Numerical Linear Algebra

NUMERICAL LINEAR ALGEBRA
W. Kahan
Table of Contents
The Solution of Linear Equations
0. Introduction.
1. The Time Needed to Solve Linear Equations.
2. The Time Needed to Solve a Linear System with a
Band Matrix.
3. Iterative Methods for Solving Linear Systems.
4. Errors in the Solution of Linear Systems.
5. Pivoting and Equilibration.
The Solution of the Symmetric Eigenproblem *
6. The General Eigenproblem and the QR Method.
7. Iterative Methods for Symmetric Eigenproblems.
8. The Reduction to Tri-diagonal form.
9. Eigenvalues of a Tri-diagonal Matrix.
10. Eigenvectors of a Tri-diagonal Matrix.
11. Errors in the Solution of an Eigenproblem.
0. Introduction. The primordial problems of
linear algebra are the solution of a system of linear
equations
Ax = b> (i.e., E.a..x = b.) ,
and the solution of the eigenvalue problem

A = X
-k k-k
for the eigenvalues X, and corresponding
eigenvectors x, of a given matrix A . The numer-
ical solution of these problems without the aid of an
electronic computer is a project not to be undertaken
lightly. For example, using a mechanical desk-
calculator to solve five linear equations in five
unknowns (and check them) takes me nearly an hour,
and to calculate five eigenvalues and eigenvectors of
a five-by-five matrix costs me at least an afternoon
of drudgery. But any of today1s electronic computers
are capable of performing both calculations in less
than a second.
T~
Sections 6 to 11 will appear in this Bulletin at a
later date.
757
Downloaded from https://www.cambridge.org/core. 28 Mar 2021 at 21:41:49, subject to the Cambridge Core terms of use.
What is the current state of the art of
solving numerical problems in linear algebra with the
aid of electronic computers? That question is the
theme of part of this paper. The rest of the paper
touches upon two or three of the collateral
mathematical problems which have captured my attention
for the past several years. These problems spring
from the widespread desire to give the computer all
its instructions in advance. When the computer is
engrossed in its computation at the rate of perhaps a
million arithmetic operations per second, human
supervision is at best superficial. One dare not
interrupt the machine too frequently with requests
"WHAT ARE YOU DOING NOW?" and with afterthoughts and
countermands, lest the machine be dragged back to the
pace at which a human can plod through a morass of
numerical data. Instead, it is more profitable to
launch the computer on its phrenetic way while we
calmly embrace mathematical (not necessarily
computational) techniques like error-analysis to
predict and appraise the machine's work. Besides,
the mathematical problems of prediction and appraisal
are interesting in their own right.
758
1. The Time Needed to Solve Linear Equations,
On our computer (an IBM 709^-11 at the University of
Toronto) the solution of 100 linear equations in 100
unknowns can be calculated in about 7 seconds;
during this time the computer executes about 5000
divisions, 330000 multiplications and additions, and
a comparable amount of extra arithmetic with small
integers which ensures that the aforementioned
operations are performed in their correct order.
This calculation costs about a dollar. To calculate
the inverse of the coefficient matrix costs about
three times as much. If the coefficients are complex
instead of real, the cost is roughly quadrupled. If
the same problem were taken to any other appropriate
electronic computer on this continent, the time
taken could differ by a factor between 1/10 and 1000
(i.e. 1 second to 2 hours for 100 equations)
depending upon the speed of the particular machine
used. These quotations do not include the time
required to produce the equations' coefficients in
storage, nor the time required (a few seconds) to
print the answers on paper.
For the next five years it will be economically

practical to solve general systems of 1000 or so
linear equations, but not 10000. One limitation is
the need to store the equations' coefficients
somewhere easily accessible to the computer1s
arithmetic units. A general system of N equations
2
has an NxN coefficient matrix containing N
elements. When N=100 , these 10000 elements fit
with ease into current storage units. When N=1000 ,
finding the space for a million elements requires
today some attention to technical details; tomorrowTs
storage units will handle a million elements easily.
o
But when N=10000 space is needed for 10 elements,
and current storage units with that capacity are
unable to share their information with the computer
at speeds commensurate with its arithmetic units.
o
Besides, to produce, collect, and check those 10
elements is a formidable undertaking.
Today, the solution of 1000 equations is not a
simple task, even on a large computer like ours.
Our computer's immediate access store, to which
reference can be made in a fraction of the time
required for one multiplication, has a capacity of
215=32768 words, of which about 10000 would be
needed for program. The remaining space is just
759
about enough for 10 or 20 rows of the matrix. The
rest of the matrix, 980 rows, has to be kept in bulk
storage units, like magnetic tapes or disks, to which
access takes at least as long as several multiplica-
tions. Now, most of the time spent in solving a
linear system is spent thus:
Select an element from the matrix,

multiply by another,
subtract the product from a third,
which is then replaced by the difference.
It is clear that careful organization is required to
prevent storage-access from consuming far more time
than the arithmetic. Such organization is possible;
for a good example see Barron and Swinnerton-Dyer
(I960). The main idea is to transfer each row of
the matrix in turn from slow storage to fast storage
and back in such a way that, while in fast storage,
each row partakes in-as many arithmetic operations
with neighbouring rows as possible. Further time is
saved by the simultaneous execution of input, output
and calculation; while one row is being transferred
from slow to fast storage, another is being
transferred back, and arithmetic operations are
being performed upon a third row. In this way,
1000 linear equations could be solved on our machine
in a morning, not much longer than would be needed
for the arithmetic operations alone (and in much
less time than would likely be needed to collect the
data or to interpret the answer).
Let us count up those arithmetic operations.

The methods most widely used for solving linear
equations are elimination methods patterned after
that described by Gauss (1826). Here is an outline:
Given the augmented matrix {A,b_} of the system
N
E a±jXj = b ± , i-1,2,...,N ,
J ~~ -*-
we select a variable, say x T , and eliminate it
from all the equations but one. This can be done,
for example, by selecting a suitable equation, say
the I , and subtracting (the I equation, times
a. T /a TT ) from (the i equation ) for all i 7* I .
th th
After the I equation and J variable have been
760
set aside, one has just (N-l) linear equations in
(N-l) unknowns left.
This simple process is repeated until there

remains only one equation in one unknown; this can
be solved easily. The solution is substituted back
into the equation previously set aside and that one
is solved. This process of back substitution is
repeated until, at the end, x T is obtained from
the Ith equation after the substitution of the
computed values of the other N-l variables.
Gaussian elimination requires
i-N3(l + I N " 1 - | N " 2 ) additions or subtractions,
and 1 ^ -1 -2
3-NJ(1 + 3N - N ) multiplications or divisions.
In the 140 years since this method appeared in print,
many other methods, some of them very different, have
been proposed. All methods satisfy the following
new theorem of Klyuyev and Kokovkin-Shcherbak (1965):
Any algorithm which uses only rational
arithmetic operations to solve a general
system of N linear equations requires at
least as many additions and subtractions,
and at least as many multiplications and
divisions as does Gaussian elimination.
One consequence of this theorem is obtained
by setting N=10000; to solve 10000 linear equations
would take more than two months for the arithmetic
alone on our machine. The time might come down to a
day or so when machines 100 times faster than ours
are produced, but such machines are just now being
developed, and are most unlikely to be in widespread
use within the next five years. The main impediment
seems to be storage access time. (For more details,
see IPIP (1965).)
In the meantime, there are several good
reasons to want to solve systems of as many as
10000 equations. For example, the study of many
physical processes (radiation, diffusion,
elasticity,...) revolves about the solution of
partial differential equations. A powerful technique
for solving these differential equations is to
761
approximate them by difference equations over a
lattice erected to approximate the continuum. The
finer the lattice (i.e. the more points there are in
the lattice), the better the approximation. In a
20x20x20 cubic lattice there are 8000 points. To
each point corresponds an unknown and a difference
equation. Fortunately, these equations have special
properties which free us from the limitation given
by Klyuyev and Kokovkin-Shcherbak. (For details
about partial difference equations, see Smith (1965)
or Forsythe and Wasow (i960).)
2. The Time Needed to Solve a Linear System

