.
These two equations reduce to
- 2B - D= 200.0538
- Be + D= - 10.9225.

(e)
(f)

Eliminating D by adding (e) and (f), we get

B=269.1313 m.
Substituting for B in (f) we obtain
D= 269.1313 - 10,9225 = 258 2088 m.
Finally, substituting for B and D in (d), we get
C=3 (269.1313) - 258.2088 - 259.057=290.1281 m.
Thus, the least squares estimates for the elevations of the three points
are:
Elevation of B=269.131 m
Elevation of C=290.128 m
Elevation of D= 258209 m.
PROBLEMS
3-1 All angles marked by (*) in the quadrilateral shown in Fig. 3-8 are
measured. Evaluate the redundancy if the shape of the quadrilateral constitutes
the mathematical model.
3-2 In addition to the marked angles all sides of the quadrilateral in Fig. 3-8 are
measured. Evaluate the redundancy if the size and shape of the quadrilateral
constitute the mathematical model.
3-3 All sides shown in Fig. 3-9 are measured with EDM equipment, and the
angles marked by (*) are measured with a theodolite. Evaluate the redundancy
if the size and shape of the figure constitute the mathematical model.

Fig 3-9
3-4 All sides of the trilateration network shown in Fig. 3-10 are measured with
an EDM device. No angles are measured. Evaluate the redundancy if the size
and shape of the network constitute the mathematical model.
3-5 The parabola y= ax2+ b x - c is to be fitted to 14 data points. The x-coordinate
of each point is an error-free constant, they-coordinate is an observed quantity,
and a, b. and c are unknown parameters. Evaluate the redundancy.
3-6 The interior angles of a four-sided closed traverse (quadrilateral) are
measured and found to be: 1 = 11000'20'' 2=9002'15'' 3= 8005'25", and
4=7952'40". Adjust these angles according to the principle of least squares,
assuming all observations are uncorrelated and have the same precision. What
is the equivalent simple adjustment: technique?

Fig 3-10

3-7 The distances shown in Fig. 3-11 are measured. All measurements
are uncorrelated and have the same precision. The measured values are:
l1 = 100.010 m, l2 = 200.050 m, l3 =200.070 m, and l 4 = 300.090 m. Use
the principle of least squares to find the adjusted distance between A and
C.

3-8 The angles about a survey station, Fig. 3-12, are measured. The
observed values are: a=6000'00', b=6000'00', c=2400025 and
d=12000'05'. All measurements are uncorrelated and have the same
precision. If these observed angles are adjusted in accordance with the
least squares principle, what are the adjusted values of angles a and b?
3-9 With reference to Fig. 3-13, the following angles are measured.
L 1= Angle AOB = 3000'20"
L 2=Angle AOC= 5000'00"
L 3=Angle COD = 2000'00"
L 4=Angle ROD = 4000'20"
The measurements are uncorrelated and have the same precision. Find
the adjusted values for the angles in accordance with the principle of least
squares.

fig 3-12
fig -13
3-10 Use the principle of least squares to estimate the parameters of the
straight line, y =ax+b that fits the following data
x
1
2
3
4
5

y
9.50
8.85
8.05
7.50
7.15

Assume the x-coordinates are eror-free constants and the y-coordinates are
uncorrelated observations with equal precision.
3-11 Figure 3-14 shows a level net connecting three bench marks, A, B, and

C. The arrows indicate the directions of higher elevation. Tne observed

differences in elevation are: l1=20.410 m, l2=10.100 m, l 3=10.300 m, and l4=
10.315 m. All observations are uncorrelated and have equal precision. Use
the principle of least squares to find adjusted value for the four elevation
differences.
3-12 The angles shown in Fig. 3-15 one measured: 1= 4000'00",
2=10000'30',
3=5000'20'',
4=12000'00'',
5=4000'20".
All
measurements are uncorrelated and have the same precision. The angle at
A is held fixed at 9000 '00''. Adjust the measured angles according to the
principle of least squares.

fig 3-14

fig 3-15

Least Squares
Adjustment

4.1. TECHNIQUES OF LEAST SQUARES

There are several techniques for least squares adjustment, all of which yield
identical results when applied to the same problem. Some of these techniques
are simpler than others and are therefore suited for classes of survey
adjustment problems that do not require the application of the more general
techniques. We have mentioned in Section 3.3 of the preceding chapter two
relatively simple techniques which are extensively used in survey adjustment.
These are (a) adjustment of indirect observations and (b) adjustment of
observations only. The first technique, adjustment of indirect observations, has
the following characteristics:
1. The condition equations include both observations and parameters (and of
course constants whenever necessary).
2. The number of condition equations, c, is exactly the same as the number of
observations or measurements, n.
3. Each condition equation contains only one observation with a unit coefficient.
An example of the condition equations for this technique is Eqs. (3-9), for the
case of measuring a distance twice, which may be rewritten here as

l1 v1 x 1 0

(4-1)

l2 v2 x2 0

Replacing x by the more commonly used symbol for each parameter used
and rearranging Eq. (4-1) gives
v1 l1

(4-2)

v 2 l 2

We can collect the residuals in one column vector v, and the observations in
another vector l, i.e.,
v1
v
v

and

l1
l
l 2

and factor out the coefficients of in both equations into one coefficient
matrix

 1
B
1

Then Eq. (4-2) in matrix form- becomes

v1 1
l1
v 1 l
2
2

(4-3)

Or
V=B=f

(4-4)

in which we use the more general term f in place of -l,

Equation (4-4) is the general form of the condition equations for the
technique of least squares adjustment of indirect observations. In practice:
there are n condition equations and u unknown parameters (which is equal to
n 0 the minimum number of observations necessary to uniquely determine the
model). This will make the matrix B an n x u matrix, the vector a u x I
column matrix, and both the v and f vectors n x 1 column matrices. Thus, he
longhand form of the condition equations is
v1 b11 1 b12 2 b1u u d 1 l1 f 1
v 2 b21 1 212 2 b2 u u d 2 l 2 f 2

(4-5)

v n bn1 1 bn 2 2 bnu u d n l n f n

Where
v1 , v 2 , , v n
b1 , b2 , , bn

are the residuals for the n observations;

are the numerical coefficients of the unknown parameters;
1 , 2 , , n are the a unknown parameters;
d 1 , d 2 , , d n , are numerical constants, which may or may not be zero, and
exist when the condition equations are directly linear;
are the numerical values of then observations;
l1 , l 2 , , l n
f1 , f 2 , , f n are the numerical constant terms on the right-hand side of the
equations (used whether the conditions are originally linear or
are linearized).
Equation (4-5) in matrix form becomes
j1 2 b1 u 1 d 1 l1 f 1
v1 b11

d2
l2
f2
b
j

2
22
2u 2
21
(4-6)

v m bm 1 bm 2 bm u u d n l n f n
and in a more concise form becomes identical to Eq. (4-4), or

v B d l f

(4-7)
n,1 n, u u ,1 n,1 n,1 n,1
The following are examples of the matrix form of condition equations for
least squares adjustment of indirect observations.
EXAMPLE 4-1
We recall the data for the problem in Example 3-5 of fitting a straight line of
the form y - ax - b=0 to three points:
POINT
1
2
3

X (cm)
2
4
6

Y (cm)
3,2
4
5

Write the condition equations in the matrix form v + B = f, indicating the

dimensions of the matrices.
Solution
Substituting the coordinates of the points into the equation of the line gives
the following three conditions:
v1 y1 ax1 b 0

or

v1 ax1 b y1

v 2 y 2 ax 2 b 0

or

v 2 ax 2 b y 2

v3 y 3 ax 3 b 0

or

v3 ax 3 b y 3

Now using the numerical values and collecting in matrices gives

v1 2 1
3 .2
v 4 1 a 4.0
2
b

v 2 6 1 5.0

Or
v B f

with corresponding values above. The vectors v and f are each 3 x I (where
3=n= number of observations), B is 3x2, and is 2x I, containing the two
unknown parameters a (the slope of the line) and b (its y-intercept). Note
that in this example f=-l because the numerical vector d is equal to zero.
EXAMPLE 4-2
Reference is made to the level net in Fig. 3-7 and the data given in Example
3-7. It is required to give the matrix form, including the dimensions of the
constituent matrices, of the condition equations for the adjustment of the
level net by the method of indirect observations.

Solution

Since n = 6 and n 0 = 3 (see Example 3-7), there are six condition equations
including three unknown parameters. To conform to the general symbols, we
will designate the elevations of points B, C, and D by 1, 2 and 3
respectively. Again with reference to Example 3-7, the six condition
equations are
B l1 v1 A 0 or

v1 1

A l1 269 .157 f 1

D l 2 v 2 B 0 or

v 2 1 3 l 2

D l 3 v 3 A 0 or

v3 3

B l 4 v 4 C 0 or

v 4 1 2 l 4

21 .040 f 4

D l 5 v 5 C 0 or

v 5 2 3 l 5

31 .891 f 5

A l 6 v 6 C 0 or

v6 2

10 .940 f 2

A l 3 258 .198 f 3

A l 6 290 .113 f 6

The only unknowns are the parameters, which can be placed in a column
vector
1
2
3

so that the matrix form of the conditions becomes

0
v1 1 0
281 .130 11 .973 269 .157
v

10 .940 10 .940
2 1 0 1 0

v 3 0 0 1 1 281 .130 22 .932 258 .198

2

0 0
21 .040 21 .040
v 4 1 1

v 0 1 1 3 0
31 .891 31 .891
5

0
281 .130 8.983 290 .113
v 6 0 1

which is of the form of Eq. (4.7), i.e., v B d l f The vector v is 6x 1, B is

6 x 3, is 3 x 1, d is 6 x 1,l is 6 x 1, and f is 6 x 1. Note that in this problem,
where the conditions are directly linear, the d vector is not zero, unlike the
preceding example. This demonstrates the statement that d may or may not be
zero depending upon the problem at hand,
The second technique of least squares adjustment is called adjustment of
observations only. As its name implies, there are no parameters included in
the condition equations. Consequently, he number of conditions, c, is equal to
the number of redundant measurements, r. Equation (3-10) is an example of
this type of condition for two measurements of a distance. It maybe rewritten
here [see Eq. (3-9), from which (3-I0) is derived] as
l1 v1 l 2 v 2 0

Or
v1 v 2 l1 l 2 0

In matrix form this becomes

v1
1 ( l1 l 2 ) f
v 2

1
which in general is

Av = f

(4-8)

Since there are r condition equations and n observations in general, then A is

an r x n it matrix, v an n x I vector and f also an n x 1 vector. In longhand
notation, the equations are
a11 v1 a12 v 2 a1n v n f 1
a 21 v1 a 22 v 2 a 2 n v n f 2

(4-9)

a r1v1 a r 2 v 2 a rn v n f r

where

v1 , v 2 , , v n
a11 , a12 , , a1n

f1 , f 2 ,, f n

are the residuals for the n observations;

are the numerical coefficients of the unknown residuals;
are the numerical constant terms on the right-hand side of the
equations.

In matrix form Eq. (4-9) becomes

a11 a12 a1n v1 f 1

a11 a12 a1n v 2 f 2

a r1 a r 2 a rn v n f n

(4-10)

which corresponds directly :o the compact form of Eq. (4-8). When the
conditions are originally linear the vector f is usually written in terms of the
given observations as
f=dAl

(4-11)

where d, as before, is a v e c t o r o f numerical constants that may or may not

be zero.
The following are examples of the matrix form of conditions for least squares
adjustment of observations only.
EXAMPLE 4-3
Consider data for three interior angles of a plane triangle:
1= 4133 '45, 2= 7857'55'', 3= 592750',
Write the condition equation in matrix form,
Solution
Since n= 3, n0= 2, r=1, there is one condition; namely,
1+v1+ 2+v2 +3+v3=180,
which may be rearranged as
v1+ v2+ v3=180- 1- 2- 3,

or
v1
1

1 1 1v 2 180 1 1 1 2 f 30' '

v3
3

which is in the matrix form

Av=dA =f.
The matrix dimensions are: A is 1 x 3, v is 3 x 1, d is 1 x 1, is 3 x 1, and f is
1 x 1 In this example, d is not zero, but equal to l80.
EXAMPLE 44
Referring to Fig. 3-7 and Example 3-7, we repeat here the observational
data for the level net: l1 =11.973 m (B to A); l2= 10.940 m (D to B);
l3=22.932 m (D to A); l4= 21.040 m (B to C); l5= 31.391 m (D to C); and
l6=8.953 m (A to C). The observations are given as being front the low point
to the high point. Point A is a bench mark with elevation 281.130 m. Write
the matrix form of the conditions for least squares adjustment of
observations only.

Solution
Since n = 6, and it would take a minimum of 3 measured differences in
elevation to determine the elevations of points B, C, and D (i.e., n0= 3),
then the redundancy is r= 6 3 = 3. There must be then three condition
equations that can be written between the observations. In order to write
these, we refer to Fig. 3-7 and realize that if we go around any loop, starting
and closing on the same point, the adjusted differences in elevation in the
loop must add up to zero. Thus,
Loop B-A-D-B
Loop D-B-C-D
Loop D-A-C-D

l1 v1 l3 v3 l 2 v 2 0
l 2 v 2 l 4 v 4 l 5 v5 0
l 3 v3 l 6 v 6 l 5 v5 0

Rearranging, we get
v1 v 2 v 3 (l1 l 2 l 3 )
v 2 v 4 v 5 (l 2 l 4 l 5 )
v 3 v 5 v 6 (l 3 l 5 l 6 )

which, in matrix form, is

1 1 1
0 1
0

0 1
0

0
1 1
0

0 1 1
0

v1
v
2 1 1 1
v 3
0
0 1
v 4 0
0 1
v
5

v 6

0
1 1
0

0 1 1
0

l1
l
2
l 3

l 4
l
5
l 6

I.e., Av = -Al where A is 3 x 6, v is 6 x 1, and l is 6 x 1. Referring to Eq. (4-8),

we see that in this example f= -Al, which means that the vector, d, as given in
Eq. (4-11), must be zero. Since the elements of l are the observed data, f can be
evaluated as the following 3 x 1 vector:
0.019
f 0.089 m

0.024

Note that the elements of f are the so-called closing errors of the individual loops
of the level network.

4.2 THE CONCEPT OF WEIGHT

In Chapter 1 the terms standard deviation (), variance (2), and precision of a
measurement were introduced. We said that a measurement of high
precision has a small variance because repeated values would cluster closely
together, thus reflecting a high degree of procedural care and instrument
Refinement. Conversely, a low precision measurement has a large variance.
Since the value of the variance goes in opposite direction to that of the
precision, another measure of precision that is directly related to it is often
used. This measure is called the weight of an observation. Thus, for any given
measurement, or observation, the higher the weight the higher is the
precision, and vice versa. Accordingly the weight to of a single observation is
defined as a quantity that is inversely proportional to the observation 's
variance, 2, i.e.,
w= k / 2

(4-12)

where k is a constant of proportionality. If an observation has unit weight

(w=1), its variance is defined by the symbol 02 Thus,
1=k / 02 ,
From which
k= 02

(4-13)

and therefore
w 02 / 2

(4-14)

Thus, the proportionality constant is the variance 02 of an observation of unit

weight. This variance is referred to by several names, such as variance
factor, variance of unit weight, and reference variance. In the remaining
discussion of this book we will adopt the term reference variance for 02 .
When two or more observations are jointly involved (such as coordinates of
the same point), in addition to their variances there are other quantities,
known as covariances, that express the interaction, or interdependence,
between them. This interaction is called correlation. For the purposes of this
chapter we will assume all observations to be uncorrelated, i.e., all
covariances are zero.
For uncorrelated measurements, x1, x2 ..., xm, with variances 12 , 22 ,..., m2 of,
respectively, the variance matrix is defined as the following diagonal matrix:
12 0 0

0 22 0

0 0 m2

(4-15)

The corresponding weights of these uncorrelated measurements are,

according to Eq. (4-14),
w1 02 / 12 , w2 02 / 22 wm 02 / m2

(4-16)

These weights may be collected into a corresponding diagonal matrix W,

called the weight matrix:
w1
0 0

0 w2 0

(4-17)
W

0 0 wm
which in view of Eq. (4-16) becomes
02 / 12
1 / 12

0
0
0
0

0
02 / 22
0
1 / 22
0

2 0
W
0

0
0

0 1 / m2
0 02 / m2

(4-18)

When the reference variance 02 is factored out in Eq. (4-18) the remaining
matrix is the inverse of the variance matrix . (The inverse of a diagonal
matrix is another diagonal matrix, each element of which is the reciprocal of
the corresponding element in the original matrix; see Appendix A.) Hence,
Eq. (4-18) becomes
W 02 1

(4-19)

[The relationship in Eq. (4-19) is actually general and applies equally to

correlated as well as uncorrelated observations, as will be explained later.]
Uncorrelated weighed observations (i.e., uncorrelated observations with
different weights, or with unequal precision) occur frequently in surveying
practice. This is particularly the case with mixed kinds of measurements,
such as distances and angles, for the variance of a measured distance in
meters is quite different form the variance of an angle expressed in radians.
Consequently, when these two variances are placed together in a variance
matrix and the corresponding weight matrix is calculated, the two weight
values will also be different Of course, measurements of the same kind,
such as distances only or angles only, could be uncorrelated but each of
different precision or weight simply because of different observers and/or
different equipment.
The succeeding sections of this chapter will deal with least squares adjustment when observations are uncorrelated but may have varying weights.
The more general case of correlated observations will be treated in Chapter
9 after further concepts in probability, statistics, and propagation are
introduced.

4.3. LEAST SQUARES ADJUSTMENT OF INDIRECT OBSERVATIONS

We have seen in Section 4.1 how the condition equations may be written
for least squares adjustment of indirect observations. In order to get a
solution for the adjustment problem, the least squares criterion must now be
imposed. For uncorrelated observations with equal precision, the criterion
calls for minimization of the function
n

v12 v 22 v n2 vi2

(4-20)

i 1

as has already been given in Chapter 3 [see Section 3.3; Eq. (4-20) is
identical to Eq. (3-8)]. For uncorrelated observations of unequal precision,
the criterion calls for minimization of the weighted function
n

w1v12 w1v 22 wn v n2 wi vi2

(4-21)

i 1

in which w1, w2,...wn are the weights of the corresponding observations, l1,
l2,,ln respectively.
In order to facilitate understanding of the matrix derivation to follow, we shall
recall Example 4-2 on the level net and proceed to work it out gradually in
matrix form. We will assume here that the six measured differences in
elevations are still uncorrelated, but now have unequal precision, i.e., different
weight. The six weights are designated by w1, w2,...w6 Thus, the function to
be minimized for this example is

w1v12 w1v 22 w3 v32 w4 v 42 w5 v52 w6 v 62

Substituting in this equation the values of the residuals from Example 4 -2
gives

w1 ( f 1 1 ) 2 w2 ( f 2 1 3 ) 2 w3 ( f 3 3 ) 2 w4 ( f 4 1 2 ) 2
w5 ( f 5 2 3 ) 2 w 6 ( f 6 2 ) 2

(4-22)

In order for in Eq. (4-22) to be a minimum, its partial derivative with

respect to each unknown parameter 1 , 2 , 3 must be made equal to zero.
Hence,

2 w1 ( f 1 1 ) 2 w2 ( f 2 1 3 ) 2 w4 ( f 4 1 2 ) 0
1

2 w4 ( f 4 1 2 ) 2 w5 ( f 5 2 3 ) 2 w6 ( f 6 2 ) 0
2

2 w 2 ( f 2 1 3 ) 2 w 3 ( f 3 3 ) 2 w5 ( f 5 2 3 ) 0
3

Clearing and collecting terms in on the left-hand side and terms in f on the
right-hand side leads to
( w1 w2 w4 ) 1 ( w4 ) 2 ( w2 ) 3 w1 f 1 w2 f 2 w4 f 4
( w4 ) 1 ( w4 w5 w6 ) 2 ( w5 ) 3 w4 f 4 w5 f 5 w6 f 6

(4-23)

( w2 ) 1 ( w5 ) 2 ( w2 w3 w5 ) 3 w2 f 2 w3 f 3 w5 f 5

This is the set of normal equations, which are equal in number to the unknown
parameters (in this case u = 3). The normal equations make a unique solution
possible because the unknown parameters are the only unknowns in them. It
should be noted as well that the condition equations are six in number but
contain nine unknowns (three parameters and six residuals), and that by adding
the three normal equations to them toe get a total of nine independent equations
in nine unknowns, hence a unique solution.
Now, if we recall the B, , and f matrices from Example 4-2, namely,
0
0
f1
1
f
1

0 1
2

1
f3
0

0 1
B
, 2 , f
0
f4
1 1

f
3
0 1 1
5

0
0 1
f 6

And if we collect the six weights into a diagonal weight matrix W, i.e.,
w1
0 0

0 w2 0

0 0 w6
We find that we are able to construct the normal equations, Eq. (4-23), form the
following relationship
(BtWB) =BtWf,
That is,
0
0
1

0 1
w1 0 0 1
1
0 1
0 0
1 1

0
w

0
0
0

1
2
0

0
0 1 1 1
2

1 1
0

0 1 1
0 1
0

3
0

1
0
0

w
6

0
0 1

(4-24)

 f1

w1 0 0 f 2
0 1
0 0
1 1
0 w2 0 f 3

0
0
0 1 1 1

f
4
0 1 1
0 1
0

0
0

w
6 f5

f6
Multiplying the matrices out gives

(4-25)

w4
w2
( w1 w2 w3 )
1

w4
( w4 w5 w6 )
w5

w2
w3
( w2 w3 w5 ) 3

(4-26)
w1 f 1 w2 f 2 w4 f 4
w4 f 4 w5 f 5 w6 f 6
w2 f 2 w3 f 3 w5 f 5

which is obviously equivalent to Eq. (4-23). Equation (4-24) can be written

more concisely as
N =t
, in

(4-27)

which
N=BtWB

(4-28)

and
t =BtWf

(4-29)

N is the coefficient matrix of the normal equations. It is square and sym metric, as illustrated in Eq. (4-25). t is the vector of constant terms.
Equations (4-24), (4-27), (4-28), and (4-29) are all quite general, although we
have arrived at them by way of an example. They apply to any problem for
which the condition equations are or the form in Eq. (4-7). This is shown to be
true by the general derivation which now follows. As already stated, the least
squares criterion for uncorrelated observations with unequal precision is
imposed by minimizing the function in Eq. (4-21). This function can be
written in matrix form as

 w1 0 0 v1
0 w 0
2
v 2
v1 , v 2 , , v n

0 0 w6 f n

(4-30)

Since when multiplied out gives

v1
v
w1v1 w1v 2 wn v n 2 w1v12 w1v 22 wn v n2


v n
Which is the same as Eq (4-21). Since we defined the residual vector v as an
n x 1 column matrix, Eq (4-30) is concisely written as

v tWv

(4-31)

Now, recalling Eq (4-7)

v B f

And rearranging it as

v f B

(4-32)

And substituting for y into Eq (4-31) we get

( f B ) t W ( f B )
( f t B t t ) W ( f B )
( f tW B t tW ) ( f B )

Or

( f tWf t B tWf f tWB t B tWB )

(4-33)

Since is a scalar, each one of the four terms on the right-hand side of Eq.
( 4 - 3 3 ) is also a scalar. Furthermore, since the transpose of a scala r is equal to
itself, then the second and third terms on the right-hand side of Eq. ( 4 - 3 3 ) are
equal, i.e.,
t B tWf (t B tWf ) t f tWB

(4-34)

In Eq. (4-34), W t is replaced by W because the weight matrix is always symmetric and therefore equal to its transpose. Combining the-second and third
terms on the right-hand side of Eq. (4-33) yields

f tWf 2 f tWB t ( B tWB )

(4-35)

In Eq. (4-35) all matrices and vectors are numerical constants, except the
vector of unknowns. Thus, in order for to be a minimum, its partial derivative
with respect to must be equated to zero. Since the first term on the right-hand
side of Eq. (4-35) does not contain , its partial derivative is automatically zero.
The partial derivatives of the second and third terms are obtained by applying
the rules of differentiation for bilinear and quadratic forms, respectively, as given
in Appendix A. Thus, differentiating with respect to and equating the result
to zero, we have:

2 f tWB 2t ( B tWB ) 0

(4-36)

Rearranging, dividing by 2 and transposing Eq. (4-36) yields

( Bt

B)

u , n n, n n, u u ,1

Bt

u , n n, n n,1

(4-37)

which is identical to Eq. (4-24). Us concise form is given by Eq. (4-27), in which
N is a square symmetric matrix of order u and tins u x 1vector. The solution of
Eq. (4-27) is
=N-1t

(4-38)

which yields the vector of parameters. This solution requires the inversion of the
u x u normal equations coefficient matrix, N.
EXAMPLE 4-5
Recall the four measurements of distance given in Example 3-2: l1 = 32.51 m, l2=
32.46 m, l3= 32.52 to, and l4= 32., 53 m. Compute the least squares estimate of the
distance: (1) if all measurements are. uncorrelated and of equal precision; and
(2) if they are uncorrelated but have the following weights: w1 =1, w2 = 2, w3 = 1
and w4= 0.5
Solution
The number of observations is m =4, and with the minimum required to fix the
model n0 = 1, the redundancy is r= 4 - 1 = 3. Carrying the final estimate as a
parameter, i.e., u= 1, there are four condition equations, one for each
observation. Letting represent the estimate of the distance, the condition
equations are

l1 v1 x or

l 2 v2 x or

l 3 v3 x or

l 4 v4 x or

v1 x l1 32.51

v 2 x l 2 32.48

v 3 x l 3 32.52

v 4 x l 4 32.53

In matrix form these equations are (see Eq. 4-6)

v1 1
v
2 1
v3 1

v 4 1

32.51
32.48

32.52

32.53

which is of the form

v B
4,1

4,1

1,1

f
4,1

The normal equations coefficient matrix N is given by Eq. (4-28) as

N = Bt WB
(1) When the observations are assumed uncorrelated and of equal weight,
the weight matrix reduces to the identity matrix, or W = I, and N then
becomes
1
1
N B t B 1,1,1 1 4
1

1

The normal equations constant terms vector is, by Eq. (4-29),

32.51
32.48
t
t
130.04
t B Wf B f 1,1,1,1
32.52

32.53
Finally, the parameter is obtained by applying Eq. (4-38), i.e.,

x N 1t (4) 1 (130.04) / 4 32.51m

which agrees with the answer obtained in Example 3-2.

