summaryrefslogtreecommitdiff
path: root/Master/texmf-dist/doc/latex/siam/lexample.tex
blob: a8502d3788aa164add85de7640c77dc3a1396370 (plain)
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
%%
%% This is file `lexample.tex', 
%% Sample file for siam macros for use with LaTeX 2e
%% 
%% By Paul Duggan for the Society for Industrial and Applied
%% Mathematics. 
%% 
%% October 1, 1995
%%
%% Version 1.0
%% 
%% You are not allowed to change this file. 
%% 
%% You are allowed to distribute this file under the condition that 
%% it is distributed together with all of the files in the siam macro 
%% distribution. These are:
%%
%%  siamltex.cls (main LaTeX macro file for SIAM)
%%  siamltex.sty (includes siamltex.cls for compatibility mode)
%%  siam10.clo   (size option for 10pt papers)
%%  subeqn.clo   (allows equation numbners with lettered subelements)
%%  siam.bst     (bibliographic style file for BibTeX)
%%  docultex.tex (documentation file)
%%  lexample.tex (this file)
%%
%% If you receive only some of these files from someone, complain! 
%% 
%% You are NOT ALLOWED to distribute this file alone. You are NOT 
%% ALLOWED to take money for the distribution or use of either this 
%% file or a changed version, except for a nominal charge for copying 
%% etc. 
%% \CharacterTable
%%  {Upper-case    \A\B\C\D\E\F\G\H\I\J\K\L\M\N\O\P\Q\R\S\T\U\V\W\X\Y\Z
%%   Lower-case    \a\b\c\d\e\f\g\h\i\j\k\l\m\n\o\p\q\r\s\t\u\v\w\x\y\z
%%   Digits        \0\1\2\3\4\5\6\7\8\9
%%   Exclamation   \!     Double quote  \"     Hash (number) \#
%%   Dollar        \$     Percent       \%     Ampersand     \&
%%   Acute accent  \'     Left paren    \(     Right paren   \)
%%   Asterisk      \*     Plus          \+     Comma         \,
%%   Minus         \-     Point         \.     Solidus       \/
%%   Colon         \:     Semicolon     \;     Less than     \<
%%   Equals        \=     Greater than  \>     Question mark \?
%%   Commercial at \@     Left bracket  \[     Backslash     \\
%%   Right bracket \]     Circumflex    \^     Underscore    \_
%%   Grave accent  \`     Left brace    \{     Vertical bar  \|
%%   Right brace   \}     Tilde         \~}


\documentclass[final]{siamltex}

% definitions used by included articles, reproduced here for 
% educational benefit, and to minimize alterations needed to be made
% in developing this sample file.

\newcommand{\pe}{\psi}
\def\d{\delta} 
\def\ds{\displaystyle} 
\def\e{{\epsilon}} 
\def\eb{\bar{\eta}}  
\def\enorm#1{\|#1\|_2} 
\def\Fp{F^\prime}  
\def\fishpack{{FISHPACK}} 
\def\fortran{{FORTRAN}} 
\def\gmres{{GMRES}} 
\def\gmresm{{\rm GMRES($m$)}} 
\def\Kc{{\cal K}} 
\def\norm#1{\|#1\|} 
\def\wb{{\bar w}} 
\def\zb{{\bar z}} 

% some definitions of bold math italics to make typing easier.
% They are used in the corollary.

\def\bfE{\mbox{\boldmath$E$}}
\def\bfG{\mbox{\boldmath$G$}}

\title{Sample file for SIAM \LaTeX\ macro package\thanks{This 
        work was supported by the Society for Industrial and
        Applied Mathematics, Philadelphia, Pennsylvania.}}

% The thanks line in the title should be filled in if there is
% any support acknowledgement for the overall work to be included
% This \thanks is also used for the received by date info, but
% authors are not expected to provide this.

