SAMPLING THEOREM SAMPLING THEOREM: STATEMENT [1/3]

1. Statement of Sampling Theorem • Given: Continuous-time signal x(t).

2. Derivation of Sampling Theorem • That’s: Bandlimited to B Hertz.
3. Ideal Reconstruction from Samples • Means: Maximum frequency is B Hertz.
4. Determining Signal Bandwidths • Means: X(ω) = F{x(t)}=0 for |ω| ≥ 2πB.
5. Finite Pulse Width Sampling RADIAN
• Means: Bandwidth=B Hertz=2πB SECOND .
6. Undersampling and Aliasing

SAMPLING THEOREM: STATEMENT [2/3] SAMPLING THEOREM: STATEMENT [3/3]

• Given: Knowledge of only the samples {x(nT )} of x(t). • Then: x(t) can be reconstructed from its samples {x(nT )}
1 SAMPLE
SAMPLE
• Means: x(t) sampled every T seconds, at rate S= T1 SECOND . • If: Sampling rate S = T SECOND > 2B=2(bandwidth).

Signal x(t) Pulse Train p(t) Sampled Signal x(t)p(t) • Where: S > 2B Here 2B is the Nyquist sampling rate.
✬✩
× ✻✻✻✻✻✻✻✻✻ = ✻✻✻
✲ t ✲ t ✻ ✻ ✲ t • Note: Co-discovered by Claude Shannon (UM Class of 1938)
✫✪
❄
❄ ❄
❄
• Note: Digital Signal Processing is possible because of this.

SAMPLING THEOREM: EXAMPLE #1 SAMPLING THEOREM: EXAMPLE #2

SAMPLE
x(t)=cos(2π1000t) sampled at S=8000 SECOND . T = S1 = 8000
1
sec. Do we need S > 2B or only S ≥ 2B?
1000 π SAMPLE
Set t=nT =n/8000: x(nT )=cos(2π 8000 n)=cos( 4 n). x(t)=sin(2π500t) sampled at S=1000 SECOND . T = S1 = 1000
1
sec.
500
Can reconstruct x(t) from its samples {x(nT )}: (How? See below.) Set t=nT =n/1000: x(nT )=sin(2π 1000 n)=sin(πn)=0!
n ... 0 1 2 3 4 ... Can’t reconstruct x(t) from its samples {x(nT )}!
x(nT ) . . . 1.0 .71 0.0 –.71 –1.0 . . . 1

0.5
1
0

0.5
−0.5

0
−1
0 0.001 0.002 0.003 0.004 0.005 0.006 0.007 0.008 0.009 0.01

−0.5

−1
0 0.5 1 1.5 2 2.5 3 3.5 4 4.5 5
sec −3
x 10
SAMPLING THEOREM: PROOF [1/3] SAMPLING THEOREM: PROOF [2/3]

• Fourier series expansion: p(t)= pk ej2πkt/T = T1

P∞
• DEF: p(t)= n=−∞ δ(t–nT )=train (series) of impulses. ej2πkt/T .
P P

R T /2
• DEF: x(t)p(t)=sampled signal=train of impulses weighted by {x(nT )}. • Fourier coefficient formula: pk = T1 −T /2 δ(t)e
−j2πkt/T
dt= T1 .
• Note: x(t)p(t)=x(t) δ(t–nT )= x(nT )δ(t–nT ).
P P
• Then: F{x(t)p(t)}= T1 F{x(t)ej2πkt/T }= T1 X(ω–2π Tk ).
P P

• So: Only need samples {x(nT )} to create x(t)p(t). • S= T1 → F{x(t)p(t)}= T1

P
X(2π(f – Tk ))= T1
P
X(2π(f –kS)).
Signal x(t) Pulse Train p(t) Sampled Signal x(t)p(t)
✬✩ ✻
F{x(t)} F{x(t)p(t)}=Spectrum of Sampled Signal
× ✻✻✻✻✻✻✻✻✻ = ✻ ✻
✟✟❍❍ ✟✟❍❍ ✟✟❍❍ ✟✟❍❍
✲ t ✲ t ✻ ✻ ✲ t ✟✟ ❍❍
✲ ✟✟ ❍❍ ✟✟ ❍❍ ✟✟ ❍❍
✲
f f
✫✪
❄
❄❄❄ –B B –B–S B–S –B B S–B S+B

SAMPLING THEOREM: PROOF [3/3] SAMPLING THEOREM: EXAMPLE

• Can: Reconstruct x(t) from x(t)p(t) if (S–B) > B → S > 2B . • Given: Continuous-time x(t) is bandlimited to 4 kHz.
SAMPLE
• By: Low-pass filtering x(t)p(t). Cutoff frequency=B Hertz. • Sample: 10 “kHz”=10000 SECOND > 2(4 kHz).
sin(2πBt) • F{x(t)p(t)}=Spectrum of sampled signal. Repeats in f !
• Formula: x(t)= 2B
X
2B T |
x(nT )δ(t
{z
− nT )} ∗
πt
.
SAMPLED SIGNAL x(t)p(t)
| {z }
LPF h(t)
• Reconstruct: Low-pass filter with cutoff=4 kHz.
F{x(t)} F{x(t)p(t)}=Spectrum of Sampled Signal
• Interpolation Formula: x(t)= x(nT )(2BT ) sin2πB(t−nT
2πB(t−nT )
.
P
) ✟❍ ✟❍ ✟❍ ✟❍
✟✟ ❍❍ ✟✟ ❍❍ ✟✟ ❍❍ ✟✟ ❍❍
✟ ❍ ✲ f ✟ ❍ ✟ ❍ ✟ ❍✲ f kHz
–4 4 –14 –6 –4 4 6 14