(2) When the observations have unequal weights 1, 2, 1, and 0.5,

respectively, the weight matrix becomes W
1
0
W
0

0 0 0
2 0 0

0 1 0

0 0 0 .5

and
1
1
2
1

N B t WB 1,1,1,1

1
1

0.5 1

1
1
1,2,1,0.5 4.5
1

1
4

It can be seen that N turns out to be the sum of the weights, or

i 1

Next
32.51
32.48
t
4.5
t B Wf 1,2,1,0.5
32.52

32.53
(1)(32.51) ( 2)(32.48) (1)(32.52) (0.5)(32.53)
146.255

Which is equal to the sum of the products of each observation and its
4

weight or

wl

i i

. The least squares estimate of the distance in this case is

i 1

x N 1t (4.5) 1 (146.255) 146.255 / 4.5 32.50m

This answer is obviously different from that obtained when the

measurements are of equal weight. It can be seen that this estimate
(32.50m) is closer to the second observation (l1= 32.48 m) than the first
estimate (32.51 m) because l2 has a higher weight. Similarly, l4 has less
influence on the new estimate because it has less weight (w4 = 0.5).
In the, second part of the preceding example, the least squares estimate
of a quantity from n uncorrelated direct measurements of different weight
turns out to be

x wi l i
i 1

which is the weighted mean of the given

wi ,
i 1
observations.

(4-39)

EXAMPLE 4-6
Consider the data for Examples 3-5 and 4-1 I n which a straight line with the
equation y= ax + b is fitted to three points: x 1=2, y1 = =3.2; x2 =4, y2 = 4.0;
x3 = 6, y3 = 5.0. All coordinates are in centimeters. The y-coordinates are the
observations, which are assumed to be (1) uncorrelated and of equal
precision, and (2) uncorrelated with variances 0.10 cm 2, 0.08 cm2, and 0.08

cm2, respectively. Compute the least squares estimates a, b of the two

parameters for each of the two cases.
Solution
The three condition equations corresponding to the three points are (see
also Example 4-1)
v1 y1 ax1 b 0

v1 2a b y1 3.2

or

v 2 y 2 ax 2 b 0

or

v 2 4 a b y 2 4 .0

v3 y 3 ax 3 b 0

or

v 3 6 a b y 3 5 .0

and in. the form v+ B= f they become

v1 2 1
3 .2
v 4 1 a 4.0
2
b

v 2 6 1 5.0
(1) The first case is when the observations are uncorrelated and of equal
precision, for which W= I. The normal equations coefficient matrix is in
this case [see Eq (4-28)
2 1

4 1 56 12
N B tWB B t B

12 3
1 1 1

And the vector t [Eq. (4-29)] is

3 .2
2 4 6
52 .4
t
t
t B Wf B
4

12 .2
1 1 1

5
is the inverse of N is

56 12

12 3

3 12
1 / 24

12 56

and according to Eq. (4-38) the parameter vector is calculated as

a N 1t 1 / 24 3 12 52 .4
12 56 12 .2

i.e.,

a [(3)(52.4) ( 12)(12.2)] / 24 0.450

b [( 12)(52.4) (56)(12.2)] / 24 2.267

which are identical (except for rounding) to the answers obtained for the same
case in Example 3-5.
(2) For the second case, we need to construct the weight matrix from the given
variances. Since the measurements are uncorrelated, the weight matrix W
may be calculated according to Eq. (4-18). Thus, we first begin by computing
1
the inverse
of the variance matrix, or
1 / 12 0
0 1 / 0.10

2
1 / 0.08
0 1/ 2 0

0
2

1
/
0
.
08
0
1
/

10

12.5

12.5

Next, we select a value for the reference variance 12 usually to make the
numerical values of the weights smaller and more convenient. Thus, if 12 =0.10,
tne weight matrix becomes [Eq. (4-10)]

W (0.10 )

10

12 .5

12 .5

With this matrix, the corresponding N and t are

1

2 4 6

t
N B WB
1.25

1.25

2 1
4 1

6 1

2 1
5
7 .5
2
69 14 .5

4 1

1 1.25 1.25
14 .5 3.5

And
3 .2

7
.
5

4.0 63 .9
t B tWf

14 .45
1 1.25 1.25

5
.
0

The inverse of N is:

69 14 .5

14 .5 3.5

3.5 14 .5
1 / 31 .25

14 .5 69

and, finally,
3.5 14 .5 63 .9 0.452
N 1t 1 / 31 .25

14 .5 69 14 .45 2.256

Thus, in the case of unequal weights, a = 0.452 and b =2.256.

Discussion
The effect of increasing the weights for points 2 and 3 relative to point 1 is
that the determined straight line passed closer to points 2 and 3 than when
all points are of equal weight. This can be ascertained by calculating the
residuals for both cases and showing that the magnitudes of v2 and v3 are
smaller in the unequal weight case than in the equal weight case. The
residuals are computed by rearranging the condition equations and

substituting the least squares estimates a, b for the parameters, i.e.,

v f B
Thus for the equal weight case (e is used as a designator)
3 .2 2 1
0.033
0
.
450

v r 4 .0 4 1
0.067 cm

2.267

5.0 6 1

0
.
033

and for the unequal weight case (u is used as a designator)

3 .2 2 1
0.040
0.452

v m 4 .0 4 1
0.064 cm

2.256

5.0 6 1

0
.
032

It is seen that the magnitude of v 1m is larger than that of v 1r while the

magnitudes of v2m v2m are smaller than those of v 2r, v3r Thus, because points
2 and 3 have relatively larger weights than point 1, the line will fit closer to
them and farther from point 1.
EXAMPLE 4-7
Solve the level net problem of Examples 3-7 and 4-2 using the procedure of

adjustment of indirect observations for two cases, (1) when the observed
differences in elevation are uncorrelated and have equal precision, and (2) when
the weight of each observation is inversely proportional to the leveled distance
(see Fig. 3-7).
Solution
Recall from Example 4-2 the matrices for the condition equations v+ B= f.
0
1 0
1 0 1

0 0 1
B
and
0
1 1
0 1 1

0
0 1

269 .157
10 .940

258 .198
f

21 .040
31 .891

290 .113

(1) For the first case, the weight matrix W of the observations is the identity
matrix because the observations are uncorrelated and have equal precision.
Consequently, the N and t matrices are [see Eqs. (4-28) and (4-29)]
0
0
1
1
0 1

3 1 1

1
0

1
0
0

0
0 1

t
NB B 0
0
0 1 1 1
3 1
1

1 1

0
0 1 1
1 1
0 1
0
3
0 1 1

0
0 1
And
269 .157
10 .940
259 .057
0 1
0 0
1 1

258 .198
t Bt f 0
0
0 1 1 1
343 .044

21 .040

215 .367
0 1 1
0 1
0

31 .891

290 .113

The inverse of N is

And

3 1 1
2 1 1

1
3 1 1 / 4 1 2 1

1 1
1 1 2
3

 2 1 1 259 .057 269 .131

N 1t 1 / 4 1 2 1 343 .044 290 .128 m

1 1 2 215 .367 258 .209

which are the elevations of points B, C, and D when the observed differences
in elevation are assumed equal in weight.
(3) The weight of each difference in elevation is proportional to the
reciprocal of the leveled distance between the two points. These
distances are given below.
Observation

l1

l2

l3

l4

l5

l6

Distance (km)

20

12

15

28

20

26

Reciprocal of
distance

0.057

0.083 0.057

0.036

0.050 0.039

Weight

1.400

2.333 1.867

l.000

1.400 1.077

The weights are such that the smallest value (0.036) in the third line is given
a weight value of 1.0 and the rest are proportionately computed. Thus the
weight matrix is
1.400

2.333

1.867
W

1.000

1.400

1.077

With this matriz, the corresponding normal equation matrices are

0
0
1.400
1

2.333
1
0 1

0 1
0 0
1 1

1
.
867
0
0

1
N B t WB 0
0
0 1 1 1

1.000
1 1
0

0 1 1

0 1
0

0 1 1
1.400

1.077 0 1
0

4.733 1.000 2.333

1.000
3.477 1.400

2.333 1.400
5.600

And

1.400
269 .157

10 .940
2.333

0 0
1 1 0 1

1
.
867
258
.
198
t B t Wf 0
0 0 1 1 1

1.000
21 .040

0 1 1 0 1 0
31 .891
1.400

1.077 290 .113

381 .3028
378 .1391

411 .8852

The inverse of N is

0.337902
0.171085

0.183543

0.171085
0.406418
0.172880

0.183543
0.172880

0.298256

The three parameters (i.e., elevations of points B, C, and D) are then

0.337902
N t 0.171085

0.183543
1

0.171085
0.406418
0.172880

0.183543 381 .3028

0.172880 378 .1391

0.298256 411 .8852

269 .135
290 .124

258 .205

EXAMPLE 4-8
A noon sight is made on the sun with a theodolite. The following data are
obtained:
TIME
40m 45.5s
46 18.5
50 03.0
56 53.0
h
12 05m 34.5s
12 12.0
16 04.5
11h

OBSERVED ALTITUDE
515109
5642
520030
0224
0106
515545
5130

For the small time range involved, the altitude is assumed to be a parabolic
function of time. Only the altitude is considered to be the observed variable; the
time is considered to be errorless, All observations are uncorrelated and equal
in precision,
Apply the method of least squares to fit a parabola to the given data and to

estimate the maximum altitude of the sun and the time at which maximum
altitude occurs.
Solution
Let T represent the time and h the altitude in a rectangular coordinate system,
the origin of which is at T0 = 12h 00m 00s, and n, = 5150'00", Thus, the data
points are:
T 19 m14 .5 s 1154 .5 s , h 69 ' '
T 13 m 41 .5 s 821 .5 s , h 402 ' '
T 9 m 57 .0 s 597 .0 s , h 630 ' '
T 3 m 07 .0 s 187 .0 s , h 744 ' '
T 5 m 34 .5 s 334 .5 s , h 666 ' '
T 12 m12 .0 s 732 .0 s , h 345 ' '
T 16 m 04 .5 s 964 .5 s , h 90 ' '

If the fitted parabola is of the form aT2 + bT- c = h, then the condition equations
are
a ( 1154 .5) 2 b( 1154 .5) 2 c 69 v1
a ( 821 .5) 2 b( 821 .5) 2 c 402 v 2
a ( 597 .0) 2 b( 597 .0) 2 c 630 v 3
a ( 187 .0) 2 b( 187 .0) 2 c 744 v 4
a (334 .5) 2 b(334 .5) 2 c 666 v 5
a (732 .0) 2 b(732 .0) 2 c 345 v 6
a (964 .5) 2 b(964 .5) 2 c 90 v 7

In matrix notation corresponding to the form v+ B= f we have

1,332 ,870 1154 .5
674 ,862 821 .5

356 ,409 597 .0

B 34 .969 187 .0
111,890 334 .5

535,824 732 .0
930 ,260 964 .5

And
a
b

c

1
1

1
1

1
1

The weight matrix is W=I, since all observations are uncorrelated and equal
in precision. Hence,
N = BtWB =BtB,
i.e,
3.5252 1012

N 9.8563 10 4

12
3.9771 10

9.8563 10 4
3.9771 10 4

729

3.9771 10 4

729

And
t=B tWf=B tf
i.e.,
9.570 10 4

t 3.630 10 4
2.946 10 4

The inverse of N is obtained by applying the matrix inversion procedure given

in Appendix A. The result is:

N 1

8.4611 10 13

1.2394 10 10
4.6782 10 7

1.2394 10 10
2.7449 10 7
4.1830 10 5

4.6782 10 7

4.1830 10 5
4.0429 10 1

The vector of least squares estimates of the parabola parameters is thus

6.1347 10 4

N 1t 1.0426 10 1 ,
0.7585 10 3

i.e.,

a 0.00061347

b 0.10426

c 758.5

Thus, the equation of the fitted parabola is:

-0.0061347 T2 0.10426 T + 758.5 = h

How well this equation fits the observed data is reflected in the vector of
residuals (values in seconds of arc):
69 1,332 ,870 1154 .5
402 674 ,862 821 .5

630 356 ,409 597 .0

v f B 744 34 .969 187 .0
666 111,890 334 .5

345 535,824 732 .0
90 930 ,260 964 .5

1
8

28
1

4
1 6.1347 10 28

1 1.0426 10 1 13
1 0.7585 10 3 11

1
8
3
1

Now, the maximum altitude occurs when dh / dT=0. Let the value of T at

which maximum altitude occurs be Tmax. Thus, dh dT 2 a t max b 0 i.e.,

Tmax

0.10426
85.0 s.
2( 0.00061347)

2a
Hence, the time at which maximum altitude occurs is:

T0 Tmax 12 h 00 m 00 s 0 h 01m 25 .0 s 11h 58 m 35.0 s

Now
2
hmax 0.00061347 Tmax
0.10426 Tmax 758 .5

0.00061347 ( 85 .0) 2 0.10426 ( 85 .0) 758 .5 763' '

Hence, the maximum altitude of the sun is

h0 hmax 5150'00' '012'43' ' 52 02'43' '

EXAMPLE 4-9
In Fig. 4-1, A and B are horizontal control points, spaced 100.000 m
(assumed errorless) apart. A third point, C, is to be located along a line
normal to AB, as shown. Two measurements are made:
distance AC:
angle at A:

l1=131.200m
l2= 4020'00",

The standard deviation of l1 is 3.035 en; the standard deviation of is l2 21".

The two observations are uncorrelated. Calculate the least squares estimate
of x, the distance between B and C.
Solution
Since it takes only one measurement to fix the model (the plane triangle in
Fig. 4-1), the redundancy is 1. Carrying one parameter, x, leads to two
conditions,

l1 v1 x 2 100 2 0

And
l 2 v 2 arctan

x
0
100

Both equations are nonlinear in the unknown x. It is therefore necessary to

linearize them according to the techniques of Chapter 2. Let
y1 l1 v1 x 2 100 2 0

And
y 2 l 2 v 2 arctan

x
0
100

According to Eq. (2-13), the linearized forms of the functions y 1 and y2 are
y1 y10

l1 v1 x 02 100 2

y1
x
x

x0
x

x 2 100 2 1
0

and

Fig 4-1
y 2 y 20

y 2
x
x

x
1

l 2 v 2 arctan 0
x
100 100[1 ( x 02 / 100 2 )]

In which x, is an approximate value for x, and x is the correction. Thus, the

linearized Condition equations are

l1 v1 x 02 100 2

x0
x 02 100 2

x 0

x0
100
2
x 0
100 x 0 / 100 2
Rearranging these equations and writing them in the form v+B=f as in Eq. (47), we have
l 2 v 2 arctan

x 2 100 2 l

v1 x 0 x 02 100 2
0
1

v
2 100 ( x 02 100 2 )
arctan( x 0 / 100) l 2

Since 100 tan (4020'00")=84.906, a good approximation for x is x0=85.000 m.

Using x0=85.000 and angular values expressed in radians, the first
approximations for matrices B and f are
0.647648
B1

0.005806

and

0.04405
f1

0.000545

Now 1=0.005 m, and . 2=21" =0.00010 radians. Letting 0=0.005, the

respective weights of l1 and l2 are
w1 1 and w2

(0.005 ) 2
2500
(0.00010 ) 2

and the weight matrix is

1 0

0 2500

The apparent disparity in the two weights arises not so much from any
imbalance in precision as from the different kinds of units used for expressing
distances (meters) and angles (radians). Thus,
N 1 B1tWB1 [0.503722]

And
t1 B1tWf 1 [0.036440]

And so
1 [x1 ] N 11t1

0.036440
0.072m
0.503722

Thus the first estimate for x is

x1 x0 x1 85.000 0.072 84.928m

Now, since during linearization we neglect all second and higher order terms, we

must ensure that x1 is not in significant error because of this approximation.

Consequently, we must Iterate the solution by using x1 as a new approximation
and computing another correction. This procedure is repeated until the
correction is insignificant.
Proceeding then with x1=84928 m, the new B and f matrices are
0.647330
B2

0.005810

and

0.00257
f2

0.000127

Thus
N 2 B2t WB 2 [0.503722]

And
t 2 B2t Wf 2 [0.000181]

And so
2 [x 2 ] N 21t 2

0.000181
0.00036m
0.503722

It is clear that x2 is not a significant correction since it is less than 0.001 m.

Hence the final least squares estimate is x = 84.928m.

A summary of the symbols and equations used for least squa res adjustment of
indirect observations will be found in section 9.5 of Chapter 9.
4.4. LEAST SQUARES ADJUSTMENT CF OBSERVATIONS ONLY
In Section 4.1 the form of the condition equations for least squares adjustment
of observations only is given by Eq. (4-8), and is repeated again as Eq. (4-40)
with matrix dimensions included:
A

r, n

n,1

f
r ,1

(4-40)

Dimension r is the redundancy, and dimension n is the number of given

observations. Since in Eq. (4-40) there are r equations in terms of n
unknown residuals, and since r=n-n 0 , and so is less than n, then no unique
solution can be obtained from the conditions alone. Additional equations
are obtained by applying the minimum criterion, i.e., minimization of Eq. (420) when the observations are uncorrelated and have the same precision,
or minimization of Eq. (4-21) when he observations are uncorrelated but
have unequal precision.
Again as in the preceding section, we shall introduce the derivation through
an example. For this purpose, we recall the level net of Example 4-4, where
n = 6 differences in elevation are measured. The three (r= 3) condition equations are
v1 v 2 v 3 (l1 l 2 l 3 ) f 1
v 2 v 4 v 5 (l 2 l 4 l 5 ) f 2
v 3 v 5 v 6 (l 3 l 5 l 6 ) f 3

(4-41)

e.i., in matrix notation

v1
v
2
0 0
0 f1
1 1 1
v 3
0 1
0
1 1
0 f 2
(4-42)

v
4

f
0
0 1
0 1 1
v 3
5
v 6
Assuming the six measurements are uncorrelated and have unequal weight,
function
w1v12 w1v 22 w3 v32 w4 v 42 w5 v52 w6 v 62
Must be minimized.
It can be seen that it is not possible to express separately each of the six
.residuals v 1, v2,..., v6 for substitution in , as was possible in the
procedure of adjustment of indirect observations, Therefore, another
approach is needed to make sure that is a minimum and, at the same
time, that the condition equations (4-41) are satisfied. This is accomplished
by rewriting each condition equation in normal form (zero on the right -hand
side), multiplying the left-hand sides by unspecified factors -2k1 and adding
the products to . The quantities ki are called Lagrange multipliers; there are
its many multipliers as there are conditions.
The new quantity to be minimized is then
w1v12 w2 v 22 w3 v 32 w4 v 42 w5 v52 w6 v 62 2 k1 (v1 v 2 v 3 f 1 )
(4-43)
2 k 2 (v 2 v 4 v 5 f 2 ) 2 k 3 (v 3 v 5 v 6 f 3 )
The factor -2 preceding each Lagrange multiplier is introduced for convenience only, to avoid unnecessary fractions and negative signs after dif ferentiation and separation of the residuals. In Eq. (4-43) the unknowns; are
the residuals v1 v 2 ..., v6 and Lagrange multipliers, k 1, k2 and k3. Therefore,
in order for ' to be a minimum, its partial derivative with respect to each of
these unknown variables must be zero.
Considering the residuals first,
'
1
2 w1v1 2k1 0 or v1
k1
v1
w1

'
2 w2 v 2 2( k1 k 2 ) 0 or
v 2

v2

'
2 w3 v 3 2( k1 k 3 ) 0 or
v 3
'
2 w4 v 4 2k 2 0 or
v 4

v3

v6

1
( k1 k 2 )
w3

1
k2
w4

v4

'
2 w5 v 5 2( k 2 k 3 ) 0 or
v 5
'
2 w6 v 6 2k 3 0 or
v 6

1
( k1 k 2 )
w2

v5

1
k3
w6

(4-44)
1
(k 2 k 3 )
w5

Next, differentiating ' with respect to k1, k2 and k3 and equating to zero,
gives

2(v1 v 2 v3 f 1 ) 0 or
k1

2k (v 2 v 4 v 5 f 2 ) 0 or
k 2

v1 v 2 v 3 f 1
v 2 v 4 v5 f 2

And

2(v 3 v 5 v 6 f 3 ) 0 or v 3 v 5 v 6 f 3
k 3
which are identical to the condition equations (4-41). Consequently, when '
is differentiated with respect to the Lagrange multipliers, the original
condition equations result; this demonstrates that the introduction of
Lagrange multipliers ensures that the conditions will be satisfied when is
minimized.
Equation (4-44) may be written in matrix notation as

0
0
v1 1 w1
1
v

1 w2
0
2
1 1
k1
v 3
1
1 w3
0 1
(4-45)

k 2
v
1
w
0

1
0
4
4

k
v
0 1 1 3
1
w
5
5

0 1
1 w6 0
v 6
The first matrix on the right-hand side of Ec. (4-45) is a diagonal matrix
representing the inverse of the weight matrix W of the observations. Such a
matrix is called a cofactor matrix and it Is given the symbol Q; thus

Q=W-1

(4-46)

The cofactor matrix represents the relative variances of the observations.