\author{Paul Duggan\thanks{Composition Department, Society
        for Industrial and Applied Mathematics, 3600 Univeristy 
        City Science Center, Philadelphia, Pennsylvania, 
        19104-2688 ({\tt duggan@siam.org}).}
        \and Various A.~U. Thors\thanks{Various Affiliations, 
        supported by various foundation grants.}}

\begin{document}

\maketitle

\begin{abstract}
An example of SIAM \LaTeX\ macros is presented. Various
aspects of composing manuscripts for SIAM's journal series
are illustrated with actual examples from accepted
manuscripts. SIAM's stylistic standards are adhered to
throughout, and illustrated.
\end{abstract}

\begin{keywords} 
sign-nonsingular matrix, LU-factorization, indicator
polynomial 
\end{keywords}

\begin{AMS}
15A15, 15A09, 15A23
\end{AMS}

\pagestyle{myheadings}
\thispagestyle{plain}
\markboth{P. DUGGAN AND V. A. U. THORS}{SIAM MACRO EXAMPLES}


\section{Introduction and examples}
This paper presents a sample file for the use of SIAM's
\LaTeX\ macro package. It illustrates the features of the
macro package, using actual examples culled from various
papers published in SIAM's journals. It is to be expected
that this sample will provide examples of how to use the
macros to generate standard elements of journal papers,
e.g., theorems, definitions, or figures. This paper also
serves as an example of SIAM's stylistic preferences for
the formatting of such elements as bibliographic references,
displayed equations, and equation arrays, among others.
Some special circumstances are not dealt with in this
sample file; for such information one should see the
included documentation file.

{\em Note:} This paper is not to be read in any form for content. 
The conglomeration of equations, lemmas, and other text elements were 
put together solely for typographic illustrative purposes and don't 
make any sense as lemmas, equations, etc.

\subsection{Sample text}
Let $S=[s_{ij}]$ ($1\leq i,j\leq n$) be a $(0,1,-1)$-matrix
of order $n$. Then $S$ is a {\em sign-nonsingular matrix}
(SNS-matrix) provided that each real matrix with the same
sign pattern as $S$ is nonsingular. There has been
considerable recent interest in constructing and
characterizing SNS-matrices \cite{bs}, \cite{klm}. There
has also been interest in strong forms of
sign-nonsingularity \cite{djd}. In this paper we give a new
generalization of SNS-matrices and investigate some of
their basic properties.
 
Let $S=[s_{ij}]$ be a $(0,1,-1)$-matrix of order $n$ and
let $C=[c_{ij}]$ be a real matrix of order $n$. The pair
$(S,C)$ is called a {\em matrix pair of order} $n$.
Throughout, $X=[x_{ij}]$ denotes a matrix of order $n$
whose entries are algebraically independent indeterminates
over the real field. Let $S\circ X$ denote the Hadamard
product (entrywise product) of $S$ and $X$. We say that the
pair $(S,C)$ is a {\em sign-nonsingular matrix pair of
order} $n$, abbreviated SNS-{\em matrix pair of order} $n$,
provided that the matrix \[A=S\circ X+C\] is nonsingular
for all positive real values of the $x_{ij}$.  If $C=O$
then the pair $(S,O)$ is a SNS-matrix pair if and only if
$S$ is a SNS-matrix.  If $S=O$ then the pair $(O,C)$ is a
SNS-matrix pair if and only if $C$ is nonsingular. Thus
SNS-matrix pairs include both nonsingular matrices and
sign-nonsingular matrices as special cases.
 
The pairs $(S,C)$ with
\[S=\left[\begin{array}{cc}1&0\\0&0\end{array}\right],\qquad 
C=\left[\begin{array}{cc}1&1\\1&1\end{array}\right]\] and 
\[S=\left[\begin{array}{ccc}1&1&0\\1&1&0\\0&0&0\end{array}\right],\qquad 
C=\left[\begin{array}{ccc}0&0&1\\0&2&0\\
3&0&0\end{array}\right]\] are examples of SNS-matrix pairs.

