Summer 2008

Signals & Systems

S.F. Hsieh

I. Introduction to Signals and Systems

to develop mathematical models/techniques for continuous-/discrete-time signal and system analysis/synthesis(design). Signals(functions) are variables(time/space) that carry information, Ex. f (t) speech(voltage/current picked up from microphone, a computer le xxx.wav, stock market index, ECG(ElectroCardioGram); f (x, y )/f (x, y ) image xxx.gif or xxx.jpg; f (t, x, y ) video xxx.mpg. Systems(mapping) process input signals to produce output signals. They can be hardwares(DSP processors/circuits/automobile, economic system) or softwares(programs).

Classications of signals
1. Digital/analog: the amplitude is discrete(quantized) or not. 2. Continuous-/Discrete-time: (MATLAB commands: plot, stem) Continuous-time signal: x(t), < t < , where the time-variable t is continuous.

Discrete-time signal: x[n] x(nT ), n = 0, 1, 2, . . ., where T represents the sampling time. Note that x[3 1 2 ] is NOT dened. Many discrete-time systems are computer programs.
x(t) 1 0.8 0.6 0.4 0.2

8 x[n]

10

12

14

16

0.8 0.6 0.4 0.2 0

I-1

3. Deterministic/Random: Deterministic signals:

periodic(Fourier series) x(t) = x(t + T ), t minimum T is the fundamental period. aperiodic(Fourier transform)

Random signals(such as noise): probability density function, mean, autocorrelation, power spectral density. 4. Energy/Power signals: Size of a signal x(t): energy:
+

|x(t)|2 dt

Denitions:

if x(t) is the voltage applied across a dummy load of one- resistor, then x2 (t) is the power assumption (Watts). Integration of power over time is the total energy of the signal, namely, E indicates the energy that can be extracted from the signal. Continuous-time: T Total energy over the time interval t1 t t2 is tt12 |x(t)|2 dt. n2 2 Discrete-time: total energy over the time interval n1 n n2 is n =n1 |x[n]| . average power: 1 T /2 P lim |x(t)|2 dt T T T /2 If x(t) is periodic, then its average power becomes 1 T /2 P = |x(t)|2 dt T T /2 The rms(root-mean-square) value of x(t) is P . x(t) is an energy signal if it has a nite energy Ex < and its power Px = 0. x(t) is a power signal if it has a nite power Px < , and energy Ex = .
2

(Lathi Ex 1.2, p 72)

Power of a single real sinusoid x(t) = C cos(0 t + ) is Px = C 2 , where C is a positive amplitude. Power of a single exponential y (t) = Dej0 t is Py = |D|2 , where D is a complex number comprised of amplitude and phase. Power of sum of orthogonal complex exponentials z (t) = i Di eji t with distinct frequencies i s is equal to the sum of individual powers Pz = i Pi = i |Di |2 . (Q) Suppose z (t) = x(t) + y (t). Under what condition will the power of z (t) be equal to the sum of those of x(t) and y (t), i.e., Pz = Px + Py ? x(t) = e3t u(t) is an energy signal with energy 1/6. x(t) = Aej 2t is a power signal with power A2 . A rectangular pulse A (t/ ) of amplitude A and duration has energy A2 . A (t/ ) cos(0 t + ) has energy A2 /2. tnT0 2 A rectangular pulse train n: ( ), T0 > , is a power signal with P = A /T0 I-2

Examples: (a) (b) (c) (d) (e)

(f) A ramp signal f (t) = t and an everlasting exponential g(t) = eat are neither energy nor power signals. (g) Is x(t) = 10 a power or energy signal? x (h) What is the energy of a sinc function sinc(x) = sin x ? (Parsevals thm in Chp 7s Fourier transform can solve it easily.)

Basic operations on signals

(b) negation: x(t) = x(t), change of polarity (c) level-shift: x(t) = x(t) + c c > 0: shift up by c c < 0: shift down by c

(d) Note that sum of two lines is itself a line, too. x3 (t) = x1 (t) + x2 ( t)
) x1 (t x2 (t)

[Ex] Given x(t), plot y (t) = 2x(t) + 3. (multiply/divide rst, then add/subtract)

x(t)
t

-1

-1

-2

2x(t) -1 0 1

2. change of Time-variable: Assume that x(t) looks like 1

x(t)

-1

I-3

(a) time-scale: x(t) = x(t), where > 0 > 1, time-compression: 1

x(2t)

< 1, time-expansion:

1 2

1 2

t x( 2 )

-2

(b) time-reversal(inversion): x(t) = x(t) 1

This may seem to hurt your instinct. For justication, Let y (t) = x(2t). Choose a few ts 1 and plot y (t), we have y (0) = x(0)(still centered around the origin), y ( 1 2 ) = x(2 2 ) = x(1), 1 1 and y ( 2 ) = x(2 2 ) = x(1). x(t)

t -1 1 The right-hand-side is mirrored about the origin, and vice versa.

