signals and linear and time-invariant systems in discrete timegik2/teach/linear_systems-dt-2.pdf ·...
TRANSCRIPT
Signals and Linear and Time-Invariant Systems in Discrete Time
• Properties of signals and systems (di↵erence equations)
• Time-domain analysis
– ZIR, system characteristic values and modes
– ZSR, unit-pulse response and convolution
– stability, eigenresponse and transfer function
• Frequency-domain analysis
c�2016 George Kesidis
1
Time-domain analysis of discrete-time LTI systems
• Discrete-time signals
• Di↵erence equation single-input, single-output systems in discrete time
• The zero-input response (ZIR): characteristic values and modes
• The zero (initial) state response (ZSR): the unit-pulse response, convolution
• System stability
• The eigenresponse and (zero state) system transfer function
c�2016 George Kesidis
2
Discrete-time signal by sampling a continuous-time signal
• Consider a continuous-time signal x : R ! R sampled every T > 0 seconds
x(kT + t0) =: x[k] for k 2 Z,where
– t0 is the sampling time of the 0th sample, and
– T is assumed less than the Nyquist sampling period of x, and
– x[k] (with square brackets) is the kth sample itself.
• Here x[·] is a discrete-time signal defined on Z.
c�2016 George Kesidis
3
Example of sampling with t0 = 0 and positive signal x
c�2016 George Kesidis
4
Introduction to signals and systems in discrete time
• A discrete-time function (or signal) x : A ! B is one with countable (time) domain A.
• We will take the range B = R or B = C.
• Typically, we will herein take domain A = Z or Z��n for some (finite) integer n � 0.
• Some properties of signals are as in continuous time: e.g., periodic, causal, bounded, evenor odd.
• Similarly, some signal operations are as in continuous time: e.g., spatial shift/scale, super-position, time reflection, and (integer valued) time shift.
c�2016 George Kesidis
5
Time scaling: decimation and interpolation
• Time scaling can be implemented in continuous time prior to sampling at a fixed rate,
• or the sampling rate itself could be varied (again recall the Nyquist sampling rate).
• In discrete time, a signal x = {x[k] | k 2 Z} can be decimated (subsampled) by aninteger factor L 6= 0 to create the signal xL defined by
xL[k] = x[kL], 8k 2 Z,i.e., xL is defined only by every Lth sample of x.
• A discrete-time signal x can also be interpolated by an integer factor L > 0 to create xL
satisfying
xL[kL] = x[k], 8k 2 Z.
• For an interpolated signal xL, the values of xL[r] for r not a multiple of L (i.e., 8k 2 Zs.t. r 6= kL) can be set in di↵erent ways, e.g., between consecutive samples:
– (piecewise constant) hold: xL[r] = xL[L br/Lc] = x[br/Lc]
– linear interpolation:
xL[r] = x[br/Lc] +r � L br/Lc
L(x[br/Lc+1]� x[br/Lc])
c�2016 George Kesidis
6
Time scaling: decimation and interpolation - Questions
• Is the functional mapping x ! xL causal for linear interpolation?
• Is the hold causal?
• Exercise: Show that if a periodic, continuous-time signal x(t), with period T0, is period-ically sampled every T seconds, then the resulting discrete-time signal x[k] is periodic ifand only if T/T0 is rational.
c�2016 George Kesidis
7
Unit pulse �, unit step u, unit delay �, and convolution *
• Some important signals in discrete time are as those in continuous time, e.g., polynomials,exponentials, unit step.
• In discrete time, rather than the (unit) impulse, there is unit pulse (Kronecker delta):
�[k] =
⇢1 if k = 00 else
• Any discrete-time signal x can thus be written as
x[k] =1X
r=�1x[r]�[k � r] =
1X
r=�1x[k � r]�[r]
= (x ⇤ �)[k]
• or just x = x ⇤ �, i.e., the unit pulse � is the identity of discrete-time convolution.
• Define the operator � as unit delay (time-shift), i.e., 8 signals y and 8k, r 2 Z,(�ry)[k] := y[k � r].
• The discrete-time unit step u satisfies � = u��u, equivalently: 8k 2 Z,
�[k] = u[k]� u[k � 1] and u[k] =1X
r=0
(�r�)[k] =1X
r=0
�[k � r].
8
Unit pulse and unit step functions
• Exercise: For any signal causal f ({f [k], k � 0}), show that
8k � 0, (f ⇤ u)[k] =kX
r=0
f [r].
c�2016 George Kesidis
9
Exponential signals in discrete time
• Real-valued exponential (geometric) signals have the form x[k] = A�k, k 2 Z, whereA, � 2 R.
• Consider the scalar z = �ej⌦ 2 C with � > 0,⌦ 2 R, where again j :=p�1.
• Generally, complex-valued exponential signals have the (polar) form
x[k] = Aej�zk = A�kej(⌦k+�), k 2 Z,where w.l.o.g. we can take
�⇡ < ⌦,� ⇡ and real A > 0.
• Exercise: Show this complex-valued exponential is periodic if and only if ⌦/⇡ is rational.
• By the Euler-De Moivre identity,
x[k] = A�kej(⌦k+�) = A�k cos(⌦k + �) + jA�k sin(⌦k + �), k 2 Z.
c�2016 George Kesidis
10
Systems - single input, single output (SISO)
• In the figure, f is an input signal that is being transformed into an output signal, y, by thedepicted system (box).
• To emphasize this functional transformation, and clarify system properties, we will writethe output signal (i.e., system “response” to the input f) as
y = Sf,where, again, we are making a statement about functional equivalence:
8k 2 Z, y[k] = (Sf)[k].
• Again, Sf is not S “multiplied by” f , rather a functional transformation of f .
c�2016 George Kesidis
11
SISO systems (cont)
• The n signals {x1, x2, ..., xn} are the internal states of the system.
• The states can be taken as outputs of unit-delay operators, �, i.e.,
8k 2 Z, (�y)[k] = y[k � 1].
• Some properties of systems are as in continuous time: e.g., linear, time invariant, causal,memoryless, stable (with di↵erent conditions for stability as we shall see).
c�2016 George Kesidis
12
Di↵erence equation for an discrete time, LTI, SISO system
• For linear and time-invariant systems in discrete time, relate output y to input f via di↵er-ence equation in standard (time-advance operator) form:
8k � �n, y[k + n] + an�1y[k + n� 1] + ...+ a1y[k +1] + a0y[k]= bmf [k +m] + bm�1f [k +m� 1] + ...+ b1f [k +1] + b0f [k],
given
– scalars ak for 0 k n, with an := 1, and scalars bk for 0 k m,
– a0 6= 0 or b0 6= 0 (so that P,Q are of minimal degree), and
– initial conditions y[�n], y[�n+1], ..., y[�2], y[�1].
• Compact representation of the above di↵erence equation:
Q(��1)y = P (��1)f, where polynomials
Q(z) = zn +n�1X
k=0
akzk, P (z) =
mX
k=0
bkzk, and
��1 is the unit time-advance operator: (��1y)[k] ⌘ y[k+1], (��ry)[k] ⌘ y[k+r]
c�2016 George Kesidis
13
Discussion: conditions for causality and di↵erence equation in �
• Exercise: Show that the di↵erence equation Q(��1)y = P (��1)f is not causal ifdeg(P ) = m > n = deg(Q), i.e., the system is not proper.
• A not anti-causal di↵erence equation can be implemented simply using memory to store asliding window of prior values of the input f and delaying the output.
• Example: Decoding B (bidirectional) frames of MPEG video.
c�2016 George Kesidis
14
Numerical solution to di↵erence equation by recursive substitution
• Given the system Q(��1)y = P (��1)f , the input f (f [k] for k � 0), and initialconditions y[�n], ..., y[�1],
• one can recursively solve for y (y[k] for k � 0) by rewriting the system equation as
y[k + n] = �n�1X
r=0
ary[k + r] +mX
r=0
brf [k + r] for k � �n
) y[k] = �n�1X
r=0
ary[k + r � n] +mX
r=0
brf [k + r � n] for k � 0.
• For example, the di↵erence equation in standard form,
y[k +1] + 3y[k] = 7f [k +1] for k � �1,
can be rewritten as
y[k] = �3y[k � 1] + 7f [k] for k � 0.
• So, given f and y[�1] we can recursively compute
y[0] = �3y[�1] + 7f [0], y[1] = �3y[0] + 7f [1], y[2] = �3y[1] + 7f [2], etc.
• Exercise: If f = u and y[�1] = 7 then find y[3] for this example.
15
Approach to closed-form solution: ZIR and ZSR
• The total response y of P (��1)f = Q(��1)y to the given initial conditions and inputf is a sum of two parts:
– the ZSR, yZS, which solves
P (��1)f = Q(��1)yZS with zero i.c.’s, i.e., with 0 = y[�n] = ... = y[�1];
– the ZIR, yZI, which solves
0 = Q(��1)yZI with the given initial conditions.
• The total response y of the system to f and the given initial conditions is, by linearity,
y = yZI + yZS.
• We will determine the ZIR by finding the characteristic modes of the system.
• We will determine the ZSR by convolution of the input with the (zero state) unit-pulseresponse, the latter also in terms of characteristic modes.
c�2016 George Kesidis
16
Total response - example (cont)
• Consider again the di↵erence equation:
8k � �1, y[k +1] + 3y[k] = 7f [k +1],
• i.e., Q(z) = z +3 with degree n = 1, and P (z) = 7z with degree m = 1,
• Exercise: Show that the following system corresponds to this di↵erence equation.
c�2016 George Kesidis
17
Total response - example (cont)
• By recursive substitution, the total response is, 8k � �1:
y[k] = �3y[k � 1] + 7f [k]= �3(�3y[k � 2] + 7f [k � 1]) + 7f [k]= (�3)2y[k � 2]� 3 · 7f [k � 1] + 7f [k]= ...
= (�3)k+1y[�1] +kX
r=0
7(�3)k�rf [r]
=: (�3)k+1y[�1] +1X
r=0
h[k � r]f [r] =: (�3)k+1y[�1] + (h ⇤ f)[k],
• where h[k] := 7(�3)ku[k] is the (zero state) unit-pulse response,
• y[�1] is the given (n = 1) initial condition, and
• we have defined the discrete-time convolution operator withP�1
r=0(...) := 0.
c�2016 George Kesidis
18
Total response - example (cont)
• Exercise: Prove by induction this expression for y[k] for all k � �1.
• Exercise: Prove convolution is commutative: h ⇤ f = f ⇤ h.
• So, we can write the total response y = yZI + yZS starting from the time of oldest initialcondition:
8k � �1, yZI[k] = (�3)k+1y[�1]
8k � �1, yZS[k] = u[k]kX
r=0
7(�3)k�rf [r] = u[k](h ⇤ f)[k]
where yZS[k] = 0 when k < 0.
• Obviously, this example involves a linear, time-invariant and causal system as described bythe di↵erence equation above.
c�2016 George Kesidis
19
Total response - discussion
• Note that in CMPSC 360, we don’t restrict our attention to linear and time-invariantdi↵erence equations.
• We use recursive substitution to guess at the form of the solution and then verify our guessby an inductive proof.
• In this course, we will describe a systematic approach to solve any LTIC di↵erence equation,
• i.e., to solve for the output of a DT-LTIC system given the input and initial conditions.
• And again as in continuous time, we will see important insights about discrete-time signalsand LTIC systems through frequency-domain representations and analysis.
c�2016 George Kesidis
20
ZIR - the characteristic values
• Note that 8k, ��rzk = zk+r = zrzk, i.e., the r-units time-advance operator, ��r, isreplaced by the scalar zr for all r 2 Z.
• Our objective is to solve for the ZIR, i.e., solve
Q(��1)y ⌘ 0 given y[�n], y[�n+1], ..., y[�2], y[�1].
• Note that exponential (or “geometric”) functions, {zk | k 2 Z} for z 2 C, are eigenfunc-tions of time-shift operators of the form Q(��1) for a polynomial Q.
• That is, for any non-zero scalar z 2 C, if we substitute y[k] = zk 8k 2 Z we get:
8k 2 Z, (Q(��1)y)[k] = Q(��1)zk = Q(z)zk.
• So, to solve Q(z)zk ⌘ 0 for all time k � 0, when z 6= 0 we require
Q(z) = 0, the characteristic equation of the system.
c�2016 George Kesidis
21
ZIR - the characteristic values (cont)
• If z is a root of the characteristic polynomial Q of the system, then
– z would be a characteristic value of the system, and
– the signal {zk}k�0 is a characteristic mode of the system when z 6= 0, i.e.,Q(��1)zk = 0 8k � 0.