When the reference variance o is unity, Q is identical to the variance matrix .
This concept is discussed further in Chapter 5, particularly for correlated
observations.
The second matrix on the right-hand side of Eq. (4-45) can be seen to be the
transpose of the coefficient matrix A of the conditions in Eq. (4-42). Denoting
the vector of Lagrange multipliers by k, i.e.,
k1
k k 2
(4-47)
k 3
and recalling that v represents the vector of residuals, Eq. (4-45) may be
rewritten concisely as
v=W-1Atk=QAtk
(4-48)

Substituting for v in Eq. (4-40) gives

A(QAtk)=(AQAt)k=f

(4-49)

The quantity (AQAt) represents a square coefficient matrix for what may be
termed normal equations for this technique of adjustment. It is referred to by
the symbol Qc, i.e.,
Q c = AQAt
(4-50)
where the inverse of Q c is denoted by W c. (There is a reason for using
symbols Qc and W c since the respective matrices do represent cofactor and
weight matrices, but explanation of this, here, would be premature and is
deferred until Chapter 9.)
Thus
Wc Qc1 ( AQA t ) 1

(4-53)

Equations (4-48) to (4-53) are all general and apply for correlated as well as
uncorrelated observations. They can be derived without reference to an
example, as will now be demonstrated.
In matrix notation, the least squares criterion calls for minimization of the
function given by Eq. (4-31),

v tWv
With k as the vector of Lagrange multipliers is constrained by the condition
equations (4-40). Thus, the function

' v tWv 2k t ( Av f )

(4-54)

must be minimized. Setting to zero the partial derivative of with respect to

k leads back to the conditions in Eq. (4-40). Setting to zero the partial
derivative of with respect to the residual vector v yields
'
2v t W 2 k t A 0
v

(4-55)

which after dividing by two, transposing and rearranging becomes

Wv A t k

(4-56)

Note that W is a symmetric matrix, i.e., W t= W; see Appendix A. Solving for v

gives:
v=W-1Atk=QAtk
(4-57)
which is identical to Eq. (4-48). The remainder of the derivation is therefore
identical to that following E q. (4-48) and need not be repeated here.
The actual steps, or algorithm, of this least squares procedure are as follows:
the coefficient matrix: Qc is formed using Eq. (4-50), then inverted and multiplied

by the vector f to give k [Eq. (4-52)]. With the value of k, the residuals are
calculated from Eq. (4-48) and then added to the observations l to yield the

adjusted observations l . Any quantities that are functions of the estimated

observations may then be computed using l and the given functions. These
steps are demonstrated by the following examples.
EXAMPLE 4-10
Referring once more to the plane triangle in Example 4-3, the three
measured interior angles are 1=4133'45", 2= 7857'55", and 3 = 5927'50".
Compute the least squares estimates of there angles, (1) if they are
uncorrelated and have equal weight, and (2) if they are uncorrelated but
have weights w1=1, w2= 0.67, and w3 = 0.50, respectively.
Solution
The one condition equation is

1 v1 2 v 2 3 v3 180 0
or
v1 v 2 v3 180 ( 1 2 3 ) 180 17959'30' ' 30' '

In matrix notation, this condition is:

v1
1 1 1v 2 30' '
v3

which is in the general form Av = f.

(1) For the case of equal weights, W = Q = I. From Eq, (4-50),
Then, form Eq (4-52)
k Qc1 f [1 / 3][30' ' ] `[10' ' ]
Next, v is calculated from Eq (4-48):

1
10 ' '

v QA k A k 1 [10 ' ' ] 10 ' '

1
10 ' '
t

Finally, adding the residuals to the given observation, we obtain the adjusted
observations:

1 1 v1 4133'55'

2 2 v 2 78 58'05'

3 3 v 2 59 28'00 '
Sum=1800000, which checks the condition
(2) for w1=1, w2=0.67, and w3=0.50, the weight matrix is
w1
1

W w2
0.67

0
.
50
w3

And the cofactor matrix Q is

Q W

1 / w1
1

1 / w2
1 .5

2
1 / w3

Now, according to Eqs, (4-50), (4-52), and (4-48), respectively

1
1

Qc AQA 1 1 1 1.5 1 [ 4.5]

2 1
t

k Qc1 f [1 / 4.5][30 ' ' ] [6.7' ' ]

And
1
1
7 ' '

v QA k
1 .5
1 [6.7' ' ] 10 ' '

13' '
2 1
t

The adjusted observations (least squares estimates) are them:

1 1 v1 4133'52 '

2 2 v 2 78 58'05'

3 3 v 2 59 28'03'
Sum=1800000, which checks

EXAMPLE 4-11
The angles shown in fig. 4-2 are measured with a theodolite. the observed
values with weights, are listed as follows
ANGLE
l1
l2
l3
l4
l5
l6

OBSERVED VALUE
445044
461025
455512
430403
483245
422742

WEIGTH
1
3
3
3
3
1

Use the method of least squares to determine adjusted values for these angles.
Solution
It takes n0=4 measured angles to uniquely fix the two triangles in fig 4-2. thus,
with n=6, the redundancy is r=6-4=2, and the corresponding two condition
equations are

Fig 4-2
l1 v1 l 2 v 2 l 3 v 3 l 4 v 4 180
l 3 v 3 l 4 v 4 l 5 v 5 l 6 v 6 180

Where v1, v2,, v6 are the residuals in the angles l1 , l 2 ,...l 6 , respectively. Then,
v1
v
2
1 1 1 1 0 0 v 3 180 l1 l 2 l 3 l 4 24 ' '
0 0 1 1 1 1 v 180 l l l l 18' '

3
4
5
6
v
5
v 6

Which is in the form Av=f

For the given weights, the cofactor matrix Q is

Q W 1

0.333

0.333

0.333

0.333

Thus from Eq (4-50),

1

0.333
1

0.333
1
1
1
1
0
0

1
Qc AQA t
0.333

1
0 0 1 1 1 1

0.333

0
0
.
333

0.666
2

2
0.666

And
2
Wc Qc (1 / 3.56 )
0.666

0.666 0.562

2 0.187

0.187
0.562

From Eq (4-52) the Lagrange multipliers are

0.562
k Wc f
0.187

0.187 24 16 .9

0.562 18 14 .6

And from Eq (4-48) the residuals are

1

0.333
1

0.333

1
v QA t k
0.333
1

0.333

0
0.333

0
17 ' '
5' '
0

1 16 .9 1' '

1 14 .6 1' '
5' '
1

1
15' '

0
0

1
1

The adjusted values are then obtained by adding the residuals to the given
observations, i.e.,

l 1 l1 v1 44 50 '27 ' '

l 2 l 2 v 2 46 10 '20 ' '

l 3 l 3 v 3 45 55'11' '

l 4 l 4 v 4 43 04 '02 ' '

l 5 l 5 v 5 48 32 '50 ' '

l 6 l 6 v 6 42 27 '57 ' '

The reader should check to make sure that the adjusted values satisfy the
condition equations exactly.

EXMAPLE 4-12
The level net used in Example 4-7 is considered once more (see fig 3-7). For
convenience, the given data are repeated bellow

FROM
(LOWER
POINT)
B
D
D
B
D
A

TO (HIGHER
POINT)
A
B
A
C
C
C

OBSERVED
DIFERENCE IN
ELEVATION
l1=11,973
l2=10,940
l3=22,932
l4=21,040
l5=31,891
l6=8,983

LEVELED
DISTANCE
(Km)
20
12
15
28
20
26

The elevation of A is fixed at 281,130 m.

Using the procedure of least squares adjustment of observations only, calculate
the elevations of points B, C, and D for two cases:
1. When the measured differences in elevation are uncorrelated and have the
same weight.
2.When the measurements are uncorrelated but have weights that inversely
proportional to the respective leveled distances.

Solution
As discussed in Example 4-4 n0=3, and with n=6, the redundancy is r=3. The
conditions corresponding to this redundancy may be written for three

independent loops, each starting with, and closing on, the same point. Using the
same loops as used in Example 4-4, and referring to Fig. 3-7, the three
conditions are
Loop B-A-D-B
Loop D-B-C-D
Loop D-A-C-D

l1 v1 l3 v3 l 2 v 2 0
l 2 v 2 l 4 v 4 l 5 v5 0
l 3 v3 l 6 v 6 l 5 v5 0

Rearranging these equations into the matrix from Av=f we have

v1
1 1 1 0 0 0 l1 l 2 l 3 0.019
v2

0 1
0 1 1 0 l 2 l 4 l 5 0.089

0 0 1 0 1 1 l 3 l 5 l 6 0.024
v 6

(1) In this case, the cofactor matrix is Q=W -1=I. thus,

0 0
1 1 0 1

Q c AQA 0
0
0 1 1 1

0 1 1 0 1 0
A

Wc Qc1

0
1 0
1 0 1
3

0
0

1
1 1 0

1
0 1 1

0 1 0

4 0.50 0.25 0.25

8 4
(1 / 16 ) 4
8 4 0.25 0.50 0.25

4 4
8 0.25 0.25 0.50

0.50 0.25 0.25 0.019 0.0258

k Wc f 0.25 0.50 0.25 0.089 0.0433

0.25 0.25 0.50 0.024 0.0150

1 0 0
0.026
1 1 0
0.018

0.0258

1 0 1
0.011

t
t
v QA k A k
0.0433

0 1 0 0.0150 0.043

0 1 1
0.028

0 0 1
0.015

The adjusted observations are

1
3
1

1
1

l 1 l1 v1 11 .973 0.026 11 .999 m

l 2 l 2 v 2 10 .940 0.018 10 .922 m

l 3 l 3 v 3 22 .932 0.011 22 .921m

l 4 l 4 v 4 21 .040 0.043 20 .997 m

l 5 l 5 v 5 31 .891 0.028 31 .919 m

l 6 l 6 v 6 8.983 0.015 8.998 m

At this point, the least squares adjustment is technically complete. The

elevations of points B, C, and D are obtained from the adjusted
observations by reference to fig. 1-7. It should be emphasized that
regardless of which combination of adjusted differences, in elevation is used,
the calculated elevation of any point is unique, because the adjusted
observations are consistent with the model, i.e., they satisfy the condition
equations. For example, the elevation of point B may be calculated as.

B A l1 281.130 11.999 269.131m

Or as

B A l 3 l 2 281 .130 22 .921 269 .131m

Similarly

C A l 6 281 .130 8.998 290 .128 m

D A l 3 281 .130 22 .921 258 .209 m

All values computed agree exactly with the results obtained in Example 3-7
and in Example 4-7 using the technique of least squares adjustment of
indirect observations. This clearly shows that no matter which technique is
used, the least squares adjustment results for a given problem are, except
for round off, the same.
(2) In this case, the weight of each measurement is inversely proportional to
the leveled distance. Thus the weight matrix W is
a / 20

a / 12

a / 15
W

a / 28

a / 20

a / 26

where a is the constant of proportionality. Since the cofactor matrix Q is the

inverse of the weight matrix, then

 20

12

15
Q 1/ a

28

20

26

Now, any convenient value can be selected for a . Let a =12 then
1.67

1.00

1.25
Q

2.33

1.67

2.17

Thus
1.67

1.00

1 1 1 0 0 0

1
.
25
AQ 0 1 0 1 1 0

2.33

0 0 1 0 1 1

1.67

2.17

1.67 1.00 1.25

0
1.00
0

0
0
1.25

0
2.33 1.67 0

0 1.67 2.17
0

And
1.67 1.00 1.25
AQ 0
1.00
0

0
0
1.25

3.92
1.00

1.25

1.00
5.00
1.67

0
0
1
1 1
0

0
0
0

1
0
1
2.33 1.67 0

0
1
0

0 1.67 2.17
0 1 1

0
1
0

1.25
1.67

5.09

And so

Wc Q

1
c

0.1000

0.3158
0.1000

0.1104

0.2563
0.1087

0.3158
k Wc f 0.1000

0.1104

0.1000

1.67
1.00

1.25
v QA t k
0
0

0.2563
0.1087

1.00
0
2.33
1.67
0

0.1104
0.1087

0.2592
0.1104 0.019 0.01225
0.1087 0.089 0.02210

0.2592 0.024 0.00555

0
0.0205
0.0099
0
0.01225

1.25

0
.
0084

0.02210

0
0.0515

0.00555
0.0276
1.67

2.17
0.0120

The least squares adjusted observations are

l 1 l1 v1 11 .9949 m

l 2 l 2 v 2 10 .930 m

l 3 l 3 v 3 22 .924 m

l 4 l 4 v 4 20 .988 m

l 5 l 5 v 5 31 .919 m

l 6 l 6 v 6 8.995 m

The elevations of the three points are

B A l 1 281.130 11.994 269.136m

C A l 6 281.130 8.995 290.125m

D A l 3 281.130 22.924 258.206m

These values differ by no more than 0.001 m. from those obtained in Example
4-7, These small differences are attributed to rounding errors.
A summary of the symbols and equations used for least squares
adjustment of observations only is will he found in Section 9.5, Chapter 9.

PROBLEMS
4-1. In fig. 4-3, AOB is a straight line and all angle measurements, l1,
through l are shown. Give the elements of the mathematical model (n, n0, r,
u, and c) for adjustment of the angles by the method of indirect
observations, and write the appropriate condition equations in the form
v+B= f.

Fig 4-3

4-2 Figure-4-4 shows seven angle measurements, l1, through l7,, made about
a survey station A. Determine the redundancy r for a least squares
adjustment of the angles, and write the appropriate condition equations in
the form Av = f
4-3 The direction of line OA, Fig. 4-5, is known and fixed. The directions of
lines OB, OC, OD, and OE are to be determined based upon the observed
angles, 1, through 8,. Determine the redundancy and write appropriate
condition equations in the form v + B = f and in the form Av = f.
4-4 With reference to the level net shown in Fig. 4-6, the following positive
elevation differences are observed:
FROM
TO STATION
OBSERVED
STATION
DIFFERENCE (m)
5
1
42.107
2
1
12.424
3
5
42.251
3
4
8.464
4
5
4.138
4
1
46.269
4
2
33.802

Fig 4-4

Station 5 is a bench mark with an elevation that is known and fixed at

500.000 m above mean sea level.
(a) Construct the mathematical model for the least squares adjustment of
the elevations by the method of adjustment of indirect observations; i.e.,
write the appropriate condition equations in the form
v+B = f.
(b) Do the same for the method of adjustment of observations only, i.e.,
write the condition equations in the form Av=f.
4-5 A distance is measured six times. The observed values are 572.182m,
572,140m, 572.103m, 572.160m, 572.125m and 572.155m. All
measurements are uncorrelated and their standard deviations are 0.030m,
0.030m, 0.030m, 0.020, 0.020m, and 0.010m, respectively. Find the least
squares estimate for the distance.

4-6 The following angles are measured about a survey station (Fig. 4-7):
=11015'20'', = 13040'08'', = 11904'42'', and = 24055'43". All
observations are uncorrelated and have the same precision. Find least
squares estimates for the angles using the method of adjustment of indirect
observations.

4-7 If the measured angles , , , and in Problem 4-6 have weights 1, 1,

1, and 3, respectively, rework the least squares solution and compare
residuals with those obtained in Problem 4-6.
4.8 Observed elevation differences and line lengths for the level network in
Fig. 4-8 are:
LINE
A to B
B to C
C to A
C to D
D to A

OBSERVED ELEVATION DIFFERENCE (m)

-12.386
-11.740
24.101
-8.150
32.296

LENGTH (km)
18
12
20
8
22

The elevation of the bench mark at B is fixed at 192.320m above mean sea
level. The weight of each elevation difference is inversely proportional to the
length of the line. Use the method of adjustment of indirect observations to

find least squares estimates for the elevations of A, C, and D.

4-9 The position of P (x, y) in Fig. 4-9 is to be determined from uncorrelated
measurements of s, b, and . Toe observed values and their standard
deviations are:

s
b

OBSERVED VALUE
352.140m
236.765m
421520

STANDARD DEVIATION
0.030 m
0.020 m
15

Compute least squares mares for x and y using the method of adjustment
of indirect observations.
4-10 Apply the method of adjustment of observations only to the data given
in Problem 4-6 to find least squares estimates for the angles. Compare the
results with those of Problem 4-6.
4-11 Apply the method of adjustment of observations only to the data given
in Problem 4-8 to fled least squares estimates for the elevations of A, C,
and D. Compare the results with those of Problem 4-8. Which method
involves less computation?
4.12 Apply the method of adjustment of observations only to the data given
in Problem 4-9 to find adjusted values for s, b, and and least squares
estimates for x and y. Compare the results with those of Problem 4-9.
4-13 Adjust the level net of problem 4-4, assuming all observations are
uncorrelated and have the same precision.
4-14 A line-crossing technique used in a hydrographic surrey of a river to
determine the distance between. two sore stations A and B (Fig. 4-10). The
sum S of the two distances S 1, and S 2 is observed at one-minute intervals
as the boat proceeds upstream. The observed values of S in meters, are:
6137, 6075, 6020, 6015, 6029, 6072, and 6143. All observations are
uncorrelated and have the same precision. If S can he approximated by a
parabolic function of elapsed time,

Fig 4-9

Fig 4-10

find the least: squares estimates for the parameters of the parabola and
thereby obtain the least squares estimate for the distance between A and B.
4-15 For the triangle in Fig. 4-11 the following observations are made:
Angle A:
4502'13"
Angle B:
8501'48"
Angle C:
4956'19"
Side a:
241.555 m
Side b:
340.097 m
The standard deviation of each angle measurement is 10' the standard
deviation of each side is 0.020m . All measurements are uncorrelated.
(a) Write appropriate condition equations for least squares adjustment of
the observations. (b) Compute the least squares estimates for the angles
and sides of the triangle.

Fig 4-11

5
Elementary Probability
Theory

5.1. RANDOM EVENTS AND PROBABILITY

In Chapter 1, the concepts of random variable and probability were

introduced. In this chapter, we shad take a close: look at random variables and
the theory of probability.
One of the basic notions of probability theory is that of a random event. A random
event is one whose relative frequency of occurrence approaches a stable limit as
the number of observations or repetitions of an experiment is increased to
infinity. The limit of the relative frequency of occurrence of a random event is
known as the probability of the event.
The probability of a random event is a number which lies somewhere between
zero and one. If the probability of a random event is zero, the event will never
occur; if the probability of the event is one, it will always occur (the certain event).
If the probability is neither zero nor one, but some number between zero and
one, the event may or may not occur, its chance of occurring being given by the
specific value of the probability.
In mathematical language, if E represents a random event, and P[E] the
probability of the event occurring, then
0 P[E] 1

(5-1)

A distance is measured repeatedly using a carbon steel tape. With much

diligence and patience, 5000 measurements are made to 0.001 m. The
measurements are corrected for systematic error and the resulting corrected
values are rounded off to the nearest centimeter. For each resulting value, the
relative frequency of occurrence a calculated and recorded (see Table 5-1).

107 Analysts and Adjustments of Survey Measurements

ROUNDED-OFF
VALUE
OF DISTANCE
MEASUREMENT
489.51 or less
(m)
489.52
489.53

NUMBER OF
RELATIVE
MEASUREMENT
FREQUENCY
S
OF OCCURRENCE.
0 OCCURRING 0
a
a/5000
206
0.0412
0.7256
3633

489,54
489.55 or more

1161
0

0.2322
0

Total

5000

1.0000

On the basis of the large number of measurement s made, the calculated

frequencies are accepted as the limiting values; i.e., they are accepted as
probabilities.
Let .4 be the event that the rounded-off value of any one of the distance
measurements is 489.51 m or less; B the event the measurement is 489.52 en;
C the event the measurement is 489.53 m; D the event the measurement. is
489.54 m; E the event the measurement is 489.55 or more. Accepting the
relative frequencies of occurrence in Table 5-1 as the probabilities, we have:
P[A] = 0
P[B]=0.0412
P[C]=0.7266
P[D] =0.2322
P[E] = 0
A very important and useful concept in Probability theory is that of independence. Two events are said to be independent if the occurrence of one has
no influence on the occurrence of the other. If two events are independent, the
probability of the two events occurring together is equal to the product of the
probabilities of the events occurring individually. In mathematical language,
P[AB] = P[A] . P[B]
5.2.

(5-2)*

RANDOM VARIABLES.

In basic probability theory, covered at the beginning of most introductory.

Textbooks on the subject, the probability of a random event is usually de* The symbol denote intersection in set theory. A B (read as A intersection
B) can be interpreted as A and B occurring together.

108 Analysts and Adjustments of Survey Measurements

scribed with reference to a sampl e space, the collection of all possible outcomes of a random phenomenon. Events are looked upon mathematically as

subsets of the sample space In many applications of probability theory,

however, including measurements in surveying, it is more convenient and
desirable to describe the probability of a random event as a function of a
random variable.

EXAMPLE 5-2
A distance, known to be 297.500 m. is measured with a 50 m carbon steel tape.
The measurement is made over a flat, horizontal surface with the tape fully supported, so that there are no errors due to slope and sag of the tape. The ends of
the tape are very carefully marked. and appropriate corrections are made so
that the effects of errors in marking, length of tape, temperature, and tension
are negligible. However, the tape is only roughly aligned, and significant random
error due to misalignment of the tape is introduced . Since any error in
alignment causes the measured value of the distance to be too high, the error is
distance due to misalignment is always positive.
With much effort, the measurement is repeated 5000 times and the resulting
values are rounded off to the n eares cen me er The results are summarized in
Table 5-2. The errors in distance due to misalignment of the tape (x) are shown
in centimeters, and relative frequencies are calculated by dividing the number of
measurements associated with each error value by the total number of
measurements. Again, since a large number of measurements i s involved, the
relative frequencies are accepted as li miting values, i.e., as probabilities.

MEASUR
ERROR IN
ED
MEASURED
DISTANC
DISTANCE
297.50
0
E
X (cm)
(m)
297.51
297.52 2
297.53 3
297.54 4
297.55 3
297.56 6
297.57
297.53 8
Sum

NUMBER OF RELATIVE
MEASUREME FREQUEN P (x)
NTS
CY
223' n
0.0446
n/5000 0.04
50
1613
0.3226
0.32
24
1705
0.34:0
0.34
03
901
0.1302
0.18
367
0.0734
0.07
05
33
131
0.0262
0.02
60
42
0.0034
0.00
84
15
0.0030
0.00
3
0.0006
0.00
26
08
5000
1.0000
1.00
00

Includes measurements with errors less than 0.5 cm.

109 Analysts and Adjustments of Survey Measurements

If the relative frequencies in Table 5-2 can be approximated by an algebraic

function of x, such a function represents a convenient basis for evaluating

probabilities. In the particular case at hand, a function which approximates the

relative frequencies well is
p( x) 2( x 0.15) 2 e 1.5 x , x 0,1,...,8

(5-3)*

where xis the error in cm.