(c) time-shift: x(t) = x(t + )

> 0, left-shifted(advanced)

x(t + )

< 0, right-shifted(delayed):

x(t )

Again, this seems to be unbelievable. A mindful reader will not hesitate to play the same game: Let y (t) = x(t 3). y (3) = x(3 3) = x(0), y (0) = x(0 3) = x(3), etc. From which, she/he convinces herself/himself. (d) Combined operation: x(t) = x(t + ) , [method 1]: Let y (t) = x(t + ), i.e., advance/delay x(t) by to obtain y (t). Let z (t) = y (t), i.e., compress/expand y (t) by .

As a check, z (t) = y (t) = x(t + ). The above procedure can change its order. BUT, be careful! [method 2]: I-4

 As a check, w(t) = v (t + ) = x([t + ]) = x(t + ) = z (t), while w (t) = v (t + ) = x([t + ]) = x(t + ) = z (t)! Note:

Let v (t) = x(t), i.e., compress/expand y (t) by . ), INSTEAD OF w (t) = v (t + ). Let w(t) = v (t +

if you have diculty in combining w(t) and v (t), you can use more dummy variables to avoid confusion: v (s) = x(s), w(t) = v (t + ) so that w(t) = v (s)|s=t+ = x([t + ]) = x(t + ) [Method 1] is preferred due to its algebraic simplicity. [Ex] Given x(t), plot y (t) = x(2t 3), and z (t) = x(2t + 3). (subtract/add rst, then divide/multiply, a rule which is exactly opposite to the one used in the amplitude operations.) It is strongly recommended to check y (t)/z (t) at some t s where x(t) changes abruptly(and t = 0). Given x(t) 1

x(t)

-1 Right shift by 3, x(t 3) 1

x(t 3)

2 Compress by 2, x(2t 3)

1 (Check) let x(2t 3) = x(0), let x(2t 3) = x(1), let x(2t 3) = x(2), Similarly, Given x(t) 1

2 we have t = 3/2 we have t = 1 we have t = 5/2

indeed y (3/2) = x(2t 3) = x(0) indeed y (1) = x(2t 3) = x(1) indeed y (5/2) = x(2t 3) = x(2)

x(t)

-1

I-5

 Left shift by 3, x(t + 3) 1

x(t + 3)

-4

-3 -2

Compress by 2 and time-reversal(whichever comes rst), x(2t + 3)

1 (Check) let x(2t + 3) = x(0), let x(2t + 3) = x(1), let x(2t + 3) = x(2),

we have t = 3/2 we have t = 2 we have t = 1/2

indeed y (3/2) = x(2t + 3) = x(0) indeed y (2) = x(2t + 3) = x(1) indeed y (1/2) = x(2t + 3) = x(2)

Signal characteristics
1. Even/Odd Even function(signal): x(t) = x(t), t, symmetric about the origin,

Odd function(signal): x(t) = x(t), t, anti-symmetric about the origin,

(Fact) Every signal x(t) can be decomposed as a sum of even and odd signals: x(t) = x(t) x(t) xe (t) + xo (t), where xe (t) E v {x(t} x(t)+2 and xo (t) Od{x(t} x(t)2 . 2

x(t) =

xe (t)

xo (t)

2. Periodic signals: x(t) = x(t + nT ), t. T is its period.

Sum of several continuous-time periodic signals may not be period, depending on the relationship among their fundamental periods.

3. Causal signal: x(t) = 0, t < 0; anticausal signal: x(t) = 0, t > 0.

I-6

Basic Signal Models

x[n] = Cz n Aej (rej )n z rej i. > 0 or r > 1: exponentially growing, ii. < 0 or r < 1: exponentially decaying, iii. = 0 or r = 0: constant. (b) = 0,