\subsection{A remuneration list}
In this paper we consider the evaluation of integrals of the 
following forms:
\begin{equation}
\int_a^b \left( \sum_i E_i B_{i,k,x}(t) \right)
         \left( \sum_j F_j B_{j,l,y}(t) \right) dt,\label{problem}
\end{equation}
\begin{equation}
\int_a^b f(t) \left( \sum_i E_i B_{i,k,x}(t) \right) dt,\label{problem2}
\end{equation}
where $B_{i,k,x}$ is the $i$th B-spline of order $k$ defined over the
knots $x_i, x_{i+1}, \ldots, x_{i+k}$.
We will consider B-splines normalized so that their integral is one.
The splines may be of different orders and
defined on different knot sequences $x$ and $y$.
Often the limits of integration will be the entire real line, $-\infty$
to $+\infty$. Note that (\ref{problem}) is a special case of (\ref{problem2})
where $f(t)$ is a spline.


There are five different methods for calculating (\ref{problem})
that will be considered:
\begin{remunerate}
\item Use Gauss quadrature on each interval.
\item Convert the integral to a linear combination of
      integrals of products of B-splines and provide a recurrence for
      integrating the product of a pair of B-splines.
\item Convert the sums of B-splines to piecewise
      B\'{e}zier format and integrate segment
      by segment using the properties of the Bernstein polynomials.
\item Express the product of a pair of B-splines as a linear combination
      of B-splines.
      Use this to reformulate the integrand as a linear combination
      of B-splines, and integrate term by term.
\item Integrate by parts.
\end{remunerate}
Of these five, only methods 1 and 5 are suitable for calculating 
(\ref{problem2}). The first four methods will be touched on and the 
last will be discussed at length.


\subsection{Some displayed equations and \{{\tt eqnarray}\}s}
     By introducing the product topology on  $R^{m \times m} \times
R^{n \times n}$  with the induced inner product
\begin{equation}
\langle (A_{1},B_{1}), (A_{2},B_{2})\rangle := \langle A_{1},A_{2}\rangle 
+ \langle B_{1},B_{2}\rangle,\label{eq2.10}
\end{equation}
we calculate the Fr\'{e}chet derivative of  $F$  as follows:
\begin{eqnarray}
 F'(U,V)(H,K) &=& \langle R(U,V),H\Sigma V^{T} + U\Sigma K^{T} -
P(H\Sigma V^{T} + U\Sigma K^{T})\rangle \nonumber \\
         &=& \langle R(U,V),H\Sigma V^{T} + U\Sigma K^{T}\rangle \label{eq2.11} \\
&=& \langle R(U,V)V\Sigma^{T},H\rangle + \langle \Sigma^{T}U^{T}R(U,V),K^{T}\rangle.     \nonumber
\end{eqnarray}
In the middle line of (\ref{eq2.11}) we have used the fact that the range of
$R$ is always perpendicular to the range of $P$.  The gradient $\nabla F$  of
$F$, therefore,  may be interpreted as the
pair of matrices:
\begin{equation}
 \nabla F(U,V) = (R(U,V)V\Sigma^{T},R(U,V)^{T}U\Sigma ) \in
R^{m \times m} \times R^{n \times n}.       			\label{eq2.12}
\end{equation}
Because of the product topology, we know
\begin{equation}
 {\cal T}_{(U,V)}({\cal O} (m) \times {\cal O} (n)) =
{\cal T}_{U}{\cal O} (m) \times {\cal T}_{V}{\cal O} (n),  		\label{eq2.13}
\end{equation}
where  ${\cal T}_{(U,V)}({\cal O} (m) \times {\cal O} (n))$  stands for the
tangent space to the manifold  ${\cal O} (m) \times {\cal O} (n)$  at  $(U,V)
\in {\cal O} (m) \times {\cal O} (n)$  and so on.  The projection of
$\nabla F(U,V)$  onto  ${\cal T}_{(U,V)}({\cal O} (m) \times {\cal O} (n))$,
therefore, is the product of the projection of the first component of
$\nabla F(U,V)$  onto  ${\cal T}_{U}{\cal O} (m)$  and the projection of the
second component of  $\nabla F(U,V)$  onto  ${\cal T}_{V}{\cal O} (n)$. 
In particular, we claim that the
projection $ g(U,V)$  of the gradient  $\nabla F(U,V)$  onto
${\cal T}_{(U,V)}({\cal O} (m) \times {\cal O} (n))$  is given by the pair of
matrices:
\begin{eqnarray}
g(U,V) = && \left( \frac{R(U,V)V\Sigma^{T}U^{T}-U\Sigma V^{T}R(U,V)^{T}}{2}U,
\right.			\nonumber \\[-1.5ex]
\label{eq2.14}\\[-1.5ex]
&&\quad \left. \frac{R(U,V)^{T}U\Sigma V^{T}-V
   \Sigma^{T}U^{T}R(U,V)}{2}V \right).\nonumber
\end{eqnarray}
Thus, the vector field
\begin{equation}
\frac{d(U,V)}{dt} = -g(U,V) 	\label{eq2.15}
\end{equation}
defines a steepest descent flow on the manifold  ${\cal O} (m) \times
{\cal O} (n)$ for the objective function  $F(U,V)$.