Values of p (x) are listed in Table 5-2 next to the relative frequencies; the
function is plotted in Fig. 5-1. It is not necessary that this function be theoretically derivable; what is important is that it fits the observed data.
In view of what has been previously discussed in Section 5.1, the event that the
error in distance takes on a particular value x is, indeed, a random event, and
the value of p (x) represents the probability of this random event occurring. For
example, the event that the error is 2 cm is a random event, and the probability
of this random event occurring is 0.3408
.
With reference to Example 5-2, let X represent the error in distance. The
random event that X takes on the specific numerical value x is represented
mathematically by the expression X=x, and the probability of this random

e = 2.71828, the base of the natural system of logarithms

110 Analysts and Adjustments of Survey Measurements

event occurring, represented mathematically by P [ X=x ], is given by the

function p (x), i.e.,
p (x) = P [ X =x ]

(5-4)

Here, X is a radom variable, and p(x) is its probability function. The random
variable and its probability function constitute wlat is known as a probability
model, a mathematical model that describes the assignment or distribution of
probabilities to a particular class of random events. In the example at hand, the
probability of the random event that X takes on the value 0 cm is
p(0)=P[X=0]=0.0450; the probability that X takes on the value 1 cm is p (1)
P[X=1] = 03224; and so on.
As a matter of convention, random variables are represented by capital italic
letters (such as X), and the numerical values they take on are represented by
lower-case italic letters (such as x). In some literature, random variables are
known as variants.
The specific function given by Eq. (5-3) is only one of many that can say Any
thing more than an illustration of what is meant by a probability function. It is
simply a mathematical model of what happens in reality, its basic function
being to distribute the total amount of probability available among all values of
the random variable. As illustrated in Table 5-2, the sum of all values of p (x)
must be unity.
The probability function p(x) is not the only kind of function associated
with a random variable. Another function with comparable importance is
F(x) = P [ X x]

for all x

(5-5)

F(x) is known as the probability distribution function of X, or simply the

distribution function. It is to be interpreted as the probability of the event that the
random variable takes on a value that is equal to or less than x.
The following properties apply to distribution functions:
P [ a < X b ] = F (b) F (a)

(5-6)

0 F (x) 1

for all x

(5-7)

F (a) F (b)

for a < b

(5-8)

F (-) = 0

(5-9)

F () = 1

(5-10)

111 Analysts and Adjustments of Survey Measurements

lim f ( x ) F ( a )
x a

(5-11)

To Illustrate F (x) and its properties, let us return to Example 5-2. The
probabilities p (x) listed in Table 5-2, are used to determine the values of F(x) in
Table 5-3. The function F(x) is plotted in Fig. 5-2.
From Table 5-2, it should be clear that the F (x) values are simply running
accumulations of the p (x) values. For example,
F(1) = p (0) / p (1) 0,0450+0.3224 = 0.3674
F(2) =p (0) +p (1) +p (2) = 0,0450 + 0.3224 + 0,3408 = 0.7082
and so on.
Now, if a=2 cm and b= 4 cm, the probability that the error X is greater than 2 cm
but less than or equal to 4 cm is given by Eq. (5-6), i.e.,
P [ 2 < X 4 ] = F ( 4 ) F ( 2 ) = 0.9622 0.7082 = 0.2540
Equation (5-7) states that all values of F (x) must lie between zero and one,
inclusive. This must be so because the values of F (x) are probabilities of
random events, and all probabilities are numbers between zero and one. For
example, F(3)=0.8887 is the probability of the event that X is 3 cm, or less; F(5)
= 0.9882 is the probability that X is 5 cm, or less.
Equation (5-8) states that the distribution function is nondecreasing , i.e., F(x)
cannot decrease in value as x increases. This property should be evident from
Table 5-3, or from Fig. 5-2.

0
1
2
3

ERROR IN
MEASURED
DISTANCE x
cm

Table 5.3
DISTRIBUTION
FUNCTION

0.0450
=0.045
F
(x)
=
P[X
x]
0.0450
-=0.367
0
0.3574
-=0.708
0.3224
4
0.7082
-=0.888
0.3408
2
0.1805
4
0.8887
+7
=
2
0.0735
0.9622
5
0.96
2
+=
0.0260
0.9882
6
0.9882
+=
0.0084
0.9966
7
0.9966
-=0.999
0.0026
8
0.9992
+2
=1.000
0.0008
0
9
1.0000 - 0
=
1.0000
112 Analysts and Adjustments of Survey Measurements

Fig. 5-2.
Equations (5-9) and (5-10) give the two extreme values for F (x). In the case at
hand,
F (x) = 0 for any value of x less than zero, and F (x) = 1 for any value
of x greater than 8.
Equation (5-11) states that the distribution function is continuous from the right.
Referring to rig. 5-2. this means that F(2) = 0.7082, not 0.3674, and
F(3)=0.8887, not 0.7082, the higher of the two possible values is taken.
It has already been stated that a random variable and its probability function
constitute a probability model. The probability model can be given as well by the
random variable and its distribution function. In either case, the model is known
as a probability distribution.
The particular model illustrated in Example 5-2 is that of a discrete probability
distribution. The characteristic feature of a discrete probability dis-

113 Analysts and Adjustments of Survey Measurements

tribution is that the probability function p (x) is nonzero only for a distinct set of
values of x, for all other values of x, p (x) is zero. In Example 5-2, p (x) is
nonzero only for x=0, 1, 2, 3, 4, 5, 6, 7, and 8. It follows that the distribution
function F (x) of a discrete probability distribution must be a step function; that
is, a function that increases only in finite jumps. Figure 5-2 shows this

characteristic quite clearly. Indeed, p (x) is precisely the jump in F (x) at each
value of x for which p (x) is nonzero.
It is important to note that although p (x) may be zero at a particular value of x,
F (x) is not necessarily zero. In Example 5-2, p (1.5) = 0, but F(1.5 ) = 0.3674;
and p (9) = 0, but F(9) = 1.0000.

5.3. CONTINUOUS PROBABILITY DISTRIBUTIONS

The measurements in Example 5-2 were rounded off to the nearest centimeter,
yielding nine distinct values for the error in distance. This led to describing the
behavior of the error in terms of a discrete probability distribution ion. If,
however, we are to look at measurements and their errors in a more general
way, unconstrained by round off, it is more convenient to use probability
distributions that are continuous.
A continuous probability distribution is a probability model which has a
continuous distribution function; i.e., a function which has no jumps.
If a distribution function F(x) has no jumps, it must follow that the corresponding
probability function p (x) is zero everywhere, and so is meaningless to use . In
its place, another function is used the probability density function f(x). This
function was mentioned briefly in chapter 1.
The density function is mathematically defined at the first derivate of the
distribution function, i.e.,
f ( x) F ' ( x)

(5-12)

F ( x)

Since F(x) is nondecreasing, its slope must be nonnegative. Hence

F(x) 0

(5-13)

From the fundamental theorem of integral calculus, we have

f ( x ) dx F (b) F ( a )

(5-14)

Therefore, referring back to Eq. (5-6), we see that

114 Analysts and Adjustments of Survey Measurements

P[ a X b] F (b) F ( a )

f ( x)dx,
a

(5-15)

which provides the means for evaluating the probability that the random variable
X takes on a value between a and b.
Note that (a b) is a random event and that the probability of this random event
is a definite integral of the density function. Hence, the probability of this event
is represented by the area under the density function between a and b.
It is quite important to recognize that evaluation
of a density function i,e.,
does no yield probability, the density function f (x) does not give the probability
that X takes on the value x, as p (x) does. To repeat, probability is represented
by the area under the density function, not by the ordinate value.
Figure 5-3 illustrates the relationship between a continuous distribution function
and the corresponding density function. Note carefully how P[a < X b] is
represented in each case. in Fig. 5-3 (a), P[a<X b] is represented by a
difference in ordinate values, F(b) -F(a); in Fig. 5-3 (b), P[a < X b] is
represented by the hatched area. Note also that F (x) is indeed continuous (i.e.,
no jumps) and it lies entirely between zero and one, and that f(x) is
nonnegative.
Two important and very useful relationships can be obtained as special cases of
Eq. (5-I5). First, by making a = - and b = ,and taking note of the properties of
the distribution given in Eqs. (5-9) and (5-10), we have

f ( x ) P[ X ] F ( ) F ( ) 1,

(5-16)

which states that t the iota: area under the density function must equal unity.
Second , by making a =- and b=x, noting again the property expressed by Eq.
(5-9), and recognizing that the event (X x ) is equivalent to the event (- <X
x) we have.
F ( x ) P[ X x ] P[ X x ] F ( x ) F ( )

f (u ) du. (5-17)

* Note that it is necessary to change the variable of integration (to u, in this

case), since x is one of the limits of integration.

115 Analysts and Adjustments of Survey Measurements

116 Analysts and Adjustments of Survey Measurements

Thus, if we Know the density function we can obtain the distribution function;
and vice versa. The probability model is specified by either function.
Perlaps the simplest example of a continuous probability distribution is the
uniform distribution. A random variable X has a uniform distribution if the non
zero portion of its density function is constant. Specifically,
f ( x)

1
ba

for a < x < b

=0

elsewhere

(5-18)

The numbers a and b are called the parameters of the distribution. The
corresponding distribution function is
for x a

F (x) = 0

1
xa
du
ba
ba

for a < x < b

=1

(5-19)

for x b

The two functions are plotted in Fig. 5-4. Note that the probability that X is less
than or equal to c is given by the hatched area in Fig. 5-4 (a) and by the
ordinate F(c) in Fig. 5.4 (b). Also note that the nonzero portion of the density
function must be 1/(b- a) in order to make the total area under the density
function equal to unity.
EXAMPLE 5-3
Let X be a uniformly distributed random variable with parameters a = 2 and b=
8, The density function of X is
f ( x)

1
1
for 2 < x < 8

82 6

=0

elsewhere

The distribution of function of x is

F (x) = 0

for x 2

x2
6

for x < 2 < 8

=1
for x 8
117 Analysts and Adjustments of Survey Measurements

Evaluate the probability that: (a) X is less than or equal to 3; (b) X is greater
than 4; (c) X lies between 5 and 7; (d) X lies between 6 and 10; (e) X equals 4.
Solution
(a) The probability that X is less than or equal to 3 is
P[ X 3] F (3)

32 1

6
6

(b) The probability that X is greater than 4 is

42 2
P[ x 4] P[ 4 X ] F ( ) F ( 4) 1

6 3
118 Analysts and Adjustments of Survey Measurements

(c) Tae probability that X lies between 5 and 7 is

72 52 5 3 1
P[5 X 7] F (7) F (5)

6 6 6 6 3

(d) The probability that X lies between 5 and 10 is

4 1
62
P[6 X 10 ] F (10 ) F (6) 1
1
6 3
6

(e) The probability that X equals 4 is

P[ X 4] P[4 X 4] F (4) F (4) 0.

5.3.

THE NORMAL DISTRIBUTION

Of all existing probability distributions, none is more important than the normal
distribution. The normal distribution has widespread application in science,

technology, and industry; it is used as the basic model for all physical
measurements, includin measurements in surveying. The density function of the
normal distribution is:
x 2
F ( x)
exp
for - < x < .
2 2
2

(5-20)*

The quantities and are the parameters of the distribution, and are called the
mean and standard deviation, respectively (see Chapter 1). They will be
discussed in more detail in Section 5.5.
The distribution function of the normal distribution is
u 2
F ( x ) exp
du.

2 2

(5-21)

These two functions are plotted in Fig. 5-5. It is quite clear from Fig. 5-5 (a) that
the normal distribution is symmetric about . The density function s points of
inflection are located at x = u and x = u + . Maximum density occurs at x =

If X is a random variable with a normal distribution, then P[X c], the probability
that X is less than or equal to c, is represented by the hatched area in Fig. 5-5
(a) and by the ordinate, F (c), in Fig. 5-5 (b).
* The notation exp (t) is another way of expressing ea, where e=2.71828, the
base of the natural system of logarithms.
119 Analysts and Adjustments of Survey Measurements

EXAMPLE 5-4
Let X be a normally distributed random variable with parameters = 12 and =
2 . i.e., the density function of X is
x 122
1
2
f ( x )
exp
0.1947 exp 0.1250 x 12
2
22
2 2

and the distribution function of X is:*

120 Analysts and Adjustments of Survey Measurements

F ( x ) 0.19947 exp 0.1250 u 12 du

x

Specific values of these two functions are listed in Table 5-4.

Evaluate the probability that: (a ) X is less than or equal to 10; (b) X lies
between 11 and 15; (c) X is greater than 16; (d) X is precisely 10.
Solution
(a) The probability that X is le ss tha n or equal to 10 is P[X 10] = F(10) =
0.1587 This probability is represented by the hatched area in Fig. 5-6 (a).
( b ) The probability that X has between 11 and 15 is P[ 11 < X < 15] = F (15) - F
(11) = 0.9332 - 0.3085 = 0.6247. This probability is represented by the hatched
area Fig. 5-6 (b).
Table 5-4
Evaluation of Density and Distribution
Functions for a normally Distributed Random
Variable with = 12 and = 2

* F (x) is left in the integral from because it cannot be integrated directly. To

evaluate this function, a series approximation is used.
121 Analysts and Adjustments of Survey Measurements

122 Analysts and Adjustments of Survey Measurements

(c)The probability that X is greater than 16 is P[X> 16]= P[16 <X < ] = F () F(16) = 1- 0.9772 = 0.0228 .This probability is represented by the hatched area
in Fig. 5-6 (c).
(d) The probability that X is precisely 10 is, of course, zero, since this is
represented by zero area under the density function. If, however, we interpret
10 as the interval between 9.5 and 10.5, it is then possible to evaluate a
probability that is nonzero:
P [9.5 < X < 10.5]= F(10.5) F(9.5) = 0.2266 0.1056 = 0.1210

A random variable Z is said to have a normal standard normal distribution if its

density function is
z2
1
f ( z )
exp
2
2

for - < z < ,

(5-22)

This function is obtained from the normal density function by setting = 0 and
= 1; it is shown in fig. 5-7.
The standard normal distribution is an important special case of the normal
distribution because it provides a convenient way of evaluating probabilities
associated with any normal distribution. Since the density function of the normal
distribution cannot integrated directly, it presents a problem whenever
probabilities have to be evaluated for specific values of and . Fortunately,
the problem can be circumvented by first transforming.

123 Analysts and Adjustments of Survey Measurements

The normal random variable X into the standard normal random variable Z and
the evaluating probabilities for Z.
The transformation, known as standardization, is given by

(5-23)

Even though it is just as impossible to integrate f (z) directly as it is f (x) for the
normal distribution there is but one f (z) to integrate. The resulting function is

P[ Z z ] ( z )

2
exp 2 du
2
1

for - < z <.

(5-24)

The integral can be evaluated by approximate means once and for all, and
values of (z) can be tabulated. Such tabulation is given in a table I of
appendix B.
To show how standardization and table I are applied, let us consider the
following example
EXAMPLE 5-5
As in Example 51 let X be a normally distributed random variable with
parameters =12 and = 2. Standardize X and use Table I of Appendix B to
evaluate the probability that: (a) X is less than or equal to 10; (b) X lies between
11 and 15; (c) X is greater than 16.
Solution
According to Eq. (5-23), the standard normal random variable is
z

x 12
.
2

(a) The probability that X is less than or equal to 10 is evaluated by first no that
if X 10 then (X 12)/2 (10 12)/2, i.e., Z -1. Thus
X 12 10 12
P[ X 10 ] P

PZ 1,
2
2

and from Table I of Appendix B, P [Z 1] = (1) = 0.1587.

124 Analysts and Adjustments of Survey Measurements

(b) The probability that X lies between 11 and 15 is evaluated in similar fashion,
i.e.,
11 12 X 12 15 12
P[11 X 15] P

P[ 0.5 Z 1.5]
2
2
2

1.5 0.5 0.9332 0.3085 0.6247 .

(c) The probability that X is greater than 16 is:

16 12 X 12

P X 16 P16 X P

2
2

P2 Z ( ) ( 2)
1 0.9772 0.0228

The suitability of the normal distributed as the probability model for

measurements is firmly based upon the behavior of actual observations. In
effect, the model is chosen to fit t he observed data. Choice of the normal
distribution has a theoretical basis as well, however. If the resulting error in a
measurement is the sum of se, oral component errors, each of which has its
own probability distribution, can be shown that the total error has approximately
a normal distribution. whether or not the component errors are normally
distributed themselves.
To illustrate this interesting characteristic let us consider three independent,
uniformly distributed random variables (the component errors). each with
parameters a= -1 and b= 1. The density function for each one of the random
variables is plotted in Fie. 5-8 (a). The sum of any two of the random variables
will have the density function plotted in Fig. 5-8 (b). (Derivation of this density
function is beyond the scope of this textbook). The sum of all three random
variables will have the density function plotted in Fig. 5-8 (c). Note the shape of
the curve, in Fig. 5-8 (c) particularly its similarity to the density function of the
normal distribution. The sum of four uniformly distributed random variables will
have a distribution that is even closer to the normal distribution. Indeed, as :he
number of random variables (error components) increases, the distribution of
their sum gets closer and closer to the normal distribution. (Theoretical proof of
this is given by the Central Limit Theorem, discussion of which is found in most
textbooks on probability theory). Thus, it is not unreasonable at all to use the
normal distribution as the probability model for an error which is assumed to
have many components.
125 Analysts and Adjustments of Survey Measurements

5.4.

EXPECTATION

Let be X be a discretely distributed random variable, with possible values X1,

X2, X3 ..., Xn. The probability function of X is given by p (Xi), i =1, 2, , n.
The weighted sum of all possible values, where the weights are the corresponding probabilities, is called the expectation or expected value of X, denoted
EX]:

E X x i p ( x i ) x1 p x1 x 2 p x 2 ... x n p x n

(5-25)

i 1

126 Analysts and Adjustments of Survey Measurements

E[X] is also the mean of X, commonly represented by the symbol , as already
indicated in Chapter I and in Section 5.4:

E X .

(5-26)

E[X], or , locates the centroid of the probability distribution. It can be positive,

zero, or negative.

If we now consider a function of X, g (X), the expectation of this function is given

by
n

E g X g x i p x i .

(5-27)

i 1

In particular, if g(x) = (X- ), we obtain

E X x i p x i .
2

(5-28)

i 1

E [(X- ) ] is called the variance of X, and it is commonly represented by the

symbol var(X), or by (see Chapter 1).

var X 2 E X .
2

(5-29)

The variance is a measure of the spread or dispersion of the probability

distribution. It is analogous to the moment of inertia in mechanics and is never
negative.
The positive square root of the variance, denoted by a, is called the standard
deviation, as already indicated in Chapter and Section 5.4. The standard
deviation is also a measure of the spread or dispersion of the probability
distribution.
EXAMPLE 5-6

Consider once more the distance that was measured in Example 5-2. Values of
the resulting error caused by misalignment of the tape and their corresponding
probabilities were listed in Table 5-2; for convenience, they are repeated in the
first two columns of Table 5-5. Evaluate the mean, variance and standard
deviation of the probability distribution of the measurements.
Solution
The mean is

127 Analysts and Adjustments of Survey Measurements

E X x1 px1 x 2 px 2 ... x n px n

0 0.0450 10.3224 ... 80.0008

2.0445

This calculation is given in the third column of Table 5-5. The fourth column of
the table lists the deviations from the mean, for use in the calculation of the
variance. The variance is

2 E X x1 px1 x2 px2 ... x n pxn

2

2.0445 0.0450 1.0445 0.3224 ... 5.9555 0.0008

2

1.4371cm 2 .

This calculation is given in the last column of Table 5-5. The standard deviation
is:

1.4371 1.1988cm.
For continuous probability distributions, the mean and variance are defined in
terms of integrals instead of sums. Thus, if X is a random variable with a
continuous probability distribution,

E X xf x dx

And

2 E X
2

x 2 f x dx.

(5-30)
(5-31)

128 Analysts and Adjustments of Survey Measurements

For symmetrical distributions, such as the normal distribution, the mean is the
central value about which he symmetry exists. Thus, the parameter found in
the density function of the normal distribution is indeed the mean of the
distribution; i.e., if X is normally distributed,
x 2
1
E X x
exp
dx .

2 2
2

(5-32)

Furthermore, the parameter found in the density function of the normal

distribution is precisely the standard deviation, i.e.. its square is the variance,

E X
2

x 2

x 2
2
exp
dx . (5-33)
2
2
2

Expectations have certain properties, among which are:

E k K 1 , where K is a constant
E Kg X KE g X
E g 1 X g 2 ( X ) E g 1 X E g 2 X .

(5-34)
(5-35)
(5-36)

These properties can be used to derive an important relationship:

2 E X 2

E X 2X 2
2

E X 2 2E X 2

EX

(5-37)

If X is a random variable, and a and b are constants, then

Y aX b

(5-38)

is the also a random variable . Furthermore, It can be shown, by applying

properties expressed in Eqs. (5-34), (5-35), and (5-36), that:
(5-39)
y ax b
And
(5-40)
2 y a 2 2 x ,
where x and y are the means of X and Y, and x and y are the variances
of X and Y, respectively.
129 Analysts and Adjustments of Survey Measurements

5.5.

MEASURES OF PRECISION AND ACCURACY

Under the assumption that the normal distribution is the acceptable probability
model for a survey measurement, we can represent the measurement by X, a
random variable with density function given by Eq. (5-20). If we set the mean
equal to zero, Eq. (5-20) reduces to
f x

x2
exp
.
2
2
2
1

(5-41)

In this form, the normal distribution is the probability model for the random error
component of a survey measurement. This density function is shown in Fig. 5-9.

The precision of a measurement, as already defined in Chapter 1, is the degree

of closeness or conformity of repeated measurements of the same quantity to
each other. Accordingly, the spread of the probability distribution of a
measurement, or of its random error component, is an indication of the
measurement's precision. A small spread indicates high precision; a large
spread low precision, as illustrated by the density functions in Fig. 5-10. Note
that because the area under each density function must be unity, the high
precision curve must peak above the low precision curve.
For purposes of analysis, a quantitative assessment of the precision of a
measurement is required. An obvious measure of precision is the standard
deviation itself. The smaller the standard deviation, the higher is the precision.
In fig. 5-10, for example, 1 < 2, where 1 is the standard deviation of

Fig. 5-9 Probability model for the random error of a survey measurement.

130 Analysts and Adjustments of Survey Measurements

the high precision measurement, and 2, is the standard deviation of the low
precision measurement.

Applying Eq. (5-15), the probability that the measurement X lies between -
and + is given by
x 2
exp
dx. (5-42)
2

2

P X

To evaluate this probability, it is best to standardize X according to Eq. (5-23),

and then use table I of appendix B. thus
X
P x X x P

P 1 Z 1
1 1
0.8413 0.1587 0.6826

(5-43)

which means that the shaded area in Fig. 5-11 is 0.6826 of the total area under
the density function.
Multiples k of the standard deviation are also used as measures of precision.
The probability that a survey measurement lies between - k and + k is
131 Analysts and Adjustments of Survey Measurements

P k X k

x 2
exp
dx.
2
2

2 2

(5-44)

Again, standardizing X according to Eq. (5-23), we get

P k X k P k Z k k k . (5-45)

Noting from the symmetry of the distribution that (k)+ (- k) = 1, Eq. (5-45)
reduces to

P k X k 2 k 1

(5-46)

Values for 0 (k) can be obtained from Table 1 of Appendix B for specific values
of k. Table 5-6 lists 0 (k) and P [- k < X<+ k] for several values of k.