• Since Q has degree n, there are n roots of Q in C, each a system characteristic value.
c�2016 George Kesidis
22
ZIR - the characteristic values (cont)
• Let n0 n be the number of non-zero roots of Q, i.e., Q(z) = Q(z)/zn�n0is a
polynomial satisfying Q(0) 6= 0.
• Though there may be some repeated roots of the characteristic polynomial Q, there willalways be n0 di↵erent, linearly independent characteristic modes, µk, i.e.,
8k � �n,
n0X
r=1
crµr[k] = 0 , 8r, scalars cr = 0.
• When n = n0, by system linearity, we will be able to write
8k � �n, yZI[k] =nX
r=1
crµr[k],
for scalars cr 2 C that are found by considering the given initial conditions
y[k] =nX
r=1
crµr[k] for k 2 {�n, ...,�2,�1},
i.e., n equations in n unknowns (cr).
• The linear independence of the modes implies linear independence of these n equations incr, and so they have a unique solution.
23
ZIR - the case of di↵erent, non-zero, real characteristic values
• If there are n di↵erent non-zero roots of Q in R, z1, z2, ..., zn, then there are n character-istic modes: for r 2 {1,2, ...n},
8 time k, µr[k] = zkr .
• Therefore,
8k � �n, yZI[k] =nX
r=1
crzkr .
• The n unknown scalars ck 2 R can be solved using the n equations:
y[k] =nX
r=1
crzkr , for k 2 {�n,�n+1, ...,�2,�1}.
c�2016 George Kesidis
24
ZIR - the case of di↵erent, non-zero, real characteristic values
• Example: Consider the di↵erence equation:
8k � �3, 2y[k +3]� 10y[k +2] + 12y[k +1] = 3f [k +2],
with y[�2] = 1 and y[�1] = 3.
• That is, Q(z) = z2 � 5z + 6 = (z � 3)(z � 2) and n = 2, P (z) = (3/2)z andm = 1.
• So, the n = 2 characteristic values are z = 3,2 and the ZIR is
8k � �n = �2, yZI[k] = c13k + c22
k
• Using the initial conditions to find the scalars c1, c2:
1 = y[�2] = c13�2 + c22
�2 and3 = y[�1] = c13
�1 + c22�1.
• Exercise: Now solve for c1 and c2.
• Note: When a coe�cient c is worked out to be zero, it may not be exactly zero in practice,and the corresponding characteristic mode zk will increasingly contribute to ZIR yZI overtime if |z| > 1 (i.e., an “unstable” mode in discrete time).
c�2016 George Kesidis
25
ZIR - the case of not-real characteristic values
• The characteristic polynomial Q may have non-real roots, but such roots come in complex-conjugate pairs because Q’s coe�cients ak are all real.
• For example, if the characteristic polynomial is
Q(z) = (z � 1)(z2 � 2z � 2)
then the characteristic values (Q’s roots) are
�1, 1± jp3 again recalling j =
p�1.
• Because we have three di↵erent characteristic values 2 C, we can specify three correspond-ing characteristic modes,
(�1)k, (1 + jp3)k, (1� j
p3)k, 8k � 0,
and construct the ZIR as
8k � �n = �3, yZI[k] = c1(�1)k + c2(1 + jp3)k + c3(1� j
p3)k
= c1(�1)k + c22kekj⇡/3 + c32
ke�kj⇡/3
where
– c1 2 R and c2 = c3 2 C so that yZI is real-valued, and again,
– these scalars are determined by the n = 3 given (real) initial conditions: y[�3], y[�2], y[�1].
26
ZIR - not-real characteristic values with real characteristic modes
• By the Euler-De Moivre identity for the previous example,
8k � �3, yZI[k] = c1(�1)k + (c2 + c3)2k cos(k⇡/3) + j(c2 � c3)2
k sin(k⇡/3)
= c1(�1)k +2Re{c2}2k cos(k⇡/3)� 2Im{c2}2k sin(k⇡/3)
• Again, because all initial conditions are real and Q has real coe�cients, yZI is real valuedand so c3 = c2 ) c2 + c3, j(c2 � c3) 2 R.
• In general, consider two complex conjugate characteristic values v ± jq corresponding totwo complex-valued characteristic modes |z|ke±jk\z, where |z| =
pv2 + q2 and
\z = arctan(q/v).
• One can use Euler’s identity to show that the corresponding real-valued characteristic modesare
|z|k cos(k\z), |z|k sin(k\z)
c�2016 George Kesidis
27
ZIR - the case of repeated characteristic values
• Consider the case where at least one characteristic value is of order > 1, i.e., there arerepeated roots of the characteristic polynomial, Q.
• For example, Q(z) = (z+0.75)3(z�0.5) has a triple (twice repeated) root at �0.75and a single root at 0.5.
• Again, {(�0.75)k} is a characteristic mode because Q(��1)(�.75)k ⌘ 0 follows from
(��1 + .75)(�.75)k = ��1(�.75)k + .75(�.75)k
= (�.75)k+1 + .75(�.75)k
= 0.
• Similarly, (0.5)k is a characteristic mode since (��1 � 0.5)(0.5)k ⌘ 0.
• Also, {k(�.75)k} is a characteristic mode because Q(��1)k(�.75)k ⌘ 0 follows from
(��1 + .75)2k(�.75)k
= (��2 + 1.5��1 + (.75)2)k(�.75)k
= ��2k(�.75)k +1.5��1k(�.75)k + (.75)2k(�0.75)k
= (k +2)(�.75)k+2 + 1.5(k +1)(�.75)k+1 + (.75)2k(�0.75)k
= (�.75)k+2((k +2)� 2(k +1)+ k)= 0.
28
ZIR - the case of repeated characteristic values (cont)
• Similarly, {k2(�.75)k} is also a characteristic mode because(��1 + .75)3k2(�.75)k = 0.
• Note that without three such linearly independent characteristic modes
{(�.75)k, k(�.75)k, k2(�.75)k ; k � 0}for the twice-repeated (triple) characteristic value -.75, the initial conditions will create an“overspecified” set of n equations involving fewer than n “unknown” coe�cients (ck) ofthe linear combination of modes forming the ZIR.
• For this example,
yZI[k] = c0(�0.75)k + c1k(�0.75)k + c2k2(�0.75)k + c3(0.5)
k, k � �4.
• If the given initial conditions are, say,
y[�4] = 12, y[�3] = 6, y[�2] = �5, y[�1] = 10,
the four equations to solve for the four unknown coe�cients ck are:
yZI[�4] = (�.75)�4c0 + (�4)(�.75)�4c1 + (�4)2(�.75)�4c2 + (.5)�4c3 = 12yZI[�3] = (�.75)�3c0 + (�3)(�.75)�3c1 + (�3)2(�.75)�3c2 + (.5)�3c3 = 6yZI[�2] = (�.75)�2c0 + (�2)(�.75)�2c1 + (�2)2(�.75)�2c2 + (.5)�2c3 = �5yZI[�1] = (�.75)�1c0 + (�1)(�.75)�1c1 + (�1)2(�.75)�1c2 + (.5)�1c3 = 10
29
ZIR - general case of repeated, non-zero characteristic values
• In general, a set of r linearly independent modes corresponding to a non-zero characteristicvalue z 2 C repeated r � 1 times are
kr�1zk, kr�2zk, ..., kzk, zk, for k � 0.
• Also, if v ± jq are characteristic values repeated r � 1 times, with v, q 2 R and q 6= 0,we can use the 2k real-valued modes
ka|z|k cos(k\z), ka|z|k sin(k\z), for a 2 {0,1,2, ..., r � 1},
where |z| =p
v2 + q2 and \z = arctan(q/v).
c�2016 George Kesidis
30
ZIR - when some characteristic values are zero
• Again let n0 n be the number of non-zero roots of Q (characteristic values),
• i.e., r := n� n0 � 0 is the order (1+repetition) of the characteristic value 0, and
• r � 0 is the smallest index such that (the coe�cient of Q) ar 6= 0.
• So, there is a polynomial Q such that Q(z) = zrQ(z) and Q(0) 6= 0.
• Because the constant signal zero cannot be a characteristic mode, we add r = n � n0
time-advanced unit-pulses:
8k � �n, yZI[k] =rX
i=1
Ci�[k + i] + yN[k]
= Cr�[k + r] + Cr�1�[k + r � 1] + ...+ C1�[k +1] + yN[k]
where yN is a “natural response” (linear combination of n0 characteristic modes).
• The n initial conditions are then met by the r coe�cients Ci of the advanced unit pulsestogether with the n0 = n� r coe�cients of the characteristic modes in yN.
c�2016 George Kesidis
31
ZIR - when some characteristic values are zero - example
• Consider a fourth-order system with characteristic polynomialQ(z) = z2(z +1)2.
• Thus the poles are 0,�1 each repeated and the (non-zero) characteristic modes are(�1)k, k(�1)k.
• So, the ZIR is, for k � �4:
yZI[k] = C2�[k +2] + C1�[k +1] + c1(�1)k + c2k(�1)k
• That is, the ZIR has four unknown coe�cients C2, C1, c1, c2 to account for the four (given)initial conditions y[�4], y[�3], y[�2], y[�1].
c�2016 George Kesidis
32
Zero State Response - the unit-pulse response
• Recall the LTIC system
nX
r=0
ar��ry =: Q(��1)y = P (��1)f :=
mX
r=0
br��rf
with an ⌘ 1, a0 6= 0 or b0 6= 0, m n.
• We can express any input signal
f [k] =1X
r=0
f [r]�[k � r] 8k � 0, i.e., 8f , f = f ⇤ �.
• So the unit pulse � is the identity of the convolution operator in discrete time.
• Thus, by LTI, the ZSR yZS is the convolution of input f and ZSR h to unit pulse �,
yZS[k] =1X
r=0
f [r]h[k � r] = (f ⇤ h)[k], 8k � 0,
• h is called the unit-pulse response of the LTIC system, i.e.,
Q(��1)h = P (��1)� s.t. h[k] = 0 8k < 0.
33
Computing an LTIC system’s unit-pulse response, h
• For the LTIC system in standard form, if a0 6= 0 then
h = (b0/a0)� + yNu
where yN is a natural response of the system (linear combination of characteristic modes).
• Note that h[k] = 0 for all k < 0 owing to the unit step u.
• The n scalars of the natural response yN component of h are solved using
(Q(��1)h)[k] = (P (��1)�)[k] for k 2 {�n,�n+1, ...,�2,�1}
c�2016 George Kesidis
34
Unit-pulse response when zero is a characteristic value
• If r � 0 is the smallest index such that ar 6= 0 (0 is a char. mode of order r), then mayneed to add r delayed unit-pulse terms to h:
h =r�1X
i=0
Ai�i� + (b0/ar)�
r� + yNu,
where
– by definition of the standard form of the di↵erence equation, if r > 0, a0 = 0 sob0 6= 0, and
– r n since 0 6= an := 1.
• So if r = 0 (i.e., a0 6= 0), then A0 = b0/a0 as above, whereP�1
i=0(...) := 0.
• Exercise: Prove Ar = b0/ar for 0 r n.
• Thus, zero is a characteristic value of degree r � 0, and
• there are r characteristic modes that will all be zero.
• The additional unit-pulse terms introduce r degrees of freedom in the form of the coe�cientsA0, A1, ..., Ar�1 to accommodate the n = r + n0 initial conditions of the unit-pulseresponse: h[�n] = h[�n+1] = ... = h[�2] = h[�1] = 0.
35
Computing the ZSR - example 1
• Recall that the di↵erence equation y = 7f � 3�y corresponds to the above system; instandard form:
8k � �1, y[k +1] + 3y[k] = 7f [k +1].
with Q(z) = z +3, P (z) = 7z and n = 1 = m.
• Since the system characteristic value is�3 and b0 = 0, the (zero state) unit-pulse responsehas the form h[k] = c(�3)ku[k].
• The scalar c is solved by evaluating the above di↵erence equation at time k = �1:
(Q(��1)h)[�1] = (P (��1)�)[�1]i.e., h[0] + 3h[�1] = 7�[0]
) c+3 · 0 = 7 · 1, c = 7
36
Computing the ZSR - example 1 (cont)
• So, h[k] = 7(�3)ku[k].
