JUNE 1985 GRACE WAHBA 125 Design Criteria and Eigensequence Plots for Satellite-Computed Tomography GRACE WAHBA Department of Statistics, University of Wisconsin Madison, Madison, WI 53705 (Manuscript received 29 December 1983, in final form 13 December 1984) ABSTRACT ‘The use of the “degrees of freedom for signal” is proposed as a design criteria for comparing different designs for satelite and other measuring systems. It i also proposed that cerain eigensequence plots be ‘examained at the design stage along with appropriate estimates of the parameter playing the role of noise to signal ratio. The degrees of freedom for signal and eigensequence plots may be determined using preseatly available prior information in the special domain along witha description of the system, and simulated data {or estimating A. This study extends work by previous investigators, Implicit is an approach to combining direct and indirest measurements of atmospheric temperature ina spectral forecast model. 1. Introduction Recently Fleming (1982) has suggested that im- proved temperature retrievals from satellite soundings may be obtained by use of data from a sensor which scans forward and back along the satellite track, and thus “looks” at @ particular point in space from several directions as well as directly down. This idea was suggested by analogy with well-known results from computed tomography techniques in use in medicine. Fleming constructed a model temperature field and simulated noisy data from three different ray configurations, one looking straight down only, ‘one having in addition one forward and one rearward angle and the third having two forward and two rearward angles (see the solid lines in Fig. 1). He then recovered the model temperatures on a two- dimensional grid with one axis vertical and one axis along the satellite track by a numerically efficient iterative procedure for solving large linear systems. He performed the necessary regularization in this ill- posed problem by stopping the iteration (see also Fleming, 1977; Wahba, 1980). Similar methods are common in medical applications. Fleming's (1982) results in the example tried were: two additional angles are better than straight down only and four are better than two from the point of view of mean- square retrieval error. We are interested in the choice of angles, spacing of observations, selection of channels and other ques- tions concerning the design of measuring systems. Westwater and Strand (1968) and Weinreb and Crosby (1972) discussed design criteria which can be used to make an evaluation of alternative satellite designs and they applied these criteria to the selection of radiometer channels. In this paper, we begin with what is essentially the design criteria proposed by {© 1985 American Meteorological Society Weinreb and Crosby. However, we propose using prior information concerning meteorological fields in the frequency or spectral domain, rather than the spatial domain, leading to details which can be dif- ferent. This approach uses information that is available at the present time (but was not in 1972!) and is particularly appropriate for the evaluation and com- parison of potential satelite systems that simulta. neously use three-dimensional information, as well as the evaluation of systems that use combined satellite and radiosonde data. Implicit in the proce- dures described here is an algorithm for combining satellite and radiosonde data. Our approach also makes clear the role of possibly variable bandwidth arameter(s) in system design, a point which has traditionally been ignored. In Section 2 we derive the design criteria in our form (as opposed to the form used by Weinreb and Crosby) and also note how data from different systems can be combined. In Section 3 we describe the idea of the “effective rank” of a system, which is roughly equivalent to the “degrees of freedom for signal” associated