 
\section{Main results} 
 
Let $(S,C)$ be a matrix pair of order $n$.  The determinant
\[\det (S\circ X+C)\] 
is a polynomial in the indeterminates of $X$ of degree at
most $n$ over the real field. We call this polynomial the
{\em indicator polynomial} of the matrix pair $(S,C)$
because of the following proposition.
 
\begin{theorem} 
\label{th:prop} 
The matrix pair $(S,C)$ is a {\rm SNS}-matrix pair if and
only if all the nonzero coefficients in its indicator
polynomial have the same sign and there is at least one
nonzero coefficient.
\end{theorem} 
 
\begin{proof}
Assume that $(S,C)$ is a SNS-matrix pair.  Clearly the
indicator polynomial has a nonzero coefficient.  Consider a
monomial
\begin{equation} 
\label{eq:mono} 
b_{i_{1},\ldots,i_{k};j_{1},\ldots,j_{k}}x_{i_{1}j_{1}}\cdots
x_{i_{k}j_{k}}
\end{equation} 
occurring in the indicator polynomial with a nonzero
coefficient.  By taking the $x_{ij}$ that occur in
(\ref{eq:mono}) large and all others small, we see that any
monomial that occurs in the indicator polynomial with a
nonzero coefficient can be made to dominate all others.
Hence all the nonzero coefficients have the same sign. The
converse is im-\linebreak mediate. \qquad\end{proof} 
 

For SNS-matrix pairs $(S,C)$ with $C=O$ the indicator
polynomial is a homogeneous polynomial of degree $n$. In
this case Theorem \ref{th:prop} is a standard fact about
SNS-matrices.
 
\begin{lemma}[{\rm Stability}]
\label{stability}
Given $T>0$, suppose that $\| \epsilon (t) \|_{1,2} \leq h^{q-2}$
for $0 \leq t \leq T$ and $q \geq 6$. 
Then there exists a positive number $B$ that depends on
$T$ and the exact solution $\pe$ only such that for all $0 \leq t \leq T$,
\begin{equation}
\label{Gron}
\frac {d}{dt} \| \epsilon (t) \| _{1,2}  \leq B
   ( h^{q-3/2} + \| \epsilon (t) \|_{1,2})\;.
\end{equation}
The function $B(T)$ can be chosen to be nondecreasing in time.
\end{lemma}
 
 
\begin{theorem} 
\label{th:gibson} 
The maximum number of nonzero entries in a {\rm SNS}-matrix
$S$ of order $n$ equals \[\frac{n^{2}+3n-2}{2}\] with
equality if and only if there exist permutation matrices
such that $P|S|Q=T_{n}$ where
\begin{equation} 
\label{eq:gibson} 
T_{n}=\left[\begin{array}{cccccc} 1&1&\cdots&1&1&1\\
1&1&\cdots&1&1&1\\ 0&1&\cdots&1&1&1\\ 
\vdots&\vdots&\ddots&\vdots&\vdots&\vdots\\ 
0&0&\cdots&1&1&1\\ 0&0&\cdots&0&1&1\end{array}\right]. 
\end{equation} 
\end{theorem} 
 
We note for later use that each submatrix of $T_{n}$ of
order $n-1$ has all 1s on its main diagonal. 
 