with a Band Matrix. The systems of linear equations
which arise from the discretization of boundary
value problems frequently have matrices {a..} with
the following "band property":
a. . = 0 if i-j| M
A Band-Matrix,
Although N equals the number of lattice points in

the discretization, and can therefore be quite
large whenever a fine lattice is needed for high
accuracy, the half-bandwidth M is usually much
smaller than N . For a boundary value problem in
6 dimensions M is usually very near the number of
points in one or two (6-1) dimensional sections of
the lattice, and hence the quotient
M/N 1-1/6
is frequently between 1 and 3 . With care, the
matrix corresponding to k coupled boundary value
problems over the same lattice can often be put in a
band form for which the quotient above lies between
k and 3k .
The advantage of a band structure derives from
the fact that it is preserved by the row-operations
involved in Gaussian elimination. This is obvious
762
when we select, for I = 1,2,..., N in turn, the
I equation to eliminate the I unknown from all
subsequent equations. It is true also when any other
row-selection rule is used, provided the width of
that part of the band above the main diagonal is
allowed to increase by M . Consequently, far less
time and space are needed to solve band-systems than
to solve general systems. The following table
summarizes the dependence of time and space
requirements upon the parameters M and N . For
the sake of simplicity, constants of proportionality
have been omitted, and terms of the order of 1/M
and 1/N or smaller are ignored.
Type of Storage required Time required for

Matrix (Total) arithmetic alone
Band MN M2N
Full N2 N3
Incidentally, there is no need to find space

for all MN elements of the band-matrix in the
immediate-access store. Instead, it suffices to
store the rows of the matrix in slow bulk storage
2
(tape or disk) and find space for 2M elements
(in some cases fewer) in the immediate access store,
Then as the I row of the matrix is transferred
rec
from slow to fast storage, a transformed (I-M) th
row can be transferred from fast to slow.
The economies that result from band structure
permit the solution of two dimensional boundary
value problems with thousands of points, and one
dimensional problems with tens of thousands of
points, to be carried out in times measured in
minutes. (For more details about the solution of
band systems, see Cayless (1961), Fox (1957)
pp. 150-155, Walsh's ch. 22 in Fox (1962), or
Varga (1962) pp. 194-201.)
3. Iterative Methods for Solving Linear

Systems. Iterative methods for solving linear
systems, due originally to Gauss (1823), Liouville
(1837) and Jacobi (1845), embody an approach which
is very different from that behind the direct
methods like Gaussian elimination. The difference
can be characterized as follows:
763
Direct methods apply to the equation Ax. = b.
a finite sequence of transformations at whose
termination the equations have a new form, say
U:x = c_ , which can be solved by an obvious and
finite calculation. For example, in Gaussian
elimination U is an upper triangular matrix which,
with £ , can be shown to satisfy
{U-,£} = iT-'-PUjb}
where P is a permutation of the identity and L
is a lower triangular matrix with diagonal elements
all equal to 1 ; and the obvious calculation that
solves Ux = c_ is back substitution. In the
absence of rounding errors, the computed solution is
exact. (For more details see Faddeev and Faddeeva
(1964) ch. II, Householder (1964) ch. 5, or Fox
(1964) ch. 3-5.)
On the other hand, an iterative method for
solving Ax = b_ begins with a first approximation
z_n , to which a sequence of transformations are
applied to produce a sequence of iterates
z, , z_p , z~ , . . . which are supposed to converge
toward the desired solution x_ • ^-n practice the
sequence is terminated as soon as some member z,
of the sequence is judged to be close enough to x_ •
An example of an iterative method is the
Liouvilie-Neumann series which is produced by what
numerical analysts call nJacobiTs Method":
Suppose Ax = b_ can be transformed
conveniently into the form
x. = Bx_ + c_
where the matrix B is small in some sense. To be
more precise, we shall assume that ||B || = 8 < 1 .
(The symbol ||. ..|| represents a matrix norm about
which more will be said later.) We begin with a
first approximation z_n , for which 0_ will do if
nothing better is known, and iterate thus:
764
z . k + 1 = Bz^ + £ for k = 0,1,2,... ,
= (I+B+B 2 + . . . +B k )c + B k + 1 z r
= x + Bk+1(z,0-x) •
Hence, the error is bounded by
l|zk-x!<Sk|lz0-x|l .
The practicality of this scheme depends upon

three considerations :
i) The smaller is 3 , the fewer are the
iterations required to effect a given
factor of reduction in the error bound.
Therefore, small values of & are desired
for rapid convergence.
ii) The better is z_0 , the fewer are the
iterations required to bring the error
bound below a given tolerance. Therefore,
good initial approximations are desired for
early termination of the iteration.
iii) If each matrix-vector multiplication Bz_,
is cheap enough that we can afford a large
number of them, then the two previous
considerations will carry less weight.
The last consideration is quite important in

many applications. For example, if the NxN matrix
B is "sparse", which means that most of B T s
elements are zeros, then the number of arithmetic
operations required to compute Bz, may well be a
p
small multiple of N instead of N . Such sparse
matrices are frequently encountered during the study
of trusswork bridges, electric networks, economists1
input-output models, and boundary value problems.
In the case of some large two-dimensional boundary
value problems, and most three dimensional ones,
it may be more economical to exploit the sparseness
of the matrix than to exploit its band properties.
Despite a sparse matrix and a fast computer,
the simple iteration described above is usually too
slow to be practical. This fact has spurred the
765
^generalization and improvement of iterative methods
in a vast diversity of ways.
One generalization begins with the iteration
= + C
-k+1 -k k-k
where
Z k = b - Az k
is z, t s "residual" in the equation Ax. = b_ .

(The Jacob! iteration is obtained formally by
letting C ~ be A T s diagonal.) For simplicity
suppose C, = YkC a with scalars y, and the matrix
C to be chosen later in accordance with the
subsequent analysis. We find that
z^ = x + Pk(CA)(z_0-x)
where P, (w) is a k degree polynomial defined

by the recurrence
P Q (w) = 1 , P k + 1 (w) = (l-ykw)Pk(w) .
To simplify the exposition now I shall assume

that CA!s elementary divisors are all linear;
otherwise what follows would be complicated by the
appearance of some derivatives of P^Cw) • The
matrix CA can be decomposed into its idempotent
elements E(x) defined by
(CA) n = z XnE(x) for all n

X
where the summation is taken over the values of the
eigenvalues X of CA . (Cf. Dunford and Schwartz
(1958) pp. 558-9.) Then
P k (CA) = Z Pk(X)'E(A) ,
A
whence comes the following theorem:

A necessary and sufficient condition that
z, -> x as k •> 00 3 no matter how zn
—k — * —0
766
be chosen, is that ïV(*) "* 0 f ° r every
eigenvalue X of CA .
This theorem is not applied directly in

practice because the eigenvalues X are generally
not known. But if all the eigenvalues X are known
to be contained in some region R in the complex
plane, then it suffices that p k ( w ) ~* ° f or all
w in R . To satisfy this requirement is not
trivial, because ^ ( O ) = ! f ° r a 1 1 k •
Given R and k , one might seek a polynomial

