This preview shows pages 1–7. Sign up to view the full content.
This preview has intentionally blurred sections. Sign up to view the full version.
View Full DocumentThis preview has intentionally blurred sections. Sign up to view the full version.
View Full DocumentThis preview has intentionally blurred sections. Sign up to view the full version.
View Full Document
Unformatted text preview: Oct. 12. 2001 8:43AM IDA CCRP N0. 7255 P. 5/7 78 Linear operators and matrix inverses IfA=[}}]then "Allz = 2, and maxlaiJl = 1
‘1] so “Aug = Zmaxa'j a.,j[.
IfA== 31] then IlAllz = V5, and “Allow = 2 5° ﬁll/1N2 = "Alloc
lfA = {3] then A2 = V5, and "Allow = 1 3° "Allz = ﬁll/“loo
lfA = {‘1’} then llAllz = V5, and IIAIII = 2 5° ﬁll/H12 = “All:
IfA = (1, (1,] then “Alb = ‘5. and “Alli = 1 SO "A": = ﬁll/lili
4.2—15 The (j, k)th element of A” A is (AHAL‘Jc = 25.70in
'i so that
(AHAM = 2560M = Z laijl2
Then
“(AH/1) = ZMHAM = ZZI‘MIZ = “Alli?
j j a
4216
xx" xx" xx"
“Au — min: = u [(I— F; AMI — mﬂ
 tr(A”A) — 2tr(5§iA”A) + tr( 1 xxHAHAxxH)
_ xHx (it”x)2
_ 2 __ _2_ H H 1 2 H
_ xHx til(x A Ax) + (xgx)2iAxi2tr(xx )
2 1
= “All?  ﬁrzAxlli + Ell4345
_ 2 _ AK 2
— llAllr xﬂx . since (st(xxHAHA) = tr(x”AHAx) = KHAH Ax. 4217 A submatrlx B can be obtained by pre and post multiplying A by matrices which select the desired rows and columns of
B'. i.e.. that are identity matrices in the desired rows/columns, and zero elsewhere. Then B = 0114in
where 01 selects the desired columns of A and Q2 selects the desired rows of A. That is, Q; and Q; contain identity
matrices (possibly permuted), with zero rows and columns in columns and/or rows not selected. Hence “01" = 1 and "02" = 1. Then
"B"? = lIQlAQallp s iIQxllp “All? "02",: s HAMP _Oct. 12. 2001 8:44AM IDA CCRP No. 7255 P. 6/7 79 4.2l8 Let V = span(P) (the space that P projects onto), and let S =: V + W, where V .L W. Let x = v +w be a decomposition
of x into the orthogonal spaces, v E V, w E W. Then P = max Px = max Px
ll "xnﬂ l“Hulk!” = max Pvl = max = 1.
Ilv+wll=l llv+wll=l 4.219 Let d = vec(D), the vector obtained by stacking the columns of D. Furthermore, let e be a vector which selects from d
those elements that were diagonal in D. For example, 4111 r112 413 T
D = 421 dzz dze d = [du d21 4:1 tin (in dsz 413 423 435]
dial €152 €133 and
e = {1,0,0,o, 1, 0, 0,0, 1]T
Then tr(D) = dTe. Now apply CS:
mm)? = IdTeI’ s lldllzllell’ = dem = IIDII’m
4.220 Let b} be the jth column of B. I ' llABll’r = ZZIZWMV
i j I:
= 220;,kﬁi,mbk,jzmd
i j k m = Zbg’AHAb;
j where bj is the jth column of B. By the deﬁnition of the 2nonn, b,” AHAb, 2
by,“ 3 "Aug SO IIABlllr S A2 2b? bi = llAllillBll’p 1
4.221 lellw = IIleI. so llAllw = "$311 llellw = "flail llWAxll = "‘Rﬁl llWAxll. Lety = Wx,sox =W*1y. Then “An.” = Im :IWAW"yu = IIWAW‘n. 4.222 If Q is unitary, then
“0!”: = xHQHQx = I”): = "x"? The converse is somewhat more tricky. Suppose that Q is isometric, so that Qxz = llxllz for all x. Then in particular.
when x = 6;, the ith unit vector, "ox" = ql'qr =1.
Thus the diagonal elements of Q” Q are 1. We can write
Q”Q = r + B
where B is a hermitian matrix that is zero along the diagonal. Now let x = e; + ej, 2' 5t j. Then
llQXlli = 2 + (ed + ej)TB(e€ + 8i) = HI“: = 2
Thus (a: + e; )TB(et + ej) = bij + by, = 0. We thus have bij = "bji, which, by the symmetry of B, must be 0. ——_———.——————w——————w_v , 1%
i.