• If the input is f [k] = 4(0.5)ku[k], the system ZSR is, for all k � 0,
yZS[k] =kX
r=0
h[r]f [k � r] =kX
r=0
7(�3)r4(0.5)k�r
= 28(0.5)kkX
r=0
(�6)r = 28(0.5)k(�6)k+1 � 1
�6� 1u[k]
= (24(�3)k +4(0.5)k)u[k].
• Note how the ZIR yZI has a term that is a characteristic mode (excited by the input f)and a term that is proportional to the input f (this forced response is an eigenresponse).
• Exercise: For the di↵erence equation, y[k + 1] + 3y[k] = 7f [k] 8k � �1: draw theblock diagram, show that h[k] = 21(�3)k�1u[k] + (7/3)�[k], and find the ZSR tothe above input f .
• Exercise: Read “sliding tape” method to compute convolution in Lathi, p. 595.
c�2016 George Kesidis
37
Computing the unit pulse response - example 2
• Find the ZSR of the following system to input f [k] = 2(�5)ku[k]:
• Exercise: show the di↵erence equation for this system (in direct canonical form) is:
8k � 0, y[k +2]� 5y[k +1] + 6y[k] = 1.5f [k +1]
• That is, Q(z) = z2 � 5z + 6 = (z � 3)(z � 2) and n = 2, P (z) = 1.5z andm = 1.
• So, the n = 2 characteristic values are z = 3,2 and b0 = 0 so the unit-pulse response
h[k] = (c13k + c22
k)u[k].
c�2016 George Kesidis
38
Computing the unit pulse response - example 2 (cont)
• To find the constants, evaluate the di↵erence equation at k = �1:
2h[1]� 10h[0] + 12h[�1] = 3�[0]) 2h[1]� 10h[0] = 3
) (2 · 3� 10 · 1)c1 + (2 · 2� 10 · 1)c2 = 3) �4c1 +�6c2 = 3
and at k = �2:
2h[0]� 10h[�1] + 12h[�2] = 3�[�1] ) 12h[0] = 0 ) h[0] = 0) c1 + c2 = 0.
• Thus, c2 = �1.5 = �c1 so that h[k] = (�1.5(3)k +1.5(2)k)u[k] and for k � 0
yZS[k] = (h ⇤ f)[k] =kX
r=0
h[r]f [k � r].
• Exercise: Write the ZSR as a sum of system modes 2k and 3k and a (force) term like theinput, here taken as f [k] = 4(�5)ku[k].
c�2016 George Kesidis
39
Convolution - other important properties
• Again, for a LTI system with impulse response h and input f , the ZSR is yZS = f ⇤ h,where
(f ⇤ h)[k] =1X
r=�1f [r]h[k � r]
• By simply changing the dummy variable of summation to r0 = h�r, can show convolutionis commutative: f ⇤ h = h ⇤ f .
• One can directly show that convolution f ⇤ h is a bi-linear mapping from pairs of signals(f, h) to signals (yZS), consistent with convolution’s commutative property and the (zerostate) system with impulse response h being LTI;
• that is, 8 signals f, g, h and scalars ↵,� 2 C,(↵f + �g) ⇤ h = ↵(f ⇤ h) + �(g ⇤ h)
• By changing order of summation (Fubini’s theorem), one can easily show that convolutionis associative, i.e., 8 signals f, g, h,
(f ⇤ g) ⇤ h = f ⇤ (g ⇤ h).
c�2016 George Kesidis
40
Convolution - other important properties (cont)
• We’ll use these properties when composing more complex systems from simpler ones.
• By just changing variables of integration, we can show how to exchange time-shift withconvolution, i.e., 8 signals f, h : Z ! C and times k 2 Z,
(�kf) ⇤ h = �k(f ⇤ h);
recall how convolution represents the ZSR of linear and time-invariant systems.
• By the ideal sampling property, recall that the identity signal for convolution is the unitpulse �, i.e., 8 signals f ,
f ⇤ � = � ⇤ f = f
• Exercise: Adapt the proofs of these properties in continuous time to this discrete-timecase.
• Exercise: In particular, show that if f and h are causal signals, then y = f ⇤ h is causal;i.e., if the unit-pulse response h of a system is a causal signal, then the system is causal.
c�2016 George Kesidis
41
System stability - ZIR - asymptotically stable
• Consider a SISO system with input f and output y.
• Recall that the ZIR yZI is a linear combination of the system’s characteristic modes, wherethe coe�cients depend on the initial conditions, possibly including some initial unit-pulseterms if zero is a characteristic value (system pole).
• A system is said to be asymptotically stable if for all initial conditions,
limk!1
yZI[k] = 0.
• So, a system is asymptotically stable if and only if all of its characteristic values havemagnitude less than 1.
c�2016 George Kesidis
42
System stability - ZIR - asymptotically stable: Example
• If the characteristic polynomial Q(z) = (z � 0.5)(z2 + 0.0625), then
• the system’s characteristic values (roots of Q) are 0.5,±0.25j each with magnitude lessthan one,
• and the ZIR is of the form,
yZI[k] =�c1(0.5)
k + c2(0.25j)k + c2(�0.25j)k
�u[k]
=�c1(0.5)
k +2Re{c2}(0.25)k cos(k⇡/2)� 2Im{c2}(0.25)k sin(k⇡/2)�u[k],
• recalling that jk = ejk⇡/2.
• So, yZI[k] ! 0 as k ! 1 for all c1, c2 (i.e., for all initial conditions), and
• hence is asymptotically stable.
43
System stability - bounded signals
• A signal y is said to be bounded if
9 M < 1 s.t. 8k 2 Z, |y[k]| M ;
otherwise y is said to be unbounded.
• For example, y[k] = 0.25(1+jp3
2)ku[k] is bounded (can use M = 0.25).
• Also, 3cos(5k) is bounded (can use M = 3).
• But both 2k cos(5k) and 3 · (�2)k are unbounded.
c�2016 George Kesidis
44
System stability - ZIR - marginally stable
• A system is said to be marginally stable if it is not asymptotically stable but yZI is always(for all initial conditions) bounded.
• A system is marginally stable if and only if
– it has no characteristic values with magnitude strictly greater than 1,
– it has at least one characteristic value with magnitude exactly 1, and
– all magnitude-1 characteristic values are not repeated.
• That is, a marginally stable system has
– some characteristic modes of the form cos(⌦k) or sin(⌦k),
– while the rest of the modes are all of the form kr|z|k cos(⌦k) or kr|z|k sin(⌦k),with |z| < 1 and integer degree r � 0.
– Exercise: Explain why we can take ⌦ 2 (�⇡,⇡] without loss of generality.
– Note: the dimension of frequency ⌦ is [⌦] = radians per unit time.
c�2016 George Kesidis
45
System stability - ZIR - marginally stable: Example
• The characteristic polynomial is Q(z) = z(z2+1)(z�0.25) gives characteristic values0,0.25,±j,
• then the system is marginally stable with modes (0.25)k, cos(k⇡/2), sin(k⇡/2),
• the last two of which are bounded but do not tend to zero as time k ! 1.
c�2016 George Kesidis
46
System stability - ZIR - unstable
• A system that is neither asymptotically nor marginally stable (i.e., a system with unboundedmodes) is said to be unstable.
• For example, the system with Q(z) = (z2 � 0.5)(z + 3) is unstable owing to thecharacteristic value �3 with unbounded mode (�3)k.
• For another example, if the characteristic polynomial is Q(z) = (z2+1)2(z�0.5) thenthe purely imaginary characteristic values ±j are repeated, and hence the two additionalmodes k sin(k⇡/2), k cos(k⇡/2) are unbounded, so this system is unstable.
• Similarly, if Q(z) = (z2 � 1)2(z � 0.5) then the characteristic values ±1 are repeatedand the modes k and k(�1)k are unbounded, so this system is unstable too.
c�2016 George Kesidis
47
ZIR stability - stability of poles
c�2016 George Kesidis
48
System stability - ZSR - BIBO stable
• A SISO system is said to be Bounded Input, Bounded Output (BIBO) stable if8 bounded input signals f , the ZSR yZS is bounded.
• A su�cient condition for BIBO stability is absolute summability of the unit-pulse response,
1X
k=0
|h[k]| < 1.
• To see why: If the input f is bounded (by Mf with 0 Mf < 1) then 8k � 0:
|yZS[k]| = |(f ⇤ h)[k]|
=
�����
kX
r=0
f [k � r]h[r]
�����
kX
r=0
|f [k � r]h[r]| (by the triangle inequality)
kX
r=0
Mf |h[r]|
Mf
1X
r=0
|h[r]| =: My < 1,
49
System stability - ZSR - BIBO stable
• The condition of absolute summability of the unit-pulse response,
1X
r=0
|h[r]| < 1,
is also necessary for, and hence equivalent to, BIBO stability.
• If any component characteristic mode of h is unbounded, then h will not to be absolutelysummable.
• Thus, if the system (ZIR) is asymptotically stable it will be BIBO stable; the converse isalso true.
c�2016 George Kesidis
50
ZSR - the transfer function, H
• Recall that for any polynomial Q and z 2 C (including s = jw, w 2 R),
Q(��1)zk = Q(z)zk, 8k � 0.
• So, if we guess that a “particular” solution of the system Q(��1)y = P (��1)f withinput f [k] = Azku[k] is of the form yp[k] = AH(z)zk = H(z)f [k], k � 0, then weget by substitution that 8k � 0, z 2 C,
(Q(��1)yp)[k] = (P (��1)f)[k] ) AH(z)Q(z)zk = AP (z)zk
) H(z) = P (z)/Q(z).
• The “rational polynomial” H = P/Q is known as the system’s transfer function and willfigure prominently in our study of frequency-domain analysis.
• So, the ZSR (forced response + characteristic modes) would be of the form:
yZS[k] = (AH(z)zk + linear combination of char. modes)u[k].
• Recall that for the example with Q(z) = z + 3 and P (z) = 7z, we computed theunit-pulse response h[k] = 7(�3)ku[k] and the ZSR to input f [k] = 4(0.5)ku[k] asyZS[k] = (24(�3)k +4(0.5)k)u[k].
• Here, note that H(0.5) = P (0.5)/Q(0.5) = 1, i.e., the forced response componentof yZS is H(0.5)f [k] = 1 · 4(0.5)ku[k] = 4(0.5)ku[k].
51
ZSR - unit-pulse response h, transfer function H, and eigenresponse
• yZS[k] = (H(z)Azk + linear combination of char. modes )u[k] is the ZSR to inputf [k] = Azku[k], where H(z) = P (z)/Q(z).
• The eigenresponse is a special case of the forced response for exponential inputs.
• If |z| = 1, i.e., z = ej⌦ for some⌦ 2 (�⇡,⇡] (w.l.o.g.), and the system is asymptoticallystable, then the ZSR tends to the steady-state eigenresponse of the system:
y[k] ! AH(ej⌦)ej⌦k as k ! 1.
• Since y = h ⇤ f , we get that as k ! 1 for a LTIC and asymptotically stable system,
yZS[k] =kX
r=0
h[r]Aej⌦(k�r)
= Aej⌦kkX
r=0
h[r]e�j⌦r ! Aej⌦kH(ej⌦),
)1X
r=0
h[r]e�j⌦r = H(ej⌦), 8⌦ 2 (�⇡,⇡].
52
ZSR - transfer function H and eigenresponse (cont)
• The LTIC system transfer function H is the z-transform of the system unit-pulse responseh:
H(z) =1X
k=0
h[k]z�k,
where z 2 C is in H’s “region of convergence”.
• Note that H(ej⌦) is periodic since H(ej⌦) ⌘ H(ej⌦+2⇡k) for any integer k.
• For the z-transform (and DTFS) we will use this notation for H, but for the DTFT we willwrite H(⌦) instead of H(ej⌦).
c�2016 George Kesidis
53
Frequency-domain methods for discrete-time signals
• Discrete-Time Fourier Series (DTFS) of periodic signals
• Discrete-Time Fourier Transform (DTFT)
• sampled data systems
* DFT & FFT
• z-transform for (complete) transient response
• eigenresponse
• canonical system realization of a di↵erence equation
c�2016 George Kesidis
54
Discrete-time Fourier series of periodic signals
• For all r,N 2 Z, note that the signal {exp(jr2⇡Nk) | k 2 Z} “repeats itself” everyN > 0
units of (discrete) time k, in particular
8r 2 Z, ejr2⇡
Nk|k=0 = 1 = ejr
2⇡
Nk|k=N
• Also the signals {exp(jr2⇡Nk) | k 2 Z} ⌘ {exp(jr02⇡
Nk) | k 2 Z} whenever r0 = r
mod N .