Pk(w) which is "best" in the sense that, for
example, of all polynomials of degree k for which
P,(0) = 1 , pi<-(w) h a s the smallest value of
max |P, (w) | over w in R

There is no general rule known for finding such best
polynomials. The following theorem helps in some
cases :
Let <j£,(r) be the lemniscate in the complex
w-plane defined by
|L k (w)| £ r < L k (0)
for some polynomial L,(w) of degree >_ k .
Then every k degree polynomial P^Cw)
with P k (0) = 1 satisfies
max |P k (w)| _> r/Lk(0) for w in ^ ( r ) .
Proof. Apply the maximum modulus theorem on

the exterior of <^f(^) to the rational function
L k (0) Pk(w)/Lk(w)
to conclude that this quotient has magnitude at

least 1 somewhere on the boundary of <>£ (r) .
Then apply the same theorem to P^Cw) inside
4(r)
767
The simplest application of this theorem is
to the circle
|L k (w)| = |(l-w) k | 1 Bk ,
which shows that, if A = I-B and we know only

that ||B|| == 6 < 1 (so that all eigenvalues X of
A must satisfy 11-X | <_ 3 < 1) , then the simple
Jacobi iteration described above is the best that
can be done.
Having chosen a polynomial *V( W ) > the

numbers y . are defined as the reciprocals of the
P
zeros of k^ w ^ ' This relation amounts to an
inconvenient restriction on the sequence of
polynomials P.(w) for j < k , and is also a
source of possible numerical instability. To
illustrate this point, suppose all the eigenvalues
X of CA lie in an interval on the real axis,
say 0 < a o < _ X < _ a . The following theorem of
Markoff is applicable:
Of all k degree polynomials p k ( z ) with
P,(0) = 1 , the one for which
max |P,(z)| on ag <_ % <_ a
is smallest is just the Tchebycheff polynomial
Tk(L(z))/Tk(L(0))
where
T,(cos 0) = cos k Q and
L(z) = (am + a 0 - 2z)/(am - a 0 ) .
Proof. If |P k (z)| <_ l/Tk(L(0)) on

a z a
o £ £ m 9 then the difference
Tk(L(z)) - Tk(L(0))Pk(z) has at least one zero
between each pair of adjacent extrema of T,(L(z))
on the interval, and another at z = 0 , making k+1
zeros altogether.
Now, the zeros \."" of T,(L(z)) include
768
some numbers quite near ao > which may be very small
in cases of slow convergence. But then the term
Y-tCr. , when its turn comes in the iteration, can be
so large and so much magnify the effects of rounding
errors that the convergence of the iteration is
jeopardized (Young (1956)).
Fortunately, the Tchebycheff polynomials
satisfy a three-term recurrence
T k+1 (L(z)) = 2L(z)Tk(L(z)) - T^ 1 (L(z))
which can be implemented conveniently and is

numerically stable. A suitable form for the
iteration is
+ + 6
^k+i • ^k Vïk k^k - s*-i> •
and an appropriate choice of y , , 6, , and
z k = x + Tk(L(CA))(zQ - x)/ T k ( L (°))
for all k >_ 0 . One great convenience of this

iteration is that the polynomial ^ ( w ) that
appears in the relation
zk = x + Pk(CA) (z_Q - x)
is the "best" such polynomial for each k , so that

there is no need to choose a degree k in advance.
This convenience persists whenever the sequence of
polynomials P, (w) are orthogonal polynomials
2
which minimize some weighted mean value of |P(w)|
over an interval of interest, because these
polynomials also satisfy a three-term recurrence.
For details see Stiefel (1958) and Faddeev and
Faddeeva (1964) ch. 9.
A very different approach to iterative
methods can be illustrated by the Jacobi method
once again. We note that, since
*k+i - 2 - = B(
^k - 1 > >
769
where 6 = ||B|| < 1 . In other words, each iteration
reduces the norm of the error by a factor 3 < 1 .
More generally, given a norm for the error

||z, - x|| or for the residual ||b_ - A^J| , one seeks
a direction £, and distance \, such that the
norm associated with
is smaller than for z, . In Gauss-Seidel

iterations the direction g. is one of the
coordinate directions; in gradient iterations the
direction g, is that in which the norm is
decreasing most quickly. For further discussion see
Householder (1964) sec. 4.2-3. I shall elaborate
upon only one such method, called "the method of
conjugate gradients".
Suppose A is symmetric and positive definite^
and let us use || r» || = (r, T r k ) as a norm for the
residual r, = b_ - Az, . Then each iteration step
= + y + 6
Sk+i ^k <A k(^k " S t - i >
looks formally just like the iteration that was used
above to construct the Tchebycheff polynomials, but
now the constants -y, and 6, must be chosen to
minimize IK,.-J I i-n that step. This choice of "y,
and 6, has the stronger property that no other
choice of v Q , 6 Q , \-., 6-,, ... , Y, , 6, could yield
a smaller value for IIUV+TJI • I n particular,
ÎIM = 2. > the iteration converges in a finite number
of steps. An excellent exposition of this powerful
technique is given by Stiefel (1953) and (1958).
Another approach to iterative methods is

embodied in the relaxation methods. The basic idea
770
here is to adjust some unknown Xj to satisfy
("relax") the I equation I. a^. x. = b. ,
even though in doing so some other equation may be
dissatisfied. The next step is to relax some other
equation, and so on. Gauss (1823) claimed that this
iteration could be performed successfully
"... while half asleep, or while thinking
about other things".
Since his time the method has been systematized and
generalized and improved by orders of magnitude,
especially where its applications to discretized
boundary value problems are concerned. The best
survey of this development is currently to be found
in Vargafs book (1962). Nowadays some of the most
active research into iterative methods is being
conducted upon those variants of relaxation known
as Alternating Direction Methods; see for example
Douglas and Gunn (1965)* Gunn (1965)* Murray and
Lynn (1965), and Kellog and Spanier (1965).
The result of the past fifteen years of
intense mathematical analysis concentrated upon
iterative methods has been to speed them up by
factors of ten and a hundred. Some idea of the times
involved can be gained from surveys by Engeli et al
(1959), Martin and Tee (1961), and Birkhoff, Varga
and Young (1962). For example, the difference
analogue of Dirichlet?s problem (Laplace's equation
with specified boundary values) in a two-dimensional
region with about 3000 lattice points can be solved
to within about 6 significant decimals in about 300
iterations of successive over-relaxation, requiring
about 30 seconds on our machine. This is one third
as long as would be needed to apply Gaussian
elimination to the corresponding band matrix. A
three-dimensional problem with 10000 equations and
unknowns could be solved on our machine in less than
5 minutes by iteration; here Gaussian elimination
takes hundreds of times as long, so the value of
iteration is well established. But the time required
for iterative methods generally cannot easily be
predicted in advance except in special cases (which
are fortunately not uncommon). Furthermore, the
choice of one out of several possible iterative
methods is frequently a matter of trial and error.
Even if the iterative method is chosen on rational
grounds, there will be parameters (like the constants
Y k and 6, above) which must be chosen carefully
for maximum efficiency; but to choose their values
771
will frequently require information that is harder
to obtain than the solution being sought in the
first place, (A welcome exception is the method of
conjugate gradients.) Evidently there is plenty of
room for more research, and especially for the
consolidation of available knowledge.
4. Errors in the Solution of Linear Systems.

