-
Notifications
You must be signed in to change notification settings - Fork 118
/
midtermsol.lyx
791 lines (622 loc) · 15.1 KB
/
midtermsol.lyx
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407
408
409
410
411
412
413
414
415
416
417
418
419
420
421
422
423
424
425
426
427
428
429
430
431
432
433
434
435
436
437
438
439
440
441
442
443
444
445
446
447
448
449
450
451
452
453
454
455
456
457
458
459
460
461
462
463
464
465
466
467
468
469
470
471
472
473
474
475
476
477
478
479
480
481
482
483
484
485
486
487
488
489
490
491
492
493
494
495
496
497
498
499
500
501
502
503
504
505
506
507
508
509
510
511
512
513
514
515
516
517
518
519
520
521
522
523
524
525
526
527
528
529
530
531
532
533
534
535
536
537
538
539
540
541
542
543
544
545
546
547
548
549
550
551
552
553
554
555
556
557
558
559
560
561
562
563
564
565
566
567
568
569
570
571
572
573
574
575
576
577
578
579
580
581
582
583
584
585
586
587
588
589
590
591
592
593
594
595
596
597
598
599
600
601
602
603
604
605
606
607
608
609
610
611
612
613
614
615
616
617
618
619
620
621
622
623
624
625
626
627
628
629
630
631
632
633
634
635
636
637
638
639
640
641
642
643
644
645
646
647
648
649
650
651
652
653
654
655
656
657
658
659
660
661
662
663
664
665
666
667
668
669
670
671
672
673
674
675
676
677
678
679
680
681
682
683
684
685
686
687
688
689
690
691
692
693
694
695
696
697
698
699
700
701
702
703
704
705
706
707
708
709
710
711
712
713
714
715
716
717
718
719
720
721
722
723
724
725
726
727
728
729
730
731
732
733
734
735
736
737
738
739
740
741
742
743
744
745
746
747
748
749
750
751
752
753
754
755
756
757
758
759
760
761
762
763
764
765
766
767
768
769
770
771
772
773
774
775
776
777
778
779
780
781
782
783
784
785
786
787
788
789
790
791
#LyX 2.3 created this file. For more info see http://www.lyx.org/
\lyxformat 544
\begin_document
\begin_header
\save_transient_properties true
\origin unavailable
\textclass article
\begin_preamble
\renewcommand{\vec}[1]{\mathbf{#1}}
\renewcommand{\labelenumi}{(\alph{enumi})}
\renewcommand{\labelenumii}{(\roman{enumii})}
\newcommand{\tr}{\operatorname{tr}}
\end_preamble
\use_default_options false
\maintain_unincluded_children false
\language english
\language_package default
\inputencoding auto
\fontencoding global
\font_roman "times" "default"
\font_sans "default" "default"
\font_typewriter "default" "default"
\font_math "auto" "auto"
\font_default_family default
\use_non_tex_fonts false
\font_sc false
\font_osf false
\font_sf_scale 100 100
\font_tt_scale 100 100
\use_microtype false
\use_dash_ligatures true
\graphics default
\default_output_format default
\output_sync 0
\bibtex_command default
\index_command default
\paperfontsize default
\spacing single
\use_hyperref false
\papersize default
\use_geometry true
\use_package amsmath 2
\use_package amssymb 2
\use_package cancel 1
\use_package esint 0
\use_package mathdots 1
\use_package mathtools 1
\use_package mhchem 1
\use_package stackrel 1
\use_package stmaryrd 1
\use_package undertilde 1
\cite_engine basic
\cite_engine_type default
\biblio_style plain
\use_bibtopic false
\use_indices false
\paperorientation portrait
\suppress_date false
\justification true
\use_refstyle 0
\use_minted 0
\index Index
\shortcut idx
\color #008000
\end_index
\topmargin 1in
\secnumdepth 3
\tocdepth 3
\paragraph_separation indent
\paragraph_indentation default
\is_math_indent 0
\math_numbering_side default
\quotes_style english
\dynamic_quotes 0
\papercolumns 1
\papersides 2
\paperpagestyle default
\tracking_changes false
\output_changes false
\html_math_output 0
\html_css_as_file 0
\html_be_strict false
\end_header
\begin_body
\begin_layout Section*
18.335 Midterm Exam Solutions: Spring 2019
\end_layout
\begin_layout Subsection*
Problem 1: (10 points)
\end_layout
\begin_layout Standard
The general formula for
\begin_inset Formula $\kappa(A)$
\end_inset
, from the book, is the supremum of the condition number
\begin_inset Formula $\Vert A\Vert\cdot\Vert x\Vert/\Vert Ax\Vert$
\end_inset
for all
\begin_inset Formula $x$
\end_inset
, i.e.