• Suppose N is the period of periodic signal x = {x[k] | k 2 Z} and⌦o = 2⇡/N be the fundamental frequency of x (recall [⌦o] = radians/unit-time).
• We can write x as a Discrete-Time Fourier Series (DTFS):
8k 2 Z, x[k] =N�1X
r=0
Drejr⌦ok.
where r indexes x’s N harmonics.
• Note that the DTFS can also be written for any discrete-time signal x : A ! R definedover any finite interval of time, e.g., A = {0,1,2, ..., N � 1} orA = {�N,�N +1, ...,�1} for integer N < 1.
c�2016 George Kesidis
55
Discrete-time Fourier series of periodic signals (cont)
• Consider the N signals ⇠r[k] := ejr⌦ok over any time-interval A of length N .
• Equivalently consider these N signals ⇠r as N -vectors in RN , i.e., the kth entry of vector⇠r is ⇠r[k].
• If these signals/vectors {⇠r}N�1r=0 are linearly independent, then they will form a basis span-
ning all other signals x : A ! R, equivalently all other vectors x 2 RN ,
• i.e., any such x can be written as a linear combination of the {⇠r}N�1r=0 giving the DTFS of
x:
xr =N�1X
r=0
Dr⇠r.
• If we show that these signals/vectors {⇠r}N�1r=0 are orthogonal then
– linear independence follows
– the rth coordinate Dr (DTFS coe�cients) is found by simply projecting x onto thevector ⇠r:
Dr = hx, ⇠ri/||⇠r||2.
c�2016 George Kesidis
56
DTFS - coe�cients (cont)
• Consider any period of x : Z ! R, say {0,1,2, ..., N � 1}.
• First note that for any v 2 Z that is not a multiple of N (so ejv⌦o = ejv(2⇡/N) 6= 1), thegeometric series
N�1X
k=0
ejv⌦ok =N�1X
k=0
⇣ejv2⇡/N
⌘k=
ejv(2⇡/N)·N � ejv(2⇡/N)·0
ejv2⇡/N � 1= 0.
• Thus, for any r 6= v 2 Z such that N 6 |(v � r), the inner product h⇠r, ⇠vi =
h{ejr(2⇡/N)k}, {ejv(2⇡/N)k}i :=N�1X
k=0
ejr(2⇡/N)kejv(2⇡/N)k =N�1X
k=0
ej(r�v)(2⇡/N)k = 0,
recalling that the inner product is conjugate-linear in the second argument so that< x, x >= ||x||2 when x is C-valued.
• So, these signals are orthogonal and the DTFS coe�cients of N -periodic x are
Dr =1
N
N�1X
k=0
x[k]e�jr⌦ok =hx, {ejr⌦ok}i||{ejr⌦ok}||2
, where ⌦o =2⇡
N.
c�2016 George Kesidis
57
DTFS - checking coe�cients
• Let’s now compute the inner product of ⇠v, for any v 2 {0,1, ..., N �1}, with the DTFSof N -periodic x:
hx, {ejv⌦ok}i =N�1X
k=0
x[k]e�jv⌦ok =N�1X
k=0
N�1X
r=0
Drejr⌦oke�jv⌦ok
=N�1X
r=0
Dr
N�1X
k=0
ej(r�v)⌦ok =N�1X
r=0
DrN�(r � v) = DvN
• Again, we have verified the DTFS coe�cients is
Dr =1
N
N�1X
k=0
x[k]e�jr⌦ok =hx, {ejr⌦ok}i||{ejr⌦ok}||2
, ⌦o =2⇡
N
c�2016 George Kesidis
58
DTFS - summary for N-periodic signal x
c�2016 George Kesidis
59
DTFS - example
Problem:Identify the DTFS coe�cients (if they exist) for
x[k] = 7 sin(5.7⇡k) + 2cos(3.2⇡k), k 2 Z.Solution:
• First note that the two components of x are periodic, so their sum is periodic.(Why is this so in discrete time?)
• Since sin and cos have period 2⇡, we can subtract integer multiples of 2⇡ to get
x[k] = 7 sin(1.7⇡k) + 2cos(1.2⇡k).
• 1.7⇡k is an integer multiple of 2⇡ when (integer) k = 20, and when k = 5 for 1.2⇡k,so least common multiple of these periods is k = 20.
• (Show that one can alternatively find the greatest common divisor of the component fre-quencies.)
c�2016 George Kesidis
60
DTFS - example (cont)
• Thus, the period of x is N = 20 and the fund. frequ. is ⌦o = 2⇡/N = 0.1⇡.
• By Euler’s identity and adding 2⇡k to the negative exponents,
x[k] =7
2jej1.7⇡k �
7
2je�j1.7⇡k + ej1.2⇡k + e�j1.2⇡k
= �3.5jej1.7⇡k +3.5jej0.3⇡k + ej1.2⇡k + ej0.8⇡k.
• So, the DTFS of x[k] =P19
r=0Drejr0.1⇡k with
D17 = �3.5j = 3.5e�j⇡/2, D3 = 3.5j = 3.5ej⇡/2, D12 = 1, and D8 = 1;
else Dr = 0 (incl. the fundamental r 2 {1,19} & DC r = 0 components).
c�2016 George Kesidis
61
DTFS - example and exercise
• Example: The DTFS of an even rectangle wave with period N = 6 and duty cycle 3:
x[k] =1X
`=�1
�6`(��1u��2u)[k] =1X
`=�1
(u[k +1� 6`]� u[k � 2� 6`])
is =5X
r=0
Drejr⌦ok,
where the fund. freq. ⌦o = 2⇡/6 and, 8r 2 Z,
Dr =1
6
2X
k=�3
x[k]e�jr⌦ok =1
6
1X
k=�1
1 · e�jr(2⇡/6)k =1
6(1+ 2cos(r(2⇡/6)k)).
• Exercise: Plot x[k] as a function of time k and plot its (periodic) spectrum:8r 2 {0,1,2, ...,5}, ` 2 Z,
X(r2⇡/6+ 2⇡`) = Dr.
62
DTFS - Parseval’s theorem
• The average power of the N -periodic discrete-time signal x is
Px =1
N
N�1X
k=0
|x[k]|2 =1
N
N�1X
k=0
x[k]x[k];
equivalently, the sum could be taken over any interval of length N 2 Z>0.
• Substituing the Fourier series of x separately for x[k] and x[k] (using a di↵erent summation-index variable for each substitution), leads to Parseval’s theorem
Px =N�1X
r=0
|Dr|2.
• Parseval’s theorem can be used to determine the amount of periodic signal x’s power residesin a given frequency band [⌦1,⌦2] ⇢ [0,2⇡] radians/unit-time:
1. determine the harmonics r⌦o of x that reside in this band, i.e., integers r 2 [⌦1/⌦o,⌦2/⌦o]where x’s fundamental frequency ⌦o = 2⇡/N .
2. sum just over these harmonics to get the answer,P
d⌦1/⌦oerb⌦2/⌦oc |Dr|2.
c�2016 George Kesidis
63
DTFS - Parseval’s theorem example
• Find the fraction of x’s average power in the frequency band [0.4⇡,1.1⇡] where
8k 2 Z, x[k] =1X
v=�1(3�[k � 4v]� 4�[k � 1� 4v])
• Solution: x has period N = 4 and average power
Px =1
N
N�1X
k=0
|x[k]|2 =1
4
3X
k=0
|x[k]|2 =1
4(32 + (�4)2 + 02 + 02) =
25
4
• x has fundamental frequency ⌦o = 2⇡/N = ⇡/2 radians/unit-time and discrete-timeFourier coe�cients
Dr =1
N
N�1X
k=0
x[k]e�j(r⌦o)k =1
4
⇣3� 4e�jr⇡/2
⌘, 0 r N � 1 = 3.
• The harmonics r of x that reside in [0.4⇡,1.1⇡] satisfy 0.4⇡ r⌦o = r⇡/2 1.1⇡,i.e., r 2 {1,2}.
• So, by Parseval’s theorem, the answer is (|D1|2 + |D2|2)/Px.
c�2016 George Kesidis
64
Periodic extensions
• Consider signal x : Z ! R having finite support {�M,�M + 1, ...,0, ...,M � 1,M}for 0 < M < 1; i.e., 8|k| > M, x[k] = 0.
• For N � M , define 2N -periodic x(N) such that
x(N)[k] =
⇢x[k] if |k| M0 if M < |k| N
• x(N) is a periodic extension of the finite-support signal x, where again x(N)’s period is 2Nand
limN!1
x(N) = x.
c�2016 George Kesidis
65
DTFS of periodic extension leading to DTFT
• For r 2 {�N +1,�N +2, ..., N � 1, N}, the DTFS of x(N) has coe�cients
D(N)r =
1
2N
NX
k=�N+1
x(N)[k]e�jr 2⇡
2Nk
=1
2N
MX
k=�M
x[k]e�jr 2⇡
2Nk
=1
2N
1X
k=�1
x[k]e�jr 2⇡
2Nk
=:1
2NX
✓r2⇡
2N
◆,
where the Discrete-Time Fourier Transform (DTFT) of (aperiodic) x : Z ! R isX : R ! C:
X(⌦) :=1X
k=�1
x[k]e�j⌦k =: (Fx)(⌦), ⌦ 2 R
• Note that Fourier integrals (spectra of discrete-time signals) are periodic, repeating them-selves every 2⇡ radians: 8⌦ 2 R, ` 2 Z,
X(⌦) = X(⌦+ `2⇡).
66
Inverse DTFT by Fourier Integral
• Thus, 8k 2 Z,
x[k] = limN!1
x(N)[k]
= limN!1
NX
r=�N+1
D(N)r ejr
2⇡
2Nk
= limN!1
NX
r=�N+1
X
✓r2⇡
2N
◆ejr
2⇡
2Nk 1
2N·2⇡
2⇡
=1
2⇡
Z ⇡
�⇡
X(⌦)ej⌦kd⌦
where the last equality is by Riemann integration with 2⇡/(2N) ! d⌦.
• Thus, we have derived the inverse DTFT by Fourier integral of X giving (aperiodic) x,
8k 2 Z, x[k] =1
2⇡
Z ⇡
�⇡
X(⌦)ej⌦kd⌦ =: (F�1X)[k].
c�2016 George Kesidis
67
DTFT Examples - exponential signal
• If x = � then obviously X ⌘ 1.
• The geometric signal x[k] = �ku[k] for scalar � s.t. |�| < 1 has DTFT
X(⌦) =1X
k=0
�ke�j⌦k =1X
k=0
(�e�j⌦)k
=1
1� �e�j⌦=
1
(1� �) cos(⌦) + j� sin(⌦)
• Note that
|X(⌦)| =1
(1� � cos(⌦))2 + �2 sin2(⌦)=
1
1+ �2 � 2� cos(⌦)
\X(⌦) = � arctan
✓� sin(⌦)
1� � cos(⌦)
◆
c�2016 George Kesidis
68
DTFT Examples - exponential signal (cont)
• The plots above are for � = 0.5.
• Note how X has period 2⇡.
• Exercise: What are the maximum and minimum values of |X|, i.e., how would this plotdepend on � > 0? Plot x and \X. How do these plots di↵er when �1 < � < 0?
• Exercise: Find the DTFT of anticausal signal x[k] = �ku[�k] for scalar � s.t. |�| > 1.
• Exercise: Find the DTFT of x[k] = �|k|, k 2 Z, for scalar � s.t. |�| < 1.
c�2016 George Kesidis
69
DTFT Examples - Square and Triangle Pulse
• For T 2 Z>0, the even rectangle pulse with support 2T +1,x = ��Tu��T+1u (i.e., x[k] = u[k + T ]� u[k � (T +1)]), has DTFT
X(⌦) =TX
k=�T
1e�j⌦k = 1+ 2TX
k=1
cos(k⌦), ⌦ 2 R.
• Exercise (even rectangle pulse in frequency domain):Show that for fixed ⌦0 s.t. 0 < ⌦0 < ⇡,
F�1{��⌦0 u��⌦0 u}[k] =⌦0
⇡sinc(⌦0k), k 2 Z.
