v2009.01.01 - Convex Optimization
v2009.01.01 - Convex Optimization v2009.01.01 - Convex Optimization
618 APPENDIX D. MATRIX CALCULUS By simply rotating our perspective of the four-dimensional representation of gradient matrix, we find one of three useful transpositions of this quartix (connoted T 1 ): ⎡ ∇g(X) T 1 = ⎢ ⎣ ∂g(X) ∂X 11 ∂g(X) ∂X 21 . ∂g(X) ∂X K1 ∂g(X) ∂X 12 · · · ∂g(X) ∂X 22 . · · · ∂g(X) ∂X K2 · · · ∂g(X) ∂X 1L ∂g(X) ∂X 2L . ∂g(X) ∂X KL ⎤ ⎥ ⎦ ∈ RK×L×M×N (1677) When the limit for ∆t∈ R exists, it is easy to show by substitution of variables in (1674) ∂g mn (X) g mn (X + ∆t Y kl e Y kl = lim k e T l ) − g mn(X) ∂X kl ∆t→0 ∆t ∈ R (1678) which may be interpreted as the change in g mn at X when the change in X kl is equal to Y kl , the kl th entry of any Y ∈ R K×L . Because the total change in g mn (X) due to Y is the sum of change with respect to each and every X kl , the mn th entry of the directional derivative is the corresponding total differential [190,15.8] dg mn (X)| dX→Y = ∑ k,l ∂g mn (X) ∂X kl Y kl = tr ( ∇g mn (X) T Y ) (1679) = ∑ g mn (X + ∆t Y kl e lim k e T l ) − g mn(X) (1680) ∆t→0 ∆t k,l g mn (X + ∆t Y ) − g mn (X) = lim (1681) ∆t→0 ∆t = d dt∣ g mn (X+ t Y ) (1682) t=0 where t∈ R . Assuming finite Y , equation (1681) is called the Gâteaux differential [36, App.A.5] [173,D.2.1] [311,5.28] whose existence is implied by existence of the Fréchet differential (the sum in (1679)). [215,7.2] Each may be understood as the change in g mn at X when the change in X is equal
D.1. DIRECTIONAL DERIVATIVE, TAYLOR SERIES 619 in magnitude and direction to Y . D.3 Hence the directional derivative, ⎡ ⎤ dg 11 (X) dg 12 (X) · · · dg 1N (X) →Y dg(X) = ∆ dg 21 (X) dg 22 (X) · · · dg 2N (X) ⎢ ⎥ ∈ R M×N ⎣ . . . ⎦ ∣ dg M1 (X) dg M2 (X) · · · dg MN (X) ⎡ = ⎢ ⎣ ⎡ ∣ dX→Y tr ( ∇g 11 (X) T Y ) tr ( ∇g 12 (X) T Y ) · · · tr ( ∇g 1N (X) T Y ) ⎤ tr ( ∇g 21 (X) T Y ) tr ( ∇g 22 (X) T Y ) · · · tr ( ∇g 2N (X) T Y ) ⎥ . . . tr ( ∇g M1 (X) T Y ) tr ( ∇g M2 (X) T Y ) · · · tr ( ∇g MN (X) T Y ) ⎦ ∑ k,l ∑ = k,l ⎢ ⎣ ∑ k,l ∂g 11 (X) ∂X kl Y kl ∑ k,l ∂g 21 (X) ∂X kl Y kl ∑ k,l . ∂g M1 (X) ∑ ∂X kl Y kl k,l ∂g 12 (X) ∂X kl Y kl · · · ∂g 22 (X) ∂X kl Y kl · · · . ∂g M2 (X) ∂X kl Y kl · · · ∑ k,l ∑ k,l ∑ k,l ∂g 1N (X) ∂X kl Y kl ∂g 2N (X) ∂X kl Y kl . ∂g MN (X) ∂X kl ⎤ ⎥ ⎦ Y kl (1683) from which it follows →Y dg(X) = ∑ k,l ∂g(X) ∂X kl Y kl (1684) Yet for all X ∈ domg , any Y ∈R K×L , and some open interval of t∈ R g(X+ t Y ) = g(X) + t →Y dg(X) + o(t 2 ) (1685) which is the first-order Taylor series expansion about X . [190,18.4] [128,2.3.4] Differentiation with respect to t and subsequent t-zeroing isolates the second term of expansion. Thus differentiating and zeroing g(X+ t Y ) in t is an operation equivalent to individually differentiating and zeroing every entry g mn (X+ t Y ) as in (1682). So the directional derivative of g(X) : R K×L →R M×N in any direction Y ∈ R K×L evaluated at X ∈ domg becomes →Y dg(X) = d dt∣ g(X+ t Y ) ∈ R M×N (1686) t=0 D.3 Although Y is a matrix, we may regard it as a vector in R KL .