We now obtain a bound on the number of nonzero entries of
$S$ in a SNS-matrix pair $(S,C)$ in terms of the degree of
the indicator polynomial. We denote the strictly upper
triangular (0,1)-matrix of order $m$ with all 1s above the
main diagonal by $U_{m}$. The all 1s matrix of size $m$ by
$p$ is denoted by $J_{m,p}$.
 
 
\begin{proposition}[{\rm Convolution theorem}]
\label{pro:2.1}  Let
\begin{eqnarray*}
a\ast u(t) = \int_0^t a(t- \tau) u(\tau) d\tau, \hspace{.2in} t \in
(0, \infty).
\end{eqnarray*}
Then
\begin{eqnarray*}
\widehat{a\ast u}(s) = \widehat{a}(s)\widehat{u}(s).
\end{eqnarray*}
\end{proposition}

\begin{lemma}
\label{lem:3.1}
For $s_0 >0$, if
$$
\int_0^{\infty} e^{-2s_0 t}v^{(1)}(t) v(t) dt \; \leq 0 \;,
$$
then
\begin{eqnarray*}
\int_0^{\infty} e^{-2s_0 t} v^2(t) dt \; \leq \; \frac{1}{2s_0} v^2(0).
\end{eqnarray*}
\end{lemma}

{\em Proof}. Applying integration by parts, we obtain
\begin{eqnarray*}
\int_0^{\infty} e^{-2s_0 t} [v^2(t)-v^2(0)] dt
&=&\lim_{t\rightarrow \infty}\left (
-\frac{1}{2s_0}e^{-2s_0 t}v^2(t) \right ) +\frac{1}{s_0}
\int_0^{\infty} e^{-2s_0 t}v^{(1)}(t)v(t)dt\\
&\leq& \frac{1}{s_0} \int_0^{\infty} e^{-2s_0 t} v^{(1)}(t)v(t) dt \;\;
\leq \;\; 0.
\end{eqnarray*}
Thus
$$
\int_0^{\infty} e^{-2s_0 t} v^2(t) dt \;\;\leq v^2(0) \int_0^{\infty}
\;\;e^{-2s_0 t} dt\;\;=\;\;\frac{1}{2s_0} v^2(0).\eqno\endproof
$$

\begin{corollary}\label{c4.1}
Let $ \bfE $ satisfy $(5)$--$(6)$ and
suppose $ \bfE^h $ satisfies $(7)$ and $(8)$
with a general $ \bfG $.  Let $ \bfG= \nabla \times {\bf \Phi} + \nabla p,$
$p \in H_0^1 (\Omega) $. Suppose that $\nabla p$ and $ \nabla \times 
{\bf \Phi} $ satisfy all the assumptions of Theorems $4.1$ and  
$4.2$, respectively. In addition suppose all the regularity
assumptions of Theorems $4.1$--$4.2$ are satisfied.  Then 
for $ 0 \le t \le T $ and $ 0 < \epsilon \le \epsilon_0 $ there exists a 
constant $ C = C(\epsilon, T) $ such that
$$
\Vert (\bfE - \bfE^h)(t) \Vert_0 \le C h^{k+1- \epsilon},
$$
where $ C $ also depends on the constants given in Theorems 
$4.1$ and $4.2$.
\end{corollary}


\begin{definition}
Let $S$ be an isolated invariant set with isolating neighborhood $N$.
An {\em index pair} for $S$ is a pair of compact sets $(N_{1},N_{0})$
with $N_{0} \subset N_{1} \subset N$ such that:
\begin{romannum}
\item $cl(N_{1} \backslash N_{0})$
is an isolating neighborhood for $S$.
\item $N_{i}$ is positively invariant relative to $N$ for $i=0,1$,
i.e., given
$x \in N_{i}$ and $x \cdot [0,t] \subset N$, then $x \cdot [0,t] \subset
N_{i}$.
\item $N_{0}$ is an exit set for $N_{1}$, i.e. if $x \in N_{1}$,
$x \cdot [0, \infty ) \not\subset N_{1}$, then there is a $T \geq 0$ such
that $x \cdot [0,T] \subset N_{1}$ and $x \cdot T \in N_{0}$.
\end{romannum}
\end{definition}