(a) = 0 and = 0, x(t) = Aet is real exponential(non-oscillating),

i. = 0 or r = 1: x(t) = Aej ejt = A cos(t + ) + jA sin(t + ) or x[n] = Aej ejn is a non-damping cisoid(complex sinusoid). will see in Fourier transforms. ii. > 0 or r > 1: x(t) = Aet ej (t+) is an exponentially increasing(unstable) cisoid iii. < 0 or r < 1: x(t) = Aet ej (t+) is an exponentially damped(decreasing) cisoid [Eulers formula] ejt cos t + j sin t and ejt cos t j sin t 1 jt 1 cos t = (e + ejt ) and sin t = (ejt ejt ) 2 2j
10 5 0 5 10 15 5 t=5:0.01:5; x=exp(0.5*t).*cos(2.7*pi*t); subplot(2,1,1); plot(t,x) damping factor =0.5 frequency=2.7pi

10 5 0 5 10 15 5 smaller damping factor =0.3 larger frequency=4.2pi

y=exp(0.3*t).*cos(4.2*pi*t); subplot(2,1,2);plot(t,y);

On the left-hand-side, < 0, est 0 as t , and the magnitude | | controls the envelope (damping factor); as | | increases, the envelope decays faster. Along the j -axis, controls the angular frequency of the rotating phasor. I-7

2. Rectangular pulse: (t/ ) 1 (t/ )

2 3. Triangular pulse: (t/ )

(t/ )

4. Unit step function: u(t) =

0
t ()d

t 1, t 0 0, t < 0

0 5. Unit ramp function: r (t) =

t = tu(t):

0 [Ex] tu(t) 2(t 1)u(t 1) + (t 2)u(t 2)

I-8

6
6.1

Impulse function
Unit impulse (Kronecker delta function) in discrete time
[n] 0, n = 0 1, n = 0 1

[n]

-1 1. unit step function: u[n] 0, n < 0 1, n 0

u[n] 1

0 1 2 3

-1

running sum: u[n] = 2. sampling property:

dierence equation: [n] = u[n] u[n 1]

x[n][n] = x[0][n] similar to dierentiation in CT x[n][n n0 ] = x[n0 ][n n0 ] similar to integration in CT 3. sifting property: x[n] =
k

x[k][n k] convolution of x[n] and [n]

= + x[2][n + 2] + x[1][n + 1] + x[0][n] + x[1][n 1] +

any sequence can be expressed as a sum of scaled/shifted impulses .

6.2

Unit impulse function(Dirac delta function): (t), in continuous time

1 t d ( ) = u(t), a pulse with unit area and zero width, located at t = 0 dt

reducing

(t) 0

area = 1 0 t

I-9

Properties of (t): 1. sifting:

t0 )dt = x(t0 )

(t t0 ) x(t)

x(t0 ) (t t0 ) t

t0 x(t) (t t0 )dt =

t0

x(t0 ) (t t0 )dt (t t0 )dt

= x(t0 ) = x(t0 ) Example:

3)dt = e6

2. Convolution of x(t) with (t t0 ) is x(t t0 ). which is useful in proving the sampling theorem and modulation/demodulation. x(t)

(t t0 ) = 0

x(t t0 )

0 Pf. x(t) (t t0 ) = 3. Other properties: x(t)(t) = x(0)(t).

t0

t0

x( )(t t0 )d =

x( )( t + t0 )d = x(t t0 ).

Question: What is x(t) [(t t0 ) + (t + t0 )] [(t t0 ) + (t + t0 )]? (demodulation)

x(t)(t t0 ) = x(t0 )(t t0 ). (t) =

x(t)(t t0 ) = x(t0 )(t t0 ).

(t) = (t).

(t).

I - 10

Systems

process an input x(t) to produce an output y (t): cause system eect 1. Physical models: an electric circuit 2. Block diagrams continuous-time: using integrators, adders, and gains (and multipliers) discrete-time: using delays, adders, and gains (and multipliers) 3. Mathematical system equations: integro-dierential equation: y (t) = 4. Interconnection of systems: (a) series(cascade):
S1 S2
dx(t) t [4x( ) dt

dierence equation: y [n] = ay [n 1] + bx[n]

y ( )]d .

(b) parallel:
S1 S2 S 2

(c) feedback:
S 1

S3

Classication of Systems
1. Continuous-time or Discrete-time, 2. Analog or Digital, 3. Memoryless: the o/p of a memoryless system at time t0 depends only on its input at the same time instant t0 . Memory is associated with storage of energy in physical systems and storage registers in digital computers. Delay, capacitor, and inductor elements are not allowed for a memoryless system. Thus, a resistive voltage divider is a memoryless system, while an RC lowpass circuit has memory. [Ex] y [n] = x2 [n] + x[n], y (t) = 10x(t): memoryless. 2 A voltage divider: vo (t) = R1R +R2 vi (t) is also memoryless
t [Ex] y (t) = x( )d, y [n] = x[n 2], y [n] = (x[n 1] + x[n] + x[n 1])/3: with memory. An electric example is a current source i(t) connected to a capacitor, then the voltage across the 1 t capacitor is v (t) = C i( )d .