132 Analysts and Adjustments of Survey Measurements

The probabilities for k=0.674, 1.645, and 1.960 are represented in Fig. 5-12.
The quantity 0.674 has been called the probable error. This term is obsolete
and should no longer be used. A more appropriate term is 50% uncertainty.
The quantity 1.645 is called the 90% uncertainty The probability tlat a
measurement is within 1.645 of its mean value is 0.900. Similarly, 1.960
is called the 95 % uncertainty, and 2.576 is called the 99% uncertainty.
EXAMPLE 5-7
A survey measurement is assumed to lave a normal distribution with mean
value 394.625m and standard deviation 0.023 m Evaluate the 90% and 95%
uncertainties for the measurement, and make appropriate probability
statements.
Solution
90% uncertainty = 1.645 (0.023) = 0.038 m.
95% uncertainty = 1.960 (0.023) = 0.045 m

133 Analysts and Adjustments of Survey Measurements

The probability is 0.90 that the measurement lies within 0.038 m of 394,625 m,
i.e., between 394.587 m and 394.663 m . The probability is 0.95 that the
measurement lies within 0.045 m of 394.625 m, i.e., between 394.580 m and
394.670 m.
The accuracy of a measurement has been defined in Chapter 1 as its degree of
conformity or closeness to the true value. Accuracy is influenced, not only by
the random error component of the measurement, but also by the bias created
by uncorrected systematic error.
The mean square error M is defined as the following expectation:

M 2 E X ,
2

(5-47)

Where is the "true " value. The mean square error is used as a measure of
accuracy.
The bias is defined as

(5-48)

By applying the properties of expectation, it can be easily shown that

M 2 2 2.

(5-49)

Just as an increase in o indicates a decrease in precision, an increase in M

indicates a decrease in accuracy.
It should be obvious from Eq. (5-49) that high precision does not necessarily
imply high accuracy. If o is small and 3 is large, the measurement in question
has high precision, but low accuracy.
If = 0 (i.e., no bias in the measurement), the mean square error is identical to
the variance, and any measure of precision is also a measure of accuracy. It is
therefore important to reduce the effect of systematic error to negligible level, if
measures of precision are to be also used as measures of accuracy.
5.6.

COVARIANCE AND CORRELATION

If two random variables X and Y are included in the same probability model,
they are said to be jointly distributed, their joint distribution junction is
F x, y P X x, Y y ,

(5-50)

Which is to be interpreted as the probability of the event that X is less than

134 Analysts and Adjustments of Survey Measurements

Or equal to some specified value x and, at the same time, Y is less than or
equal to y.
The distribution functions of the individual random variables are obtainable from
the joint distribution function as follows:
F x P X x P X x , Y F x ,

(5.51)

And
F y PY y P X , Y y F , y (5-52)

When F (x) and F (y) are derived from the joint distribution function in this way,
they are called the marginal distribution functions of X and Y, respectively.
If X and Y are continuous random variables, their joint density function is given
by
f x, y

2
F x, y
xy

And their joint distribution function can then be expressed as follows:

(5-53)

x y

F x, y

f u, v dudv

(5-54)

Where f (u, v) is the joint density function.

It can be shown as well that the marginal density functions of X and Y are
respectively
f x

f x, y dy

(5-55)

And
f y

f x, y dx

(5-56)

In Section 5.1 it was stated that two events are independent if the occurrence of
one has no influence on the occurrence of the other. In similar fashion,

135 Analysts and Adjustments of Survey Measurements

Two random variables are said to be independent if the value taken on by one
has no influence on the value taken on by the other. The relationship for
independent events, expressed by Eq. (5-2), can be extended to the distribution
functions by simply recognizing that (X Y y ) are events. Thus
P X x, Y y P X x PY y

(5-57)

Or, alternatively,
F x, y F x F y

(5-58)

For Independent random variables X and Y.

If X and Y have a joint continuous distribution and are independent,
differentiation of both sides of Eq. (5-58) yields
f x, y f x f y

(5-59)

This product relationship of independence can be futher extended to expectations. In particular, it can be shown that
E XY E X E Y xy

For independent X and Y.

(5-60)

Now the marginal distributions of X and Y have their own separate means and
variances; these are x and y respectively, and x and y, respectively. In
addition to these four parameters, there exists a fifth parameter which measures
the degree of correlation between the two random variables. This parameter is
known as the covariance: it is designated by the symbol cov (X, Y), or by xy,
and it is defined as the following expectation.

cov X , Y xy E X x Y y .

(5-61)

It can be positive, zero, or negative. When xy is positive, random variables X

and Y are said to be positively correlated; when xy is zero, the random
variables are said to be uncorrelated, or correlation-free; and when xy is
negative, they are said to be negatively correlated.
For the continuous case,

xy

x y f x, y dxdy

xs

(5-62)

136 Analysts and Adjustments of Survey Measurements

If the expectation on the right-hand side of Eq. (5-61) is expanded and
simplified using the expectation properties expressed by Eqs. (5-34), (5-35),
and (5-36), the following useful relationship results:

xy E X x Y y

E XY x Y X y x y

E XY x E Y E X y x y

E XY x y x y x y

E XY x y

(5-63)

If X and Y are independent. Eq. (5-60) holds, and it must necessarily follow from
Eq.
(5-63) that

xy 0

(5-64)

i.e., X and Y are uncorrelated.

Although independence leads to zero covariance, as just shown, the converse
is not necessarily true; i .e., zero covariance does not generally imply
independence. Fortunately, when X and Y have a joint normal distribution, zero
covariance is a sufficient condition for independence. Therefore, for
measurements that are normally. Distributed, zero covariance and independence are equivalent conditions.

If the random variables are standardized by the transformation given by Eq. (523), the expectation of :..e product of the standardized variables is a
dimensionless number known as the correlation coefficient, designated xy thus
X Y y
x

x y

xy E

(5-65)

which simplifies to

xy

xy
x y

(5-66)

Since x and y are always positive, xy takes on the sign of xy. It can be
demonstrated that the absolute value of xy never exceeds unity, i.e.,
1 xy 1. (5-67)

137 Analysts and Adjustments of Survey Measurements

EXAMPLE 5-8

Calculate the correlation coefficient of the X and Y coordinates of a survey

point, given x = 1.72 cm, y = 1.18 cm, and xy = 0.32 cm.
Solution.

x 1.72 1.31cm
y 1.18 1.09 cm
Thus,

xy

xy
x y

0.32
0.22
1.311.09

The sum of two jointly distributed random variables is also a random variable.
Depending on the joint distribution involved, the sum may or may not have the
same kind of distribution as its components. For example, if X and Y are
uniformly distributed, X+ Y does not have a uniform but a triangular distribution,
as already shown in Fig. 5-8 (b); however, if X and Y are normally distributed. it
can be demonstrated that X+ Y has a normal distribution. In any case, the mean
of the sum of two jointly distributed random variables is equal to the sum of their
means, i.e.,

E X Y E x E Y

(5-68)

Or, alternatively,

x y x y .

(5-69)

The variance of the sum can be determined as well from the variances and
covariance of its components, quite apart from the kind of distribution involved.
According to Eq. (5-29), the variance of the sum is the following expectation:

x2 y E X Y x y ,
2

(5-70)

Which reduces to

138 Analysts and Adjustments of Survey Measurements

x2 y E X x Y y

E X 2 X x Y y Y y
2

E X 2 E X x Y y E Y y
2

x2 2 xy y2 .

(5-71)

Example 5-9
Calculate the mean and variance of the sum of the X and Y coordinates in
Example 5-8, given 2 = 43.00 cm, x = 1.72 cm, y = 27.00 cm, y =1.18
cm, and xy= 0.32 cm.
Solution
The mean of X + Y is

x y x y 43.00 27.00 70.00cm

The variance of X+ Y is

x2 y x2 2 xy y2 1.72 20.32 1.18 3.54cm 2

To complete this section, we note that if X and Y are independent, xy= 0 and
Eq. (5-71) reduces to

 x2 y x2 y2 ,

(5-72)

Which is a very important and useful relationship.

5.8. COVARIANCE, COFACTOR, AND WEIGHT MATRICES

In Section 5.7, the joint distribution of two random variables was discussed and
the term covariance was introduced. Covariance is basically a relationship
between two random variables only. If three or more random variables are
jointly distributed, we must consider the covariances for all possible pairs. Thus,
when we speak of the covariance of three jointly distributed random variables,
X, Y, and Z, we are referring specifically to three
covariances ___ xy for X and Y, yz for Y and Z, and xz for X and Z.
In dealing with m jointly distributed random variables, It is convenient to collect
them into a single vector. A vector whose components are jointly
139 Analysts and Adjustments of Survey Measurements
Distributed random variables is known as a random vector. Thus, if X1, X2 , ...,
Xm. are m jointly distributed random variables, the vector
X 1
X 2

.
X

.
.

Xm

(5-73)*

Is a random vector.
If we let
X 1 1
X 2
2

.
x
,
.

Xm m

(5-74)

where 1, 2, ..., m, are the mean values of X1, X2, ..., Xm, respectively then.

x x t

X 1 1
X 2
2

X 1 1 X 2 2 ... X m m
.

Xm m


X 1 1 2
X 1 1 X 2 2 ... X 1 1 X m m

X 2 X 1 1
X 2 2 2
... X 2 2 X m m
2
. (5-75)
...................
..........................
...
.....................

X m m 2
X m m X 1 1 X m m X 2 2 ...
By taking expectations of the elements in be matrix on the right-hand side of Eq.
(5-75), and noting from Eq. (5-29) that E[(Xi- i)] = i (the variance of Xi), and
from Eq. (5-61) that E[(Xi - i) (Xj - j )] = i (the covariance of Xi and Xj), and
that ij= ji,we obtain the symmetric matrix
* A boldface lower case letter is used to represent a random vector in keeping
with established vector and matrix notation; the elements , however, arc in
capital letters to represent random variables.
140 Analysts and Adjustments of Survey Measurements

12

xx ...12

1m

12
22
...

2m

... 1m

... 2 m
.
... ...

... m2

(5-76)

The variances of the individual random variables form the main diagonal of xx;
the covariances of all possible pairs of the random variables are the off-diagonal
elements of E-, with each covariance appearing twice (in one position above
and in one position below the main diagonal).
The matrix xx is called the variance-covariance matrix of x, or simply the
covariance matrix of x, since the variance of a random variable can be looked
upon as its covariance with itself (i = ii). The term dispersion matrix is also
applied to xx.
If the random variables in x are uncorrelated, all covariance (off-diagonal)
elements of xx are zero, and the matrix is diagonal [see Eq. (4-15), Chapter 4].
In such a case, xx maybe called a variance matrix, as it was in Chapter 4.
However, it is quite acceptable refer to a diagonal xx as a covariance matrix,
even though all covariances in it are zero.
The relationship between the weight matrix W and the corresponding variance
matrix (or covariance matrix, as we will now call it) was also given in Chapter
4 by Eq. (4-19 ). This relationship, with subscripts added to indicate reference to
random vector x, is restated as Eq. (5-77):
W xx 02 xx .
1

(5-77)

(The quantity o was introduced in Chapter 4 as the reference variance.)

Equation (5-77) is general, applying to correlated as well as uncorrelated

random variables, measurements, or observations, i.e., Eq. (5-77) is not

restricted to diagonal matrices.
Caution: If Wxx is nondiagonal, the simple weights calculated in Eq. (4-16)
are not to be used as diagonal elements of Wxx. Only when Wxx i s diagonal
(i.e., when the observations are uncorrelated) are the weights calculated in Eq.
(4-16) identical to the diagonal elements. This restriction is illustrated in the
following example.
Example 5-10
Two observations are represented by the random vector
X 1
X .
X 2
141 Analysts and Adjustments of Survey Measurements

The variances of X1 and X2 are 1 and 2; respectively; the covariance of X1

and X2 is 12, and the correlation coefficient is 12.
(a) For a selected reference variance derive the weight matrix of x in terms of
the given parameters. (b) Show that weights calculated in Eq. (4-16), Chapter 4,
are identical to the diagonal elements of the weight matrix only when 12=0.
Solution
(a) According to Eq. (5-77), the weight matrix of x is:
W xx

2
0

1
xx

12

12

12

22

2
0

02
2 2
1 2 122

22

12

12
.
12

From Eq. (5-66), we see that 12 = 12 1 2 and so 12 - 12 = 12 (1

- 12). Thus,
22
12 1 2
02
W xx 2 2

1 2 1 122 12 1 2
12

o2

12 1 122

o2 12

1 2
1 2
12

o2 12

2
1 2 1 12
.
o2

22 1 122

(b) from Eq. (4-16) we get

w1

o2
02
andw

.
2
12
22

Obviously, the diagonal elements of Wxx are o/1 of and o/3 only when
12=0. When 120, the weights w1 and w2 cannot be used as diagonal
elements of Wxx.
Each element of xx can be divided by o to yield a scaled version of , This
new matrix is designated Qxx and it is known as the cofactor matrix of x. Thus,

Q xx

2
0

xx

12
2
0
12
2
0
...
1m
2
0

12
02
22
02
...

2m
02

1m

02
2m
...
02
...

...
...

(5-78)

...

m2
02

142 Analysts and Adjustments of Survey Measurements

Or, equivalently,
xx 02 Q xx .

(5-79)

The term cofactor matrix. was introduced in Chapter 4 with respect to

uncorrelated observations only. In this chapter, it is general, taking into account
covariances as well as variances, or more specifically, their relative magnitudes.
As such, Qxx is sometimes called the relative covariance matrix, Only when
Qxx is multiplied by al to give xx is an absolute scale provided.
EXAMPLE 5.11
The position and elevation of a survey station are given by the random vector
X 1
x X 2 ,

X 3

In which X1 and X2 are the planimetric coordinates, and X3 is the elevation.

The standard deviations of X1, X2,and X3 m, 0.050 m, and 0.015 en,
respectively. The correlation coefficient between X1 and X2 is -0.50. There is no
correlation between X1 and X3, or between X2 and X3.
(a) Evaluate the covariance matrix of x.
Solution

12 0.025 0.000625 m 2
2

22 0.050 0.002500 m 2
2

32 0.015 0.000225m 2
2

12 12 1 2 0.500.0250.050 0.000625m 2
13 23 0

Thus,
xx

0.000625
0.000625

0.000625
0.002500
0

m 2
0

0.000225
0

(b) If a reference variance of 0.000625 m' is chosen, evaluate Solution

By Eq. (5-78), the cofactor matrix is
143 Analysts and Adjustments of Survey Measurements

Q xx

0.000625
1
0.000625

0.000625

0.000625
0.002500
0

0.000225
0

0
1.00 1.00

1.00 4.00
0 .

0
0
0.36

In Example 5-11, the elements of Qxx are easily manageable, dimensionless

numbers which portray the relative magnitudes of the variances and
covariances more readily than the elements of xx. The choice of reference
variance is arbitrary; it is chosen for convenience.
When the units of the components of x are mixed, it is not possible to construct
a cofactor matrix entirely of dimensionless numbers. This is demonstrated in the
following example.
EXAMPLE 5-12
The position of a survey station is given in polar coordinates, R and . R is
expressed in meters; in radians. The variance of R is 1.0 10 4 m 2 , the
variance of is 2.0 10 9 rad 2 , and the covariance of R and is 1.8 10 7 m-rad.
Thus, if x is a random vector whose components are R and ,
1.0 10 4 m 2

xx 1.8 10 7 m rad

1.8 10 7 m rad
.
2.0 10 9 rad 2

(a)If a reference variance of 1.0 10 4 m 2 chosen, evaluate Qxx.

Solution
Q xx

1
1.0 10 4 m 2

1.0 10 4 m 2

7
1.8 10 m rad

1.8 10 7 m rad
.
2.0 10 9 rad 2


1 .0
1.8 10 3 rad / m

.
3
5
2
2
1.8 10 rad / m 2.0 10 rad m

In this case, only the first element is dimensionless.

(b) If a reference variance of 2.0 x 109 rad is chosen, evaluate Qxx.
144 Analysts and Adjustments of Survey Measurements

Solution
5.0 10 4 m 2 / rad 2
Q xx
90 m / rad

90 m / rad

1 .0

In this case, only the last element is dimensionless.

(c) If a reference variance of 1.0 10 2 (no units) is chosen, evaluateQxx
Solution
1000 m 2
Q xx
1.8m rad

1.8m rad
.
0.02 rad 2

Example 5-12 shows that the tams of the elements of Qxx can vary depending
on the choice of units for the reference variance. Since it is impossible to select
a reference Qxx variance that will give a completely dimensionless when the
units of the components of x are mixed, there is merit in selecting a
dimensionless reference variance. as was done in part (c) of Example 5-12, so
that the elements of Qxx are the same units as their counterparts in xx .
Now, by inverting both sides Eq. (5-78), we get
Q xx1 o2 xx1 .

(5-80)

Comparing Eq. (5-80) with Eq. (5-77). it is obvious that

W xx Q xx1 ,

(5-81)

i.e., the weight matrix is the inverse of the cofactor matrix. Equation (5-31) is
applicable to correlated as well as uncorrelated observations. The same
relationship was given previously by En. (4-46) in Chapter 4 for uncorrelated
observations only,
5.9, INTRODUCTION TO SAMPLING

We have seen that if the probability distribution of a random variable is known

we are able to make probability statements about the random variable, as
demonstrated in Example 5-3, 5-4, and 5-5. We can also calculate the mean
and variance of the random variable, as shown in Example 5-6. To say that a
probability distribution is known implies that the distribution type (uniform,
normal, etc.) is known and that the distribution parameters are fully specified.
145 Analysts and Adjustments of Survey Measurements

In some cases, it is possible to construct a useful probability distribution from

theoretical considerations by making appropriate assumptions. Probability
distributions for games of chance can be satisfactorily determined in this
manner. More often, however, it is not possible to determine a probability
distribution based only on theory. The alternative approach that must be taken
is to determine the distribution empirically; that is, by making observations. This
is wlat las to be done for the probability distribution of a survey measurement.
To say that a probability distribution can be determined from the observations is
not quite correct, if by the term determine" we mean exact evaluation. It is more
appropriate to say tlat the probability distribution is statistically estimated.
The most common observational technique used as basis for statistically
estimating the probability distribution of a survey measurement is to repeat the
measurement several times. If the repeated measurements are independent of
each other and if they are made under the same conditions, they are known
collectively as a random sample. The number of measurements in the sample is
known as the sample size.
For convenience, the notion of population is introduced. The population is the
total collection of objects or values from which a sample can be drawn. In many
cases, the population is finite and well defined, such as the people in a city, a
deck of cards, or a collection of chips in a bowl. For a survey measurement,
however, we must imagine the population to be the infinite collection of all
possible values of the measurement.
Sample size plays an important role in the estimation of a probability
distnbution. Generally speaking, the larger the sample size the better the
estimate. More on sample size will be discussed in Chapter 8. A sample of
infinite size is equivalent to the entire population. Although it is impossible in
practice to obtain a sample of infinite size, we can, nevertheless, conceptualize
such a sample as the complete set of measurements that would permit precise
determination of the probability distribution,

PROBLEMS
5-1 Random variable X las the following probability function:

px

120
0.7 x 0.35 x for x = 0, 1, 2, 3, 4, 5.
x!5 x !

(a) Evaluate and plot p (z) for all given values of z (b) Evaluate and plot F(x)
=P [X x] for -1 < x < 7. (c) Evaluate P [1 < X 3].
146 Analysts and Adjustments of Survey Measurements

5-2 A continuous random variable X has the following density function:

f x

x
18
0

for x

0<x<6

elsewhere

Derive the distribution function of X and evaluate P [X< 1], P [1 < X <3], andP [X
> 5].
5-3 X is a random variable with the following probability density function:
f x
0

2
x 1
9

for 1< X < 4

elsewhere

Derive the distribution function of X and evaluate P [X<2], P[X> 3], and P [2.5 <
X< 3.5].
5-4 The density function of a random variable X is
f x

sin x
2

for 0 < x <

Derive the distribution function of 5 and evaluate P[X< /3], P [X> /2], and
P [/3 < X < /2].
5-5 X is a uniformly distributed random variable with parameters a = 5 and b =
9. Evaluate P[X<5], P[X<8], P[X<10] , P[6<X<7], P[4<X<6 ].
5-6 Given a random variable Z with standard normal distribution, determine
from Table I of Appendix B the probability that Z takes on a value: (a) between
0.35 and 1.65; (b) between -1.96 and 42; (c) between -1.50 and 1.50; (d) less
than 0.50; (e) greater than zero; (f) greater than 3.00; (g) between -0.674 and
0.674.
5-7 A random variable X is normally distributed with mean 72 and variance 25.
Determine P [X< 77], P [X < 84 ], P [X> 84], P [67 <X< 77], P [72 < X< 73], and
P [X= 68].
5-8 Evaluate the mean and variance of the random variable in Problem, 5-1.
5-9 Evaluate the mean and variance of the random variable in Problem 5-2.

5-10 Evaluate the mean and variance of the random variable in Problem 5-3.
5-11 Evaluate the mean and variance of the random variable in Problem 5-4.
5-12 X is a normally distributed random variable. E[X] = 20, and var [ X] =4. (a)
What is the probability that X is greater than 23? (b) What is the probability
that )(lies between 19 and 21? (c) What is the probability that X is less than 18?
5-13 A random variable X is normally distributed. If E[X] =40 and P [X<8]=
0.8413, determine the mean, variance and standard deviation of X.

147 Analysts and Adjustments of Survey Measurements

5-14 A measurement is normally distributed with mean 110.156 m and standard

95,7% deviation 0.022 m. (a) Evaluate the 50% uncertainty, 90%
uncertainty,95% uncertainty and 97% uncertainty of the measurement. (b)
Evaluate the probability that the measurement lies between 110.150m and
110.170 m.
5-15 The residual error in a measurement is composed of a bias of +0.08 en
and a random component which has the following density function:

f x

x 2 1
1
exp
m .
0.15
0.0072

(a) Evaluate the precision of the measurement, as expressed by the standard

deviation, 50% uncertainty, and 90% uncertainty. (b) Evaluate the accuracy of
the measurement, as expressed by the mean square error.
5-16 X and Y are two jointly distributed, independent, normal random variables.
The mean and variance of X are 10 and 16, respectively; the mean and
variance of Y are 15 and 9, respectively. (a) Evaluate P [X 5, Y 16.5]. (b)
write the density function of Z, where Z= X+ Y. (c) evaluate P [25 Z 30].
5-17 The X and Y coordinates of a survey point have standard deviations 0.040
m and 0.030 m, respectively. (a) Evaluate the correlation coefficient of X and Y
If the covariance of X and Y is 0.0009 m. (b) Evaluate the covariance of X and
Y if the correlation coefficient is -0.30.
5-18 Two measurements are normally distributed with standard deviations 0.52
m and 0.38 m, respectively. Evaluate the standard deviation of the sum of the
two measurements if the correlation coefficient of the two measurements is: (a)
0.5: (b) 0; (c) -0.5; (d) 1.0.

5-19 Select a reference variance of 0.0009 m and evaluate the cofactor and
weight matrices of the coordinates in Problem 5-17 for each case, (a) and (b).
5.20 The position of a survey station is given by an angle and distance S. The
standard deviations of and S are 20" and 0.10 m, respectively, and the correlation coefficient is 0.50. (a) Evaluate the covariance matrix for and S, using
radians as the units for . (b) Select a reference variance of 0.0010 m and
evaluate the cofactor and weight matrices for and S. Assign appropriate units
to all elements.