• For T 2 Z>0, the odd triangle pulse with support 2T +1,x[k] ⌘ k(��Tu[k]��T+1u[k]) has DTFT
X(⌦) =TX
k=�T
ke�j⌦k = � 2jTX
k=1
k sin(k⌦), ⌦ 2 R.
c�2016 George Kesidis
70
DTFT Examples - exponential sinusoid
• For fixed time K0, clearly
F{�[k �K0]}(⌦) = ejK0⌦,
where here � is the unit pulse.
• Note that ejK0⌦ is a sinusoidal function of ⌦ with period 2⇡/K0 radians.
• Exercise: For fixed frequency ⌦0 radians/unit-time, show that
F{e�j⌦0k}(⌦) = 2⇡1X
v=�1�(⌦�⌦0 + 2⇡v),
where here � is the Dirac impulse (in the frequency domain ⌦ 2 R). Hint: work with F�1.
• So, the DTFT of a N -periodic signal with Fourier series
N�1X
r=0
Drejr2⇡
Nk F! 2⇡
1X
v=�1
N�1X
r=0
Dr�(⌦� r2⇡
N+2⇡v)
c�2016 George Kesidis
71
DTFT - Time shift and frequency shift properties
• If fixed K0 2 Z and X = F{x} then
F{�K0x}(⌦) =1X
k=�1
(�K0x)[k]e�j⌦k
=1X
k=�1
x[k �K0]e�j⌦k
=1X
k0=�1
x[k0]e�j(k0+K0)⌦
= e�jK0⌦X(⌦),
i.e., shift in time by K0 corresponds to product with sinusoid of period 2⇡/K0 (linearphase shift) in frequency domain.
• Exercise: Prove the dual property that if fixed ⌦0 2 R and X = F{x} then
F{x[k]ej⌦0k}(⌦) = X(⌦�⌦0),
i.e., modulation (multiplication by a sinusoid) in time domain results in frequency shift.
c�2016 George Kesidis
72
DTFT - convolution properties
• Let Xr = F{xr} for r 2 {1,2}.
F{x1 ⇤ x2}(⌦) :=1X
k=�1
(x1 ⇤ x2)[k]e�j⌦k
:=1X
k=�1
1X
l=�1
x1[l]x2[k � l]e�j(k�l)⌦e�jl⌦ i.e., ⇥ejl⌦e�jl⌦ = 1
=1X
k0=�1
1X
l=�1
x1[l]x2[k0]e�jk0⌦e�jl⌦ where k0 = k � l
=1X
l=�1
x1[l]e�jl⌦1X
k0=�1
x2[k0]e�jk0⌦ =: X1(⌦)X2(⌦)
• Exercise: Prove the dual property that
F{x1x2}(⌦) =1
2⇡(X1 ⇤X2)(⌦) :=
1
2⇡
Z
2⇡X1(v)X2(⌦� v)dv.
• Exercise: Use the convolution properties to prove the time and frequency shift properties.Hint: (�Ko�) ⇤ x = �Kox.
• Exercise: Show that DTFT is a linear operator.
73
DTFT - Parseval’s Theorem
• The energy of a signal DT x is
Ex :=1X
k=�1
|x[k]|2 =1X
k=�1
x[k]x[k] =1X
k=�1
(F�1X)[k](F�1X)[k]
=1X
k=�1
1
2⇡
Z
2⇡X(⌦0)ej⌦
0kd⌦0 1
2⇡
Z
2⇡X(⌦)ej⌦kd⌦
=1X
k=�1
1
2⇡
Z
2⇡X(⌦0)ej⌦
0kd⌦0 1
2⇡
Z
2⇡X(⌦)e�j⌦kd⌦
=1
2⇡
Z
2⇡X(⌦0)
Z
2⇡X(⌦)
1
2⇡
1X
k=�1
ej⌦0ke�j⌦k
!d⌦d⌦0
=1
2⇡
Z
2⇡X(⌦0)
Z
2⇡X(⌦)
1
2⇡
�2⇡�(⌦�⌦0)
�d⌦d⌦0
=1
2⇡
Z
2⇡X(⌦0)X(⌦0)d⌦0 =
1
2⇡
Z
2⇡|X(⌦0)|2d⌦0, recalling that for
fixed ⌦0: F�1{2⇡�(⌦�⌦0)}[k] = e�j⌦0k &R2⇡ X(⌦)�(⌦�⌦0)d⌦ = X(⌦0).
c�2016 George Kesidis
74
DTFT - Parseval’s Theorem - example
• The even rectangle pulse with support 2T +1, x = ��Tu��T+1u has energy
Ex =1X
k=�1
|x[k]|2 =TX
k=�T
12
= 2T +1.
• Recall its DTFT is X(⌦) =PT
k=�T e�j⌦k, so
1
2⇡
Z
2⇡|X(⌦)|2d⌦ =
1
2⇡
Z
2⇡X(⌦)X(⌦)d⌦ =
1
2⇡
Z
2⇡
TX
k=�T
e�j⌦kTX
k0=�T
e+j 0⌦kd⌦
=1
2⇡
Z
2⇡
TX
k=�T
1+X
k<k0
ej(k0�k)⌦
!d⌦
=TX
k=�T
1
2⇡
Z
2⇡1d⌦+
X
k<k0
1
2⇡
Z
2⇡ej(k
0�k)⌦d⌦ =TX
k=�T
1 + 0
= 2T +1.
75
DTFT - Parseval’s Theorem - exercises
• Exercise: Repeat this calculation using X(⌦) = 1+ 2PT
k=1 cos(⌦k).
• Exercise: Compute amount of energy of x in the frequency band [�⇡/6,⇡/6] radians/unit-time, i.e.,
1
2⇡
Z ⇡/6
�⇡/6|X(⌦)|2d⌦
c�2016 George Kesidis
76
Analysis of Stable DT LTI Systems in Steady-State
• Consider a SISO, DT-LTIC system described by the di↵erence equation
Q(��1)y = P (��1)f,
where f is the input and y is the ZSR (output).
• Recall that by the time-shift property,
Q(ej⌦)YZS(⌦) = P (ej⌦)F (⌦) ) YZS(⌦) = H(⌦)F (⌦).
• We now re-derive from first principles the eigenresponse by first recalling that the ZSRyZS = f ⇤ h where h is the unit-pulse response.
• Taking DTFTs, YZS = HF where H = Fh is the transfer function.
• Suppose the system is BIBO/asymptotically stable, i.e., the n roots of Q (system char.modes/poles) z all have modulus |z| < 1.
• The ZSR will consist of a forced response plus characteristic modes, where the latter will! 0 over time (our stability assumption) so that the forced response becomes the steady-state response.
c�2016 George Kesidis
77
Analysis of Stable DT LTI Systems in Steady-State (cont)
• The forced response to a persistent sinusoidal input
f [k] = Afej(⌦ok+�f)
will be of the form
yss[k] = Ayej(⌦ok+�y)
where (for k � 0),
Q(ej⌦o)yss[k] = (Q(��1)yss)[k] = (P (��1)f)[k] = P (ej⌦o)f [k].
) yss[k] =P (ej⌦0)
Q(ej⌦0)f [k]
• Also, the ZSR yZS = h ⇤ f , i.e., for all time k � 0:
yZS[k] =kX
v=0
h[v]Afej(⌦o(k�v)+�f) = f [k]
kX
v=0
h[v]e�j⌦ov
! f [k]H(⌦o) =: yss[k] as k ! 1.
c�2016 George Kesidis
78
Transfer Function and Eigenresponse in Discrete Time (cont)
• Equating the forced responses (steady-state response for a stable system), we again getthat the system transfer function is
H(⌦) =P (ej⌦)
Q(ej⌦)= (Fh)(⌦).
• Note that 8k 2 Z, H(⌦) = H(⌦+2⇡k).
• Also, we write H(⌦) not H(ej⌦) for the DTFT.
• So, the eigenresponse of a BIBO/asymptotically stable SISO, DT-LTIC system is the steady-state response to a sinusoid:
f [k] = Afej(⌦ok+�f) ! H(⌦o)f [k] = Ayej(⌦ok+�y) =: yss[k]
• The system magnitude response (gain) is |H(⌦)| = |P (ej⌦)|/|Q(ej⌦)|,i.e., Ay = Af |H(⌦0)|.
• The system phase response is \H(⌦) = \P (ej⌦)� \Q(ej⌦),i.e., �y = �f + \H(⌦0).
c�2016 George Kesidis
79
Eigenresponse - example
• Problem: For the system 2y[k] = 0.6y[k � 1]� 7f [k] find the steady-state response(if it exists) to f [k] = 4cos(5k)u[k].
• Solution: The di↵erence equation in standard form is(Q(��1)y)[k] = y[k + 1] � 0.3y[k] = �3.5f [k + 1] = (P (��1)f)[k], whereQ(z) = z � 0.3 and P (z) = �3.5z.
• The sole system characteristic value (root of Q, system pole) is 0.3, hence the system isBIBO/asymptotically stable.
• By Euler’s identity f [k] = (2ej5k +2ej(�5)k)u[k]
• By linearity, the eigenresponse is therefore
2H(5)ej5k +2H(�5)ej(�5)k,
where H(⌦) = P (ej⌦)/Q(ej⌦) = �3.5ej⌦/(ej⌦ � 0.3) = H(�⌦), so that
|H(⌦)| =3.5p
(cos(⌦)� .3)2 + sin2(⌦), \H(⌦) = ⇡ +⌦� arctan
✓sin(⌦)
cos(⌦)� .3
◆
• Exercise: Show that the eigenresponse is also simply |H(5)|4cos(5k + \H(5)).
80
2D Image Processing Example
• Apply 1-dimensional filtering to a 2-dimensional (2D) image by separately performing rowand column operations.
• For 256⇥ 256 pixel (2D) image,
f =
2
664
f [1,1] f [1,2] ... f [1,256]f [2,1] f [2,2] ... f [2,256]
... ... ... ...f [256,1] f [256,2] ... f [256,256]
3
775
• If f [k, i] represents the 8-bit (grey) intensity of the pixel in row k and column i (i.e., 8bits per pixel or bpp), then the “raw” image size will be 2563bits = 16Mb = 2MB.
• Each of f ’s rows of pixels can be processed by a system with unit-pulse response h toobtain a new row of pixels, and thus a new image y:
8k, f [k, ·] ! h ! y[k, ·]
• Alternatively, each of f ’s columns of pixels can be processed by a system with unit-pulseresponse h to obtain a new column of pixels, and thus a new image y:
8i, f [·, i] ! h ! y[·, i]
c�2016 George Kesidis
81
Image Processing: High-Pass and Low-Pass Filtering
• The system h may have a specific signal processing objective.
• The output pixels y[k, i] may be quantized to fewer bpp than those of the input, thusachieving image compression.
• The simple low-pass filter (L)
h[k] =1
2(�[k] + �[k � 1])
✓) y[k] =
1
2(f [k] + f [k � 1])
◆
can capture shading and texture in the image.
• The simple high-pass filter (H)
h[k] =1
2(�[k]� �[k � 1])
✓) y[k] =
1
2(f [k]� f [k � 1])
◆
can capture edges in the image.
• Typically more compression possible in higher-frequency bands (H).
c�2016 George Kesidis
82
Image Processing: Tandem Row and Column Filtering
f ! rowfiltering
! columnfiltering
! y
• Define yLH as the output of
f ! L ! H ! yLH
• Similarly define yLL, yHH and yHL.
• The y images are downsampled by a factor of four (two in each direction).
• The yLL image will have a lot of energy while yHH will have the least energy.
• This motivates non-uniform quantization (bit allotment per pixel) of these images.
• Together with a coding strategy for the quantized images (particularly for the regions of zeropixel-values), this is the basic approach used in JPEG leading to very good compression,e.g., from 8 bpp to 0.2-0.5 bpp.
c�2016 George Kesidis
83
Sampling Continuous-Time Signals (A/D)
• Consider continuous-time signal x with X = Fx.
• Recall that by sampling at period T with impulses in continuous time t 2 R, we get
xT(t) :=1X
k=�1
x(kT )�(t� kT )F!
1X
k=�1
x(kT )e�jkTw =: XT(w),
equivalently, XT(w) =1
T
1X
v=�1X
✓w � v
2⇡
T
◆.
• Now define the sampled process in discrete-time k 2 Z and its DTFT,
x[k] := x(kT )F! X(⌦) =
1X
k=�1
x[k]e�j⌦k.
• Substituting w = ⌦/T we get
X(⌦) = XT
✓⌦
T
◆=
1
T
1X
v=�1X
✓⌦� v2⇡
T
◆.