I - 11

4. Invertible: A system is invertible if the cascade of this system with its inverse system yields an output which is the input to the rst system. x[n] system y [n] inverse system w[n] x[n] Examples:
t (a) System: y (t) = 2x(3t); its inverse system: w(t) = 1 2 y( 3 ) n k = x[k ];

(b) A running-sum system: y [n] = tion: w[n] = y [n] y [n 1].

its inverse system satises a dierence equa-

Inverse processing is important in equalization, lossless coding, etc. 5. Causal(nonanticipative) and noncausal: A causal systems output y (t0 ) can only depends on the present and past inputs: {x( ), t0 }, i.e., the system cannot anticipate the future input. o/p: y1 (t)

causal

i/p: x(t)

0 o/p: y2 (t)

Noncausal

-2

Later, we will show that, for a causal LTI system, its impulse response h(t) = 0, t < 0.

[Ex] y [n] = x[n] x[n + 1], and y (t) = x(t + 1) are noncausal, because the output y (2) at present time t = 2 depends on the future input x(3) at t = 3. A memoryless system is also causal.

All physical systems with time as the independent variable are causal. There are practical systems that are not causal: Physical systems for which time is not the independent variable e.g., the independent variable is (x, y ) as in an image. y [n] = (x[n 1] + x[n] + x[n + 1])/3, take the average of three neighboring data. Processing of signals is not in real time, e.g., the signal has been recorded or generated in a computer. 6. Stable: A system is stable in the sense of bounded-input, bounded-output(BIBO) if the output y (t) is bounded for a bounded input x(t), i.e., if |x(t)| B1 , t, then B2 < , |y (t)| B2 , t. For a stable linear-time-invariant system, its impulse response must be absolutely integrable: |h(t)|dt < . (to be shown later) 7. LINEAR: linear combination of inputs lead to linear combination of corresponding outputs (superposition): x1 (t) Linear y1 (t) If and x2 (t) Linear y2 (t) then ax1 (t) + bx2 (t)
Linear ay1 (t) + by2 (t)

I - 12

Suppose y1 (t) = H[x1 (t)], and y2 (t) = H[x2 (t)]. If the output due to the input a1 x1 (t) + a2 x2 (t) is equal to a1 y1 (t) + a2 y2 (t) then the system H is linear, i.e., check if H[a1 x1 (t) + a2 x2 (t)] = a1 y1 (t) + a2 y2 (t) Linearity does not allow x2 (t), |x(t)|. 8. TIME-INVARIANT(TI): a delayed input leads to a corresponding delayed output: If x(t) Time Invariant y (t) then x(t ) Time Invariant y (t )
?

for all x(t) and . We need to verify if the following is true: H[x(t )] = {H[x(t)]}|t:t = y (t ), (a) A system is time-invariant if its system parameters are xed over time. + x(t)

?

y (t)

d y (t) + y (t) = An RC lowpass lter with constant resistance R and C is time-invariant: RC dt x(t). If the resistance R(t), changes with time, then the system becomes time-varying: d R(t)C dt y (t) + y (t) = x(t), because its system parameters are not constant over time. (b) Time-invariance does not allow x(2t), x(t), tx(t). (c) Ex. A compressor with y [n] = x[M n] is not TI because y [n n0 ] = x[M (n n0 )] = x[M n n0 ]. Be careful in distinguishing x[M (n n0 )] and x[M n n0 ].

LTI systems
1. Many man-made and naturally occurring systems can be modeled as linear time-invariant (LTI) systems. Ex. resistors(R), inductors(L), capacitors(C). 2. A xed-coecient linear dierential equation (made up from RLC components) are LTI. (CT) 3. A xed-coecient linear dierence equation (delay, gain, adder) are LTI. (Discrete-Time) 4. (Ex) Is the following system linear-time-invariant? y (t) = x(t)g(t) where x(t) and y (t) denote the input and output, respectively. Linear, yes. Let y1 (t) x1 (t)g(t) and y2 (t) x2 (t)g(t). Suppose x(t) = a1 x1 (t) + a2 x2 (t) H[x(t)] = x(t)g(t) then = a1 x1 (t)g(t) + a2 x2 (t)g(t) = a1 y1 (t) + a2 y2 (t) I - 13

 Time-varying. Let y (t) = x(t)g(t). Suppose xd (t) = x(t ). H[xd (t)] = x(t )g(t) = y (t ) = x(t )g(t )

If g(t) = cos c t, we call it linear modulation.