with a design. The degrees of freedom for signal is related to, but not exactly the same as, one of the criteria’ used by Weinreb and Crosby, and is analogous to the usual degrees of freedom for signal in analysis of variance. ‘We suggest the use of eigensequence plots along with the GCV (generalized cross validation) estimate of the bandwidth (or signal to noise ratio) parameter on these plots to evaluate and compare different systems from the point of view of degrees of freedom for signal The details of the approach described here are perfectly general and can be used to make a prelim- inary evaluation of combinations of measuring sys- tems; for example, the use of several satellites simul- taneously and the combined use of direct and indirect126 ‘measurements. In this paper we describe the design criteria in context of a global system for measuring the atmospheric temperature. The design criteria is, however, applicable to much more modest design problems. 2. Design criteria For simplicity we will make assumptions similar to those made by Fleming (1982); that is, that surface temperature is known accurately, an initial (either first-guess or climatology) value To of the temperature field is known and that it is adequate! to linearize the Planck function about To. With these approxi- mations, given a particular design, the data may be ‘modeled as follows after subtracting out the mean: Yer = a Reload + eens where k indexes the “look” direction and subsatellite point of the sensor, v is the central frequency of the spectral window, x is a point in the atmosphere titude, longitude, pressure), and K represents the instrumental spectral response function convolved with the atmospheric transmittance along the Ath ray. ‘The integral is taken along the Ath ray. Figure 1 depicts several rays from an orbiting satellite looking, forward and back along the satellite track, as well as a single ray in a general position looking to the side of the satellite track (dashed line). Here 7%x) = T(x) = Tox) and the «, represent measurement, quadra- ture and modeling errors. (See, c.g., Wark and Flem- ing, 1966; Fritz er al, 1972). We shall assume that the observations have been normalized so that Ed, is roughly constant and the ¢, are roughly indepen dent. Next, we shall assume that 7° possesses a (gener- alized) Fourier series expansion in some appropriate basis functions in latitude, longitude and pressure, for example: T* (latitude, longitude, pressure) =D TusYlPib0), @) where P = (latitude, longitude), p = pressure (or other vertical coordinate), the Yi, are spherical har- monics and the }, are appropriate (continuous) func~ tions in the vertical, possibly represented in terms of splines. In (2) we are assuming (for maximum gen- erality) that a global three-dimensional temperature analysis will be carried out and the design problem may include assessing the efficiency of overlapping rays from sidelooking satellites. However, the ap- " Fora more careful approach to the nonlinearity, the linearization jin O'Sullivan (1983) p. 78 and O'Sulivan and Wah (1988) may e used, JOURNAL OF ATMOSPHERIC AND OCEANIC TECHNOLOGY Vouume 2 Flo, |. Satelite viewing geometry proach may be modified so that x ranges in a region either swept out by the line from the satellite perpen- dicular to the ground as the satellite moves along its track (then sines and cosines could replace the Y;,) or in a limited three-dimensional volume of space. ‘The choice would depend on whether side-looking detectors are allowed and multiple satellites are being studied possibly in conjunction with the radiosonde network. ‘The observations are now modeled as Ye = Z Tiny { KulVide 3) By dae where Vis) = YAP IAP). We can rewrite (3) as y=XB+e (4) where y is the (rearranged) vector of the observations, ‘Yees B is the (rearranged) vector of the Ti, and ¢ is ithe (rearranged) vector of the «x... Letting i'stand for , », and j