It is possible in principle to perform a variant of
Gaussian elimination using integer arithmetic
throughout in such a way that no rounding errors
are committed (see Aitken's method described by
Fox (1964) on pp. 8 2 - 8 6 ) . But the integers can
grow quite long, as much as N times as long as
they were to begin with in the given NxN matrix A.
Whenever N is large, one is easily persuaded to
acquiesce when the computer rounds its arithmetic
operations to some finite number of significant
digits specified in advance. Consequently, it comes
as no surprise when the calculated value z^ of the
solution of
Ax = b
exhibits a small residual
r = b_ - A^ .
How small must r be to be negligible? The
following example shows that this question can have
a surprising answer.
Example 1.
A _ /.2161 .l44l\ /.l44o\ v _ /.9911\
A
" U.2969 -8648/ > 5 . " (.8642/ > £ ~ (-.4870/ •
Then the residual is
r = b - Az = (--0000000l\ eexxaaccttl± y .
r D_ - AZ_ - ^ .00000001/ y >
no other vector z_ specified to 4 dec. can have

a smaller residual r unless r = 0_ . But z_ does
not contain a single correct digit I The"correctft
solution is
Linear systems with this kind of pathological

behaviour are often called "ill-conditioned".
Precisely what does "ill-conditioned" mean?
772
This example and other problems of error
analysis are easier to discuss using the language of
matrix norms, which I digress to introduce here.
A common vector norm is the Holder norm
||x||p = (Zjlxjl?) 1 ^ for 1 I P < - .
T h i s norm i s e a s i l y shown t o have t h e p r o p e r t i e s

t h a t one e x p e c t s of a v e c t o r norm:
||x|| > 0 except t h a t ||0J| = 0 .
a = a f o r a 1 1
|| 2L|| I I II^LII scalars a .
+
||2L Z.|| £ | | £ | | + ||JL|| (The T r i a n g l e I n e q u a l i t y ) .
Any l i n e a r t r a n s f o r m a t i o n A from one normed l i n e a r
s p a c e t o a n o t h e r can be normed i n t h e n a t u r a l way:
||A|| = max ||AxJ|/||x|| o v e r x + _0 .
(I use "max" instead of "sup" because they amount to
the same thing for finite dimensional spaces,)
Among the matrix norms most often used are
||A|| = max ||Ax|| / | | x | L with s u i t a b l y chosen

p and q ; e . g .
ooôo = max llAxllyHxJI^ = ma X l | ^ | a i J | (max r o w - s u n )

, _ = max ||Ax|J,/||x|L = max. | z . | a - . | (max c o l u m n - s u n )
IIAIL,I = m a x iML/yii =max ij i a iji

||A|| 2 2 = max ||AxII 2 /||xJ| 2 = (max e i g e n v a l u e of A T A ) 1 / 2
T 1/2
( A n o t h e r m a t r i x norm, J|AjJE = ( t r a c e A A) ' , is
widely used but cannot be defined as a maximum of a
ratio of two vector norms. Its main value is as an
easily computed estimate of |JAJJ2 2 , because
Ul\E L M2$2 > WE/^rank (A) . )
U n t i l f u r t h e r n o t i c e , t h e m a t r i x norm u s e d below
w i l l be assumed t o be one of t h e norms IIAII
773
Finally, the notion of a dual linear
functional should be mentioned.. If the row-vector
£T is regarded as a linear operator from a normed
vector space to the space of complex numbers normed
as usual, then
Bz.T||a max lz.Tï.|/||x|l over

2L * £ •
e
-g- B l i p = m ax lzTx|/||x||p = ||ï||q where p" 1 + q"1 = 1
And the Hahn-Banach theorem guarantees that to each

x ^ 0 corresponds a dual functional £ T such that
Z\ = IIJLT| 11*11 = 1 •
(For more details, see Householder (1964) or any text
on normed linear spaces; e.g. Day (1962), Kantorovich
and Akilov (1964).)
Now it is possible to discuss the meaning of
"ill-condition". To each matrix A , regarded as a
linear operator from one normed space to another, can
be assigned its condition number K(A) associated
with the norms and defined thus:
K(A) -= S2L M/l over a l l x ï 0_ and £ ï jO

Wll HI
In other words, if the vectors Ax, and Ab are
regarded as errors correlated by A(x^ + AxJ = b_ + Ab^ ,
where x_ a n d 5. satisfy A^c = b_ , then
1/K(A) £ (i|Axtl/||x|l)/(llAb||/||b||) <K(A) .
This means that a small change Ab_ in b_ causes a
change Ax in x_ which has, relatively speaking, a
norm that can be K(A) times as big. When K(A) is
very large, we say that "A is ill-conditioned" .
It is easy to prove that when A is a square
matrix,
K(A) - |A| HA"1! .
The matrix A of the numerical example is very ill-
conditioned indeed;
774
/-86480000. Il*4l0000.>
A" - UonEnnnnn
\i2969OOOO. o-i£-innnrj J
-21610000. and
K(A) * 2 x 10 8 .
If we apply the inequality, associating r with Ajb
and z_ - x with Ax , we verify that
(fl£ - xJl/fl^l)/(|IlJ|/IlbJ|) * (l/2)/(l(T 8 ) < K(A) .

Had K(A) been known in advance, the example would
not have come as a surprise.
Another important property of the condition
number is given by the following:
THEOREM: A differs from a singular matrix
by no more in norm than |]A||/K(A) , (Gastinel)
i.e., given A , ||A||/K(A) = min ||AA|| over all
singular (A + AA) .
P r o o f . Of c o u r s e , i f (A + AA) i s s i n g u l a r ,
t h e n t h e r e i s some x ^ 0_ f o r which (A + AA) x. = Q.
Therefore || AA|| >_ ||AAX)| / ||x ||
• rai/M
= IIAXII/IIA- 1 AXII
i I/HA"1)! = ||A||/K(A) .
To find a AA which achieves the inequality we
for
consider that vector £ which
II A " d = II A " ! ! ||Z|J ¥ 0 .
1 1
T -1
Then let w be dual to A £
T 1
i.e. w A" Z = JJw j| jJA" 1 2|| - 1 ,
T
and set AA = - £ w T .
We have (A + AA) A~ £ = 0 , so A + AA is singular.
And
|AA[| = max JJ^ w T xJJ/j|x|| o v e r 2L ^ H
= BjJ max |w T x|/iixii
775
-lid I*TI MMI/lK^Jl = i/IJA-1!
= ||A||/K(A) .
Let us return to the example again. If the

elements of A have been rounded to 4 decimal
places, then the uncertainty introduced by the
rounding is 10000 times larger than the difference
between A and the nearest singular matrix! Under
these circumstances it is reasonable to ask whether
the system Ax_ = b_ deserves to have a solution.
The pathological behaviour of ill-conditioned

matrices seems to have preyed upon the minds of the
early analysts of the error committed during
Gaussian elimination. Certainly the conclusions of
von Neumann and Goldstine (19^7, 1951 ) are
incredibly pessimistic; for example they concluded
that on a machine like ours there were substantial
risks taken in the numerical inversion of matrices
of orders much larger than 20, although their
error-analysis was correct in other respects.
(Their trouble arises from an attempt to compute
A" from the formula
A - 1 = ( A ^ r 1 AT ,
an attempt which we know now to have been ill-
conceived. )
A more nearly modern error-analysis was

provided by Turing (1948) in a paper whose last few
paragraphs foreshadowed much of what was to come,
but his paper lay unnoticed for several years until
Wilkinson (I960) began to publish the papers which
have since become a model of modern error-analysis.
WilkinsonTs main result about Gaussian