\subsection{Numerical experiments} We conducted numerical experiments 
in computing inexact Newton steps for discretizations of a  
{\em modified Bratu problem}, given by  
\begin{eqnarray} 
{\ds \Delta w + c e^w + d{ {\partial w}\over{\partial x} } } 
&=&{\ds f \quad {\rm in}\ D, }\nonumber\\[-1.5ex]
\label{bratu} \\[-1.5ex]
{\ds w }&=&{\ds 0 \quad {\rm on}\ \partial D , } \nonumber
\end{eqnarray} 
where $c$ and $d$ are constants. The actual Bratu problem has $d=0$ and  
$f \equiv0$. It provides a simplified model of nonlinear diffusion  
phenomena, e.g., in combustion and semiconductors, and has been 
considered by Glowinski, Keller, and Rheinhardt \cite{GloKR85}, 
as well as by a number of other investigators; see \cite{GloKR85} 
and the references therein. See also problem 3 by Glowinski and  Keller  
and problem 7 by Mittelmann in the collection of nonlinear model 
problems assembled by Mor\'e \cite{More}. The modified problem  
(\ref{bratu}) has been used as a test problem for inexact Newton 
methods by Brown and Saad \cite{Brown-Saad1}.  
 
In our experiments, we took $D = [0,1]\times[0,1]$, $f \equiv0$, 
$c=d=10$, and discretized (\ref{bratu}) using the usual second-order 
centered differences over a $100\times100$ mesh of equally 
spaced points in $D$. In \gmres($m$), we took $m=10$ and used fast  
Poisson right preconditioning as in the experiments in \S2. The computing  
environment was as described in \S2. All computing was done  
in double precision.  
 
 

\begin{figure}[ht] 
\vspace{2.5in}
\caption{{\rm Log}$_{10}$ of the residual norm versus the number of
{\rm GMRES$(m)$} iterations for the finite difference methods.} 
\label{diff} 
\end{figure}  
 

 
In the first set of experiments, we allowed each method to  
run for $40$ {\gmresm} iterations, starting with zero as the initial  
approximate solution, after which the limit of residual norm  
reduction had been reached. The results are shown in Fig.~\ref{diff}.  
In Fig.~\ref{diff}, the top curve was produced by method FD1. 
The second curve from the top is actually a superposition of  
the curves produced by methods EHA2 and FD2; the two curves are 
visually indistinguishable. Similarly, the third curve from  
the top is a superposition of the curves produced by methods EHA4 
and FD4, and the fourth curve from the top, which lies barely above  
the bottom curve, is a superposition of the curves produced by  
methods EHA6 and FD6. The bottom curve was produced by method A. 
 
In the second set of experiments, our purpose was to assess the 
relative amount of computational work required by the methods  
which use higher-order differencing to reach comparable levels  
of residual norm reduction. We compared pairs of methods EHA2  
and FD2, EHA4 and FD4, and EHA6 and FD6 by observing in each of 
20 trials the number of {\gmresm} iterations, number of $F$-evaluations,  
and run time required by each method to reduce the residual norm 
by a factor of $\e$, where for each pair of methods $\e$ was chosen  
to be somewhat greater than the limiting ratio of final to  
initial residual norms obtainable by the methods. In these trials,  
the initial approximate solutions were obtained by generating random  
components as in the similar experiments in \S2. We note that for every  
method, the numbers of {\gmresm} iterations and $F$-evaluations required  
before termination did not vary at all over the 20 trials. The {\gmresm}
iteration counts, numbers of $F$-evaluations, and means and standard  
deviations of the run times are given in Table \ref{diffstats}. 
 