stand for ls, y, we have that the i, jth entry of X is X= [ BiG ends, © Letting 8; = Tiny» if Tp has been obtained from climatology and the ¢ have been chosen appropriately, a fair amount of information may be constructed or assumed concerning the prior distribution of 6,. (See, for example, Baer, 1980; Stanford 1979; Kasahara and Puri 1981; Smith and Woolf 1976.) An illustration Of the explicit use of Stanford's results in this context can be found in Wahba (1982b). We shall suppose that the 6; have a prior distribution which is multi- variate normal with means zero, and a prior covari- ance matrix given by EBB = box, = (oy). © In the sequel we will be assuming that on is known, but the scale factor b may not be. We suppose that the errors can be modeled (approximately) as inde-June 1985 pendent Gaussian random variables with a common (possibly unknown) variance o*. Then a regularized estimate of @ is 8, given by the minimizer of ly — XB? + 8'="'8, ” where the prime denotes a transpose, n is the number of data points (dimension of y), and 2 is the smoothing or regularization parameter that controls the trade- off between the goodness of fit to the data as measured by (I/n)ly — XBI? and the “smoothness” of B as measured by 8'2~'8. The minimizer 6, is given by By = BX(XEX’ + md 'y, 8) where I is the 1 X 7 identity matrix and the temper- ature estimate is To + T;! where TAP, p) = E Tray¥idPybs(0) lb and the T,j., are the components of 8,. This can also be shown to be the Bayes estimate (Gandin estimate) of 8 with the choice 4 = o7/nb. That is, is the conditional expectation of @ given the data, This result is found in a more general setting in Kimeldorf and Wahba (1971); see also Wahba (1978a) In practice the estimate can be extremely sensitive to the choice of \ and not “robust” to misspecification of o*/nb or other modeling assumptions (see Appendix B) and so X should be chosen either from experience (by “eyeball”) or by a good data-based method such as generalized cross validation (GCV) (see e.g. Craven and Wahba, 1979; Golub, Heath and Wahba, 1979; Halem and Kalnay, 1983; Wahba and Wendelberger, 1980). We will, for the moment, however, leave ) as parameter. Now, suppose our criteria for preferring one design over another is to minimize the mean (M) integrated square error, (ISE) where 0 ISE = [1'0) — TPs. Jaen finer (0) Expanding (10) in the Wi,, gives ISE = © (8; ~ Ba dau(Bx — Bra) ay K where, if j = (ls, 7) and k = (J, 9) then af erimeen YUP DY Yu(P)64( pI Pa. We now take the expected value of (11), over the distribution of both the 8, and the «. Substitution of By = DX(XEN' + mA '(KB + &) (12) into (11) gives the mean ISE (MISE), MISE) = E@™OM6 + 26'M\OMye + &MSQM39) (13) GRACE WAHBA 127 where Q is the matrix with jkth entry qq and My = T~ EX(XEX' + aN", Mz = BX(XEX' + mY, Carrying out the expectation operation in (13), after, assuming that E48; = 0 gives MISE = Trace (6M\QM,2 + 0°M3QM;). (14) Letting Q? and E' be the symmetric square roots of Q and , itis shown in Appendix A that rearranging (14) results in 1 7 MISEUX, ») = Trace Q'7[5 — EX(WEX' + mu XEIO!? Trace QIK" + m\ly2X310" (1s) It can be shown that the right-hand side of (15) is minimized over \ for n\ = 02/b. Making this choice for ) gives yas,» = tase o”[2 2x xe 2 + 1) x20". (16) 6 With Q = / this is the same design criteria as that proposed by Weinreb and Crosby (1972). ‘We want to choose X’ so that the right-hand side of (16) is as small as possible. We have the following. Theorem: Let X; and X2 be two design matrices of the same dimension and suppose that A°X;X,6 > A'X2X26 for all 6 (that is, XX, — X3X2 is non- negative definite), then MISE(X;) < MISE(X2) a7) for any nonnegative definite Q and ¢ >0. Proof: See Appendix A. Unfortunately this provides only a partial ordering. ‘We would like to find a more graphic way of evalu- ating a design, or comparing two designs, independent of Q. We will do this in Section 3. Note that if radiosonde information is to be com- bined with satellite information, then one just in- creases the dimension of the data vector y in (4). If Ye is a direct measurement of temperature at a point x, = (Pj, p) then this simply adds a row to the X matrix with entries Xj = ¥iAP)b(7). 