148 Analysts and Adjustments of Survey Measurements

6
Variance-Covariance Propagation

6.1. INTRODUCTION

In Chapter 2 the basic technique of error propagation was introduced. Given the
errors in a set of measurements, the technique of error propagation is used to
evaluate the resulting errors in quantities that are calculated from the
measurements.
Although error propagation, as dealt with in Chapter 2, is very useful in studying
the combined influences of errors in measurements on the quantities computed
from these measurements, it is nevertheless based upon the assumption that
the measurement errors are known and given in the first place. In practice, if the
errors are known, they are invariably eliminated by applying appropriate
corrections, leaving nothing to propagate. Such is the case when we have
known systematic errors, i.e., the measurements are corrected for known
systematic. Error before they are used to calculate the desired quantities. If
random errors are considered, however, the specific values of the errors are not
known, and it is impossible to correct for them or to apply the technique of error
propagation given en in Chanter 2.
Even though the specific values of the random errors are not known, it is still
possible to study, the effects of their propagation. However, instead of working
with the actual error values, as was the case in Chapter 2. We must now work
with the probability distributions of the errors, or, equivalently, with the

probability distributions of the corresponding measurements and computed

quantities.
If the random vector x represents the set of measurements, and the random
vector y represents the set of calculated quantities such that
y f x ,

(6.1)

149 Analysts and Adjustments of Survey Measurements

Then the propagation problem basically consists of finding the joint probability
distribution of y, given the joint probability distribution of x. However derivation
of the joint distribution of y from the joint distribution function of x can be quite
involved mathematically, even when y is a relatively simple function of x, and it
is not within the scope of this text to attempt such derivation. Fortunately, for our
purposes, there is an effective alternative if we limit consideration to linear
functions of x, or to linearized forma nonlinear functions of x. This alternative
involves propagation of and covariances only.

6.2. DERIVATION OF THE PROPAGATION LAWS

Since variances and covariances are expectations, the usual procediment

deriving their propagation laws is to use the properties of expectation, another
approach is used here instead, with the hope of provide better insight into
variance-covariance propagation. This approach defining variances and
covariances in terms of sums calculated from sample of infinite size.
The concept of a sample was introduced in Chapter 5. Let x be a rata, vector
composed of two jointly distributed random variables X1 and X2. And let
X 11 X 12 X 1q
X , X ,..., X
21 22 2 q

Be a sample of size q taken from the probability distribution of x. If x1 and x2

are the mean values of X1 and X2, respectively, the deviation of the sample
vector component from its mean value is
X 1k X 1k x1

And
X 2 k X 2 k x 2 ,

k = 1, 2, , q.

Let us now define the variances of X1 and X2 as

x21 lim q

1 q
X 12k

q k 1

And

150 Analysts and Adjustments of Survey Measurements

2
x2

lim q

1 q
X 22k

q k 1

(6-4)

The covariance of X1 and X2 as

x1x 2 lim q

1 q
X 1k X 2 k
q k 1

(6-5)

The definition of variance expressed by Eqs. (6-3) and (6-4) is consistent the
definition given by Eq. (5-28) if we bear to mind that the probability of any value
xi occurring is reflected in the sample by the frequency of occurrence of that
value. Indeed, if the sample is infinite in size (q), the relative frequency of
occurrence of a particular value is precisely its probability (see Section 5,1).
Similarly, the definition of covariance expressed by Eq. (6-5) is consistent with
the discrete equivalent of the definition given by Eq. (5-62).
Let us next consider a random vector y composed of two jointly distributed
random variables Y1 and Y2 such that
y Ax b,

(6-6)

Which A is a 2 x 2 coefficient matrix, and b is a vector composed of two

constants.
Written in full, Eq. (6-6) is
Y1 a11 X 1 a12 X 2 b1
Y2 a 21 X 1 a 22 X 2 b2

Applying this relationship to the k th sample component, we have

Y1k a11 X 1k a12 X 2 k b1

(6-7)

Y2 k a 21 X 1k a 22 X 2 k b2

(6-8)

Now, if the deviations of the calculated values of Y1 and Y2 from their

respective means values, y1 and y2 are
Y1k Y1k y1

And

Y2 k Y2 k y 2 ,

K= 1, 2, , q

(6-9)

It can be show that.

151 Analysts and Adjustments of Survey Measurements
Y1k a11 X 1k a 21 X 2 k

And
Y2 k a 21 X 1k a 22 X 2 k ,

k= 1, 2, , q

(6-10)

Defining the variances and covariance of Y1 and Y2 in the same way as the
variances and covariances of X1 and X2 are defined in Eqs. (6-3), (6-4), and (65), we have

y21 lim q

1 q
Y12k

q k 1

y22 lim q

1 q
Y22k

k 1
q

(6-11)

(6-12)

And

y1 y 2 lim q

1 q
Y1k Y2 k .
q k 1

(6-13)

Substituting the right-hand side of Eq. (6-10) for Y1k in Eq. (6-11), expanding,
and following the basic rules of summation, we get
1 q
( a11 X 1k a12 X 2 k ) 2

k 1
q
1 q
lim q k 1 ( a112 X 12k 2a11 a12 X 1k X 2 k a122 X 22k )
q

y21 lim q

lim q

q
q

1 q 2
2
a

2
a
a

a122 X 22k

2k
11 1k 11 12 1k
q k 1
k 1
k 1

 a112 lim q

1 q
1 q 2
2

a
a
lim
a12 X 1k X 2 k
1k 11 12 q q
q k 1
k 1

a122 lim q X 22k

K 1

a
2
11

2
x1

2a11 a12 x1x 2 a122 x22 .

(6-14)

152 Analysts and Adjustments of Survey Measurements

Similarly, it can be shown that

2
2
y22 a 21
x21 2a 21 a 22 x1x 2 a 22
x22

And that

y1 y 2 a11 a 21 x21 a11 a 22 a12 a 21 x1x 2 a12 a 22 x22 .

(6-15)
(6-16)

Let xx and yy be the covariance matrices of the random vectors x and y,

respectively, i.e.,
2
xx x1
x1 x 2

x1 x 2

x22

(6-17)

y21
yy
y1 y 2

y1 y 2

y22

(6-18)

It is easily shown that Eqs. (6-14), (6-15), and (6-16) can be expressed in the
matrix Form
yy A xx A t ,

(6-19)

Where
a11
A
a 21

a12
.
a 22

It will be found that Eq. (6-19) holds, net only for vector x with n=2 components
and vector y with m=2 components, but for all values of m and n. As such
represents the general law of propagation of variances and covariances for all
linear functions,
y =Ax+ b.
For linearized functions, the coefficient matrix A is replaced by the Jacobian
matrix

Y1
Y1
Y1
...
X
X 2
X n
1

Y2 Y2 ... Y2
J yx X
X 2
X n .
1
...
... ... ...
Y
Y m
Y m
m

...
X n
X 1 X 2
and the general law of propagation of variances and covariances becomes

153 Analysts and Adjustments of Survey Measurements

t
yy J yx xx J yx
.

The relationships expressed by Eqs . (6-14), (6-15), and (6-16), and the Eqs. (619) and (6-20) are valid no matter what the probability of x is.
When the components of the random vector x are independent (uncorrelated),
function the covariance matrix xx is diagonal. Thus, for the linear function.
Y a1 X 1 a 2 X 2 ... a n X n

in which X1, X2, ..., Xn are independent (uncorrelated), application (6-19) yields

2
y

x21
0

0 x22

a1 , a 2 ,..., a n
...
...

0
0

.... 0
,
... ....
2
... xn

...

which results in

y2 a12 x21 a 22 x22 ... a n2 xn2 .

For the nonlinear function represented by
Y f X 1 , X 2 ,..., X n

In which X1, X2, ..., Xn, are independent (uncorrelated), application (6-20)
yields

Y Y
Y 0

,
,...,

Xn ...
X 1 X 2

2
x1

2
y

2
x2

...
...

...

...

...

0 X 1
Y
0
X
... 2
...
xn2 Y

X n

Which results in
2

Y
Y 2 Y 2
x1
x 2 ...

X 1
X 2
X n
154 Analysts and Adjustments of Survey Measurements
2
y

2
xn

Equations (6-22) and (6-24) constitute the linear and nonlinear cases,
respectively, of what is known as the special law of propagation of variances. In
Chapter 5, Eq. (5-79) gave the relationship between the covariance matrix xx
and the corresponding cofactor matrix Qxx of a random vector x.
This relationship is
xx 02 Q xx

(6-25)

in which 02 a is the reference variance.

For the random vector y, a similar relationship between the covariance matrix
yy. and cofactor matrix Qyy can be written, i.e.,
yy 02 Q yy .

(6-26)

Substituting the right-hand sides of Eqs. (6-25) and (6-26) for xx and yy ,
respectively, in Eqs. (6-19) and (6-20), and dividing by 02 we get
Q yy AQ xx A t

for the linear case

(6-27)

And
t
Q yy J yx Q xx J yx

for the nonlinear case (6-28)

Where Jyx is the Jacobian matrix.

6.3. EXAMPLES
The propagation laws derived in Section 6.2 are now illustrated by several
examples.
EXAMPLE 6-1
A distance is measured in four legs. The measured legs are 461.825 m,
215.623 m, 117.221 m, and 512.338 m, and the corresponding standard
deviations are 0.021 m, 0.017 en, 0.010 m, and 0.025 m, respectively. All
measurements are independent. Evaluate the standard deviation of the total
measured distance.
Solution
Let the random variables X1, X2, X3, and X4 represent the four individual
measurements and Y the total distance. Thus

Y X1 X 2 X 3 X 4 ,

with x1 = 0.021 m, x 2 = 0.017 m, x 3 = 0.010 m, and x 4 = 0.025 m.

155 Analysts and Adjustments of Survey Measurements
Since the measurements are independent, the special law of propagation of
variances applies. Obviously, Y is a linear function of the X's, with coefficients
a1 = a2 = a3 = a4 = 1. Thus, applying Eq. (6-22), we get

y2 a12 x21 a 22 x22 a32 x23 a 42 x24

1 0.021 1 0.017 1 0.010 1 0.025
0.001455
2

And the standard deviation of the total Measured distance is

y 0.001455 0.038 m.
EXAMPLE 6- 2
Two measurements are independent and have standard deviations 1 = 0.20 m
and 2=0.15 m, respectively. Evaluate the standard deviations of the sum and
difference of the two measurements. Also evaluate the correlation between the
sum and difference.
Solution
Let X1 and X2 represent the two measurement, and y1 and Y2 their sum and
difference respectively, i.e.,
y1 X 1 X 2 and Y2 X 1 X 2 .
In matrix farm (y= Ax b), the functions are
Y1 1 1 X 1 0
Y 1 1 X 0.
2
2

Since X1 and X2 are independent, Eq , (6-22 ) can be used to evaluate the

variances. Thus,

y21 1 12 1 22 12 22
2

0.20 0.15 0.0625 m 2 .

2

And

y22 1 12 1 22 22 0.0625 m 2 .
2

Note deviation that the sum and difference have the same variance. Thus, the
standard of both the sum and the difference is

 y1 y 2 0.0625 0.25m
156 Analysts and Adjustments of Survey Measurements
To find the correlation between Y1 and Y2, we must use the general law of
propagation of covariance, Eq. (6-19). Thus,
yy A xx A t
1 1 12 0 1 1

2
1 1 0 2 1 1
2 22 1 1 12 22
12
2
2
2
1 2 1 1 1 2

12 22
.
12 22

The diagonal elements of yy, are the variances y1 and y2, as previously
determined. Each off-diagonal element i s the covariance of Y1 and Y2, i.e.,

y1 y 2 12 22 0.20 0.15 0.0175 .

2

Thus, the correlation coefficient of y1 and Y2 is

y1 y 2
y1 y 2

0.0175
0.28.
0.250.25

EXAMPLE 6-3
If a quantity is independently measured n times, and the n measurements,
represented by X1, X2, ..., have different weights, w1, w2, wn, respectively,
then the weighted mean of the measurements is defined as
__

X0

w1 X 1 w2 X 2 ... wn X n
.
w1 w2 ... wn

Show that the weight of the weighted mean is equal to the sum of the weights of
the individual measurements.
Solution
Let w = w1 + W2 + ... + Wn In matrix form (y= Ax + b), the weighted mean is

__

X1

W W
W X2
1 , 2 ,..., 2 0.
w ...
w w

X n

Now,
157 Analysts and Adjustments of Survey Measurements

Q xx W

1
xx

1
W
1
0

...

...

0
1
W2
...
...

... 0
.
... ...
1
...

W n
...

Thus,
yy A xx A t
1
0
W
1

1
Wn 0
W1 W 2

,
,...,

W2
w
w w
..
...

0
0

W W2 ... Wn 1
1

.
w 2

W
0 1

w
W
... 0 2
w
... ... ...
1 Wn

...

W n w
...

Hence Wyy = Qyy = [w], i.e., the weight of Xw (X medium), is the sum of
the weights, w.
EXAMPLE 6-4
A distance is measured twice using two different methods. The two
measurements are X1=321.643 m and X2= 321.618 m, with the standard
deviations 1= 0.030m and 2=0,015 m, respectively. Evaluate the weighted
mean and the standard deviation of the weighted mean of the two
measurements, Assume that X1 and X2 are uncorrelated.
Solution
Let 0=0.030 m (selection of a 0 is arbitrary). Then the weights of the
measurements are

02 0.0302
w1 2
1
1 0.0302

and

02 0.0302
w2 2
4,
1 0.0152

And the weighted mean (see Example 6-3) is

158 Analysts and Adjustments of Survey Measurements

__

Xw

w1 X 1 w2 X 2 1321.643 4 321.618

321.623m
w1 w2
1 4

Now, according to Example 6-3, the weight of the weighted mean is w1 + w2 =

1 + 4 = 5, Thus, the variance of the weighted mean is

w2

02
w1 w2

0.0302
5

0.00018m 2

And the standard deviation of the weighted mean is

w 0.00018 0.013m
Note: The variance w2 can also be evaluated as follows:
w1
X
w1 w2

___

w2
X 1

w1 w2

1
4
X 2 X 1 X 2 .
5
5

Thus
2

1
4
12 22
5
5
2
w

1
4
2
2
0.030 0.015 0.00018m 2 .
5
5

EXAMPLE 6-5
The angles and sides of a right-angled plane triangle are shown in Fig. 6-1.
Sides a and b are measured and the measurements are 416.050 m and
202.118 m, respectively. The measurements are independent, the standard
deviation of a is 0.020 m, and the standard deviation of b is 0.012 m. Evaluate
the elements c and of the triangle and their standard deviations. Also
determine the correlation, if any, between c and .

Fig. 6-1.
159 Analysts and Adjustments of Survey Measurements
Solution.
The functions which express c and in terms of a and b are
c a2 b2
b
sin 1
a
Using these functions to evaluate c and , we get
c

416 .050 2 202 .018 2

363 .656 m.

 sin 1

202.018
29 03"54.
416.050

Since the functions are nonlinear, the Jacobian matrix Jyx must be evaluated,
noting that
a
x
b

c
y

and

Now
c
2

a b2
a a

1 2
a b2
2

c
2

a b2
b b

1 2
a b2
2

1/ 2

1/ 2

2a

2b

1/ 2

1/ 2

1
b
b

sin 1
2
2
a a
a
1 b / a a

1
b
1

sin 1

2
b b
a
1 b / a a

a
2

2 1/ 2

b
b

2 1/ 2

a
c

b
c

b
b
2
ac
a b a
2

1 1

a b a c
a

Thus,
J yx

c a
b c
b

b ac

b

1.144
c
1 1.336 10 3

0.556

.
2.750 10
3

Now,
0.0202
0
xx

0.0122
0
160 Analysts and Adjustments of Survey Measurements

Therefore, applying the general law of propagation of variances and

covariances, we get
yy J yx xx J yx
0.556 0.0202
1.144

3
2.750 10 3 0
1.336 10
5.680 10 4

7
8.315 10

1.144 1.336 10 3

0.0122 0.556 2.750 10 3

0

8.315 10 7
.
1.803 10 9

Thus

c 5.680 10 4 0.024m.
1.803 10 9 4.246 10 5
And the correlation coefficient of c and is

radians = 8.8

8.315 10 7
0.816 .
0.024 4.246 10 5

6.4. STEPWISE PROPAGATION

In general, field measurements are considered to be independent or.
Uncorrelated. Thus, when the random vector x represents field measurements,
the special law of propagation of variances, given by Eqs. (6-22) and (6-24), is
applicable.
As long as the derived quantities are calculated directly from the uncorrelated
field measurements, application of the special law presents no problem.
However, many surveying calculations are made in more than one step. This
means that intermediate quantities are calculated as steps in the overall
process of evaluating the final quantities, and it is very likely that these
intermediate quantities are correlated. If the special law of propagation of
variance is applied in stepwise fashion as well, and the intermediate quantities
are correlated, incorrect values for the variances of the final quantities will be
obtained, This point is illustrated by the following example.
EXAMPLE 6-6
Let X1 and X2 represent two field measurements. The quantity Z is calculated
from X1 and X2 in two steps using two intermediate quantities, Y1 and Y2. The
calculations are:

161 Analysts and Adjustments of Survey Measurements

Step 1: Y1 3 X 1 X 2 2 ,
Y2 X 1 2 X 2 3
If X1 and X2 are independent (uncorrelated), and their standard deviations are
5.0 cm and 3.0 cm, respectively, evaluate the standard deviation of Z.
Solution I
Express Z directly in terms of X, and X2, and apply the special law of variance
propagation in one step.
Z Y1 Y2 3 X 1 X 2 2 X 1 2 X 2 3
4 X 1 3 X 2 5.

Thus,

z2 4 5.0 3 3.0 481cm 2

2

And

z 481 21 .9cm.

Solution II
Apply the special law of variance propagation to each step of the calculation.
Step 1.
Y1 3 X 1 X 2 2

y21 3 5.0 1 3.0 234 cm 2

2

Y2 X 1 2 X 2 3

y22 1 5.0 2 3.0 61cm 2

2

Step 2.
Z Y1 Y2

y2 y21 y22
234 61 295cm 2
y 295 17 .2cm.

This solution is Incorrect; see the discussion following the example. Solution III
Apply the general law of variance and covariance propagation to each step of
the calculation,

162 Analysts and Adjustments of Survey Measurements

Step 1.
Y1 3 1 X 1 2
Y 1 2 X 3
2
2
yy

2
3 1 5.0

1 2 0

3 1 234 93

3.02 1 2 93 61
0

Step2.

Z 1

Y1
1
Y2

234
zz 1 1
93

93 1
481
61 1

z2 481cm 2
z 481 21.9cm
Solutions I and III agree, and are correct. Solution II is incorrect because it fails
to take into account the correlation between Y1 and Y2. The offdiagonal

elements of yy in Solution III (93 cm ) are the covariance of Y1 and Y2. The
correlation coefficient calculated from this covariance is

93

0.78.

234 61

Note that the variances in yy of Solution III (234 and 61) agree with the
variances calculated for Y1 and Y2 in Solution II. The problem with Solution II is
not in the calculation of the variances. but its failure to take the covariance into
account.
The fore going example illustrates the danger in using the special law of
propagation in steps. If the propagation can be done in one step using field
measurements that are independent, the special law can be applied, as it was
in Solution I. If the propagation is to be done in two or more steps, the general
law should be used, as it was in Solution III.
EXAMPLE 6-7
Stations A and C in Fig. 6-2 are at the same elevation. Independent field
measurements are made of the base b, horizontal angles and , and vertical
angles and .

163 Analysts and Adjustments of Survey Measurements

The measurements and their standard deviations are:

b=127,552m
=8241'13"
= 8020`42
=1215'37"
=1220'23"

b=0.008m
= 10 =4.85 x (10exp -5)
radians
= 10 =4.85
x (10exp -5)
radians
= 15 =7.27
x (10exp -5)
radians
= 15" =7.27
x (10exp -5)
radians.

Evaluate the height h, and the standard deviation of h using: (a) computation in
one step; (b) computation in two steps.
Solution
(a) Computation in one step. Referring to Fin. 6-2, sides a and c can be
calculated from base b and angles and :
a

b sin
b sin
,c
sin
sin

And height h can be calculated in two ways:

h1 a tan

and

h2 c tan

If the average of h1 and h2 is taken as the final value for h we have

164 Analysts and Adjustments of Survey Measurements
h 1 / 2 h1 h2 h
1 / 2 a tan c tan
b

.
2 sin

This function is the basis for a one-step computation of h; thus

h

127.552 sin 82 4113"tan 12 1537" sin 80 2042" tan 12 2053"

2 sin 163 0155"

94 .232 m.

The function is also the basis or application of the special law of propagation of
variance. Since it is a nonlinear function of b, , , , and , it is first necessary
to evaluate the partial derivatives of h with respect to the five measured
quantities. Thus,
h h 94.232

0.739
b b 127.552
h b sin tan cos sin tan sin tan cos

2
sin 2

b tan cos
h cot 315m / rad
2 sin
h b tan cos

h cot 317m / rad

2 sin

h b sin sec 2

227m / rad

2 sin

h b sin sec 2

226m / rad

2 sin
Applying the special law of propagation of variance, we lave
2

h
h 2 h 2 h 2 h 2
b2


b




2
2
2
2
5
0.739 0.008 315 4.85 10 317 4.85 10 5
2

2
k

227 7.27 10 5
2

=0.00105.
Thus k 0.0105 0.0032m.
165 Analysts and Adjustments of Survey Measurements
(b) Computation in two steps. The height h can be computed in the following
two steps:
Step 1:

b sin
127 .552 sin 82 4113"
a

433 .508 m
sin
sin 163 0155"

b sin
127 .552 sin 80 2042"
c

430 .873 m
sin
sin 163 0155"
Step 2:
h 1 / 2 a tan c tan

1 / 2 433.508 tan 12 1537" 430.873 tan 12 2023"

94 .232 m.

In this computation, a and c are intermediate quantities that are very likely
correlated because they are functions of the same measured quantities, b, ,
and . Thus, if k is to be evaluated in steps, the general law of propagation of
variance and covariance should be used.
Let x be the vector of field measurements, y be the vector of intermediate
quantities, and h the height vector, consisting of the single element, h. Thus,
b
a

c

x , y , and h=[h].






Applying the general law of propagation of variances and covariances, we have:
Step 1: yy J yx xx J yxt ;

Step 2: kk J ky yy J kyt ;
Where

 a a a a a
b

c c c c c
J yx b

b

b
166 Analysts and Adjustments of Survey Measurements

And
h h h h
J ky , ,
, .
a c

The nonzero elements of Jyx are:

Thus:

J yx

3.399 1476 1421 0 0

3.378 1412 1485 0 0

0
0
0
1 0

0
0
0 1
0

The elements of Jky are:

h 1
tan 0.1087
a 2
h 1
tan 0.1094
c 2

167 Analysts and Adjustments of Survey Measurements

h a
sec 2 227
2
h c
sec 2 226
2

Thus Jky = [0.1087, 0.1094, 227, 226].

Now, the covariance matrix of the yield measurements is

And kk J yx yy J kyt 0.00104. Since the single element of kk is k. The

standard deviation of h is k 0.00104 0.032m.
In this particular example, the one-step calculation is the preferred solution
because it obviously takes less work than the two-step calculation. However,
there are many problems in which it is very difficult or awkward to express the
final desired quantities as single functions of the field measurements. In these
cases, stepwise computation is the preferred procedure, and stepwise variancecovariance propagation must follow. If the propagation must be done in steps, it
is essential that it be done using complete covariance matrices, i.e., using the
general law of propagation of variances and covariances.
6.5. PROPAGATION FOR LEAST SQUARES ADJUSTMENT OF INDIRECT
OBSERVATIONS
In Chapter 4, the condition model for adjustment of indirect observation was
given by the matrix equation.
v B f .

168 Analysts and Adjustments of Survey Measurements

(6-29)

In this model, v is the vector of residuals, is the vector of unknown parameters, B is the matrix of numerical coefficients of the unknown parameters, and
f d l,

(6-30)

Where l is the vector of observations and d is a vector of numerical constants.