\begin_inset Formula
\[
\kappa(A)=\Vert A\Vert\left(\sup_{x\ne0}\frac{\Vert x\Vert}{\Vert Ax\Vert}\right)=\left(\sup_{x\ne0}\frac{\Vert Ax\Vert}{\Vert x\Vert}\right)\left(\sup_{x\ne0}\frac{\Vert x\Vert}{\Vert Ax\Vert}\right).
\]
\end_inset
Since
\begin_inset Formula $A$
\end_inset
(
\begin_inset Formula $n$
\end_inset
columns) is a subset of the columns of
\begin_inset Formula $B$
\end_inset
(
\begin_inset Formula $n'\ge n$
\end_inset
columns), then for every
\begin_inset Formula $x\in\mathbb{C}^{n}$
\end_inset
there is an
\begin_inset Formula $x'\in\mathbb{C}^{n'}$
\end_inset
such that
\begin_inset Formula $Ax=Bx'$
\end_inset
— that is,
\begin_inset Formula $x'$
\end_inset
is simply
\begin_inset Formula $x$
\end_inset
padded with zeros for the extra columns of
\begin_inset Formula $B$
\end_inset
.
Furthermore, in any of our
\begin_inset Formula $L_{p}$
\end_inset
norms we have
\begin_inset Formula $\Vert x\Vert=\Vert x'\Vert$
\end_inset
.
So, if
\begin_inset Formula $x_{*}$
\end_inset
is a vector where
\begin_inset Formula $\frac{\Vert Ax\Vert}{\Vert x\Vert}$
\end_inset
achieves its supremum, there is an
\begin_inset Formula $x'$
\end_inset
such that
\begin_inset Formula $\frac{\Vert Ax_{*}\Vert}{\Vert x_{*}\Vert}=\frac{\Vert Bx'\Vert}{\Vert x'\Vert}$
\end_inset
, and hence
\begin_inset Formula
\[
\sup_{x\ne0}\frac{\Vert Ax\Vert}{\Vert x\Vert}\le\sup_{x'\ne0}\frac{\Vert Bx'\Vert}{\Vert x'\Vert}.
\]
\end_inset
Similarly for
\begin_inset Formula $\frac{\Vert x\Vert}{\Vert Ax\Vert}$
\end_inset
.
Hence
\begin_inset Formula $\kappa(A)\leq\kappa(B)$
\end_inset
.
\end_layout
\begin_layout Subsection*
Problem 2: (5+5 points)
\end_layout
\begin_layout Standard
For a diagonalizable
\begin_inset Formula $m\times m$
\end_inset
matrix
\begin_inset Formula $A=X\Lambda X^{-1}$
\end_inset
, the matrix square root is
\begin_inset Formula
\[
A^{\frac{1}{2}}=X\Lambda^{\frac{1}{2}}X^{-1}=X\left(\begin{array}{cccc}
\sqrt{\lambda_{1}}\\
& \sqrt{\lambda_{2}}\\
& & \ddots\\
& & & \sqrt{\lambda_{m}}
\end{array}\right)X^{-1}.
\]
\end_inset
\end_layout
\begin_layout Enumerate
\begin_inset Formula $A$
\end_inset
may be nearly defective, in which case
\begin_inset Formula $X$
\end_inset
is badly conditioned and multiplying by
\begin_inset Formula $X$
\end_inset
or
\begin_inset Formula $X^{-1}$
\end_inset
will be inaccurate.
(Being exactly defective is exceedingly rare — a set of measure zero among
all matrices, so you might ignore this case, but you can't ignore the possibili
ty of being nearly defective.)