\begin{table}
\caption{Statistics over $20$ trials of {\rm GMRES$(m)$} iteration numbers,  
$F$-evaluations, and run times required to reduce the residual norm by  
a factor of $\e$. For each method, the number of {\rm GMRES$(m)$} iterations  
and $F$-evaluations was the same in every trial.} 

\begin{center} \footnotesize
\begin{tabular}{|c|c|c|c|c|c|} \hline  
&& Number of & Number of & Mean Run Time & Standard \\ 
Method & $\e$ & Iterations & $F$-Evaluations& (Seconds) & Deviation \\ \hline 
\lower.3ex\hbox{EHA2} & \lower.3ex\hbox{$10^{-10}$} & \lower.3ex\hbox{26} &  
\lower.3ex\hbox{32} & \lower.3ex\hbox{47.12} & \lower.3ex\hbox{.1048} \\ 
FD2 & $10^{-10}$ & 26 & 58 & 53.79 & .1829 \\ \hline  
\lower.3ex\hbox{EHA4} & \lower.3ex\hbox{$10^{-12}$} & \lower.3ex\hbox{30} &  
\lower.3ex\hbox{42} & \lower.3ex\hbox{56.76} & \lower.3ex\hbox{.1855} \\  
FD4 & $10^{-12}$ & 30 & 132 & 81.35 & .3730 \\ \hline  
\lower.3ex\hbox{EHA6} & \lower.3ex\hbox{$10^{-12}$} & \lower.3ex\hbox{30} &  
\lower.3ex\hbox{48} & \lower.3ex\hbox{58.56} & \lower.3ex\hbox{.1952} \\ 
FD6 & $10^{-12}$ & 30 & 198 & 100.6 & .3278 \\ \hline  
\end{tabular}
\end{center} 
\label{diffstats} 
\end{table}  

In our first set of experiments, we took $c=d=10$ and used right 
preconditioning with a fast Poisson solver from {\fishpack}  
\cite{Swarztrauber-Sweet}, which is very effective for these  
fairly small values of $c$ and $d$. We first started each method 
with zero as the initial approximate solution and allowed it  
to run for 40 {\gmresm} iterations, after which the limit of residual  
norm reduction had been reached. Figure \ref{pdep} shows plots  
of the logarithm of the Euclidean norm of the residual versus  
the number of {\gmresm} iterations for the three methods. We note
that in  Fig.~\ref{pdep} and in all other figures below, the plotted
residual norms were not the values maintained by {\gmresm}, but rather
were computed as accurately as possible ``from scratch.''  That is, 
at each {\gmresm} iteration, the current approximate solution was 
formed and its product with the coefficient matrix was subtracted 
from the right-hand side, all in double precision.  
It was important to compute the residual norms in this way because  
the values maintained by {\gmresm} become increasingly untrustworthy  
as the limits of residual norm reduction are neared; see \cite{Walker88}.  
It is seen in Fig.~\ref{pdep} that Algorithm EHA achieved  
the same ultimate level of residual norm reduction as the FDP  
method and required only a few more {\gmresm} iterations to do
so.  
 
\begin{figure}[t] 
\vspace{3in}
\caption{{\rm Log}$_{10}$ of the residual norm versus the number of 
{\rm GMRES}$(m)$ iterations for $c=d=10$ with fast Poisson
preconditioning. Solid curve: Algorithm {\rm EHA}; dotted 
curve: {\rm FDP} method; dashed curve: {\rm FSP} method.} 
\label{pdep} 
\end{figure} 
 
 
In our second set of experiments, we took $c=d=100$ and carried out  
trials analogous to those in the first set above. No preconditioning  
was used in these experiments, both because we wanted to compare 
the methods without preconditioning and because the fast  
Poisson preconditioning used in the first set of experiments is 
not cost effective for these large values of $c$ and $d$. We first  
allowed each method to run for 600 {\gmresm} iterations, 
starting with zero as the initial approximate solution, after which  
the limit of residual norm reduction had been reached. 
 

\section*{Acknowledgments}
The author thanks the anonymous authors whose work largely
constitutes this sample file. He also thanks the INFO-TeX mailing
list for the valuable indirect assistance he received.
 
 
\begin{thebibliography}{10} 
\bibitem{bs} {\sc R.~A. Brualdi and B.~L. Shader}, 
{\em On sign-nonsingular matrices and the conversion of the
permanent into the determinant}, in Applied Geometry and
Discrete Mathematics, The Victor Klee Festschrift, P.
Gritzmann and B. Sturmfels, eds., American Mathematical
Society, Providence, RI, 1991, pp. 117--134.
 