elimination can be summarized thus:
Provided Gaussian elimination is carried out
in a reasonable way (about which more later), the
computed approximation z_ to the solution of
Ax = b_
will satisfy instead an equation of the form
(A + AA) z = b
776
where, although AA is not independent of b^ and
z, , AA satisfies an inequality of the form
||AA||< c g N NP 3 " S ||A||

where N is the order of A ,
3~s represents "1 unit in the last place"
-s -8
e.g. e =10 on an 8 dec. digit machine,
c is a modest positive constant, usually
less than 10,
p is a small positive constant always
smaller than 3.
g N is the pivot-ratio, about which more later.
The constants c and p depend upon the details of

the arithmetic and the norm; they will not be
discussed here. (See Wilkinson (1963) and references
cited therein.)
In short, JJAAjl is comparable to rounding

errors in JJAJJ ; and if the data in A is already
uncertain by more than a few hundred units in the
last place carried then the perturbation AA
attributable to the process of elimination will be
quite negligible by comparison. Indeed, in many
cases the perturbation AA will amount to less than
one unit in the last place of each element of A !
So, a small residual

r = b_ - Az_ = AA z
is just what might be expected from Gaussian
elimination. But the error z, - x. is another
matter;
£ - x =-A AA z_ , where JJ AA|) <_ e |}AJ[ , say .
Therefore \\z - x|l 1 IIA"1 II I|AA|| || z j |
< K(A) e 11 M
,l
- 1 - K(A) e ^"
where K(A) is A f s condition number,
e = c g N N p B~S and is very small,
and we assume that K(A) e < 1 .
777
In other words, although the residual r is always
small, the error £ - x. may be very large if A is
ill-conditioned; but this error will be no worse than
if A were in error by about c g N N^ units in its
last place to begin with.
The constant g N has an interesting history
It is connected with the rate of growth of the
"pivots" in the Gaussian elimination scheme. The
pivot is the coefficient a-j-j by which the I
equation is divided during the elimination of x, .
This term "pivot" will be easier to explain after
the following example has been considered:
Example 2.
-10 -1
A = 1
1
À A. .1
0 -2 2 x.
,-1 h { -2 0.9999999998 1.0000000002
. „ «.1,x.
-2
2 1.0000000002 0.9999999998/ x Q
Clearly, A is not ill-conditioned at all.

But suppose we apply Gaussian elimination to solve
the equation Ax_ = b_ . Our first step could be to
eliminate x, from equations 2 and 3 by subtracting
suitable multiples of equation 1 from them. The
reduced matrix should be
x
__1 13
10
2 x lu" -1 1
"1
0 -4999999999 5000000001
0 5000000001 -^999999999/
778
But if we have a calculator whose capacity is limited
to 8 decimal digits then the best we could do would
be to approximate the reduced matrix by
-10
22.
-1 l u
l
-5 x 10- 5 x 10: bf.
5 x 10* -5 x 10'
but this is precisely the reduced matrix we should

have obtained without rounding errors if A had
originally been
_2
-10 -1
A + AA = 0
0
In other words, the data in A f s lower right hand

2x2 submatrix has fallen off the right hand end
of our calculator's 8-digit register, and been lost.
The result is tantamount to distorting our original
data by the amount of the loss, and in this example
the distortion is a disaster.
These disasters occur whenever abnormally
large numbers are added to the moderate sized
numbers comprising our data. To avoid such disasters
it is customary to choose the variable Xj to be
eliminated, and/or the row I with which it is to be
eliminated from all other rows, in such a way that
the pivot aIJ is the largest available element a
ij
in its row, or column, or both. Since the typical
computation replaces a by
ij
for all (i,j) Ï (I,J)
ij = aij " iJ Ij IJ
a a /a
we see that max.â 1 ..! <_ 2 max 1

ij ij
so that no abnormally large numbers should appear.
In the example above we might choose l21 as the
pivot to obtain the reduced matrix
779
1
°2
•1.0000000 1.0000000 b'
2 2
(working to 8 significant digits) .

This reduced matrix is what would have resulted if
no rounding errors had been committed during the
reduction of
-1
A + AA = 1
1
which differs negligibly from the given matrix A .

Wilkinson's error bound, quoted above, assumes
that each pivot a T T is the largest in its row or
else in its column of the reduced matrix, and then
g
N is the ratio of the largest of the pivots to
the largest element in A . We can see that, since
the largest elements of each reduced matrix never
exceed twice those of the previous one,
N-l
% ±2
This bound is achieved for X = 1 by the matrix
Example 3:
1 0 0 0 0 1
-1 1 0 0 0 1 a±J = -1 if i > j ,
-1 -1 0 1
1 0 a±± = 1 if i < N ,
-1 -1 -1 1
A
x •
a±1 • 0 if i < j < N,
a.j, = 1 if i < N, and
1 0 1 = X
-1 1 1
-1 -1 -1 -1 -1 X N x N
780
if each pivot is chosen on the diagonal as one of
the largest elements in its column, and the columns
are chosen in their natural order 1,2,3, ..., N
during theêlimination. But when we repeat the
computation with X = 2 and sufficiently large N ,
an apparent disaster occurs -because the value of
X = 2 gets lost off the right-hand side of our
computing register. On a binary machine like our
7094 (using truncated 27-bit arithmetic), X = 2
is replaced by X = 1 if N > 28 . An example like
this was used by Wilkinson (1961, p. 327) as part of
the justification for his recommendation that one
use both row and column interchanges when selecting
pivot a T T to ensure that it is one of the largest
elements in the reduced matrix. This pivot-selection
strategy is called "complete pivoting" to distinguish
it from "partial pivoting" in which either row
exchanges or column exchanges, but not both, are
used.
The other justification for complete pivoting

was Wilkinson1s proof of a remarkable bound for the
ratio g N of the largest pivot to the largest
element of A :
g N < iX.Z1^1'2.*1'3 ... N 1 / ( N - 1 } ) 1 / 2 < 2N** 1O

«N ,
which is certainly far smaller than 2

(This bound is worth a small digression. It
is known to be unachievable for N > 2 ; and the
bound
% £N
has been conjectured for complete pivoting when A
is real. The conjectured bound is achieved whenever
A is a Hadamard matrix, and L. Tornheim has shown
that the conjecture is valid when N s 3 • He has
also shown that when A is complex the larger bound
g 3 £ 16/3 3 7 2
can be achieved.)
Despite the theoretical advantages of complete
pivoting over partial pivoting, the former is used
much less often than the latter, mainly because
781
interchanging both rows and columns is far more of a
nuisance than interchanging, say, rows alone.
Moreover, it is easy to monitor the size of the
pivots used during a partial pivoting computation,
and stop the calculation if the pivots grow too
large; then another program can be called in to
recompute a more accurate solution with the aid of
complete pivoting. Such is the strategy in use on
our comDUter at Toronto, and the results of using
this strategy support the conviction that intolerable
pivot-growth is a phenomenon that happens only to
numerical analysts who are looking for that
phenomenon.
Despite the confidence with which the computed

vector z_ , produced by Gaussian elimination or some
other comparable method, can be expected to have a
residual
r = b_ - Az_
which is scarcely larger than the rounding errors

committed during the calculation of r_ ,
i . e . Dr | | - lib - AzJI ~ N6- S (||b|| + ||A|| ||Z||) ,