i
e 101 __________________________..__.—————————— 5.3.19 (a) Let y’” = £4. and let J = [1,0, . .. ,0]. We want to maximize yTe = xTAe subject ton? = 1. Setting Ar J = xTAe  2x x
wherexisalagnngemultiplier,weﬁnd bax—J = Ae  Ax = 0
so that x = «iAe = ﬁat, where a; is the ﬁrst column ofA. Then xTx = :lglflt = §2IIIiII2. so that A =a1l
Wehave 1 T 1 T
= ———A Ae = ——A 81,
” "mu "an $01!“) = "31" (b) Since y(1) = Ila; II. the nonzero value in the ﬁrst column of HA is the same as the value obtained by the optimization. 5.320 If Px = (I  2wT/vTv) 6 span(y) then we must have v E span(x, y). Let us set therefore v = x + try for some a to be found. We have r T
Px=x__ T2v(x xIax y) 2
x x+ 2m: y+ayl
_z__ 2xxTx—2axx7y _ (_2 xxT—xTy
£3: + 2ax7'y + allylP y xTx + ZaxTy + dilly2 Tx — ZanTy + alelyll2 — 2x3th — mTy xxT — xTy — _________———————————— + y(—2————————— — xTx + ZaxTy + ally"2 XTX + 2037'! + allyll’ To obtain a result proportional to y, the ﬁrst term must be zero. This means that we must have
tnzllyll2 = llxll’, I]! lell2
v = x + —y.
llyll2 Another v which is proportional to this (and hence has the same overall result) is llxll2 Ilyll" A geometric interpretation is given in the ﬁgure below: v is chosen so that VJ” is midway between x and y (normalized). so
that the reﬂection through v‘ is proportional to y. V 5.3—2l AH = A — 2Avv” /v”v = A + ﬁwv” where [3 = —2/vv” and w = Av. 5322 A segment of code to compute the results is shown here: Algorithm 5.7 Compare leastsquares solutions _____________________________.__————————— 102 Someimpomntmatrixfactoriutions
%
A = [10000 10001; 10001 10002; 10002 10003; 10003 10004; 10004 10005];
b = [20001; 20003; 20005; 20007; 20009];
cond(A)
cond(A'*A) x1 = i’nv(A"A) *A'*b % QR method [V,R] = qrhouse(A),
Q = ququ); C Q'*b; c c(1:2); R1 = R(1:2.1:2); x2 = backsub(R1.c) A"A:
choleskwx);
A'*b;
forsub(L,p) ;
x3 = backsub(L’,y) _____________________________—————————— 5.323 inf. The matrix (a) Using the Mom cond command. we ﬁnd that cond (A) = 1.41 x 10‘, and com! (A’ «11)
is poorly conditioned! The rest of the results follow: x1 =
1.6258
1.4550 x2 =
1.0000
1.0000 x3 =
1.3334
0.6667 We see that the QR method worked the best! (a)
G6 __ c039 sin9 cos9 sin9 _ c0329+sin29 cosOsinO—cosOsinO __I
9 ‘0'— sin9 c059 —sin9 cos9 _ cosOsinOcosasina c0529+sin29 —_ '
Interpretation: ﬁrst rotating an angle of 9, then an angle of 9 is the same as no rotation at all.
(b)
ca _ c059 —sin0 cosd: —sin¢ _ cosOcos¢—sin95in¢ cos€sin¢—cos¢sin9
a ‘ sin0 c039 sin¢ c054: _ cos¢sin0+cosesin¢ cochosdasinﬂsind)
= cos(9+¢) —sin(¢+0)
sin(¢+9) cos(9+¢) ' Since there is an identity operation and each element has an inverse operation. and the operation is associative. the set of rotations forms a group. 107 628 LetA‘ beaneigenvalue ofA.witheollespondingeigenvectorx. Then
Ax = Xx.
Then det(A‘I  A) = 0.
NowdetQI (A+rI)) =det((Ar)I—A).whichiszerowhenA—r= A‘,orA = A‘ +r. Forthisvalue,letybe
the eigenvector:
(A + rI)y = (A‘ + r)y
implies that
Ay = X)!
That is. y is the eigenveetor of A corresponding to A', so y = x.
6.29 lf Ax = Ax. then, multiplying both sides by A we have
A’x = AAx = MAX) = A’x.
Thus A“ is an eigenvalue of A”. and x is an eigenvector. The process repeats for arbitrary powers of A.
6.210 Ifo = Ax. multiply both sides by A":
A'le = AA"): x = AA"1x
i—x = Alx so 1/) is an eigenvalue of A", and x is the eigenvector. 62]! Let
g(:l:) = z" +,<),.—1z""l +    + go
Then
9(A) = A" + sin—1.4”" +   +goA
Then i
i g(A)x = A”): + g.._1A""1x +    + goAx But ij = ij, so we have
g(A)x = X'x + g..1A”"x +    + goAx = g(A)x.
So if A is an eigenvalue of A. then 9(A) is an eigenvalue of g(A). 6.212 With P2 = P, we must have A2 = A, so ,\ = 0 or 1.
6.243 (a) Let AB): = Ax. Then (mull. both sides by B) we have (BA) (Bx) = ABx “owl...” i Let y = Bx. Then we have
BAy = Ay so A is an eigenvalue of BA.