\end_layout
\begin_layout Enumerate
One possible answer is that all her matrices are Hermitian (or anti-Hermitian/sk
ew-Hermitian).
\begin_inset Newline newline
\end_inset
\begin_inset Newline newline
\end_inset
For the
\begin_inset Formula $X\Lambda^{\frac{1}{2}}X^{-1}$
\end_inset
formula to be accurate, you need
\begin_inset Formula $X$
\end_inset
to be well-conditioned, and the best case for this is if
\begin_inset Formula $A$
\end_inset
is normal (
\begin_inset Formula $AA^{*}=A^{*}A)$
\end_inset
, in which case
\begin_inset Formula $X$
\end_inset
can be chosen unitary (condition number 1).
The only cases where you can typically see that
\begin_inset Formula $A$
\end_inset
is normal by inspection are the Hermitian or anti-Hermitian cases.
(Another possibility would be diagonal matrices
\begin_inset Formula $A$
\end_inset
, but you were told that the matrices were non-sparse.)
\end_layout
\begin_layout Subsection*
Problem 3: (10 points)
\end_layout
\begin_layout Standard
If one of the
\begin_inset Formula $x_{i}$
\end_inset
values is sufficiently large and positive (
\begin_inset Formula $\gtrsim710$
\end_inset
in double precision), then
\begin_inset Formula $e^{x_{i}}$
\end_inset
will overflow and you will get +Inf.
Alternatively, if
\emph on
all
\emph default
of the
\begin_inset Formula $x_{i}$
\end_inset
values are sufficiently large in magnitude and negative (
\begin_inset Formula $\lesssim-745$
\end_inset
in double precision), then
\begin_inset Formula $e^{x_{i}}$
\end_inset
will underflow to
\begin_inset Formula $+0.0$
\end_inset
and the
\begin_inset Formula $\log$
\end_inset
will give you
\begin_inset Formula $-\text{Inf}$
\end_inset
.
To start with, we want to avoid both of these cases.
\end_layout
\begin_layout Standard
\series bold
8/10 points:
\series default
A simple solution is to compute
\begin_inset Formula $X=\max_{i}x_{i}$
\end_inset
, and then use the identity
\begin_inset Formula
\[
f(x)=\log\left(\sum_{i=1}^{n}e^{x_{i}}\right)=\log\left(e^{X}\sum_{i=1}^{n}e^{x_{i}-X}\right)=X+\log\left(\sum_{i=1}^{n}e^{x_{i}-X}\right).
\]
\end_inset
This solves the overflow problem, because
\begin_inset Formula $x_{i}-X\le0$
\end_inset
and hence
\begin_inset Formula $e^{x_{i}-X}$
\end_inset
can only be small, not large.
What about underflow? Without loss of generality, let's suppose that
\begin_inset Formula $X=x_{1}$
\end_inset
.
Then we have
\begin_inset Formula
\[
f(x)=X+\log\left(1+\sum_{i=2}^{n}e^{x_{i}-X}\right).
\]
\end_inset
Notice that
\begin_inset Formula $e^{x_{i}-X}$
\end_inset
in the sum may underflow to zero, but we will never get zero as the argument
of the
\begin_inset Formula $\log$
\end_inset
because we have
\begin_inset Formula $1+\cdots\geq1$
\end_inset
.
So we won't get
\begin_inset Formula $-\text{Inf}$
\end_inset
even if the
\begin_inset Formula $x_{i}$
\end_inset
are large negative numbers.
\end_layout
\begin_layout Standard
\series bold
10/10 points:
\series default
However, there is still a subtle problem: if
\begin_inset Formula $\sum_{i=2}^{n}e^{x_{i}-X}\ll1$
\end_inset
, then in floating-point arithmetic we may get
\begin_inset Formula
\[
X+\log\left(1\oplus\sum_{i=2}^{n}e^{x_{i}-X}\right)=X+\log(1)=X,
\]
\end_inset
so the contribution of the
\begin_inset Formula $\sum_{i=2}^{n}e^{x_{i}-X}$
\end_inset
is lost.