an important problem remains. How large is the
error z_ - x. with which z_ approximates the "true
solution" x. of Ax = b_ ? This question is
meaningful even if A and b_ are not known
precisely; we can interpret z_ to be the solution of
a perturbed system
(A + AA) z, = ;b + Ab_
i n which |] AA JJ and ||Ab_|| a r e bounded i n some g i v e n
way, and h e n c e so i s v_ = AAz_ - Ab_ . A p r e c i s e
answer t o t h e q u e s t i o n i s
x - z = K-\ , || z - x|| < || A" 1 !! | | r | | ,

but here we must know || A" || in order to complete
the answer. If we try to compute A~ , we shall
instead obtain some approximation, say Z , and
once again we shall have to ask the question
How large is the error Z - A" ?
This question can be answered fairly easily if Z

is accurate enough, as shall now be shown.
782
Each column z. of Z = {z.-,,z_p> ... > Zvr)
can be regarded as an approximation to the
corresponding column X- °f
A = A = I X_n y X_P * * * * * ~Kf *
the solution of AX = I . A reasonable way to solve

the last equation is to use Gaussian elimination, in
which case each column z_. will be computed
separately and will satisfy
(A + A ± A) z_. = (the ith column of I)
where ||A.A|| <_ e || A || for some small e which

depends upon the details of the program in a way
discussed by Wilkinson (1961).
Now let R be the residual
R = I - AZ .
It is not necessary to compute R ; we can write
R = I-A {z^z^, ... 3 z_N> = {A1Az_1,A2Az_2, ..., A ^ ^ } ,
i n which each column A.Az_. of R i s bounded i n

norm by
KAzJ < cflAfllzJ .

Therefore
P = llHJI i n ||A|||jZ||
where r\/e depends upon N and the norms; usually
n/e 1 N 1 / 2 .
Since e can be predicted in advance, so can n ;
and it is possible to check whether
n IIAJUJZJI < 1 ,
in which case p < 1 and the following argument is
valid:
783
H A-^fl^Jlzll+H A-^-zj^nzii^ IIA-^-CI-AZ)!! =!IZ/|+|| A - :L R||
£l|z|M|Ai|p ;
llA^llillZll/d-p) .
Then IIA"1-ZIU lU""1!! p £ ||z||p/(l-p) .
The last inequality says that, neglecting
modest factors which depend upon N , the relative
error || A~1-Z||/||z || is at worst about K(A) times
as large as the relative errors committed during
each arithmetic operation of ZTs computation. In
other words, if A is known precisely then one must
carry at least l°g-iQ K(A) more decimal guard
digits during the computation of A~ than one
wants to have correct in the approximation Z , and
one can verify the accuracy of Z at the end of
the computation by computing n ||A || ||Z || .
If the method by which an approximation Z
to A~ was computed is not known, there is no way
to check the accuracy of Z better than to compute
R and p directly, and this calculation is not
very attractive for two reasons. First, the
computation of either residual
R = I-AZ or I-ZA
costs almost as much time as the computation of an
-1 3
approximate A ; both computations cost about N
multiplications and additions. Second, if K(A) is
large then ||l-AZ|| and || I-ZA || can be very
different, and there is no way to tell in advance
which residual will give the least pessimistic
over-estimate of the error in Z .
(p=||l-AZ|J = ||A(I-ZA)A""1|| <_ K(A)||I-ZA|| etc.)
Both residuals can be pessimistic by a factor like
K(A) . Finally, although a better approximation to
A" than Z is the matrix
Z 1 = Z + Z(I-AZ) = Z + (I-ZA)Z
(because [[I-AZj = |[(I-AZ)2[| <_ ||l-AZ||2 ) ,
784
the computation of Z., is in most cases more costly
and less accurate than a direct computation of an
approximate A" using Gaussian elimination with
double precision arithmetic. For example, on our
7094 it takes less than twice as long to invert A
to double precision (carrying 16 dec.) than to do
the same job in single precision (8 d e c ) , and the
double precision computation has almost 8 more
correct digits in its answer. But Z., has at most
twice as many correct digits as Z . Therefore, if
Z comes from a single precision Gaussian elimination
program, it will have about 8-log K(A) correct
digits. Z, will have 16-2 log K(A) digits at
best. The double precision elimination will produce
about 16-log K(A) correct digits. Thus does
engineering technique overtake mathematical
ingenuity !
The solution of Ax_ = b_ for a single vector
x_ is not normally performed by first computing
A" and then x = A" b_ for four reasons. First,
the vector Zb_ , where Z is an approximation to
A~ , is frequently much less accurate than the
approximation z_ given directly by Gaussian
elimination. Second, the direct computation of the
vector £ by elimination costs about 1/3 as much
time as the computation of the matrix Z . Third,
if one wants only to compute a vector z_ which
makes r = b_ - Az_ negligible compared with the
uncertainties in b_ and A , then Gaussian
elimination is a satisfactory way to do the job
despite the possible ill-condition of A , whereas
b_ - A(Zb_) = Rb_ can be appreciably larger than
negligible. Fourth, Gaussian elimination can be
applied when A is a band matrix without requiring
the vast storage that would otherwise be needed for
A" . The only disadvantage that can be occasioned
by the lack of an estimate Z of A~ is that
there is no other way to get a rigorous error-bound
for z^ - x . This disadvantage can be partially
overcome by an iterative method known as
re-substitution.
To solve Ax_ = b_ by re-substitution, we

first apply any direct method, say Gaussian
elimination, to obtain an approximation z_ to x •
785
This vector z^ will be in error by e_ = x - z_ , and
Ae_ = A(x-z) = b - A^ = r ,
which can be computed. (It is necessary to compute
r carefully lest it consist of nothing but the
rounding errors left when ID and A_z nearly cancel.
Double precision accumulation of products is
appropriate here.) Clearly, the error e_ satisfies
an equation similar to x T s except that r replaces
b_ . Therefore, we can approximate e_ by f , say,
obtained by repeating part of the previous calcula-
tion. If enough intermediate results have been
saved during the computation of z_ , one obtains f_
by repeating upon r the operations that transformed
b_ into z_ . The cost of f_ in time and storage is
usually negligible.
Now z_! = z_ + f_ is a better approximation to

x than was £ , provided _z was good enough to
begin with. We shall see that this is so in the case
of Gaussian elimination as follows:
(A + A-jÂ) z. = b , where || A ^ || <_ e||A || .

r = b_ - A;z , say exactly for simplicity.
Il£ll = llA3_Az.ll <. s||A|| |(z.|| ,
and this inequality is not normally a wild over-