9.1
1.

Examples
d dt y (t)

+ 10y (t) = x(t) is a causal LTI system. y [n] = 0.9y [n 1] + x[n] is causal LTI, too.
dn dn1 dtn y (t) + an1 dtn1 y (t) + + a0 y (t) d d = bm dt m x(t) + + b1 dt x(t) + b0 x(t) is a linear
m

2. In general, system. 3. y (t) =

x( )h(t

)d is an LTI system. reads as y (t) is the convolution of x(t) and h(t).

[Pf] The proof for linearity is omitted. Let x (t) = x(t ), then y (t) : =

is the response to the input x (t) x ( )h(t )d x( )h(t )d x(s)h(t s)ds x( )h(t )d, from the denition of y (t) = x( )h(t )d

let s , sorry for so many dummy variables =

y (t ) = 4. (a) y (t) = (b) y [n] (c) y (t) y [n] (d) y (t)

[Pf] = x ( u ) du , which means that y (0) is the integral of future input x (0+) upto x ( ). Non0 causal! 5. (a)
d dt y (t)

x(t )h( )d is LTI, too. (commutative law for convolution) = m: x[m]h[n m] = m: x[n m]h[m] is LTI. (discrete-time convolution) t = x( )h(t )d is a causal LTI system. = n m: x[m]h[n m] is a causal LTI system. t = x(t )h( )d is a noncausal, linear, time-varying system. t 0 Say h(t) = 1, then y (t) = x(t )d we can see that y (0) = x( )d

+ ty (t) = x(t) is a linear system. d d y1 (t) + ty1 (t) = x1 (t) y2 (t) + ty2 (t) = x2 (t) dt dt

d d y1 (t) + 2 y2 (t) + t[1 y1 (t) + 2 y2 (t)] = 1 x1 (t) + 2 x2 (t) dt dt

d [1 y1 (t) + 2 y2 (t)] + t[1 y1 (t) + 2 y2 (t)] = 1 x1 (t) + 2 x2 (t) dt Thus the response to the input 1 x1 (t) + 2 x2 (t) is 1 y1 (t) + 2 y2 (t). (b)
d dt y (t)

+ 10y (t) + 5 = x(t) is a nonlinear system.

I - 14

(d) y (t) = sin[x(t)] is time-invariant.

(c) y (t) = x(t2 ) is linear, noncausal, and time-varying. y (2) depends on x(4); thus noncausal. Let x (t) = x(t ), then y (t) = x (t2 ) = x(t2 ) = y (t ) = x((t )2 ); thus time-varying. (e) y (t) = x(2t) is NOT time-invariant. [Pf] Suppose x(t) = y (t) = x(2t). Let x (t) = x(t ), then the output from x (t) is y (t) = x (2t) = x (s)|s=2t = x(s )|s=2t = x(2t ) Compared with the delayed output y (t ): y (t ) = y (s)|s=t = x(2s)|s=t = x(2(t )) = x(2t 2 ) (f) ny [n] + a1 y [n 1] = x[n] is a linear and time-varying(because of n before y [n]). We nd that y (t) = y (t ), it is NOT TI!

(h) y [n] = x[n2 ] is linear, noncausal, and time-varying. (i) y [n] = sin(x[n]) is time-invariant.

(g) y [n] + 0.8y [n 1] + 5 = x[n] is a nonlinear system(because of the constant 5).

9.2

Questions

1. If some system: x(t) y (t) is linear, and another system: y (t) z (t) is linear, too, Is the cascaded system x(t) z (t) linear? 2. If some system: x(t) y (t) is linear, and another system: x(t) z (t) is linear, too, Is the parallel system x(t) w(t) = y (t) + z (t) linear? 3. Redo the above two questions, when the systems are time-invariant. 4. Is the cascaded connection of two nonlinear systems is nonlinear? 5. (Oppenheim et al.) Consider three systems with the following input/output relationships: system 1 : y [n] = x[n/2], n : even 0, n : odd

1 1 system 2 : y [n] = x[n] + x[n 1] + x[n 2] 2 4 system 3 : y [n] = x[2n] Suppose these systems are cascaded in series. Find the input/ouput relationship for the overall cascaded system. Is it linear? Is it time-invariant? 6. Can you generalize the denitions of linearity and time-invariance for a system processing 3D signals with 2 independent variables: s(t1 , t2 ), where s = [s1 , s2 , s3 ]?

I - 15