(b) bet Ax. = xix. fori = 1,2, . .. ,n, and let By. = My; fori = 1,2,... ,n. If, as the problem suggests. there are n ,wwsswm . ,, § common eigenvectors, order them so that x. = y. fori = 1, 2, . . . , 11. With n independent eigenvectors. we can write
A = TAIT"
_ where T = [x1 , . . . ,xnl and A; is diagonal. We can also write B = TAzT"
Then AB = TAiT‘1TA2T" = Timur" = BA. 110 Eigenvalues and Eigeuvectors stW 6325 In the Schur lemma, the triangularization T = UHAU has the same eigenvalues as A. Similarity preserves tank. so T has 6.4
the same rank as A. When rank(T) = r < m, only the ﬁrst 1' columns of T have new information in them. so the lower 7
right (m — r) x (m — r) comer of T must be zero. But since the eigenvalues of T are on the diagonal. we must have at least at — r zero eigenvalues.
6326 (a) letAbeunitary.ThenA”A =AA” =1.
LetAbesymmetric. ThenATA = AAT.
LetAbe Hermitian. Then AHA = AA”.
Let A be skew symmetric. Then ATA = —AA = (—AAT) = AAT.
LetAbeskewHermitian. Then A”A= AA = (—AA”) = AA”.
(b) Let A be normal. Then the triangulan'zation UHAU = T satisﬁes (UH/1U)” = UAHU = T”.
Multiplying these we have
(U" AU)(U" AU)” = U” AA”U = v” A" AU
so TT" = T” T. But the only way this can happen is if T is diagonal. 6.327 If A2 = A, then the eigenvalues are either 1 or 0. Then tr(A) is the number of nonzero eigenvalues, which is the rank of the
matrix. 6328 (a) The (111, n)th element of FF” is N—l N—l N_1
(FFHLM' = Z fijnk = z e—ermkejzrrnk = Z e—j2r(m—n)k/N
k=0 k=0 k=0 6
.4
Whenm—n.=0,wehave __
Nl
Z e—j21r(m—n)k/N = N —
Is=o
and when m  n aé 0 we have
N24 —j21r(mn)k/N ej2'(m_")k " 1
e = —.———— = 0.
e—121r(mn) __ 1 k=0 We also observe that this behavior is periodic with period N.
Hence (FFH)m,. = N6...,,., so l/NFFH = I. (b) Let (m, n)th element of C is cm... = c((,._,,.))~. (The notatiorr ((z))N here means :1: modulo N.) Let f. be the lth
column of F: n =[e'1'2*“‘/”], k = 0,1,... ,N — 1. Let us consider the product of C with f}. The pth component of the product is "'1 N—l = Z cp‘ffl'i = Z C((i_p))Ne—)21rh/N i=0 i=0
Nl N—l
_ _ —j21rl(iP+P)/N _ —’21rlp/N '21rl(i—p)/N
 E :Ca—pe —e ’ C((imue J 
i=0 i=0 The latter sum is the same for all values of 1). Hence we see that the pth component of Cf; is proportional to the pth
component of f1 , so f; is an eigenvector. The eigenvalue is Nl
_' 1'
I" = Z C((apmve ’2” (' “M, i=0 which is the DFI' of the sequence {ci}. lfx = Uzy. then (mu? + 2,1152: = 0, so
N04") = span(Uz) _______________________________————————————————
7.34 Here is some MATLAB code to ﬁnd the solution: [U.S.V] = svd(A); xhat = v*pinv(S)'U'*b giving
0.544248
2.40265 * = 3.09292
3.73009 The norm of this solution is "it": = 5.4359. while x = 5.4772. The solution of smallest norm is selected. 7.35 "Air — bu: = IIUEV" vz'b  bu:
= "(22'  I)U"bu§ NowEEl I=diag(0,0,... ,0,1,1,...1).so
(22* — 1w” = U," and the result is established. The quantity Uf' b is that portion of b which is orthogonal to the range of A, and the norm
simply ﬁnds the length of that vector. _, _ 8.97059
7.4.5 (11) A 1:. [4.67647]. 9.85
condition of A is 100.) 0.0294118 0.117647 .
i...
W A " [0.0294118 0.117647]"“‘““°s°'““°“w°
A, 1 _ 0.147059 A, 1 _ 0.15
1 ’ 0.147059 1.1 ‘ 0.15 Now there are only slight changes in the solution. corresponding to slight changes in the RHS. The solution vector is
a multiple of [1, l]T. which points nearly orthogonal to the sensitive direction.
7.57 Let 2 = 71w. where (b) A‘lb = [10'15]. Even though an element of b only changed by .l, the solution changed by essentially 1. (The II
.S
s ‘75 Then, since z = 1, we must have k+l S/JI’WI2 = 1 i=1 k+l "(A — 3):": = Zaﬂvfz)’
i=1 k+l k+1 = Zan‘zl‘ﬁlz 2 “H‘ 2 I‘Y‘I2 = n+1, i=1 i=1 where the inequality follows since the singular values are ordered 01 2 oz . . . 2 0,... Equalin can be achieved if z = V“. 1, which occurs when B = [v1, . . . ,vk]. ...
View Full
Document
 Fall '09
 Swindlehurst
 Digital Signal Processing, Signal Processing

Click to edit the document details