\bibitem{djd} {\sc J. Drew, C.~R. Johnson, and P. van den Driessche}, 
{\em Strong forms of nonsingularity}, Linear Algebra Appl.,
162 (1992), to appear. 
 
\bibitem{g} {\sc P.~M. Gibson}, {\em Conversion of the permanent into the 
determinant}, Proc. Amer. Math. Soc., 27 (1971),
pp.~471--476.
 
\bibitem{klm} 
{\sc V.~Klee, R.~Ladner, and R.~Manber}, {\it
Signsolvability revisited}, Linear Algebra Appl., 59
(1984), pp.~131--157.
 
\bibitem{m} 
{\sc K. Murota}, LU-{\em decomposition of a matrix with
entries of different kinds}, Linear Algebra Appl., 49
(1983), pp.~275--283.

\bibitem{Axelsson}
{\sc O.~Axelsson}, {\em Conjugate gradient type methods for unsymmetric and
  inconsistent systems of linear equations}, Linear Algebra Appl., 29 (1980),
  pp.~1--16.

\bibitem{Brown-Saad1}
{\sc P.~N. Brown and Y.~Saad}, {\em Hybrid {K}rylov methods for nonlinear
  systems of equations}, SIAM J. Sci. Statist. Comput., 11 (1990), 
  pp.~450--481.

\bibitem{DES}
{\sc R.~S. Dembo, S.~C. Eisenstat, and T.~Steihaug}, {\em Inexact {N}ewton
  methods}, SIAM J. Numer. Anal., 19 (1982), pp.~400--408.

\bibitem{EES}
{\sc S.~C. Eisenstat, H.~C. Elman, and M.~H. Schultz}, {\em Variational
  iterative methods for nonsymmetric systems of linear equations}, SIAM J.
  Numer. Anal., 20 (1983), pp.~345--357.

\bibitem{Elman}
{\sc H.~C. Elman}, {\em Iterative methods for large, sparse, nonsymmetric
  systems of linear equations}, Ph.D. thesis, Department of Computer
  Science, Yale University, New Haven, CT, 1982.

\bibitem{GloKR85}
{\sc R.~Glowinski, H.~B. Keller, and L.~Rheinhart}, {\em Continuation-conjugate
  gradient methods for the least-squares solution of nonlinear boundary value
  problems}, SIAM J. Sci. Statist. Comput., 6 (1985), pp.~793--832.

\bibitem{Golub-VanLoan}
{\sc G.~H. Golub and C.~F. Van~Loan}, {\em Matrix Computations}, 
  Second ed., The Johns  Hopkins University Press, Baltimore, MD,  1989.

\bibitem{More}
{\sc J.~J. Mor\'e}, {\em A collection of nonlinear model problems}, in
  Computational Solutions of Nonlinear Systems of Equations, E.~L. Allgower and
  K.~Georg, eds., Lectures in Applied Mathematics, Vol. 26, American
  Mathematical Society, Providence, RI, 1990, pp.~723--762.

\bibitem{Saad}
{\sc Y.~Saad}, {\em Krylov subspace methods for solving large unsymmetric
  linear systems}, Math. Comp., 37 (1981), pp.~105--126.

\bibitem{Saad-Schultz}
{\sc Y.~Saad and M.~H. Schultz}, {\em {\rm GMRES}: A generalized minimal
  residual method for solving nonsymmetric linear systems}, SIAM J. Sci. Statist.
  Comput., 7 (1986), pp.~856--869.

\bibitem{Swarztrauber-Sweet}
{\sc P.~N. Swarztrauber and R.~A. Sweet}, {\em Efficient {\rm FORTRAN}
  subprograms for the solution of elliptic partial differential equations}, ACM
  Trans. Math. Software, 5 (1979), pp.~352--364.

\bibitem{Walker88}
{\sc H.~F. Walker}, {\em Implementation of the {\rm GMRES} method using
  {H}ouseholder transformations}, SIAM J. Sci. Statist. Comput., 9 (1988),
  pp.~152--163.

\bibitem{Walker89}
\sameauthor, {\em Implementations of
  the {\rm GMRES} method}, Computer Phys. Comm., 53 (1989),  pp.~311--320.

\end{thebibliography} 

\end{document}