Associated with the vector of observations l is a covariance matrix ll as well as
a cofactor matrix Qll and a weight matrix Wll. In dealing with propagation in a
least squares solution, it is convenient to work with cofactor matrices.
Accordingly, propagation in this section is presented with respect to cofactor
matrices.
In keeping with the symbols used in Chapter 4, the subscript "ll" is dropped from
Qll , and Wll, i.e., for the vector of observations l, Q represents the cofactor
matrix, and W (the inverse of Q), represents the weight matrix.
Now, for the linear function y = Ax+ b, the law of propagation of cofactors, as
given by Eq. (6-27), is
Q yy AQ xx A t .

Expressing Eq. (6-30) in the y= Ax + b form we have

f I l d .

(6-31)

Applying the law of propagation of cofactors to Eq. (6-31) we get

Q ff I Q I Q.
t

(6-32)

Thus the cofactor matrix of l is also the cofactor matrix of f.

The steps of the least squares adjustment of indirect observations, as provided
in Chapter 4, are
N B tWB
t B tWf
N l t
v f B

(6-33)
(6-34)
(6-35)
(6-36)

And
^

l lv

(6-37)

169 Analysts and Adjustments of Survey Measurements

Here N is the coefficient matrix of the normal equations, and ^l is the vector of
adjusted observations.

Restating Eqs. (6-34) through (6-37) in the y = Ax + b form, we have

t B tW f

(6-38)
(6-39)

N l t
v f B

F BN l t
F BN l B tWf

I BN l B tW f

(6-40)

And
^

l lv

l f B
B d .

(6-41)

In each case, what is in the parenthesis ( ) constitutes the A matrix. Applying

Eq. (6-27), the law of propagation of cofactors, to Eqs. (6-38) through (6-41), we
get

All intermediate steps in Eqs. (6-42) through (6-45) are Left as an exercise for
the student.
From Eqs. (6-44) and (6-45), it is easily seen that
Q ^ ^ Q Qvv
ll

(6-46)

Covariance matrices for t, , v, and ^l are obtained by multiplying the corresponding cofactor matrices by the reference variance o.
170 Analysts and Adjustments of Survey Measurements
EXAMPLE 6-8
For the level net in Example 4-7, calculate the covariance matrix for the
adjusted elevations of points B, C, and D, for the case of weighted
observations, given a reference variance of 0.0016 m. Also calculate the
standard deviations in the adjusted elevations of B, C, and D.
Solution

According to Eq. (6-43), and using the data in Example 4-7, we have:

Q N

0.337902

symmetric

0.171085
0.406418

0.183543
0.172880 .

0.298256

The required covariance matrix is thus

0.0016 N
2
0

5.406 2.738 2.936

10
6.502 2.766 m 2

4.773
4

The standard deviations in the adjusted elevations are

B 10 2 5.406 0.023 m

C 10 2 6.502 0.025m

D 10 2 4.773 0.022 m .
A summary of the symbols and equations used in propagation for least squares
adjustment of indirect observations will be found in Section 9.5, Chapter 9.
6.6. PROPAGATION FOR
OBSERVATIONS ONLY

LEAST

SQUARES

ADJUSTMENT

OF

The condition model for adjustment of observations only was also given in
Chapter 4. In matrix notation, this model is
(6-47)
Av f .
In this model, v is the vector of residuals, A is the matrix of numerical coefficients of the residuals, and
f d Al , (6-48)

171 Analysts and Adjustments of Survey Measurements

where l is a vec tor of observations and d is a vector of numerical constants.
Note that the f vector defined by Eq. (6-48) is different from the f vector defined
by eq . (6-30).
Propagation will again be presented with respect to cofactor matrices, with Q
and W representing the cofactor and weight matrices, respectively, of l.
Expressing Eq. (6-48) in the y = Ax+ b form, we have
f Al d

(6-49)

Applying the law of propagation of cofactors to Eq. (6-49), we get

Q ff AQ A AQA t .
t

(6-50)

The steps of the least squares adjustment of observations only, as given in

Chapter 4, are

And
^

l l v,

(6-55)

where Qc is a special cofactor matrix (see Chapter 9 for explanation), Wc is the

corresponding weight matrix, and k is the vector of Lagrange multipliers.
Obviously, from Eqs. (6-50) and (6-51), the cofactor matrix of f is equal to Q,,
i.e.,
Q ff Qc

(6-56)

Applying the law of propagation of cofactors to Eqs. (6-53) and (6-54), we get
Qkk Wc Q ff Wc Wc
t

And

Q xx QA t Qkk QA t

(6-57)

QA tWc AQ, (6-58)

where all cofactor and weight matrices are symmetric.

To get the cofactor matrix of vector ^l , we must first express ^l as a linear
172 Analysts and Adjustments of Survey Measurements
function of only. Thus, From Eqs . (6-55), (6-54), (6-53), and (6-48), respectively, we obtain

from which we get

(6-59)

Applying the law of propagation of cofactor to Eq. (6-59) we get

which reduces to
Q ^ ^ Q QA t Wc AQ .
ll

(6-60)

The reduction is left as an exercise for the student. From Eqs. (6-58) and (6-60
), it is obvious that
(6-61)
Q ^ ^ Q Qvv ,
ll
which is the same relations hip as given by Eq. (6-46). Again, covariance
matrices for k, v, and ^l are obtained by multiplying the corresponding cofactor
matrices by the reference variance, 0
EXAMPLE 6-9
Calculate the covariance matrix and the standard deviations of the adjusted
angles in the plane triangle of Example 4-10, for case (2), unequal weights,
given ~o= 10".
Solution
From Eq. (6-53) and the dam in Example 4-10,

173 Analysts and Adjustments of Survey Measurements

Then, from Eq. (6-61), we get

Thus, the covariance matrix of the adjusted angles is

and the standard deviations of the adjusted angles are

1 78 8.8"
2 100 10"

3 111 10.5"
A summary of the symbols and equations used in propagation for least squares
adjustment of observations only will be found in Section 9.5, Chapter 9,
PROBLEMS
6-1 The interior angles of an n-sided closed traverse arc measured. All
measurements are uncorrelated. If the standard deviation of each measurement
is 4.0 evaluate the standard deviation of the sum of the measured angles for:
(a) n = 5; (b) n =10; (c) n = 20.
6-2 Two interior angles of a plane triangle are measured and the third angle Is
with computed from them. If the two angle measurements are uncorrelated,
standard deviations 4,5 and 6.0, respectively, evaluate the standard deviation
of the computed angle.
174 Analysts and Adjustments of Survey Measurements
6-3 A distance is measured in three segments. Each segment is measured
more than once and in each case a mean segment length (simple average) is
obtained. The distance is then computed as the sum of the mean segment
lengths. Following are the measurements, all uncorrelated, in meters:

Compute the distance and determine the standard deviation of the computed
distance if the standard deviation of a 100 m measurement is 0.010 an and the
variance of each measurement is directly proportional to its magnitude.
6-4 A distance is measured using three different methods. The observed values
and their standard deviations (all in meters) are shown below:

Determine the weighted mean of the observed distances and the standard
deviation of this weighted mean, assuming the three observations are independent.
6-5 Six independent determinations of the elevation of a point are made. These
values and their corresponding weights are shown below.

Compute the weighted mean of the six elevations and evaluate the standard
deviation of this weighted mean if a weight of 2 corresponds to a standard
deviation of 0.030 m.
6-6 The area of a trapezoidal parcel of land is computed as follows:
175 Analysts and Adjustments of Survey Measurements
a a2
Area 1
b1
2

where a1, a2, and b are independently measured dimensions. If measured

values for a1, a2, and b are 319,414 m, 481,112 m, and 502.307 no,
respectively; and their standard deviations arc 0.030 m, 0.042 m, and 0.020 m,
respectively, compute the area of the parcel and the standard deviation of this
computed area.
6.7 Figure 6-3 shows a small triangulation scheme in which side m is calculated
from independent measurements of base b and angles , , , and , as follows;
mb

sin sin
.
sin sin

If the measurements and their standard deviations are

6-8 compute or and its standard deviation.

The two components of vector x are related to the three components of vector u
as follows:
x1 2u1 3u 2 u 3 and x 2 u1 u 2 u 3

If the covariance matrix of u is

xx

1 1 0
1 2 1

0 1 2

176 Analysts and Adjustments of Survey Measurements

Evaluate the covariance matrix of x.
6-9 The position of a survey station, P (x, y) in Fig. 6-4, is computed from independent measurements of the distance r and angle a. If the measured values
are r=250.000 m and =3000'00" and their standard deviations are 0.050 m
and 20", respectively, compute x and y and their standard deviations. Determine
also the correlation coefficient for x and y.
6-10 The plane coordinates of P, Fig. 6-4, and their standard deviations are: x=
300.000 m, y=4000.000 to. x, =0.040 m, y=0.050 m. The correlation
coefficient for x and y is -0.70. Compute the polar coordinates r and of P and
their covariance matrix. Then compute the standard deviations of rand a (in
seconds of arc) and their correlation.
6 - 11 The azimuth and length s of line CD in Fig. 6-5, are measured. The
observed values and their standard deviations are:

130 0000" 10"

s 2000 .00 m s 0.08 m.

177 Analysts and Adjustments of Survey Measurements

The position of C is x= 4000.00 m, y = 5000.00 m; and x= y = 0.10 m.
Assuming all given data are uncorrelated, determine the position (x, y) of D and
the covariance matrix associated with this position.
6-12 X1, X2, and X3 are independent measurements with standard deviations
2.0 cm, 4 .0 cm, and 2.0 cm, respectively. Y1 and Y2 are computed from X1,
X2, and X 3
as follows:
Y1 X 1 X 2 2X 3
Y2 X 1 2 X 2

And Z1 and Z2 are computed from Y1 and Y2 as follows:

Z 1 0.5Y1 Y2
Z 2 Y1 0.5Y2

Evaluate the covariance matrices for the random vectors

Y1
Z1
Y and Z
Y2
Z 2

6 - 13 The coordinates of P1 in Fig. 6-6 are x1 =1000.00 m, y1 = 1000.00 m; the

coordinates of P2 are x2 = 1100.00 m, y2 = 1600.00 m. The covariance matrix
for the vector (x1, y1, x2, y2) t is

178 Analysts and Adjustments of Survey Measurements

Evaluate the covariance matrix for x and y, where x=x2 - x1and y= y2 y1. Then evaluate the covariance matrix for the length s and azimuth of P1 P2
and determinate the standard deviations matrix for and correlation coefficient.
6-14 the error R In centering theodolite A t over station A in Fig, 6.7 produces
an error = 2 - 1 in the angle measured between targets T1 and T2. The
distances A t to T1 and T2 are D1 and D2 respectively; the distance between T1
and T2 is D3 Let X and Y be rectangular components of R, as shown. If X and Y
are independent random variables, each with standard deviation , show that
the resulting standard deviation in is

D3
radians.
D1 D2

(1) Derive and use the relationship U= Y cos - X sin , where U is the
displacement of A' from AT, as shown. (2) erros 1 and 2 , in ,radians, can be
approximated by U/D1 and Y/D2 , respectively.
6-15 Determinate the standard deviations of and ^b computed in example 4-6,
case (2). Chapter 4.
6-16 Determinate the standard deviations of the least square estimate for x in
example 4-9, Chapter 4.
6-17 With the reference to Example 4-11, Chapter 4, determinate the standard
deviation of each adjusted angle in the reference standard deviation 0, is 12.
6-18 With reference to example 4-8, evaluate the standard deviations for the
computed values of Tmax and hmax if the standard deviation in observing
altitude with the theodolite is 10.

179 Analysts and Adjustments of Survey Measurements

6-19 Refer to the level network in Problems 4-8 and 4-11, Chapter 4. Use
Problem
4-8 solution data and Eq. (6-45) to evaluate Qjl, tle cofactor matrix for tle
adjusted elevation differences. Then use Problem 4-1l solution data and Eq. (660) to evaluate Qjl . Compare the two evaluations.
6-20 With reference to Problem 4-9, Chapter 4, evaluate the covariance matrix
for tle least squares estimates for x and y and the covariance matrix for the
adjusted observations. Check the evaluation of the covariance matrix for the
adjusted observations by using Problem 4-12 solution data.

180 Analysts and Adjustments of Survey Measurements

7
Preanalysis of Survey
Measurements

7.1 PREANALYSIS PROCEDURE

Preanalysis
of survey measurements is analysis of the component
measurements of a survey project before the project is actually undertaken.
Preanalysis is very helpful overall design of the survey project because it
provides a basis for evaluation of the accuracies of the survey measurements,
for meeting tolerances that may be imposed on these measurements, and for
selection of suitable instrumentation and measurement procedures.
In this introduction to preanalysis, we shall assume that all components of a
survey measurement are free of bias caused by systematic error. This means
that variances, or standard deviations, or multiples of standard deviations can
be used as measures of accuracy as well as measures of precision, and that
the propagation laws of Chapter 6 can be applied. We shall further assume that
all measurement components are independent , so that the special law of
propagation of variances, as expressed by Eqs. (6-22) and (6-24) in Chapter 6,
can be used. For convenience, the special law of propagation of variances is
expressed once more below.
Linear functions. If Y is a linear function of the independent measurements
X1, X2,..., Xn, i.e.,
Y a1 X 1 a 2 X 2 ... a n X n

Then

y2 a12 x21 a 22 x22 ... a n2 xn2 ,

(7-1)

where a1 , a2, ..., an, are constant coefficients; x1, x2, ..., xn are the
respective variances of X1, X2, ...,Xn; and y is the variance of Y.

181 Analysts and Adjustments of Survey Measurements

Nonlinear functions. If Y is a nonlinear function of the independent
measurements X1, X2, ..., Xn, then

(7-2)
Preanalysis not only makes use of Eqs. (7-1) and (7-2) directly, but inversely as
well in order to evaluate the required accuracies of the input measurements X,
given the required accuracy of the resultant quantity Y. If we assume that each
input measurement contributes equally to the total accuracy of the end result,
then for the linear case we have

(7-3)
from which we get

(7-4)
where l ai l denotes the absolute value of ai.
For the nonlinear case, again assuming an equal contribution from each input
measurement, we have

(7-5)
from which we get

(7-6)
where a Y

X i

is the absolute value of Y X i .

When each input measurement contributes equally to the accuracy of the end
result, the measurements are said to have balanced accuracies.
EXAMPLE 7.1
The dimensions of the base of a large rectangular reservoir are 85 no and 60 m,
to the nearest meter, If the area of the reservoir 's base is to be determined with
a standard
182 Analysts and Adjustments of Survey Measurements

deviation of 0.6 m, evaluate the required standard deviations of the

measurements of the reservoir's dimensions, assuming the measurements
have balanced accuracies.
Solution
The area of the reservoir's base is
A LW ,

where L and W are the length and width of the base, respectively.
Obviously, n =2, and for purposes of analysis, L = 85 in and W = 60 in.
Now A = W and A L . and so from Eq. (7-6), we have
w
L

And

Thus, for balanced accuracies, the length and width of the reservoir's
base should be measured with standard deviations of 7 mm and 5 mm,
respectively, if a standard deviation of 0.6 m 2 in the area is to be
achieved.
Often it is not possible to have a set of input measurements with
accuracies that are exactly balanced, because of limitations in the
instruments and/or procedures used in making the measurements. In
such cases, trade-offs are made in which the standard deviations of one
or more components are necessarily increased to reflect the imposed
instrument/procedure limitations, while the standard deviations in the
remaining components are correspondingly decreased.
EXAMPLE 7-2
The height h of a survey station A (Fig. 7-1) above the instrument center
at B is to be determined with a standard deviation of 0.010 m from
measurements of the slope distances, the vertical angle a. and the
target height t. The function used is
h =s sin t.
For the purpose of preanalysis, estimated values for s and are 400 in
and 30, respectively.
(a) Evaluate the standard deviations in measuring s, and t, assuming
balanced
183 Analysts and Adjustments of Survey Measurements

Accuracies. (b) if the standard deviation in measuring is limited by the

used to 5.0", reevaluate the standard deviations it, measuring s and t to
accommodate this limitation in .
Solution
(a) Since h is a nonlinear function of three input measurements, Eq. (7-6)
is used, with n= 3. For s=400 m and =30, the partial derivatives are
h
=sin = 0.500; h
= s cos = 346 m; h
= -1. If 1 =0.010 m
t
s

and balanced accuracies are required, we have

(b) From Eq. (7-2),

If = 5.0" =2.4 x 10 5 radians, and k, must still be 0.010 m, then

184 Analysts and Adjustments of Survey Measurements

Balancing the accuracies of the two remaining measurements, s and t, we

obtain

It is convenient to summarize the results of the preanalysis in tabular

form, thus providing an accuracy budget for the measurements. An
accuracy budget for part (b) of , Example 7-2 is presented in Table 7- 1

In surveying fieldwork, the principal measured quantities are angles,

distances, and differences in elevation. In the subsequeat adjustment and
analysis of a survey project, prior knowledge of the accuracies, or relative
accuracies, of the principal measurements is needed so that appropriate
covariance or cofactor matrices can be constructed. Te get realistic values
for the accuracies, a preanalysis of the measurements m, have to be
made.
If the measurement of angle, distance, or elevation difference is analyzed
into its component parts , we find that many things contribute to the
overall accuracy of the measurement, This is ilustrated in each of the
three sections which follow,
7.2. HORIZONTAL ANGLE MEASUREMENT WITH A THEODOLITE

The measurement of a horizontal angle with a theodolite is influenced by

ma ny sources of error. These include: axial errors of the, instrument,
errors

185 Analysts and Adjustments of Survey Measurements

in centering and leveling the theodolite over the survey station, errors in
aligning or centering the survey targets, errors in pointing at the survey
targets., and errors in reading the horizontal circle. The overall accuracy
of the measured angle, therefore, depends on the magnitudes of these
component errors, as well as on other important influences, such as tripod
instability and atmospheric refraction.

Axial errors of the theodolite are essentially eliminated by following the

standard procedure of observing in paired setsonce with the telescope
in the direct position (face left), and once with the telescope in the reverse
position (face right)and taking a mean. It is assumed that this procedure
is allways, followed, without exception.
The error in centering the theodolite over the survey station is basically a
two-dimensional position error that is assumed to have a bivariate normal
distribution (see Chapter 8 for discussion of the bivariate normal
distribution) the coordinate components of which are independent and
have the same standard deviation, c. It can be shown (see Problem 614) that the standard deviation of the error in the measured angle
caused by the error in centering is

(7-7)
where D1 and D2 are the distances from the theodolite station A (Fig. 7
2), to targets T1 and T2, respectively, and
(7-8)

186 Analysts and Adjustments of Survey Measurements

is the distance between T1 and T2, D1, D2 , D3, and c in Eq. (7-7) used
be expressed in the same units. For tripod-mounted c ranges typically
from 0,5 to 3 mm.
In general, the error in leveling the theodolite has negligible e ffect on the
measurement of the angle, so long as the instrument is leveled carefully
and the inclinations of the lines of sight are not excessive.
It is easily seen (Fig. 7-3) that the error in the measured angle caused
by

an error in target alignment et is

(7-9)
If both targets arc aligned independently and with the same accuracy
(represented by standard deviation t) at distances D1 and D2
respectively, from the theodolite, application of Eq. (7-1) yields the
following expression for the standard deviation ot of the error in the
angle caused by errors in target alignment:

(7-10)
D1, D2, and t in Eq. (7-10) must be expressed in the same units; a,
ranges typically from 0.5 to 5 mm.
The standard deviation p in pointing at a survey target is influenced by
the telescope optics, the target design, and the atmospheric conditions
between the telescope and target. Four pointings are made for each
paired set of observations (two pointings at each target). Thus, for n sets
of observations, 4n pointings are made.
Under the assumption that all pointings are independent, and the mean of
the n sets is taken, application of Eq. (7-1) results in the following
standard deviation op for the error in the measured angle caused by
pointing error:

(7-11)

187 Analysts and Adjustments of Survey Measurements

The range of values for p is typically 1" to 4".
Error in reading the horizontal circle is introduced only two times in the 2n
observations of the angle if a repeating instrument is used, and 4n times if
a direction instrument is used. Accordingly, if r is the standard deviation
in reading the horizontal circle, the following formulas for the standard
deviation op of the error in the measured Angle caused by reading error
are easily derived from Eq. (7-1)

For repeating theodolites;

(712)

For direction theodolites;

(713)

The range of values for r is typically 1" to 10 depending upon the particular instrument used.
Since the errors in centering the instruments, allinging the targets,
pointing, and reading the circle are additive, the combined standard
deviation in measuring the horizontal angle can be expressed as follows,
in accordance with Eq. (7-1):
(7-14)
Care must be taken to express all terms in the same units, usually
seconds of arc.
EXAMPLE 7-3
A repeating theodolite is to be used to mea sure the angles of a traverse,
the sides of which are essentially in a straight line. The length of each
traverse side is approximately 800.m. The standard deviations in
i nstrument centering, target alignmen t, pointing and circle reading are,
respectively c = 2. 0 mm; t = 4.0 mm p = 2.0"; and r=6.0". Evaluate
the expected standard deviation in . measuring each angle of the traverse
if: (a) one set of observations is taken; (b) four sets are taken.
Solution
Since the traverse is in a straight line and each side is 800 m in length, D1
=D2 = 800 m, and D3= 1600 m.
(a) For one set of observations (n =1),

188 Analysts and Adjustments of Survey Measurements

Thus

(b) For four sets of observations (n = 4),

as before, and

Thus,

7.3. DISTANCE MEASUREMENT BY EDM

The slope distance S between an electro-optical distance me t er (EDM )
and its reflector is given by the function

(7-15)
Where . Is the wavelength of modulation;
189 Analysts and Adjustments of Survey Measurements
m is the whole number of wavelengths in the measurement (assumed
errorless);
u is the fractional part of a half-wavelength (~/2 ) obtained by measuring
phase difference;
k is the zero correction, which includes the instrument and reflector
constants.
The wavelength of modulation, in turn, can be expressed as follows:

(7-16)
Where
c is the velocity of light in vacuum;
n is the index of refraction of the atmosphere;
f is the modulation frequency.
Under the assumption that the errors in , u, and k are independent, the
special law of propagation of variance, Eq. (7-1), can be applied to Eq. (715) to yield

(7-17)
Similarly, assuming that errors in c, n, and f are independent, Eq. (7-2) can
he applied to Eq. (7-16) to yield

(7-18)
Combining Eqs, (7-17) and (7-18), we get

(7-19)
For the purpose of evaluating variance, we can approximate the slope
distance by using the first term (by far the largest) of Eq. (7-15) only, i.e.,

(7-20)
Then
(7-21)
190 Analysts and Adjustments of Survey Measurements
Letting

And taking the square root of both sides of Eq. (7-21), we get
(7-22)

Factors a and b are generally given as accuracy parameters for a

particular EDM instrument. The ratios (relative accuracies) , , and
c

f
f

, and the resulting factor b are customarily given in parts per million

(ppm).
EXAMPLE 74
The velocity of light is known to be 299792.5 km/sec with a standard
deviation of 0.1 km/sec. If the index of refraction of the atmosphere can be
determined with a relative accuracy n/n of 2.0 ppm, the modulation
frequency of a particular EDM instrument can be determined with a
relative accuracy f/f of 1.0 ppm, phase difference can be measured with
a standard deviation of 5.0 mm, and the zero correction can be
determined with a standard deviation of 2.5 mm, evaluate the factors a
and b, and determine the standard deviation of a distance measurement of
2 km.
Solution
We are given that c= 299792.5 km/sec and c= 0.1 km/sec. Then

and since n/n=2.0ppm, and f/f=1.0

Also, m = 5.0 mm, and k =2.5 mm. Therefore,

For S=2 km=2.0x 10 6 mm

7.4. ELEVATION DIFFERENCE BY DIRECT LEVELING

The direct leveling procedure involves observing a vertically he ld rod with
a surveyor's level that has been suitably "leveled" before each
observation,
191 Analysts and Adjustments of Survey Measurements
Either by centering a spirit bubble or by using a pendulum compensator.
The instrument leveling operation amounts to placing the axis of
collimation of the instrument into a horizontal position (if the instrument is
perfectly adjusted) or into a position that has a very small inclination (the
usual situation).
Figure 7-4 illustrates the leveling geometry. Sight distances are
represented by D (backsight distance) and Db (foresight distance), rod
readings by r (backsight) and rb (foresight), and the small inclinations
(due to imperfect adjustment) of the axis of collimation by and b. If

proper leveling procedure is followed, and b should be equal, within

the accuracy of the spirit bubble or pendulum compensator. Furthermore,
the sight distances D and Db should be relatively short and
approximately equal, thus eliminating the effects of earth curvature and
atmospheric refraction.
With reference to Fig. 7-4, the difference in elevation between A and B is
(7-23)
Where , and b are expressed in radians.
For equal sight distances (D =Db =D), Eq. (7-23) reduces to
(7-24)
Obviously, if = b as is intended in the leveling procedure, Eq. (7-24)
reduces to the simple relationship
(7-25)

192 Analysts and Adjustments of Survey Measurements

However, for analysis purposes, and b must be treated as
independently observed quantities, whether they are equal or not, and so
Eq. (7-24) must be used.
If the same care is taken each time in leveling the instrument, we can
assume that and b have the same accuracy, represented by standard
deviation . It is further assumed that rod readings r and rb are made
with the same accuracy, represented by standard deviation r and that r
is proportional to sight distance, i.e.,
(7-26)

where is the standard deviation per unit length of sight distance.