5 = (ls, 7). If different measuring systems are being combined it is appropriate to scale the observations in units chosen so that the ¢, are about the same size; for example, if the standard deviation of a measurement is known,128 JOURNAL OF ATMOSPHERIC then y; can be scaled to units of this standard devia- tion. We will assume that this can be done, at least approximately, from properties of the measuring equipment, 3. Eigensequence plots, effective rank and degrees of freedom for signal Letting the dimension of X be n X r, we have not discussed the relative size of n and r. In meteorological work it is frequently reasonable that r > n, since ‘meteorological fields contain information at all scales. Certainly in the design phase one should allow r to be as large as computationally feasible consistent with the availability of measured, theoretical or conjectured prior variances. One does not expect to get very good estimates of individual 8, with r > m; however, it is TP, p) that is actually desired and good estimates of T'(P, p) may be obtainable even though some of the individual coefficient estimates appear poor. In- spection of (8) shows that the number of linearly independent pieces of information in y available for estimating 8 (and hence 7°) is limited by the number of eigenvalues of XZ" which are at least not negligible compared to mA, where 7A is appropriately chosen, see Appendix B. The “signal” along an eigenvector with eigenvalue much less than n\ will be down in the “noise.” Proceeding under the assumption that 7 < 1, it is nevertheless typical in ill-posed problems that the “effective rank” of matrices playing the role of XEX’ is much less than nt, when 7 is large. The effective rank of XEX’ can be roughly defined as the number of eigenvalues of ¥ZX" not small compared to the noise (relative to b) in the system. (See Wahba, 1980). This “noise,” in practice, includes not only the measurement error, but the errors in modeling the atmospheric transmittance functions, linearizing Planck’s function and computing the integrals in (5) using quadrature formulae. The effective rank of XEX’ can easily be studied by plotting the eigenvalues of X22" on a log-log plot. Figure 2 gives an eigensequence plot of the eigen- values reprinted from Nychka, Wahba, Goldfarb and Pugh (1983; hereinafter referred to as NWGP). The mildly ill-posed problem in NWGP is concerned with the recovery of three-dimensional tumor-sized distri butions from tumor radii observed from two-dimen- sional slices. This is a tomographic problem of a somewhat different form than the one under study. Nevertheless, there are some common aspects. There were n = 80 observations; 68 of the 80 eigenvalues appear on this plot. The precipitous drop-off of the last few eigenvalues has been attributed to artifacts of the quadrature procedure. Data from an active experiment using the design behind this plot were actually analyzed and n} estimated by GCV appears on the figure. In practice \ would appear instead of AND OCEANIC TECHNOLOGY ‘Voume 2 Fic, 2. Eigensequence plot from NWGP. nd in (8) where, in the design phase, { would be obtained by simulating realistic examples. One can see that there are only six eigenvalues at least as large as n\. Strictly speaking, comparing the eigensequence plots for X,2X} and X,2X5 does not necessarily provide enough information for choosing between X; and Xz on the basis of criteria (16); nevertheless, these plots can be quite informative. A measure of comparison between X; and Xy which depends only on the respective eigenvalues and 2is the degrees of freedom for