SAMPLING AND RECONSTRUCTION: SUMMARY DETERMINING SIGNAL BANDWIDTH: EXAMPLES

p(t) p(t)
↓ ↓ 1. x(t)= 3t+7
t
sin(2π5t): Rewrite: x(t)=3 sin(2π5t)+7π sin(2π5t)
πt
.
→ {x(nT )}} → → x(t)p(t) → LOWPASS
O O
x(t) → → x(t)p(t) FILTER → x(t)
| {z
SAMPLING
| {z } 2. x(t)=sin(2π3t) sin(2π2t): Rewrite: x(t)= 12 [cos(2π1t)–cos(2π5t)].
RECONSTRUCTION
✬✩ × =
✻✻✻✻✻✻✻✻✻
✻
✻ ✻ 3. x(t)= sin(2π3t)t2sin(2π2t) : Then X(ω)= 2π
1
[rect( ω6 ) ∗ rect( ω4 )].
✲ t ✲ t ✻ ✻ ✲ t x(t)p(t)
Convolve length=6 with length=4→length=10.
✫✪
❄
❄ ❄
❄
✟✟❍❍ ✟✟❍❍ ✟✟❍❍ ✟✟❍❍ F{x(t)p(t)} SAMPLE
✟✟ ❍❍ ✲
f ✟✟ ❍❍ ✟✟ ❍❍ ✟✟ ❍❍✲
f For all 3: Bandlimited to ω = 10π or f =5 Hertz. Need 10 SECOND .
–B B –B–S B–S –B B S–B S+B
 SAMPLING WITH FINITE-WIDTH PULSES [1/2] SAMPLING WITH FINITE-WIDTH PULSES [2/2]

• Problem: The ideal impulse train p(t)= δ(t–nT ) doesn’t exist! • Fourier series expansion: pq (t)= pq,k ej2πkt/T . Pulse width: ∆.
P P

R ∆/2
• Sol’n: Use pq (t)= q(t–nT ) where q(t)=short pulse (does exist). • Fourier coefficient formula: pq,k = T1 −j2πkt/T 1
sin( kπ∆
P
−∆/2 e dt= kπ T ).

• Then: pq (t)= q(t–nT )= δ(t–nT ) ∗ q(t)=p(t) ∗ q(t):

P P
• Then: F{x(t)pq (t)}= pq,k F{x(t)ej2πkt/T }= pq,k X(ω–2π Tk ).
P P

q(t) p(t) pq (t)

• S= T1 → F{x(t)p(t)}= pq,k X(2π(f – Tk ))= pq,k X(2π(f –kS)).
P P
* ✻ ✻ ✻ ✻ ✻ =
✲ ✲ ✲ F{x(t)} F{x(t)p(t)}
–2T 0 2T t –2T 0 2T t –2T 0 2T t ✟❍ ✟❍
✟✟ ❍❍ ✏✏PPP ✟✟ ❍❍ ✏✏PPP
✟ ❍ ✲ f ✏✏ P ✟ ❍ ✏✏ P✲f
–B B –B–S B–S –B B S–B S+B

UNDERSAMPLING AND ALIASING [1/4] UNDERSAMPLING AND ALIASING [2/4]

SAMPLE MAX.
• What if RATE
= S < 2B = 2 FREQ =Nyquist frequency? • Image spectrum impersonates actual spectrum! Overlap!
• Aliasing: image spectrum masquerades as actual spectrum! • Low-pass filtering this→400 Hz, not 600 Hz, sinusoid!
SAMPLE
• EX: 600 Hz sinusoid sampled at 1000 SECOND . 1000<2(600). 0.5

1 n 0.4
• Sample: x(t)=cos(2π600t) with T = 1000 . Set t=nT = 1000 : 0.3

• Get: Samples x(nT )=cos(2π 600n

1000 )=cos(1.2πn)=cos(0.8πn).
0.2

0.1

• Since: cos(1.2πn)=cos(1.2πn–2πn)=cos(−0.8πn)=cos(0.8πn). 0
−1500 −1000 −500 0 500 1000 1500

UNDERSAMPLING AND ALIASING [3/4] UNDERSAMPLING AND ALIASING [4/4]

SAMPLE SAMPLE
• 500-Hertz sinusoid sampled at 450 SECOND . Samples: circles. • x(t)=cos(2π300t+1) sampled at 500 SECOND . Reconstructed=?
n
• Did samples come from 450 Hertz or 50 Hertz sinusoid? • Sample: t= 500 → x(nT )=cos(2π 300
500 n+1)=cos(1.2πn+1).

1
• But: cos(1.2πn+1)=cos(1.2πn–2πn+1)=cos(−0.8πn+1)=cos(0.8πn–1).
0.5
• Reconstruct: n=500t → cos(0.8π(500t)–1)= cos(2π200t–1)
0

−0.5
• 300 Hertz aliased down to 200 Hertz. Phase also changed.
−1
0 0.005 0.01 0.015 0.02