• Exercise: Read decimation (downsampling) and interpolation (upsampling) of Lathi Figs.8.17 & 10.9.
84
Sampling Continuous-Time Signals - example
• We are particularly interested in the case where
– the continuous-time signal x is band-limited, i.e., 9w0 > 0 s.t. X(w) = 0 for|w| > w0, and
– the sampling frequency is greater than Nyquist’s, i.e., 2⇡/T > 2w0 ) w0T < ⇡.
• Example: For fixed w0 > 0, consider the cts-time signal x(t) = Asinc(w0t) with FT
X(w) =A⇡
w0 (u(w + w0)� u(w � w0)).
• Sampling x at period T < ⇡/w0 we get the discrete-time signal x[k] = Asinc(w0kT ).
• Using inverse DTFT, recall that we can easily check that the DTFT of x is,
X(⌦) =1X
v=�1
A⇡
w0T(u(⌦+ w0T � 2⇡v)� u(⌦� w0T � 2⇡v))
=1X
v=�1
1
TX
✓⌦� 2⇡v
T
◆,
noting 8T > 0, u(⌦T± w0) = u(1
T(⌦± w0T )) = u(⌦± w0T ), ⌦ := ⌦� 2⇡v.
85
Sampling Continuous-Time Signals - example (cont), w0T < ⇡
c�2016 George Kesidis
86
Sampled Data Systems: A/D (analog-to-digital conversion)
• Suppose the signal f is sampled every Ts seconds, i.e., at sampling frequencyws := 2⇡/Ts.
• Recall Poisson’s identity (the Fourier series of the picket-fence function)
pTs(t) =
1X
k=�1
�(t� kTs) =1
Ts
1X
k=�1
ejkwst
• Let’s rederive the relationship between the spectrum of a sampled continuous-time signaland its discrete-time counterpart by first defining the discrete-time signal
8k 2 Z, f [k] = f(kTs).
• We want to relate the (continuous-time) Fourier transform of f to the (discrete-time)Fourier transform of f ,
F (⌦) :=1X
k=�1
f [k]e�j⌦k =1X
k=�1
f(kTs)e�j⌦k.
c�2016 George Kesidis
87
Sampled Data Systems: A/D (cont)
• To this end, recall
f(t)pTs(t) =
1
Ts
1X
k=�1
f(t)ejkwst F!1
Ts
1X
k=�1
F (w � kws), and also
f(t)pTs(t) =
1X
k=�1
f(kTs)�(t� kTs)F!
1X
k=�1
f(kTs)e�jwkTs = F (wTs).
• Equating these two expressions for F{fpTs} we get,
F (wTs) =1
Ts
1X
k=�1
F (w � kws).
• Substituting w = ⌦/Ts we get,
F (⌦) =1
Ts
1X
k=�1
F
✓⌦� k2⇡
Ts
◆.
c�2016 George Kesidis
88
Sampled Data Systems: D/A (digital to analog conversion)
• Now consider a discrete time signal y[k].
• We implement at D/A with a Ts-second hold, i.e., construct the continuous-time signal
y(t) :=1X
k=�1
y[k]rTs(t� kTs), where
rTs(t) := u(t)� u(t� Ts)
F! Tssinc(wTs/2)e�jwTs/2 =: RTs(w).
• Note that y is in the form of a convolution, so:
Y (w) =1X
k=�1
y[k]RTs(w)e�jwkTs
= RTs(w)Y (wTs)
c�2016 George Kesidis
89
Sampled Data Systems: equivalent cts-time transfer function
• Consider a digital system H(⌦) (or H(ej⌦) depending on notation),whose (ZS) output is y when the input is f , i.e., Y = HF .
• The equivalent continuous-time transformation of the tandem system
f ! A/D (Ts-sample)f! H(⌦)
y! D/A (Ts-hold) ! y
with input f has (ZS) output
Y (w) = RTs(w)Y (wTs) = RTs
(w)H(wTs)F (wTs)
= RTs(w)H(wTs)
1
Ts
1X
k=�1
F (w � kws).
• Exercise: Show that if f is band-limited by ws/2 (i.e., ws is greater than f ’s Nyquistfrequency) and the previous sampled data system is followed by an ideal low-pass filter withbandwidth ws/2, then the equivalent (continuous-time) transfer function is
H(w) = H(wTs)T�1s RTs
(w)(u(w + ws/2)� u(w � ws/2))
c�2016 George Kesidis
90
Sampled Data Systems: equivalent cts-time transfer function -discussion
• Note that the term in the transfer function H,
T�1s RTs
(w)(u(w + ws/2)� u(w � ws/2)) = sinc(⌦/2)(u(⌦+ ⇡)� u(⌦� ⇡))
is not a constant function of ⌦ = wTs.
• This distortion due to the hold function R can be reduced by putting in tandem with Han equalizer system with transfer function approximately
dR�1(⌦) :=1X
k=�1
u(⌦+ ⇡ � k2⇡)� u(⌦� ⇡ � k2⇡)
sinc((⌦� k2⇡)/2)
i.e.,
H(⌦) ! dR�1(⌦)
c�2016 George Kesidis
91
Sampled Data Systems: equalization of hold sinc(⌦/2) by dR�1(⌦)
• the hold (at left, R) distorts the signal by attenuating its higher frequency components
• the equalizer (at right, R�1) amplifies at the higher frequencies to cancel out this distortion
c�2016 George Kesidis
92
DFT and FFT - Reading Exercise on Computational Issues
• Read Lathi Sec. 5.2 and 5.3 re. continuous-time FS, FT
• Read Lathi Sec. 10.6 re. DTFS, DTFT
c�2016 George Kesidis
93
Transient analysis in discrete time by unilateral z-transform
• z-transform definition and region of convergence.
• Basic z-transform pairs and properties.
• Inverse z-transform of rational polynomials by Partial Fraction Expansion (PFE).
• Total transient response of SISO DT LTIC systems Q(��1)y = P (��1)f .
• The steady-state eigenresponse revisited.
• System composition and canonical realizations.
c�2016 George Kesidis
94
The unilateral z-transform & region of convergence
• The z-transform of a signal x = {x[k]}k�0 is
X(z) = (Zx)(z) =1X
k=0
x[k]z�k := limK!1
KX
k=0
x[k]z�k,
where z 2 C.
• If the signal x is bounded by an exponential (geometric), i.e.,
9 M, � 2 R>0 such that 8k 2 Z�0, |x[k]| M�k (i.e., �M�k x[k] M�k)
then the series X(z) converges in the region outside of a disk centered 0 2 C,
{z 2 C | |z| > �}.
• To see why bounded by an exponential su�ces, recall absolute convergence ) convergence:
8k � 0, |x[k]z�k| = |x[k]| · |z|�k M�k|z|�k = M(�/|z|)k
)1X
k=0
|x[k]z�k| M
1X
k=0
(�/|z|)k which converges if �/|z| < 1.
c�2016 George Kesidis
95
Basic z-transform pairs and RoCs
�[k]Z! 1, z 2 C
u[k]Z!
1X
k=0
z�k =1
1� z�1=
z
z � 1, |z| > 1
�ku[k]Z!
1X
k=0
�kz�k =1
1� �z�1=
z
z � �, |z| > |�|
{�k�1u[k � 1]})(z) Z!1X
k=1
�k�1z�k = z�11X
k0=0
�k0z�k0
= z�1 1
1� �z�1=
1
z � �, |z| > |�|
ej⌦oku[k]Z!
1X
k=0
ej⌦okz�k =1
1� ej⌦oz�1, |z| > 1 (� = ej⌦o)
k�ku[k]Z!
1X
k=0
k�kz�k = �d
d�
1X
k=0
�kz�k = �d
d�
1
1� �z�1=
�z�1
(1� �z�1)2,
|z| > |�|Exercise: Find Z{A cos(⌦ok + �)u[k]} and Z{A sin(⌦ok + �)u[k]}.
c�2016 George Kesidis
96
Basic z-transform properties: linearity
• The z-transform is a linear operator: for all scalars a1, a2 2 C and all signalsx1, x2 : Z�0 ! C with respective ROCs C1, C2 ⇢ C,
(Z{a1x1 + a2x2})(z) = a1(Zx1)(z) + a2(Zx2)(z), z 2 C1
\C2.
• Note that
{z | |z| > �1}\
{z | |z| > �2} = {z | |z| > max{�1, �2}} ⇢ C.
c�2016 George Kesidis
97
Basic z-transform properties: advance time shift
• Advance time shift (no change in RoC): Let X = Zx.
��1xZ!
1X
k=0
x[k +1]z�k = �zx[0] +1X
k=�1
x[k +1]z�k
= �zx[0] + z
1X
k=�1
x[k +1]z�(k+1)
= �zx[0] + z
1X
k0=0
x[k0]z�k0
= �zx[0] + zX(z)
• Exercise: For v 2 Z>0 show by induction that
(Z{��vx})(z) = �vX
k=1
zkx[v � k] + zvX(z)
c�2016 George Kesidis
98
Basic z-transform properties: delay time shift
• Delay time shift (no change in RoC): For v 2 Z>0,
�v(xu)Z!
1X
k=0
x[k � v]u[k � v]z�k
=1X
k=v
x[k � v]z�k =1X
k0=0
x[k0]z�k0�v
= z�vX(z).
• So in the “zero-state” (input-output) context (i.e., x[k]u[k] = 0 for k < 0), we identifymultiplying by z�1 in complex-frequency domain with the unit delay � in thetime domain.
• Delay v 2 Z>0 of non-causal x:
�vxZ!
1X
k=0
x[k � v]z�k =1X
k0=�v
x[k0]z�k0�v
=�1X
k0=�v
x[k0]z�k0�v + z�vX(z).
99
Basic z-transform properties: frequency shift & convolution
• Let X = Zx with RoC C(�) := {z 2 C | |z| > �}.
�kx[k]Z!
1X
k=0
�kx[k]z�k =1X
k=0
x[k](z/�)�k = X(z/�), z 2 C(�|�|).
i.e., ⇥�k in the time-domain is dilation by � in the z-domain.
• For signals x1, x2 : Z�0 ! C (x1[k], x2[k] = 0 for k < 0), with respective ROCsC1, C2 ⇢ C,
x1 ⇤ x2Z!
1X
k=0
(x1 ⇤ x2)[k]z�k =
1X
k=0
kX
v=0
x1[v]x2[k � v]z�(k�v)z�v
=1X
v=0
x1[v]z�v
1X
k=v
x2[k � v]z�(k�v)
=1X
v=0
x1[v]z�v
1X
k0=0
x2[k0]z�k0
= X1(z)X2(z), z 2 C1
\C2.
c�2016 George Kesidis
100
Basic z-transform properties: convolution, IVT & FVT
• So convolution in the time-domain is multiplication in the frequency domain.
• The converse is also true.
• Directly by definition of X = Zx, we get the initial value theorem
limz!1
X(z) = x[0].
• There is also a “final value” theorem for limk!1 x[k].
c�2016 George Kesidis
101
Total response of SISO LTIC systems
• We now study transient analysis of LTI di↵erence equations using z-transforms.
• Recall our system is defined given polynomials P,Q, input f and initial conditions:
– Q(��1)y = P (��1)f , where y is the (total) output and
– input f [k] = 0 for k < 0,
– degree of polynomial Q = n � m = degree of polynomial P (causal system),
– Q(z) = zn +Pn�1
v=0 avzv (i.e., an = 1) and P (z) =
Pmv=0 bvz
v,
– an 6= 0 or bn 6= 0 for poly’ls Q,P of minimum degree,
– n initial conditions y[�n], y[�n+1], ..., y[�2], y[�1].
• We can restate the di↵erence equation in terms of delays by delaying both sides by ntime-units (i.e., applying with �n), to get
�nQ(��1)y = �nP (��1)f
) Q(�)y :=nX
v=0
av�n�vy =
mX
v=0
bv�n�vf =: P (�)f
102
Total response of SISO LTIC systems (cont)
• So, taking the z-transform of the (delay) di↵erence equation, we get by the (delay) time-shift and linearity properties that
nX
v=0
av
�1X
k=�v
y[k]z�k�v + Q(z�1)Y (z) = P (z�1)F (z)
• So, solving for the total response Y we get
Y (z) =P (z�1)
Q(z�1)F (z)�
Pnv=0 av
P�1k=�v y[k]z
�k�v
Q(z�1)= YZS(z) + YZI(z)
• where the ZIR and ZSR in the complex-frequency (z) domain respectively are
YZI(z) := �Pn
v=0 avP�1
k=�v y[k]z�k�v
Q(z�1)= �
Pnv=0 av
P�1k=�v y[k]z
n�k�v
Q(z)
YZS(z) :=P (z�1)
Q(z�1)F (z) =
P (z)
Q(z)F (z) = H(z)F (z) (transfer function H).