signal. We may define LE. signal (X, \*) as, .f. signal (X, X*) = trace (KEXY(XEX’ + mary! (18) +m where 2, v = 1, 2, +++ m are the eigenvalues of ZX", and \* is a good choice of X.? To understand this definition, which is analogous to similar defini: tions in analysis of variance, observe that y can be decomposed into signal and noise as follows: yortte as) where i = XBR = NEX(WEX + mT ly (estimated signal), ef = mM*(XEX' + mdtl)ly, (estimated noise), n= trace I = trace XEX(KEX' + nd*1 (GE. for signal, + trace m\*(XEX" + mda)! It is necessary, of course, that the \* used provides a 00d partition of y into signal and noise for this definition to be valid since dif. for signal varies from nto 0 as A* varies from 0 to oo. It is clear that one (Gf, for noise). ? Weinreb and Crosby’s trace M of their Ea (10) would correspond to trace XEEN(UBN' + dtl)! The present criteria are likely to be less sensitive to misspecification of ©.June 1985 GRACE wants as many eigenvalues as possible to be langer than md*, One can make a loose association of the Uf for signal with the effective rank. Thus, X; is to be preferred to X2 if Af, signal (X,, Mf) > of. signal (%2, AN). (20) We have deliberately allowed Mf and Xf to be different, and not necessarily equal to o?/nb, since in practice, as well as in Monte Carlo experiments with a small number of examples, the optimum may depend on X as well as T* and the noise in the system. We will iscuss the choice of d* in practice below. Aside from the intuitive justification for the use of the maximization of df. for signal as a design criteria, we will provide another (theoretical) justification. If we look at the expected normalized sum of squared errors in estimating almost any 7 linear funtionals of T® normalized by their prior covariance matrix, we find that the result is n — df, signal. We explain this, result below and give a proof. To explain this result we first define a linear functional L, = L,(T°) as follows. If T' = 5 BM;, 7 then L,(T%) is some linear combination of the 8); that is, Dh; LAT’) for some fixed 1; = (ly, - - + 1j)- For example LT!) = Tx) is a linear functional if we take, as W,(x)); and 17) = f acarener is a linear functional with ly = f w,(0¥%(x)dx. The (prior) covariance of L,(7*) and ‘L,(7*) is simply ELAT YLT) = EE lB X binBee em = wz, Li(T*) is estimated by L(T3!) = 1B. Let (bry + + ,) be the vector of errors in estimating the values of linear functionals, that is, & = L(7%) ~ LT), 1 = 1, 2, ..., m We have the following remarkable situation, Theorem: let Zyz, be the prior covariance matrix of [L(7%), - - = L,(7)] and suppose that the /), - > ~ fy associated with Ly, ++ + Zy are linearly independent and are linear combinations of the r > n columns of X. Then ERD. 18 = n— a for signal, Proof: We first observe that E= LB -B) WAHBA, 129 where L is the m X r matrix whose rows are (4, + ty,and ,, = bLEL'. Thus EG Dye = EB — RYLAOLELY "L(G ~ 2). (21) Setting \ = o/nb as in (15) it can be seen (starting with (11)) that the right hand side of (21) is given by the right hand side of (16) with Q replaced by L(LELY'L. Using the fact that trace BA = Trace ABY gives Sut = Trace (LELY“LEL) = Trace L(LELY'LEX(XEX' + MINX. (22) Using the assumption that /,, ..., ly are linearly independent and are linear combinations of the columns of X implies that L = BX for some invertible matrix B, this results in L(L3L)'L =. X'B" X (BYEX'BY'BN = X(NEX')'X, Substituting this into (22) results in eB = W(XEXYXEN’ + nd n~ Af. signal. ‘We will describe an experiment to compare different global systems for measuring temperature using 4.f. signal as the evaluation criteria. Assume that a rep- resentative set of global temperature analyses is avail- able from a multilevel global spectral model which provides sets of analyses of temperature at level p, in ‘the form: TWP, Py) = 2B VAP). If for example we assume the functions , satisfy dp) = 1,j = k, = 