Typical values for 0 range from 0.2 to 2.5'; typical values for range
from 0.001 to 0.015 mm/m. Applying the special law of propagation of
variance to Eq. (7-24), we get
(7-27)
which reduces to
(7-28)
Equation (7-28) is for one instrument setup only. For n setups, as
occurring in a line or circuit of leveling, the special law of variance
propagation yields
(7-29)
Since the total length of a line or circuit of leveling is given by
(7-30)
it follows from Eq. (7-29) that the standard deviation in h is:
(7-31)
EXAMPLE 7.5
A surveyor's level is equipped with a spirit bubble that can be centered
with a standard deviation of 2.0. The Instrument and accompanying rod
are designed such that the rod can be read with a standard deviation of
0.012 mm per meter of sight distance. If sight distances of 60 m are used,
evaluate the expected standard deviation in determining a differe nce in
elevation with this level and rod for a level line that is 10 km long.
193 Analysts and Adjustments of Survey Measurements

Solution

Thus, from Eq. (7-31),

7.5. SURVEY TOLERANCES

The tolerance of a measurement is its maximum acceptable error. In order
to state with reasonable assurance that the error of a measurement will be
less than its specified tolerance, it is necessary to equate the tolerance to
a range of uncertainty that carries a very high level of probability. The
most commonly used uncertainty for this purpose is 3, which, under the
assumption that the measurement is normally distributed, reflects a
probability of 0.997. Accordingly, a measurement is said to meet its
specified tolerance t if its standard deviation a is such that
3 t .

(7-32)

Obviously, if a measurement is to be acceptable, the procedure for making

the measurement must be in accord with the measurement's specified to
tolerance. This means that the accuracy of each component of the
measurement and other relevant parameters must be specified. The
following example illustrates this point.
EXAMPLE 7-6
The surveyor's level and rod in Example 7-5 are to be used to determine
the difference in elevation between two bench marks that are 1.5 km
apart. Determine the maximum sight distance that is permissible if a
tolerance of 10 mm is allowed in the measured difference in elevation.

194 Analysts and Adjustments of Survey Measurements

Solution
Since tA= 10 mm, the maximum allowable standard deviation in h is,
according to Eq. (7-32),

Now S = 1500 to, and from Example 7-5,

Thus, according to Eq. (7-31),

Hence, the maximum permissible sight distance is 30 m.

PROBLEMS
7-1 Side a of the plane triangle ABC, Fig. 7-5, is computed from
independent measurements of side b and angles and using the sine
law

If ABC is essentially an equilateral triangle with each side approximately

100 meters in length, determine the standa rd deviations with which b, ,
and must be measured in order to obtain a standard deviation of 0.020
m in the computed value of a. Assume balanced accuracies.

195 Analysts and Adjustments of Survey Measurements

7-2 If side b of the triangle in Problem 7-1 can be measured with a
standard deviation no less than 0.016 m, determine the standard
deviations with which angles and must be measured in order to
maintain a standard deviation of 0.020 m for the calculated value of side
a.
7-3 If the area of the triangle in Problem 7-1 is to be computed from
independent measurements of side b and angles and using the
function

determine the standard deviations with which b, , and must be

measured in order to obtain a standard deviation of 1.0 m in the

computed area. Explain the unusual result obtained for . (Hint: pass a
circle through points A, B, and C and note what happens when b and are
held fixed and is varied.) What is a practical value to assign to o?
7-4 Height h (Fig. 7-6), is to be computed from independent
measurements of horizontal angles and , vertical angle , and
horizontal distance c. The vertical angle can be measured with a standard
deviation no less than 15. If h must be determined with a standard
deviation no larger than 0.040 m, determine suitable values for the
standard deviations of the measurements, given = = 85, =15, and
c=50 as. Summarize the results in tabular form similar to Table 7-1.
7-5 A direction theodolite is to be used to measure the angles of a
traverse. The length of each traverse side is approximately 5DO en. The
standard deviations in instrument centering, target alignment, pointing,
and circle reading are, respectively: c= 1.5 mm; t= 2.5 mm; p=2.5`;
and r=2.0 Evaluate the expected standard deviation in measuring
traverse angle in n paired sets for: (a) = 45 , n = 2; (b) 6 = 45 , n =
6; (c) = 135', n = 2; (d) =135 , n = 6.
7-6 The angles at the corners of a rectangular parcel of land, 100 m by
200 m, are to be measured with a repeating theodolite. For the particular
instrument used,

196 Analysts and Adjustments of Survey Measurements

the standard deviations in pointing and circle reading are 2.0" and 3.0", re spectively. If the standard deviation in each angle measurement must be
no greater than 5.0", evaluate suitable standard deviations for instrument
centering and target alignment, assuming two paired sets are to be used
in each measurement and targets are to be placed at the parcel corners.
7-7 The velocity of light is known with a relative accuracy c/c of 0.3 ppm,
and the index of refraction of the atmosphere can be determined with a
standard deviation of 2.0x 10 6 . The wavelength of modulation of an
electro-optical distance meter (EDM) is 10 m and nit value is known with a
standard deviation of 8 m. Phase difference can be measured with a
standard deviation that is 1/3000 of the modulation wavelength, and the
zero correction has a standard deviation of 3.0 mm. Evaluate the a and b

factors for this EDM and determine the standard deviations for
measurements of 500 m, 2.0 km, and 5.0 km.
7-8 The compensator of a surveyor's level allows the instrument to be
leveled with a standard deviation of 1.5. The rod can be read with a
standard deviation of 0.010 mm per meter of sight distance. If sight
distances of 80 m are used, express the expected standard deviation in
determining elevation difference with this level and rod as a function of K1
the length of the level line in kilometers.
7-9 Two bench marks are 20 Km apart. The elevation difference between
these two bench marks must be determined with a standard deviation that
is no greater than 0.010 m. Using the level and rod of Problem 7-8, two
level lines are run between the bench marks and a mean value is taken.
Determine the maximum sight distance that should be used.
7-10 In the subtense bar method of measuring distance the horizontal
distance between the theodolite and subtense bar is computed as follows:

where b is the length of the bar and is the subtended angle (Fig. 7-7). If
b and are uncorrelated, show that the standard deviation in the
computed value of D is

197 Analysts and Adjustments of Survey Measurements

Assume cot (/2) csc (/2) for small .
7-11 Use the relationship in Problem 7-10 to evaluate the standard
deviations in measuring distances of 50 m, 100 m, 200 m, and 400 m, with
a 2 m subtense bar and precise theodolite, given b=0.1 mm and o=
1.0. Comment on the relative effects of b and o for short distances and
for long distances.
7-12 A distance of 1200 m is to be measured with the 2 m subtense bar
and precise theodolite of Problem 7-11 (b=0.2 mm and o= 1.0"). If the
distance is to be measured in n legs of equal length (e.g., one possibility
is 2 legs of 600 m), determine n if the standard deviation in the measured
distance must be no greater than 0.50 m.

7-13 A taped distance is influenced by many sources of error, two of which

are: (1) errors in marking the ends of the tape, and (2) error in calibrating
(or standardizing) the length of the tape. Assume for this problem that
these two sources of error are the only ones that are significant. If D is the
measured distance, corrected for length of tape, show that the variance i n
D is

where n is the number of tapelengths in (he measured distance, m is the

standard deviation in marking each end of the type, and is the
standard deviation in calibrating the length of the tape.
7-14 With reference to Problem 7-13, a distance of 2 km is measured with
a 50 m steel tape; m =3.0 mm and =0.5 mm. Evaluate the standard
deviation in the corrected distance.
7-15 Another source of error in taping is the error in measuring the tape
temperature. In measuring a distance with a steel tape, the temperature
correction for a single tapelength is: ct= Cl (T- T,) where C is the
coefficient of thermal expansion (11.6 x 10 6 m/m/C for carbon steel), l is
the nominal length of the tape, T is the field temperature, and T, is the
standardization temperature (assumed errorless). Assuming the standard
deviation in measuring field termperature is 1.0C, calculate the standard
deviation in the temperature correction for a 2 :km distance measured with
a 50 m steel tape: (a) if the field temperature is observed once only for the
entire distance; (b) if the field temperature is observed once every four
tapelengths.
7-16 The error in applying pull to the tape also affects a taped distance. In
measuring a distance with a steel tape supported at its ends only, the
combined tapelength correction for tension and sag is given by

where P is the pull applied to the tape In the field, P, is the pull applied
during tape standardization, A is the cross-sectional area of the tape, E is
the modulus of elasticity of the tape material (B=2.1 x 10 4 kg/mm for
carbon steel), I is the nominal length of the tape, and w is the mass of the
tape per unit length. A
198 Analysts and Adjustments of Survey Measurements
2 km distance is measured with a 50 m steel tape that is supported at its
ends only. The cross-sectional area and mass of the tape are 3.8 mm and
6.030 kg/m, respectively, P1 =10.0 kg (assumed errorless) and P= 8.0 kg.
Calculate the combined tension and sag correction for the entire 2 km
distance and determine the standard deviation of this correction if the
standard deviation In P is 0.5 kg. (Note: Each tapelength has its own
independent pull.)

7-17 A 50 m steel tape is to be used to measure a distance of 800 m. The

tape parameters (C, E, A, and w) are those given in Problems 7-15 and 716. The tape is to be supported at its ends only with a pull of 10.0 kg, and
the tape temperature is to be observed once every four tapelengths. If the
standard deviations in marking the ends of the tape, standardizing the
tape, measuring tape temperature and applying pull are 3.0 mm, 0.5 mm,
1.0C, and 0.5 kg, respectively, evaluate the standard deviation of the
taped distance corrected for length of tape, temperature, tension, and sag.
Summarize the analysis in the form of an accuracy budget, as shown in
Table 7-1.
7-13 if the 800 m distance in Problem 7-17 must be measured with a
standard deviation of 0.010 m, evaluate the required standard deviations
in marking, standardization, temperature measurement and ;nil, assumin g
a balanced system (i.e., the four sources of error contribute equally).
7-19 The interior angles of a 16-sided closed trace:se are measured
independently. If the tolerance (maximum acceptable angular error of
closure) for the sum of the interior angles is set at 60", determine the
standard deviation with which each angle should be measured.
7-20 if the standard deviation in the length of a 2 m subtense bar is 0.20
mm, determine the required standard deviation in the measurement of the
sobtended angle in order to measure a distance of 100 m in one leg if
the tolerance in the measured distance is 0.10 m. (See Problem 7 -10.)

199 Analysts and Adjustments of Survey Measurements

8
Introductory
Statistical Analysis

8.1. SAMPLES AND STATISTICS

The concept of sampling was introduced in Chapter S. Particular
reference was made to the random sample consisting of n independent,
identically distributed observations. The notion of population was
introduced as well. When n independent measurements of a quantity are
made under the same conditions, they are considered collectively as a
random sample of size n it drawn from a population of infinite size.
More than one sample can be drawn from a population. In general, the
samples will not be identical, i.e., the observed data will vary from sample
to sample.
It is possible to construct new quantities that are functions of the observa tions (random variables) X1, X2,..., Xn which make up the random sample
of size n. Since these new quantities are functions of random variables,
they are random variables themselves, having their own probability
distributions. These quantities are known as statistic.
Perhaps the simplest example of a statistic is the sample mean,

(8-1)
If observations X1, X2, , Xn are it independent, identically distributed
random variables, each with mean and variance then X is also a
random variable which can be shown to have a mean and variance /n
(i.e. Section 8.5). The specific probability distribution of X will depend on
the distribution of the population from which the observations are drawn; if
the population is normally distributed, X will be normally distributed. Since
the observations will vary from sample to sample, the values of X will
vary.
___

___

___

___

200 Analysts and Adjustments of Survey Measurements

sample to sample according to the probability distribution of X . Under the
assumption that the normal distribution is the basic probability model for
all survey measurements, we shall confine further discussion in this
chapter to sampling From a normal population.
___

Although X , the sample mean, is normally distributed when the population

is normally distributed, it cannot be assumed that all other statistics con ___

structed from a random sample drawn from a normal population are also
normally distributed.
Other distributions play important roles in dealing with samples drawn
from normal populations. These distributions are known as sampling distributions, the most common of which are the chi-square and Student (t) distributions. These two distributions are briefly discussed in the sections
which follow.
8.2. THE CHI-SQUARE DISTRIBUTION
If Z1, Z2, ..., Zn are n independent random variables, each with a standard
normal distribution (i.e., =0, and =1; see Section 5.4) it can be shown
(beyond the scope of this text) that the probability density function of
(8-2)
Is
(8-3)*
Y is said to have a Chi-square (x 2) distribution with n degrees of freedom.
The distribution parameter is it, and the mean and variance of Y are
() is known as the gamma function. It it defined as the integral

and is evaluated using the following relationships, derived from the

integral:

And

for any positive lrteger p (i.e., p= 1 , 2, 3, ...).

201 Analysts and Adjustments of Survey Measurements
(8-4)
And
(8-5)
EXAMPLE 8-1

The random variable Y is the sum of the squares of four standard normal
random variables, i.e., Y has a chi-square distribution with four degrees of
freedom (n= 4). Evaluate and plot the density function of Y. Evaluate the
mean and standard deviation of Y.
Solution
For n=4, (n/2) = (2) = 1! = 1. Thus, from Eq. (8-3), the density function
is

For y= 0.5,

For y = 1.0

Other evaluations off (y) are given in Table 8-1, and the density function is
plotted in Fig. 8-1. The mean and standard deviation of Y are

202 Analysts and Adjustments of Survey Measurements

Chi-square density functions for several values of is are plotted in Fig.

8-2. For small values of n, the density functions are asymmetric with long
tails to the right. For larger values of n, the curves approach the
symmetric shape of the density function of the normal distribution with
mean is and variance 2n.
If Y has a chi-square distribution with n degrees of Freedom, its
distribution function, according to Eq. (5-17), is

(8-6)
where f (u) is the density function given by Eq. (8-3).
Bear in mind that F (y) is a probability. When this probability is fixed at
some specified value, F (y) = p, the corresponding value of y is known as
the pth fractile, or pth percentile. For the chi-square distribution with n
degrees

203 Analysts and Adjustments of Survey Measurements

Of freedom, The pth percentile is represented by the symbol X p2,n values of

X p2,n are given in the table II of appendix B

8-3 THE T (STUDENT)* DISTRIBUTION

If Z is a random variable that has a standard normal distribution, and Y is
an independent random variable that has a chi-square distribution with n
degrees of freedom, then the random variable

(8-7)
is said to have a t distribution with n degrees of freedom. The probability
density function of the t distribution is

(8-8)
*"Student" was the pseudonym of W. S. Cosset, the statistician who first
derived the t-distribution .

204 Analysts and Adjustments of Survey Measurements

The parameter of the distribution is n, where is n a positive integer. The
mean and variance of T are
(8-9)
and

(8-10)
The distribution is symmetric about zero.
EXAMPLE 8-2
The random variable T has a t distribution with four degrees of freedom.
Evaluate and plot the density function of T. Evaluate the standard
deviation of T.
Solution
For n = 4,

And

Thus, from Eq. (3-8), the density function is

For t=Q, f (t)=0.375 (1) 2.5 =0.375.

Fort= 1, f(t)=0.375(1+1/4) 2.5 '=0215.
Other evaluations off (t) are given in Table 3-2, The density function Is
plotted In Fig. 8-3. The standard deviation of T is

It is obvious from Fig. 8-3 that the shape of the t distribution's density
function resembles that of the normal distribution. However, the two distributions are not the same, particularly for small values of n. As n
increases, the t distribution approaches the standard normal distribution.
Indeed, for large is, the standard normal distribution can be used to
approximate the t distribution.
205 Analysts and Adjustments of Survey Measurements

The distribution function of T is

(8-11)

where f (u) is the density function given by Eq. (8-8). If this function is
fixed at some specified value p, (i.e., F(t ) =P [T t] =p), the value oft
associated with p is designated tp,n, representing the pth percentile of the
t distribution with it degrees of freedom. Values of tp,n are given in Table
III of Appendix B.

206 Analysts and Adjustments of Survey Measurements

8-4. COMMON SAMPLE STATISTICS
In Section 8.1 it was stated that it is possible to construct new random
variables that are functions of the random variables X1, X2, ...,Xn, which
make up a random sample of size n, and that these new random variables
are called statistics.

Common statistics that are used to locate the "center" of a sample include
the following.
Sample Mean. The sample mean, X , has already been defined in Eq. (81), he value of X is defined as the simple average of all values in the
sample. Sample Medium If the values in the random sample are arranged
in order of magnitude, the value of the sample median is the value in the
middle of the arranged set if n is odd, or it is the a verage of the two middle
values if n is even.
___

___

Sample Midrange. The sample midrange is the average of the largest and
smallest values in the sample.
EXAMPLE 8-3
Following are the results of 20 independent measurements of a distance,
made under the same conditions:

In statistical language, the above 20 numbers are the values of a random

sample of size 20. Evaluate the sample mean, sample median and sample
midrange.
Solution
Sample mean,*

Arrange the values in order of magnitude (last two significant figures only):
55, 56, 58, 59, 59, 60, 60, 61, 61, 61, 61, 62, 62, 62, 63, 64, 64, 65, 68 ,
69.

* X represents a specific numerical value for the random variable X

207 Analysts and Adjustments of Survey Measurements
___

where the values averaged are the tenth and eleventh values in order of
magnitude

Common statistics that are used to measure the spread or dispersion of a

sample include the following.
Sample Range, The sample range is the difference between the largest
and s mallest values in the sample, i,e ., it is the total spread of the
sample.
Sample Mean Deviation, The deviation of an observation Xi, is defined as
the difference Xi- X , where X is the sample mean. The sample mean
deviation is the average of the absolute values of the deviations of all
components of a random sample, i.e.,

(8-12)
Sample Variance. The sample variance S is defined as the following
function of the random sample

(8-13)
where X is the sample mean. The reason for using n - 1 instead of n in Eq.
(8-13) will be taken up in Section S-6.
Sample Standard Deviation. The sample standard deviation S is defined
as the positive square root of the sample variance.
EXAMPLE 8-4
For the random sample in Example 8-3, evaluate he sample range,
sample mean. deviation, sample variance, and sample standard deviation.
Solution

The deviations and their squares are listed in Table 8-3

Sample variance,
208 Analysts and Adjustments of Survey Measurements

Sample standard deviation,

Sample Covariance. If the random sample is drawn from a twodimensional population such that pairs of observations are obtained, it is
possible to evaluate a sample covariance. Le: the random sample consist
of the paired observations (X1, Y1), ,Yn). The sample covariance is then
defined as

(8-14)

209 Analysts and Adjustments of Survey Measurements

where of X is the sample mean of X1, X2 , ..., Xn, and Y is the sample
mean
Y1, Y2, ..., Yn.
8.5. ESTIMATION OF THE MEAN
A statistic which is used to estimate the value of a parameter of a
probability distribution is called a point estimator of the parameter, or

simply an estimator. Since an estimator is a statistic, it is a random

variable. The numerical value of the estimator resulting from a specific
sample is called an estimate.
In eneral, if the parameter is represented by , the estimator is
represented by . Thus, if the parameter in question is the mean of the
distribution, the estimator is ^.
Several statistics, such as the sample mean, sample median and sample
midrange, can be used as estimators of the distribution mean. The
selection of a particular statistic is based upon certain desirable
properties, such as unbiased , consistency, efficiency, and sufficiency, as
well as ease in computation. Comprehensive discussion of the properties
of estimators is beyond the scope of this text. However, two of the
properties, unbiasedness and consistency, ill be touched upon.
Unbiasedness. If ^ is an estimator of a parameter of a distribution, such
that the expectation of is , i.e.,
(8-15)
then ^ is said to bean unbiased estimator of .
Consistency. If ^ is an unbiased estimator of and the variance of approaches zero as the sample size gets larger, then 6 is said to be a
consistent estimator of .
The most common estimator of the distribution mean is the samp mean
X . To test this statistic for unbiasedness and consistency, use is made of
the following two important relationships.
___

(1) If Y=a1X1+a2X2+...+anXn, where X1, X2, ...,Xn are random variables

(not necessarily independent) and a1, a2, ..., an are constants, then the
expectation or mean of Y is given by
(8-16)
(2) If Y= a1X1 + a2X2 + ...+ anXn where X1, X2, ..., Xn are independent
random variables, then the variance of Y is given by
(8-17)
210 Analysts and Adjustments of Survey Measurements
Equation (8-16) is easily derived from the basic properties of expectation
given by Eqs. (5-35) and (5-36). Equation (8-17) is simply a restatement of
the special law of propagation of variance for linear functions, already
given by Eq. (6-22),
Now, the sample mean, as defined by Eq. (3-1), is equivalently expressed
as

(8-18)
Applying Eqs. (8-16) and (8-17) to Eq . (8-18), and noting that E [X1] =
E
[X2] = ... = E [Xn] = , and x1= x2 = = xn we get, for the expectation
and variance of X,

(8-19)
And

(8-20)
Obviously, since E [ X ]= ; X and since them unbiased Xisan
variance of X must approach zero as n goes to infinity, X must be a
consistent estimator of . Hence, on the basis of unbiasedness and
consistency, X is a good estimator of .
___

___

___

___

___

If the random sample is drawn from a normal population, it will be found

that X is normally distributed. Even if the population is not normal, X is
still approximately normally distributed by virtue of the Central Limit
Theorem.
___

___

8.6. ESTIMATION OF THE VARIANCE

The most common estimator of the distribution variance is the sample
variance S , given by Eq. (8-13). To show that S is an unbiased estimator
of , it must be first shown that.

(8-21)