• Regarding this total transient response, note how
– the z-transform’s unilateral aspect captures the impact of initial conditions (ZIR), and
– a greater range of inputs f than under DTFT through RoC ⇢ C (not just |z| = 1).
103
Total response of SISO LTIC systems - example
• Suppose i.c. y[�1] = �1, input f [k] = 2(�3)ku[k] and output y s.t.
8k � �1, 2y[k +1] + 2y[k] = 3f [k +1] + 2f [k].
• To find the total response y, we take the z-transform of the equivalent system: 8k � 0,
2y[k] + 2y[k � 1] = 3f [k] + 2f [k � 1]) 2Y (z) + 2(z�1Y (z) + y[�1]) = 3F (z) + 2z�1F (z).
• So by the delay property for non-causal signals (y), the total response
Y (z) =3+ 2z�1
2 + 2z�1F (z) +
�2y[�1]
2 + 2z�1
= H(z)F (z) +�y[�1]
1 + z�1
= YZS(z) + YZI(z)
with RoC for Y being the intersection of those of F and H.
104
Total response of SISO LTIC systems - example
• Here F (z) = Z{2(�3)k} = 2/(1 + 3z�1), y[�1] = �1, so
Y (z) =3+ 2z�1
2 + 2z�1·
2
1+ 3z�1+
1
1+ z�1
=3+ 2z�1
(1 + z�1)(1 + 3z�1)+
1
1+ z�1
=
✓3.5
1+ 3z�1+
�0.5
1+ z�1
◆+
1
1+ z�1
where for the last equality see PFE below (here in z�1).
• Understanding that the ZIR begins at k = �1 (initial condition) and the ZSRat time k = 0, we get:
8k � �1, y[k] = (3.5(�3)k � 0.5(�1)k)u[k] + (�1)k = yZS[k] + yZI[k],
where we minded the ambiguity Zx = Zxu.
• Exercise: Verify this solution using time-domain methods, i.e.,y = yZI + yZS = yZI + h ⇤ f , where h and yZI consist of char. modes.
c�2016 George Kesidis
105
Inverse z-transform of proper rational polynomials
• We now describe how to find Z�1X of causal signal X that is rational polynomial in z,i.e., X(z) = M(z)/N(z) where M(z) and N(z) are polynomials in z.
• If deg(M) = deg(N) + 1, we perform long division to write X = c + M/N wheredeg(N) = deg(M) and Z�1X = c� + Z�1{M/N}.
• If deg(M) = deg(N) and M(0) = 0 (so z�1M(z) is a polynomial), we can factor zfrom M to get
X(z) = zz�1M(z)
N(z).
• We will find Z�1X using PFE of the strictly proper rational polynomial z�1M(z)/N(z).
• Alternatively, we could apply PFE on strictly proper rational polynomials in z�1,z�KM(z)/(z�KN(z)) where K := deg(N), as in the previous example.
c�2016 George Kesidis
106
Partial Fraction Expansion (PFE) example in z (not z�1)
• For example, suppose
X(z) :=z(3z +2)
z2 � 0.64= z
(3z +2)
(z +0.8)(z � 0.8)
= z
✓0.25
z +0.8+
2.75
z � 0.8
◆= 0.25
z
z +0.8+ 2.75
z
z � 0.8
where PFE (below) gave the numerators (residues) 0.25 and 2.75.
• So,
(Z�1X)[k] = 0.25(�0.8)ku[k] + 2.75(0.8)ku[k]
• Note that the associated RoC of X is {z 2 C | |z| > 0.8}.
c�2016 George Kesidis
107
Partial Fraction Expansion (PFE) - preliminaries
• Let K = deg(N) = deg(M) so that we can factor
N(z) =KY
k=1
(z � pk),
where the pk are the roots of N (poles of M/N).
• We assume M and N have no common roots, i.e., no “pole-zero cancellation” issue toconsider, so that the pk are the poles of M/N .
• Again, we assume M(0) = 0 (0 is a zero of M/N) and so z�1M(z) is a polynomial ofdegree K � 1.
• Note that the RoC for M(z)/N(z) is {z 2 C | |z| > maxk |pk|}.
c�2016 George Kesidis
108
PFE - the case of no repeated poles
• Suppose there are no repeated poles for M/N , i.e., 8k 6= l, pk 6= pl.
• In this case, we can write the PFE of z�1M(z)/N(z) as
z�1M(z)
N(z)=
KX
l=1
cl
z � pl
)M(z)
N(z)= z
z�1M(z)
N(z)=
KX
l=1
clz
z � pl=
KX
l=1
cl1
1� plz�1
where the scalars (Heaviside coe�cients) cl 2 C are
cl =z�1M(z)Qk 6=l(z � pk)
�����z=pl
= limz!pl
z�1M(z)
N(z)(z � pl) =
z�1M(z)
N(z)(z � pl)
����z=pl
.
• That is, to find the Heaviside coe�cient ck over the term z � pk in the PFE, we haveremoved (covered up) the term z � pk from the denominator N(z) and evaluated theremaining rational polynomial at z = pk.
• This approach, called the Heaviside cover-up method, works even when p is C-valued.
• Given the PFE of z�1M/N , (Z�1M/N)[k] =PK
l=1 clpkl u[k].
109
PFE - proof of Heaviside cover-up method
• To prove that the above formula for the Heaviside coe�cient cl is correct, note that theclaimed PFE of z�1M(z)/N(z) is
KX
l=1
cl
z � pl=
PKl=1 cl
Qk 6=l(z � pk)
N(z)
• Thus, the PFE equals z�1M(z)/N(z) if and only if the numerator polynomials are equal,i.e., i↵
z�1M(z) =KX
l=1
clY
k 6=l
(z � pk) =: M(z).
• Again, two polynomials are equal if their degrees, L, are equal and either:
– their coe�cients are the same, or
– they agree at L + 1 (or more) di↵erent points, e.g., two lines (L = 1) are equal ifthey agree at 2 (= L+1) points.
• Since z�1M(z) is a polynomial of degree< K, it su�ces to check that whether z�1M(z) =M(z) for all z = pk, k 2 {1,2, ...,K}, i.e., this would create K equations in < K un-knowns (the coe�cients of M).
110
PFE - proof of Heaviside cover-up method (cont)
• But note that any pole pr of z�1M(z)/N(z) is a root of all but the rth term in M , sothat
M(pr) = crY
k 6=r
(pr � pk)
=
0
@ z�1M(z)Qk0 6=r(z � pk0)
�����z=pr
1
AY
k 6=r
(pr � pk)
=p�1r M(pr)Q
k0 6=r(pr � pk0)
Y
k 6=r
(pr � pk)
= p�1r M(pr).
• Q.E.D.
c�2016 George Kesidis
111
PFE - the case of no repeated poles - example
• To find the inverse z-transform of a proper rational polynomial X = M/N with M(0) =0, first factor its denominator N and factor z from M , e.g.,
X(z) =z3 + 5z2
z3 + 9z2 + 26z +24= z
z2 + 5z
(z +4)(z +3)(z +2), for |z| > 4.
• So, by PFE
X(z) = z
⇣c4
z +4+
c3
z +3+
c2
z +2
⌘= z
z�1M(z)
N(z))
z�1M(z) = 1z2 + 5z +0 = c4(z +3)(z +2)+ c3(z +4)(z +2)+ c2(z +4)(z +3) =: M(z).
• We can solve for the 3 constants ck by comparing the 3 coe�cients of quadratic M andM .
• The Heaviside cover-up method suggests we try z = �2,�3,�4 to solve for c2, c3, c4:
c4 =z2 + 5z
(z +3)(z +2)
����z=�4
= �2, c3 =z2 + 5z
(z +4)(z +2)
����z=�3
= 6, c2 =z2 + 5z
(z +4)(z +3)
����z=�2
= �3
• Thus, x[k] = (Z�1X)[k] = (�2(�4)k +6(�3)k � 3(�2)k)u[k].
c�2016 George Kesidis
112
PFE - the case of a non-repeated, complex-conjugate pair of poles
• Again, recall that for polynomials with all coe�cients 2 R, all complex poles will come incomplex-conjugate pairs, p1 = p2.
• The case of non-repeated poles p1, p2 = ↵±j� (↵,� 2 R, j :=p�1) that are complex-
conjugate pairs can be handled as above, leading to corresponding complex-conjugate Heav-iside coe�cients c1, c2, i.e., c1 = c2.
• In the PFE, we can alternatively combine the terms
c1
z � p1+
c2
z � p2=
r1z + r0
(z � ↵)2 + �2
where by equating the two numerator polynomials’ coe�cients,
r0 = � c1p2 � c2p1 = � 2Re{c1p2} 2 R and r1 = c1 + c2 = 2Re{c1} 2 R.
• Exercise: Show that
2|c| · |p|k cos(k\p + \c) Z!cz
z � p+
cz
z � p
c�2016 George Kesidis
113
PFE - non-repeated complex-conjugate pair of poles - example
• To find the inverse z-transform of
X(z) =3z2 + 2z
z3 + 5z2 + 10z +12,
first factor the denominator and divide the numerator by z to get
X(z) = z3z +2
(z2 + 2z +4)(z +3).
• Note that the poles of X are �3 and �1± jp3 (so X’s RoC is |z| > 3).
• So, we can expand X to
X(z) = zr1z + r0
z2 + 2z +4+ z
c3
z +3,
where by the Heaviside cover-up method,
c3 =3z +2
z2 + 2z +4
����z=�3
= � 1.
c�2016 George Kesidis
114
PFE - non-repeated complex-conjugate pair of poles - example (cont)
• To find r1, r0, we will compare coe�cients of the numerator polynomials of X (actuallyz�1X) and its PFE, i.e.,
0z2 + 3z +2 = (r1z + r0)(z +3)+ c3(z2 + 2z +4) (⇤)
= (r1 � 1)z2 + (3r1 + r0 � 2)z +3r0 � 4.
• Thus, by comparing coe�cients
0 = r1 � 1, 3 = 3r1 + r0 � 2, 2 = 3r0 � 4
we get
r0 = 2 and r1 = 1.
• Note how z = �3 in (⇤) gives c3 = �1 as Heaviside cover-up did.
c�2016 George Kesidis
115
PFE - non-repeated complex-conjugate pair of poles - example (cont)
• Thus by substituting, we get
X(z) = zz +2
z2 + 2z +4+ z
�1
z +3
• Exercise: Show that
x[k] = (Z�1X)[k]
=⇣p
4/3 2k cos(k2⇡/3� ⇡/6)� (�3)k⌘u[k].
c�2016 George Kesidis
116
PFE - the general case of repeated poles
• If a particular pole p of z�1M(z)/N(z) is of order r � 1, i.e.,N(z) has a factor (z�p)r,then the PFE of z�1M(z)/N(z) has the terms
c1
z � p+
c2
(z � p)2+ ...+
cr
(z � p)r=
rX
k=1
ck
(z � p)k=
z�1M(z)
N(z)��(z)
with ck 2 C 8k 2 {1,2, ..., r}, where�(z) represents the other PFE terms of z�1M(z)/N(z)(i.e., corresponding to poles 6= p).
• Note that equating z�1M(z)/N(z) to its PFE and multiplying by (z � p)r gives
z�1M(z)
N(z)(z � p)r = cr +
r�1X
k=1
ck(z � p)r�k +�(z)(z � p)r
)z�1M(z)
N(z)(z � p)r
����z=p
= cr,
i.e., Heaviside cover-up (of the entire term (z � p)r) works for cr.
c�2016 George Kesidis
117
PFE - the general case of repeated poles (cont)
• To find cr�1, we di↵erentiate the previous display to get
d
dz
z�1M(z)
N(z)(z � p)r =
r�1X
k=1
ck(r � k)(z � p)r�1�k +d
dz�(z)(z � p)r
= cr�1 +r�2X
k=1
ck(r � k)(z � p)r�1�k +d
dz�(z)(z � p)r
) cr�1 =
✓d
dz
z�1M(z)
N(z)(z � p)r
◆����z=p
• If we di↵erentiate the original display k 2 {0,1,2, ..., r � 1} times and then substitutez = p, we get (with 0! := 1)
✓dk
dzkz�1M(z)
N(z)(z � p)r
◆����z=p
= k!cr�k
) cr�k =1
k!