0, j # k, then these functions provide an interpolation between levels, that is TIP, p) = 2B ise ¥iPiolP) and we can use the resulting @ vector as “ground truth” for atest case (ignoring of course any differences between true atmospheres and atmospheres produced in a model). As an approximation it will probably be necessary to assume most if not all of the nondiagonal entries in 2 are 0, particularly those corresponding to Is # I, The diagonal entries may then be approx- imated from the N sets of {Bj} oF, for example, from data such as Stanford's (1979) by the method described in Wahba (1982b). The entries in the design matrix X should then be computed by very accurate quadrature methods where the integrals in (1) have been scaled so that the ¢,, have at least approximately the same standard deviation. If b is absorbed into > and the scaling is such that Ee,,? ~ 1 then nA = o%/ b ~ 1 im theory and it remains to compute d.f. for signal for each trial X under study. Eigenvalues of symmetric nonnegative definite ma- trices of dimension up to several hundred can be computed using double precision EISPACK (Smith et al., 1976). If 2 is diagonal, it may be cheaper and more accurate to compute ‘the singular values of130 "2X" using the singular value decomposition in LINPACK (Dongarra et al, 1979). The civenvalues of XEN" are the squares of the singular values of 2'2X", Approximate information concerning very much larger matrices may be obtained using the truncated singular value decomposition in Bates and Wahba (1982), and/or Lanczos methods for large matrices described by, for example, Cullum er al (1982). In actuality the results may be very sensitive to the choice of \ and also to the approximations made in obtaining 2. A good adaptive choice of data for can partiallly compensate for scaling and misspecifi- cation errors in 3 and o”. One may use GCV in a Monte Carlo study to obtain good values of \ as follows: given a temperature analysis in the form of fa single vector 8, one generates one sample data vector y = X8 + « where the ¢ is a vector of pseudorandom numbers with appropriate variances. ‘The GCV estimate of for this sample data vector is the minimizer of 4 Ty XEX(XEX + Ty? va) E trace (= XBX EX + my] Values of should be obtained for a reasonably representative set of test (true) 6 and then a range of values of df. for signal evaluated for each candidate design. In the process other criteria for evaluating @ design, e.g., f (7%) — T;G)Pds may also be eval- uated for a set of test 8 via (11) by setting d= 4. If it is desired to compare several possibly overlapping satellite configurations (without regard to the existence of the radiosonde network) it is quite legitimate to consider rays only in a limited but typical volume of the atmosphere, It is conjectured that eigensequence plots comparing different satellite scanning designs will show that, e., combining side-looking scans from successive passes of a satellite along with data in the plane of the orbit (@s suggested by Suomi, 1983) would have highly desirable properties. We close with a few remarks. Quadrature error in, eg. evaluating the X, in (5) can be surprisingly important in ill-posed problems and should not be treated cavalierly at either the design or the data analysis stage. This point is discussed in some detail in NWGP, where the use of matched quadrature for ill-posed problems is discussed. Eigensequence plots obtained via inaccurate quadrature may present a different appearance than those from a highly accurate quadrature, and a poor quadrature procedure or unrealistic value of X may mask differences between systems. In practice observational data are generally JOURNAL OF ATMOSPHERIC AND OCEANIC TECHNOLOGY Voume 2 combined with forecast data. For approaches to doing so that are compatible with the analysis implicit in this paper, see Halem and Kalnay (1983), Wahba (1982a) and Wahba (1984), Acknowledgments. This research was supported by NASA under Contracts NAG-128 and NAG 5-316, APPENDIX A Proofs ‘The proof of Eq. (12) is as follows. Letting Q"” be the symmetric square root of Q, EB'M\OM8 = b Trace MiQM,E = b Trace Q"M,2MiQ"? = b Trace QY?