estimate.
(A + A2A) f = r , ||A 2 A|| <_ e||A|| .
r = b_ - Az/ = r - Af = A2Af , so
ll£'ll 1 e ||A|| Hill < £ ||A|| ||(A + AgA)- 1 !! ||£||
e K A
1 ( ) llllll/^ 1 - e
K(A)) i f e K(A) < 1 .
And if e K(A) << 1 then ||rf|| can be expected to
be much smaller than || r || . If z_! is renamed z_ ,
the process can be continued. We have left out
several details here; the point is that the process
of re-substitution generally converges to an
approximation z_ which is correct to nearly full
single precision, provided the matrix A is jfarther
from a singular matrix than a few hundred units in
its last place. The problem is to know when to stop.
786
The word "convergence" is well-defined
mathematically in several contexts. But the
empirical meaning of "convergence" is more subtle.
For example, suppose we consider the sequence
Z p Z p , . . ., z_ , ... of successive approximations to
x_ produced by the re-substitution iteration, and
suppose that z = z ,-, = z , 0 = . . . . We should
^ -m -m+1 -m+2
conclude that the sequence has converged. And if
z n - z
—m-1 —m is a good
° deal smaller than
z n - z
—m-2 ., j
—m-1 ' we should incline to the belief that
the convergence of the sequence is not accidental;
there is every
J reason to expect z to be the
^ -m
correct answer x_ except for roundoff in the last
place. But a surprise is possible if A is
exceptionally ill-conditioned:
Example 4. Here is an example of a 2x2
system with
A
= /.8647 .5766) = /.2885 )
ana
1.4322 .2822/ - (.1442/ *
We shall use Gaussian elimination to compute a first
approximation z_ to the solution x of Ax = b_ .
Then r = b_ - Az, is computed exactly, and the
solution e_ of Ae^ = r is approximated by £ ,
obtained again by Gaussian elimination. z_T = z. + f. ,
and rT = b_ - Az_T .
We shall try to calculate x correctly to
3 sig. fig T s. It seems reasonable to carry one
guard digit at first, since we can repeat the
calculation with more figures later if that is not
enough. We shall truncate all calculations to
4 sig. fig f s., just like our 7094 (except that it
truncates to about 8 sig. fig T s.). Intermediate
results enclosed in parentheses are obtained by
definition rather than by means of an arithmetic
operation.
787
Comment Equ'n Coef. Coef. Right hand
no. of x. of x 2 side b
1st pivotal row is . . . El .8647 .5766 .2885
.4327/.8647= .4998 E2 .4322 .2882 .1442
.4998xEl Is . . . El' (.4322) .2881 .1441
E2-E1' .. . E3 ( 0 ) .lxlO"3 .lxlO"3
3
E3/.lxl0~ Z2 ( 0 )( 1 ) 1.000
.5766xZ2 E3' ( 0 ) (.5766) .5766
E1-E3* E4 (.8647) ( 0 ) -.2881
E4/.8647 Zl ( 1 )( 0 ) -.3331
Our first approximat ion is z 333l\ •

• h .000/
Next we compute r = b - Az exactly using double
precision:
-4
Residual of El ... Rl -.6843x10
-4
Residual of E2 ... R2 -.3418x10
We get f_ by repeating upon r the operation which

transformed b into z .
.4988XR1 Rl' -.3420x10 -4

R2-R1' R3 .2xl0"7
R3/.lxlO-3 . F2 .2000x10 -3
.5766xF2 R3' .1153x10 -3
R1-R3' R4 -.1837x10 -3
R4/.8647 Fl -.2124x10 -3
-.2124
z = /-3331) f = .2000 x l 0 ~ 3 , z_
x±u z ' =zti_
33312l|
z + f =( 1#Mo002000/ | *
- (1.000/
Clearly z_! i s so close t o z_ t h a t e i t h e r i s an

a c c e p t a b l e 3 - f i g u r e approximation t o x . But, j u s t
in case t h e r e i s some doubt, we compute
= b - AzT = -.00000008772, >

,00000002072
788
which is reassuringly smaller than r_ .
to 3 sig fi
Is x = (î'Joo) ' g?s? No
> 1= /_1
The only clue to ATs ill-condition is the

cancellation in E3 . If there were time and space
available, an example could be constructed of
sufficiently high dimensionality that severe
cancellation would not occur to warn of the disaster.
Presumably this kind of disaster is rare in practice,
because none has yet been reported elsewhere. Indeed,
a prominent figure in the world of error-analysis has
said
"Anyone unlucky enough to encounter this

sort of calamity has probably already been
run over by a truck."
But being run over by a truck can hardly go

unnoticed.
Despite the risks, re-substitution is the

most reasonable and efficient way to check and
improve the accuracy of an approximation z_ when
the matrices A and b_ are known more precisely
than to within the uncertainties AA and Ab_
in the perturbed equation
(A + AA) z, = b + Ab
satisfied by the product z_ of Gaussian elimination.
For fuller detail, see Wilkinson (1963) pp. 121-126.
But if A and b_ are intrinsically uncertain in,
say, their fourth decimal place, and if Gaussian
Elimination has been carried out with about 6 to 8
sig. figTs and with a reasonable pivotal strategy,
then re-substitution may well be pointless, since
the errors committed during the elimination will be
negligible compared with intrinsic uncertainties.
5. Pivoting and Equilibration. How reliable

are the sizes of the pivots as indications of a
matrixfs ill-condition? Is it true that a matrix is
ill-conditioned if and only if some of its pivots
are small? Most people who are experienced with
hand calculations would answer "yes" to the last
question unless they have tried to test their belief
on problems of high dimensionality with the aid of
an electronic computer. When the dimensionality of a
789
problem becomes large (say > 30), much of our
experience and Intuition with small dimensionality
(say <. 5) becomes misleading. The following
examples are designed to correct misleading
impressions.
MIS-STATEMENT NO. 1. The determinant of A

is the product of the pivots encountered during
Gaussian elimination (to within a * sign); and since
a singular matrix has determinant zero, and det A
is a continuous function of A , an ill-conditioned
(nearly singular) matrix must have a small
determinant and hence must have at least one small
pivot.
The flaw in this argument is the same as that which

says that, since |x| is a continuous function of
x no matter how large N may be, |x| / must be
small when x is small. The trouble is that two
"small" numbers can still be relatively very different.
A matrix counterexample is
Example 5.
f
1 -1 -1 -1
•• •
0 1 -1 -1 a±J = 0 i f i > J ,
0 0 1 ... -1
A = • •• • a 1(J - - 1 i f i < j ,
•
1 -1 &il = 1 .
fi 1
NxN
Here det A = 1 , and every pivot can be 1 but

A can be made singular by subtracting 2 1-N from
all ail Therefore, when N is large A differs
negligibly from a singular matrix, and must be ill-
conditioned. The ill-condition of A is not
"caused" by a large number of nearly equal elements,
as some observers have suggested, because |if all
-lTs in A are replaced by +l f s then A becomes
the well conditioned inverse of
1-10 0
0 1-1 0
0 0 1 0
1 -1
0 1
790
The foregoing examples indicate that Gaussian
elimination is a poor way to determine the rank of
a matrix because a few rounding errors may suffice to
cause none of the pivots to be small despite a
theorem which says that, in the absence of rounding
errors, the rank of a matrix is the same as the
number of non-zero pivots generated during Gaussian
elimination with both row and column pivotal
interchanges to select maximal pivots.
Most other methods for determining rank fare
no better in the face of roundoff. For example, the
Schmidt orthogonalization procedure can be described
in terms of an orthogonal projection of ATs n-th
column a upon the space spanned by the previous
n-1 columns a-,, . . . , a __-, (see Householder (1964)
pp. 6-8 and 134-7). The columns can be interchanged,
if necessary, in order at each stage to maximize the
distance from —n
a to the space
^ of —a.l*
, .....
5 a -, .
—n-1
If this is done, the rank r of A will become
evident when a ,-,, a + ? , •••* and a N all have
distance zero from the space spanned by a-, , a,p, ...,
and a . However, if A is merely nearly singular,
there is no guarantee that any of the distances
mentioned above will be anywhere near as small as
the distance between A and the nearest singular
matrix. Difficulties arise whenever || Av|| << |JA|| [|v||
for some vector v whose components v. can be
ordered in such a way that they steadily decrease in
magnitude to a point where the smallest component is
negligible compared with the largest. The following
example illustrates the phenomenon:
a.. = 0 if i >J
a.. = s-i-1 for i,j = 1,2,
i-1 . «.
a. . = -cs if l
ij
2 2
Here N is large (N > 30) and s + c = 1 .
Since A is upper triangular, the n column of 1
is distant a from the space spanned by the
previous columns. Also, this example has been so
chosen that no column interchanges are needed to
maximize the distance, since a , a^,,, ... , and
a N are all equally distant from the space spanned
791
The
by a,, a 2 , ... j â n i • smallest distance
N-l
a
nn l s a NN = s
* How much sma
H e r > t h a n a„ N
is the distance || AA || between A and the nearest
singular matrix A + AA ?
By examining the vector Av , where
v. := c(l+c) " •"" except v„ = 1 ,
we can show that
||AA||/a NN = 0 ( l / ( l + c ) N " 1 ) as N - « .
- 0(l/(l+/l-aNN2/N)N) .
For fixed a N N = a > 0 , the righthand side tends to