Recall the Taylor expansion
\begin_inset Formula
\[
\log(1+y)=y-\frac{y^{2}}{2}+\frac{y^{3}}{3}-\cdots,
\]
\end_inset
so even if
\begin_inset Formula $0<y\ll1$
\end_inset
, we are not supposed to get zero from the log.
This can lead to an inaccurate result.
For example, consider the case of
\begin_inset Formula $n=2$
\end_inset
with
\begin_inset Formula $x_{1}=10^{-20}>x_{2}=\log10^{-20}\approx-46.0517$
\end_inset
.
Then the correct answer is
\begin_inset Formula
\[
x_{1}+\log(1+e^{x_{2}})=10^{-20}+\log(1+10^{-20})\approx2\times10^{-20}.
\]
\end_inset
but in floating-point arithmetic we will get
\begin_inset Formula $x_{1}\oplus\log(1\oplus e^{x_{2}})=x_{1}\oplus\log(1)=x_{1}\approx10^{-20}$
\end_inset
, which is off by a factor of 2! The solution is that we need to compute
\begin_inset Formula $\operatorname{log1p}(y)=\log(1+y)$
\end_inset
accurately even for very small
\begin_inset Formula $y$
\end_inset
, and fortunately most math libraries (including Julia's) provide a built-in
\begin_inset Quotes eld
\end_inset
\family typewriter
log1p
\family default
\begin_inset Quotes erd
\end_inset
function that does just that.
So, in summary, if we want an accurate result we really need to use a floating-
point version of the expression:
\begin_inset Formula
\[
f(x)=X+\operatorname{log1p}(\sum\nolimits'e^{x_{i}-X}),
\]
\end_inset
where
\begin_inset Formula $\sum'$
\end_inset
denotes the sum omitting a single term with
\begin_inset Formula $x_{i}=X=\max_{j}x_{j}$
\end_inset
.
If we want, we could implement this sum with pairwise summation or similar,
for even more accuracy.
If we didn't have a
\begin_inset Quotes eld
\end_inset
\family typewriter
log1p
\family default
\begin_inset Quotes erd
\end_inset
function available, to accurately compute
\begin_inset Formula $\operatorname{log1p}(y)=\log(1+y)$
\end_inset
, we could implement it ourselves using the Taylor series when
\begin_inset Formula $|y|$
\end_inset
is sufficiently small (although it turns out that there are more clever
ways to do it).
\end_layout
\begin_layout Subsection*
Problem 4: (10 points)
\end_layout
\begin_layout Standard
\series bold
8/10
\series default
: We can use the Hessenberg factorization
\begin_inset Formula $A=QHQ^{*}$
\end_inset
, which can be computed in
\begin_inset Formula $\Theta(m^{3})$
\end_inset
operations from class, and for which
\begin_inset Formula $H$
\end_inset
is tridiagonal if
\begin_inset Formula $A$
\end_inset
is Hermitian.
Then
\begin_inset Formula
\[
f(z)=\det(A-zI)=\det(QHQ^{*}-zI)=\det\left[Q(H-zI)Q^{*}\right]=\det(Q)\det(H-zI)\det(Q^{*})=\det(H-zI)
\]
\end_inset
by elementary properties of determinants.
Since
\begin_inset Formula $H-zI$
\end_inset
is tridiagonal, as mentioned in class we can find its LU factorization
in
\begin_inset Formula $\Theta(m)$
\end_inset
operations, from which the determinant is simply the product of the diagonal
entries of
\begin_inset Formula $U$
\end_inset
.
A little care is needed for the case where
\begin_inset Formula $H-zI$
\end_inset
is nearly singular, though.
\end_layout
\begin_layout Standard
\series bold
10/
\series default
10
\series bold
:
\series default
Since in neither the book nor in class did we explicitly study the LU decomposit
ion of tridiagonal matrices — I only stated in passing that it was
\begin_inset Formula $\Theta(m)$
\end_inset
— and some care is needed in the singular case, to get full marks on this
problem you need to do a bit more work to convince me of how you would
compute
\begin_inset Formula $\det H$
\end_inset
.
In particular, there are lots of ways to derive nice explicit formulas
here.