{2 — 2EX(XEX’ + md 'XE + EX(KEN + nN EN (XEN + mNy'XE}0" Ee'M3QMz« = o* Trace MQM; = & Trace Q'?M:M30'? = 0? Trace Q""{EX(XEX" + mMly?X2}0"", (an (a2) Using (XEX' + aN 'XEX(WEN’ + ny! = (XEN + md = nXEN + my? and adding (A1) and (A2) gives Trace Q'?E'{b[I — B'2X(XEX' + mY IXE'?] + (0? = AN)E"PX(XEX’ + mAI7XE!}Z"2Q'?, (a3) which gives (12). The proof of the theorem in Section 2 is: Suppose that B'X)X)8 > BAB for any B. We will show this implies that Trace QMPEX(XEX) + mM) 'X, ZO" > Trace Q'?EXYX2EX2 + m"XEQ"? (Ad) for any Q and 2. We will assume that 5 is nonsingular. Then our hypotheses imply that p'E'7X,X2!%6 = p'D'? X X13,2"8 for any 6; in other words, BAX N E> BAX EIA,JUNE 1985 GRACE where 4 > B means A ~ B is nonnegative definite. Let A= DX, where A and B are p X n. We have to show that AA’ > BB! => A(A'A + nN)’ > BUB'B + mM’. B=2'Xy, Wee first show that AIA + MAINA = AMAA’ + IY, This is equivalent to showing AIA + WMI)'A(AA' + nD) Expanding the left-hand side gives AAA + mMDA'ADA + mACA'A + TA’ = AIA + mY AA + mAD)A' — mA’ + NIA + AAA + TY’ Ad, Aa Now, let (A4’ + AJ) = C and (BB' + nd) = D. Therefore AA(AA + MY! = (C ~ nC“ = T— nC BB'(BB + nN) = (D ~ nN)D*! = I= mC“ Now Ad’ > BBY = C> D, and C> D = C"! < D"! (Gee, e.g, Marshall and Oikin, 1979, p. 464), which in turn implies that ! — n\C-! > 7 ~ nD“, so the proof is finished APPENDIX B Remarks on the Specification of o?/nb ‘Suppose that 7° has an infinite series expansion in the ¥;. Under the assumption that £8? = bay, j= 1, 2, ...5 «0 and (for mathematical convenience only) E86, = 0, j # k, then for each r= 1, 2,..., 1 1S be, 7 EB DE (Wy) => Dt = b, (BI) where (7) and 2(?) are the first rand r X r components of 6 and 3 respectively. A different modeling as- sumption is that 7° has the property that (B2) Under this assumption 8, of (8) is still an appropriate estimate of the first r components of 8, for appropri- ately chosen (see, e.g. Wahba, 197a), but 1 BIDE-(B(A) + 0 as r~ co $0 6 is not readily defined independent of p. The GCV will return a good estimate of A under WAHBA 131 either assumption (B1) or (B2) (see Wahba, 1977b) and the design criteria resulting from the assumptions of this paper (i.e. assumption B1) appear eminently plausible even if (B2) is true, A related but somewhat harder to study design criteriz under assumption (B2) appears in Wahba (19780). [REFERENCES Baer, F., 1980: Three-dimensional scaling and structure of atmo- ‘spheric energetics. J. Atmos. Si, 38, 52-68, Bates, D. M., and G. Wahba, 1982: Computational methods for Generalized cross-validation with large data sets, Treatment of Integral Equations by Numerical Methods, C. T. H. Baker and G. F. Mile, Bds, Academic Press, 283-796, Craven, P, and G, Wahba, 1979: Smoothing noisy data with spline functions: Estimating the correct degree of smoothing bY the method of generalized crosevalidation. Numer. Math, 31, 377-403. Cullum, J, R- Willoughby, and M. Lake, 1983: A Lancoos algorithm for computing singular values and vectors of large matrices SIAM J. Scientific and Statistical Computing, 4, 197-215. Dongarra, J.J. Bunch, C. Moler and G. Stewart, 1979: Linpack User's Guide, SIAM, Philadelphia. Fleming, H., 1977: Comparison of linear inversion methods by cxamination of the duality between iterative and inverse tmatrix methods. Inversion Methods in Atmospheric Remote Sounding, A. Deepak, Ed., Academic Pres, 325-355. — 1082: Satelite remote sensing by the technique of computed Tomography. J. Appl. Meteor. 