- Page 567 and 568: A.6. SINGULAR VALUE DECOMPOSITION,
- Page 569 and 570: A.7. ZEROS 569 Given symmetric matr
- Page 571 and 572: A.7. ZEROS 571 (TRANSPOSE.) Likewis
- Page 573 and 574: A.7. ZEROS 573 For X,A∈ S M + [31
- Page 575 and 576: A.7. ZEROS 575 A.7.5.0.1 Propositio
- Page 577 and 578: Appendix B Simple matrices Mathemat
- Page 579 and 580: B.1. RANK-ONE MATRIX (DYAD) 579 R(v
- Page 581 and 582: B.1. RANK-ONE MATRIX (DYAD) 581 ran
- Page 583 and 584: B.2. DOUBLET 583 R([u v ]) R(Π)= R
- Page 585 and 586: B.3. ELEMENTARY MATRIX 585 If λ
- Page 587 and 588: B.4. AUXILIARY V -MATRICES 587 the
- Page 589 and 590: B.4. AUXILIARY V -MATRICES 589 18.
- Page 591 and 592: B.5. ORTHOGONAL MATRIX 591 B.5 Orth
- Page 593 and 594: B.5. ORTHOGONAL MATRIX 593 Figure 1
- Page 595 and 596: Appendix C Some analytical optimal
- Page 597 and 598: C.2. TRACE, SINGULAR AND EIGEN VALU
- Page 599 and 600: C.2. TRACE, SINGULAR AND EIGEN VALU
- Page 601 and 602: C.2. TRACE, SINGULAR AND EIGEN VALU
- Page 603 and 604: C.3. ORTHOGONAL PROCRUSTES PROBLEM
- Page 605 and 606: C.4. TWO-SIDED ORTHOGONAL PROCRUSTE
- Page 607 and 608: C.4. TWO-SIDED ORTHOGONAL PROCRUSTE
- Page 609 and 610: Appendix D Matrix calculus From too
- Page 611 and 612: D.1. DIRECTIONAL DERIVATIVE, TAYLOR
- Page 613 and 614: D.1. DIRECTIONAL DERIVATIVE, TAYLOR
- Page 615 and 616: D.1. DIRECTIONAL DERIVATIVE, TAYLOR
- Page 617: D.1. DIRECTIONAL DERIVATIVE, TAYLOR
- Page 621 and 622: D.1. DIRECTIONAL DERIVATIVE, TAYLOR
- Page 623 and 624: D.1. DIRECTIONAL DERIVATIVE, TAYLOR
- Page 625 and 626: D.1. DIRECTIONAL DERIVATIVE, TAYLOR
- Page 627 and 628: D.1. DIRECTIONAL DERIVATIVE, TAYLOR
- Page 629 and 630: D.1. DIRECTIONAL DERIVATIVE, TAYLOR
- Page 631 and 632: D.2. TABLES OF GRADIENTS AND DERIVA
- Page 633 and 634: D.2. TABLES OF GRADIENTS AND DERIVA
- Page 635 and 636: D.2. TABLES OF GRADIENTS AND DERIVA
- Page 637 and 638: D.2. TABLES OF GRADIENTS AND DERIVA
- Page 639 and 640: Appendix E Projection For any A∈
- Page 641 and 642: 641 U T = U † for orthonormal (in
- Page 643 and 644: E.1. IDEMPOTENT MATRICES 643 where
- Page 645 and 646: E.1. IDEMPOTENT MATRICES 645 order,
- Page 647 and 648: E.1. IDEMPOTENT MATRICES 647 When t
- Page 649 and 650: E.3. SYMMETRIC IDEMPOTENT MATRICES
- Page 651 and 652: E.3. SYMMETRIC IDEMPOTENT MATRICES
- Page 653 and 654: E.3. SYMMETRIC IDEMPOTENT MATRICES
- Page 655 and 656: E.5. PROJECTION EXAMPLES 655 E.4.0.