(Outside of an exam you would just google
\begin_inset Quotes eld
\end_inset
determinant tridiagonal matrix,
\begin_inset Quotes erd
\end_inset
of course.) For example, if we write:
\begin_inset Formula
\[
H=\left(\begin{array}{ccccc}
a_{1} & \overline{b_{1}}\\
b_{1} & a_{2} & \overline{b_{2}}\\
& b_{2} & a_{3} & \ddots\\
& & \ddots & \ddots & \overline{b_{m-1}}\\
& & & b_{m-1} & a_{m}
\end{array}\right),
\]
\end_inset
then each step of Gaussian elimination transforms the
\begin_inset Formula $2\times2$
\end_inset
diagonal block
\begin_inset Formula
\[
\left(\begin{array}{cc}
d_{k} & \overline{b_{k}}\\
b_{k} & a_{k+1}
\end{array}\right)\longrightarrow\left(\begin{array}{cc}
d_{k} & \overline{b_{k}}\\
0 & a_{k+1}-\frac{b_{k}\overline{b_{k}}}{d_{k}}
\end{array}\right),
\]
\end_inset
so that the diagonal entries satisfy the recurrence relation
\begin_inset Formula
\begin{alignat*}{1}
d_{1} & =a_{1}\\
d_{k+1} & =a_{k+1}-\frac{|b_{k}|^{2}}{d_{k}}.
\end{alignat*}
\end_inset
and once it is reduced to upper-triangular form then the determinant is
simply the product of the pivots
\begin_inset Formula $\prod d_{k}$
\end_inset
.
This recurrence may look slightly dangerous at first — what if
\begin_inset Formula $d_{k}=0$
\end_inset
? However, this division by zero goes away when you multiply the entries
together — consider the term
\begin_inset Formula $d_{k}d_{k+1}$
\end_inset
— and after a little thought you can see that the the product
\begin_inset Formula
\[
p_{k}=\prod_{i=1}^{k}d_{k}
\]
\end_inset
satisfies a simpler recurrence (called the
\begin_inset Quotes eld
\end_inset
continuant
\begin_inset Quotes erd
\end_inset
in linear algebra):
\begin_inset Formula
\begin{alignat*}{1}
p_{0} & =1\\
p_{1} & =a_{1}\\
p_{k+1} & =d_{k+1}p_{k}=p_{k}a_{k+1}-p_{k-1}|b_{k}|^{2},
\end{alignat*}
\end_inset
which has no possibility of division by zero, giving
\begin_inset Formula $\det H=p_{m}$
\end_inset
in
\begin_inset Formula $\Theta(m)$
\end_inset
operations.
Finally, get
\begin_inset Formula $\det(H-zI)$
\end_inset
, we simply modify this recurrence to subtract
\begin_inset Formula $z$
\end_inset
from the diagonals:
\begin_inset Formula
\begin{alignat*}{1}
p_{0} & =1\\
p_{1} & =a_{1}-z\\
p_{k+1} & =p_{k}(a_{k+1}-z)-p_{k-1}|b_{k}|^{2}.
\end{alignat*}
\end_inset
This recurrence can also be derived in other ways, e.g.
by cofactor formulas.
For the case of real
\begin_inset Formula $b_{i}$
\end_inset
(real-symmetric
\begin_inset Formula $A$
\end_inset
and
\begin_inset Formula $H$
\end_inset
), the same recurrence is given in equation (30.9) of the Trefethen
\begin_inset space ~
\end_inset
& Bau textbook.
\end_layout
\begin_layout Standard
Another possible
\begin_inset Formula $\Theta(m)$
\end_inset
determinant algorithm is to do the QR factorization of
\begin_inset Formula $H-zI$
\end_inset
, which can be accomplished in
\begin_inset Formula $\Theta(m)$
\end_inset
operations by Givens rotations as you showed in pset
\begin_inset space ~
\end_inset
3.
Then the determinant is simply
\begin_inset Formula $\det R$
\end_inset
(since
\begin_inset Formula $\det Q=1$
\end_inset
for Givens rotations), which is the product of the diagonal entries of
\begin_inset Formula $R$
\end_inset
.
\end_layout
\end_body
\end_document