✓dk
dzkz�1M(z)
N(z)(z � p)r
◆����z=p
.
c�2016 George Kesidis
118
PFE - the general case of repeated poles - example
• To find the inverse z-transform of
X(z) =z(3z +2)
(z +1)(z +2)3,
write the PFE of X as
X(z) = z
✓c1
z +1+
c2,1
z +2+
c2,2
(z +2)2+
c2,3
(z +2)3
◆,
so clearly the RoC of causal X is |z| > 2.
• By Heaviside cover-up
c1 =3z +2
(z +2)3
����z=�1
= �1 and c2,3 =3z +2
z +1
����z=�2
= 4.
c�2016 George Kesidis
119
PFE - the general case of repeated poles - example (cont)
• Also,
c2,2 =1
1!
✓d
dz
3z +2
z +1
◆����z=�2
=1
1!
1
(z +1)2
����z=�2
= 1
c2,1 =1
2!
✓d2
dz23z +2
z +1
◆����z=�2
=1
2!
�2
(z +1)3
����z=�2
= 1
• Thus,
X(z) = z�1
z +1+ z
1
z +2+ z
1
(z +2)2+ z
4
(z +2)38|z| > 2
) x[k] = (Z�1X)[k] =
✓�(�1)k + (�2)k + k(�2)k�1 + 4
k(k � 1)
2(�2)k�2
◆u[k].
• Exercise: Show by induction and integration by parts that: 8m 2 Z>0,⇣ km
⌘�k�mu[k]
Z!z
(z � �)m
• Exercise: Find the ZSR y to input f [k] = 2jku[k] = 2ejk⇡/2u[k] of the marginallystable system H(z) = 4/(z2 + 1).
120
PFE of M/N when M(0) 6= 0
• If M(0) 6= 0 (so cannot factor z from M(z)), then just perform long division ifdeg(M) � deg(N) to get a strictly proper rational polynomial, factor N to find thepoles, and find the PFE as before.
• When taking inverse z-transform, recall the z-transform pair
�k�1u[k � 1]Z!
1
z � �, |z| > |�|
c�2016 George Kesidis
121
PFE without factoring z from the numerator first
• For example, to find the ZSR to f [k] = 2(�1)ku[k] of the system
y[k +1]� 4y[k] = 5f [k],
take the z-transform to get
YZS(z) = H(z)F (z) =5
z � 4F (z) =
10z
(z � 4)(z +1)
=8
z � 4+
2
z +1(by PFE)
) yZS[k] = 8(4)k�1u[k � 1] + 2(�1)k�1u[k � 1]
• Note that the unit-pulse response is
h[k] = Z�1(H)[k] = 5(4)k�1u[k � 1],
and that, by delaying the di↵erence equation to get
y[k] = �4y[k � 1] + 5f [k � 1],
we see that (the ZSR) yZS[0] = 0.
• Exercise: First factor z from the numerator of YZS before PFE to show that
yZS[k] = 2(4)ku[k]� 2(�1)ku[k].
Is this result di↵erent? Check for k = 0 and k > 0.
122
PFE and eigenresponse for asymptotically stable systems
• The total response of a SISO LTI system to input f is of the form
Y (z) = H(z)F (z) +P1(z)
Q(z)=
P (z)
Q(z)F (z) +
P1(z)
Q(z)= YZS(z) + YZI(z).
where P1 depends on the initial conditions and the RoC is the intersection of that of inputF = Zf and the system characteristic modes.
• Unlike for DTFT notation, here write H(z) = P (z)/Q(z) = (Zh)(z).
• Suppose the system is BIBO/asymptotically stable and the input is a sinusoid atfrequency ⌦o radians per unit time, f [k] = Aej(⌦ok+�)u[k] = Aej�(ej⌦o)ku[k] withA > 0) F (z) = Aej�z/(z � ej⌦o) with RoC |z| > 1.
• Since ej⌦o cannot be a system pole (owing to asymptotic stability all poles have modulusstrictly less than one), we can use Heaviside cover-up on
YZS(z) = H(z)F (z) = zP (z)
Q(z)(z � ej⌦o)Aej� to get
YZS(z) = zH(ej⌦o)
z � ej⌦o
Aej� + char. modes = H(ej⌦o)F (z) + char. modes.
c�2016 George Kesidis
123
PFE and eigenresponse for asymptotically stable systems (cont)
• Thus, the total response of an asymptotically stable system to a sinusoidal input f atfrequency ⌦o is
y[k] = H(ej⌦o)f [k] + linear combination of characteristic modes.
• So by asymptotic stability, the steady-state response is the eigenresponse, i.e., as k ! 1,
y[k] ! H(ej⌦o)f [k] = H(ej⌦o)Aej(⌦ok+�) = |H(ej⌦o)|Aej(⌦ok+�+\H(ej⌦o)),
where again,
• H = P/Q is the system’s transfer function,
• |H(ej⌦o)| is the system’s magnitude response at frequency ⌦o radians/unit-time, and
• \H(ej⌦o) is the system’s phase response at ⌦o.
c�2016 George Kesidis
124
PFE and eigenresponse for asymptotically stable systems (cont)
• Laplace’s approximation: the rate at which the total response converges to the eigenre-sponse response is according to the characteristic value of largest modulus,
– which will be < 1 owing to the stability assumption,
– i.e., giving the modes(s) that ! 0 slowest.
• In continuous-time systems, it’s the characteristic value of largest real part, which will benegative owing to stability assumption.
c�2016 George Kesidis
125
Canonical (ZS) system-realizations - direct form
• Consider the proper (m n) transfer function
H(z) =P (z)
Q(z)=
bmzm + bm�1zm�1 + ...+ b1z + b0
zn + an�1zn�1 + ...+ a1z + a0=
Y (z)
F (z)
• The direct-form realization employs the interior system state X := F/Q,i.e., F = QX and Y = PX where the former implies (with an = 1),
F (z) =nX
r=0
arzrX(z) ) znX(z) = F (z)�
n�1X
r=0
arzrX(z).
• For n = 2, there are two “system states” (outputs of unit delays), X and zX (respectively,x[k] and (��1x)[k] = x[k +1]):
126
Canonical system-realizations - direct form (cont)
• Now adding Y = PX, we finally get the direct-form canonical system-realization of H:
• Again, state variables taken as outputs of unit delays, here: x, ��1x, ... ��(n�1)x.
• If bn = b2 6= 0, there is direct coupling of input and output, H is proper but not strictlyso, h = Z�1H has a unit-pulse component b2�.
c�2016 George Kesidis
127
Canonical system realizations - direct form (cont)
• Note that this n = 2 example above can be used to implement a pair of complex-conjugatepoles as part of a larger PFE-based implementation (with otherwise di↵erent states); e.g., forn = 2, H(z) = P (z)/Q(z) where
Q(z) = z2 + a1z + a0 = (z � ↵)2 + �2
for ↵,� 2 R, so the poles are ↵± j�.
c�2016 George Kesidis
128
Canonical system realizations by PFE
• In the general case of a proper transfer function, we can use partial-fraction expansion
– grouping the terms corresponding to a complex-conjugate pair of poles, i.e., a second-order denominator, and
– using a direct-form realization for these terms.
• Besides the PFE-based and direct-form realizations, there are other (zero-state) systemrealizations, e.g., “observer” canonical.
• For proper rational-polynomial transfer functionsH = P/Q, all of these realizations involven (= degree of Q) unit delays, the output of each being a di↵erent interior state variableof the system.
c�2016 George Kesidis
129
Canonical system realizations by PFE - example
H(z) =.3z2 � .1
z2 � 0.1z � .3= .3+
.3z � .01
(z � 0.6)(z +0.5)= .3+
.17/1.1
z � 0.6+
.16/1.1
z +0.5
Note that one cannot factor z from the numerator of H.
Exercise: Find a realization for this transfer function H by
1. splitting/forking the input signal F ,
2. using the direct canonical form for each of these 3 terms of H found by long division andPFE, and
3. summing three resulting output signals to get the (ZS) output Y = HF .
c�2016 George Kesidis
130
Digital Proportional-Integral (PI) system
• Consider a continuous-time signal x sampled every T seconds,
8k 2 Z+, x[k] = x(kT ),
and its integral y(t) =R t
0 x(⌧)d⌧ .
• The sampled integral can be approximated, y(kT ) ⇡ y[k], by the trapezoid rule,
y[k] = y[k � 1] +x[k � 1] + x[k]
2T.
• In the complex-frequency domain,
Y (z) = Y (z)z�1 +X(z)z�1 +X(z)
2T
)Y (z)
X(z)=
T
2·1+ z�1
1� z�1.
c�2016 George Kesidis
131
Digital PI system (cont)
• So, a digital PI transfer function would be of the form,
G(z) = Kp
+K
i
T
2·1+ z�1
1� z�1.
for constants Kp
,Ki
.
• In practice, PID or PI systems G are commonly used to control a plant H, where G maybe in series with H or in the feedback branch.
• Exercises:
– Draw the direct-form canonical realization for G.
– Draw the block diagram for the closed-loop system with negative feedback:Y = HX and X = F �GY where H is the (open-loop) system.
– Find the closed-loop transfer function Y/F and recall the pole placement problem tostabilize H.
c�2016 George Kesidis
132
Recursive Least Squares (RLS) Filter - Introduction
• Consider a LTI system with input f and output y,
y[k] =KX
r=0
h[k � r]f [r] + v[k], k 2 Z,
where v is an additive noise process and K is the maximum system order.
• The system (unit-pulse response) h is not known.
• Past values of the output y are observed (known).
• At time k, the objective is to forecast the next output y[k + 1], based on the assumedknown/observed quantities:
– the next input f [k +1],
– the past R input-output pairs {f [r], y[r]}k�R+1rk.
c�2016 George Kesidis
133
RLS objective and Rth-order linear tap filter
• The output of an Rth-order RLS tap-filter at time k is,
yk[i] =iX
r=i�R+1
⌘k[i� r]f [r], i k +1.
• The objective of this filter at time k is to accurately estimate the system output y[k+1]with yk[k +1] by choosing the R filter coe�cients
⌘k[k �R+1], ..., ⌘k[k � 1], ⌘k[k]
that minimize the following sum-of-square-error objective:
Ek =kX
r=k�R+1
�k�r|y[r]� yk[r]|2 =kX
r=k�R+1
�k�r|ek[r]|2
where
– � > 0 is a forgetting factor and
– error ek[r] := y[r]� yk[r].
c�2016 George Kesidis
134
RLS filter
• So, to minimize Ek, substitute yk[r] into Ek and solve
0 =@Ek
@⌘k[i]for i 2 {k �R+1, ..., k � 1, k}.
• That is, R equations in R unknowns: for i 2 {k �R+1, ..., k � 1, k},
0 =kX
r=k�R+1
2�k�rek[r]@ek[r]
@⌘k[i]
=kX
r=k�R+1
2�k�r(y[r]� yk[r])
✓�@yk[r]
@⌘k[i]
◆
=kX
r=k�R+1
2�k�r(yk[r]� y[r])f [r � i]
• Exercise: Prove the last equality.
c�2016 George Kesidis
135
RLS filter (cont)
• Substituting yk[r], rewrite these equations to get the following R equations in R unknowns⌘k[i] that are Ek-minimizing: for i 2 {k �R+1, ..., k � 1, k},
kX
r=k�R+1
�k�rf [r � i]rX
`=r�R+1
f [`]⌘k[r � `] =kX
r=k�R+1
�k�ry[r]f [r � i]
• Exercise: Prove the last equality and write it in matrix form.
• Exercise: Research how the Ek-minimizing filter parameters ⌘k can be computed recur-sively, i.e., using ⌘k�1.
• The filter order R can also be “trial adapted” to discover the system order K so that theerror-minimizing filter parameters ⌘k “track” the system unit-pulse response h over time k.
• Note the required initial “warm-up” period of R time-units where the outputs of system hare simply observed and recorded and no estimates are made.
• Exercise: If there was no additive noise process v and the system unit-pulse response hhad finite support (i.e., a FIR system with K < 1), show how h can be deduced frominput-output (f, y) observations.
c�2016 George Kesidis
136