- Page 657 and 658: E.5. PROJECTION EXAMPLES 657 a ∗
- Page 659 and 660: E.5. PROJECTION EXAMPLES 659 E.5.0.
- Page 661 and 662: E.6. VECTORIZATION INTERPRETATION,
- Page 663 and 664: E.6. VECTORIZATION INTERPRETATION,
- Page 665 and 666: E.6. VECTORIZATION INTERPRETATION,
- Page 667 and 668: E.6. VECTORIZATION INTERPRETATION,
618 APPENDIX D. MATRIX CALCULUS<br />
By simply rotating our perspective of the four-dimensional representation of<br />
gradient matrix, we find one of three useful transpositions of this quartix<br />
(connoted T 1 ):<br />
⎡<br />
∇g(X) T 1<br />
=<br />
⎢<br />
⎣<br />
∂g(X)<br />
∂X 11<br />
∂g(X)<br />
∂X 21<br />
.<br />
∂g(X)<br />
∂X K1<br />
∂g(X)<br />
∂X 12<br />
· · ·<br />
∂g(X)<br />
∂X 22<br />
.<br />
· · ·<br />
∂g(X)<br />
∂X K2<br />
· · ·<br />
∂g(X)<br />
∂X 1L<br />
∂g(X)<br />
∂X 2L<br />
.<br />
∂g(X)<br />
∂X KL<br />
⎤<br />
⎥<br />
⎦ ∈ RK×L×M×N (1677)<br />
When the limit for ∆t∈ R exists, it is easy to show by substitution of<br />
variables in (1674)<br />
∂g mn (X) g mn (X + ∆t Y kl e<br />
Y kl = lim<br />
k e T l ) − g mn(X)<br />
∂X kl ∆t→0 ∆t<br />
∈ R (1678)<br />
which may be interpreted as the change in g mn at X when the change in X kl<br />
is equal to Y kl , the kl th entry of any Y ∈ R K×L . Because the total change<br />
in g mn (X) due to Y is the sum of change with respect to each and every<br />
X kl , the mn th entry of the directional derivative is the corresponding total<br />
differential [190,15.8]<br />
dg mn (X)| dX→Y<br />
= ∑ k,l<br />
∂g mn (X)<br />
∂X kl<br />
Y kl = tr ( ∇g mn (X) T Y ) (1679)<br />
= ∑ g mn (X + ∆t Y kl e<br />
lim<br />
k e T l ) − g mn(X)<br />
(1680)<br />
∆t→0 ∆t<br />
k,l<br />
g mn (X + ∆t Y ) − g mn (X)<br />
= lim<br />
(1681)<br />
∆t→0 ∆t<br />
= d dt∣ g mn (X+ t Y ) (1682)<br />
t=0<br />
where t∈ R . Assuming finite Y , equation (1681) is called the Gâteaux<br />
differential [36, App.A.5] [173,D.2.1] [311,5.28] whose existence is implied<br />
by existence of the Fréchet differential (the sum in (1679)). [215,7.2] Each<br />
may be understood as the change in g mn at X when the change in X is equal