21, 1538-1549. Fritz, S, D. Wark, J. Fleming, W.” Smith, H. Jacobowitz, D. Hilleary and J. Alishouse, 1972: Temperature sounding from satelite. NOAA Tech, Rep. NESS 59 Golub, G. H., M. Heath and G. ahs, 1979: Generalized cross validation as a method for choosing a good ridge parameter. Technometries, 2, 215-224, Halem, M, and E, 1983: Variational global analysis of satellite temperature soundings. Preprints Sith Conf, Numerical Weather Prediction, Omaha, Amer. Meteor. So. Kasahara, A., and K. Puri, 1981; Spectral representation of three- dimensional global data by expansion in normal mode func- tions. Mon. Wea, Rev, 109, 37-51 KimeldortG., and G. Wahba, 1971: Some results on Tehebychetian spline functions. J. Marh, Anal. Appl 38, 82-95 Marshall A. and I.Olkn, 1979: Inequalities: Theory of Majorzation ‘and ls Applications. Academie Press, NY. Nychka, DG. Wahba, S. Goldfarb and T. Pugh, 1984: Cross ‘validate spine methods fo the estimation of thre dimensional {tumor size cistrbutions from observations on two dimensional cross sections. J. Am. Stat Asse, 79, 388, 812-846, (O'Sulivan, F,, 1983: The analysis of some penalized likelihood ‘estimation schemes. Ph.D. thesis, University of Wisconsin- Madison. [TR-726} 117 pp. —— and'G. Wahba, 1984:'A cross-validated Bayesian retrieval "algorithm for non-linear remote sensing experiments. Dept. of| Statistics, University of Wisconsin-Madizon, [TR 747] To appear, J. Comput. Physics. ‘Smith, B. J. Boyle, J. Dongarra, B. Garbow, Y. Ikebe, V. Klema, and C. Moler, 1976: Mairix Eigensystem Routines-BISPACK ‘Guide. Springer-Verlag. $51 pp, Smith, W., and H. Woolf, 1976: The use of eigenvectors of Statistical covariance matrices fr interpreting sitlite sounding radiometer observations. J. Atmos. Sci, 33, 1127-1140, Stanford, J, 1979: Latitudinal. wavenumber power spectra of tato- soherié temperature fluctuations. J. mos. Se, 36, 921-931 Suomi, V., 1983: High equatorial orbit satellites. Talk, Weather “arecisting and Analysis Seminat, 1983, Madison, WL Wahba, G., 1977a: Optimal smoothing of density estimates, Clas132 sification and Clustering J. Van-Ryzin, Ed, Academic Press, London, 423-458. — 197T%: Practical approximate solutions to linear operator ‘uations when the dats are noisy. STAM J. Numerical Analysis, 14, 651-867 —, 1978a: Improper prior, spline smoothing and the problem ‘of guarding against’ model errors in repression. J. Roy. Sta Soe, B40, 364-372. —, 1978: Comments on “Curve iting and optimal design for prediction” by A. O'Hagan, J. Roy. Star Soe. BAO, 35-36. 11980: I posed problems: Numerical and statistical methods Tor mildly, moderately, and severely ill posed problems with noisy data. Statistics Dept, University of Wisconsin-Madison, TR $95} (io appear in Proceedings Int. Conf. om Ill Posed Problems, M. Z. Nashed, Ed.) —1982a: Variational methods in simultaneous optimum inter- polation and initialization. The Interaction Between Objective ‘Analyses and Initialization, D. Williamson, Eé., Meteorology 127, Atmospheric Analysis and Prediction Division, NCAR, Boulder, CO, 178-185 JOURNAL OF ATMOSPHERIC AND OCEANIC TECHNOLOGY Voume 2 ——, 1982b: Vector splines on the sphere, with application to the “stimation of vorticity and divergence from discrete, noisy data, Multivariate Approximation Theory, Vol.2,W. Schempp fand K. Zeller, Eds, Birkhauser Verlag, 407-429, 1984: Variational methods for multidimensional. inverse ‘problems, Advances in Remote Sensing Retrieval Methods, A. Deepak, Ed,, A. Deepak Pub. Co., (in pres. and J. Wendelberger, 1980: Some new mathematical methods, foe variational objective analysis using splines and cross validation. Mon. Wea, Re, 108, 36-57. Wark, D., and H. Fleming, 1966: Inditect measurements of atmo- ‘spheric temperature profiles from satellites, Mon. Wea, Rev. 94, 351-362. Weinreb, M., and D. Crosby, 1972: Optimization of spectral intervals for remote sensing of atmospheric temperature profiles ‘Remote Sens. Environ, 2, 193-20. Wesiwater, E, R, and O. N. Strand, 1968: Statistica information ‘content of radiation measurements used in indirect sensing. J Atmos. Sci, 28, 150-758,