zero like
l/[a exp/-2 (N - 1/3 log a) log a ] as N -> °° .
-N
For fixed N , it is like 2 as a ™ •> 0 . In
other words, A can be closer to singular than
a. by orders of magnitude if N is large.
No simple method is known for computing the

rank of a matrix in the face of roundoff. An
effective but complicated method has been given by
Golub and Kahan (1965).
MIS-STATEMENT NO. 2. The reason for pivotal

interchanges is to prevent incorrect answers caused
by the use of an inaccurate small pivot.
This statement seems reasonable in the light
of the 2x2 example at the end of section 4 of
this paper,, where cancellation in E3 produced a
tiny pivot .0001 whose value consisted almost
entirely of rounding error. And the computed
answer was quite wrong. Post hoc ergo propter hoc.
However, the pivot 2x10"" in example 2 is quite
accurate, yet any answer gained by its use is
likely to be wrong.
Now, what can one mean by the accuracy of a

pivot? The meaning would be clear if the object of
792
our computation were to calculate pivots, but that
is not our object. We wish to satisfy a set of
linear equations, and the computed solution £ can
come very near to satisfying Ax_ = b_ even though
almost all pivots are entirely different from what
they would have been in the absence of roundoff. In
example 4, Rl and R2 are as small as one could
reasonably expect from 4-figure working, and remain
so even when the pivot .0001 is replaced by, say,
.0002 . There are occasions when a small residual
is all that is wanted (see the discussion of
eigenvector calculations below). In such a case,
we must conclude that small pivots have not prevented
a correct answer from being produced. Besides,
pivotal interchanges do not prevent small pivots.
What is the significance of a small pivot?

In the absence of other information, none. For
example, if A is a diagonal matrix
a.. = 10" 1 exactly ,
then the system Ax^ = b_ can be solved trivially

and precisely (in decimal arithmetic) despite tiny
pivots. On the other hand, if we are given a matrix
norm such that all perturbations AA of equal norm
|| A A || are considered equally important, and if we
measure the ill-condition of A in terms of this
norm, then a small pivot tells us that A is ill-
conditioned as follows:
If the rows and columns of A are ordered

•properly to begin with, the process of Gaussian
elimination can be identified with a triangular
factorization
A = LU - E
where L is unit lower triangular(l. . = 0 if i<j and £..,=1),

U is upper triangular (u..=0 if i>j) , and
E is the contribution of roundoff.
(See Wilkinson (1963).)
If partial pivoting is used, |&. .| £ 1 for i >_ j .
If complete pivoting is used, |u. . | <_ |u..| too.
The pivots are the numbers u.. , and
II E|| 1 e ||L || ||U ||
793
where e is comparable with the relative error
associated with one rounding error (e is about
—8
10"" on our machine).
Suppose, now, that some pivot u.. is small.

Then let U+AU differ from U only in that u..
is replaced by zero. Therefore,
A+AA = A+E + LAU = L(U + AU) is singular .
And || A A || is of the order of || L|| ( e||u|| + I ^ J ) .
Since 1 <_ || L || < N for most norms of interest here,

and |f TU II < Ng„ [f"A || in most cases of interest (g„
was the pivot-growth ratio), one perturbation
AA that makes A+AA singular is of the same order of
magnitude as the smallest pivot u.. . And since
|| A A || > /K(A) , the condition number
K(A) > /(N|u..|) . A small pivot implies ill-
condition with respect to the given norm.
The foregoing argument also shows why pivotal

interchanges are necessary. They help to keep ||U||
from growing too large, thereby contributing to
keeping ||E|| small, and this last is what we want.
The error in example 2 when 2x10" is used as a
pivot illustrates the consequences of allowing ||u||
to grow too large. Had any other element of A
been chosen as a pivot, no such error could have
occurred.
MIS-STATEMENT NO. 3. If A T s condition

number K(A) is very large, and if A and b_ are
uncertain by a few units in their last place, then
no numerical method is capable of solving Ax_ = b_
more accurately than to about K(A) units in
x_!s last place.
This statement would be true if the
uncertainties in A , b_ .and x^ were measured in
the norms corresponding to K(A) . The most
appropriate norms for b_ and x_ would be such that
perturbations of equal norm were equally likely or
equally costly or otherwise practically
indistinguishable. But rarely in practice is an
appropriate norm chosen on that basis. Usually one
of the Holder norms is chosen on the basis of
794
convenience, and such norms can be terribly
inappropriate.
Example 6.
Let A = -1 10 " 10 " and b =
—8
with a relative uncertainty of 10" in each
element. In other words, A + AA is acceptable in
place of A provided | Aa.. . | <_ 10" | a.. . | . If any
of the aforementioned Holder norms are used, the
condition number of A is at least 10 because
10
there exists a AÂ with || AÂ || <_ 10" ||Â|| such
that Â+AA is singular. Therefore, when Gaussian
elimination carried out with eight sig. fig.
arithmetic gives no useful answer, one is not
surprised. "The system is ill-conditioned."
However, the true solution is
x =
_7
with a relative error smaller than 10 in each
component no matter how A and b_ are perturbed,
provided only that no element of A or b is
changed by more than 10 of itself. This system
is well conditioned! But not in the usual Holder
norm.
Example 6 can be obtained from example 2 by
a diagonal transformation;
I = DAD with D = diag (105 , 10" 5 , 10~ 5 ) .

This transformation does no more than shift a
decimal point 10 places left or right. If Gaussian
elimination is applied simultaneously to the
matrices A and A , then the results in both
cases will be identical down to the rounding errors
except for the 10 place shifts of decimal point.
But example 2 teaches us not to use 2x10" as a
795
pivot, whereas the most natural pivot in example 6
is the corresponding element 2. Any other pivot
would be far better.
This is where equilibration comes in.
Equilibration consists of diagonal transformations
intended to scale each row and column of A in such
a way that, when Gaussian elimination is applied to
the equilibrated system of equations, the results
are nearly as accurate as possible. In other words,
the system
Ax = b
is replaced by
(RAC) £ = (Rb)
where R and C are diagonal matrices. Then
Gaussian elimination (or any other method) is applied
to the array {RAC, R M to produce an approximation
to 2. anc* hence to x. = ^Z. •
How should R and C be chosen? No one has
published a foolproof method. The closest anyone has
come is in a paper by F.L. Bauer (1963) in which the
R and C which minimize K(RAC) are described in
terms of an eigenvalue and eigenvectors of certain
matrices constructed from A and A*" . But no way
is known to construct R and C without first
knowing A"
There is some doubt whether R and C
should be chosen to minimize K(RAC) . The next
example illustrates the problem; the reader should
write out the matrices involved in extenso for N = 6
to follow the argument.
Let A be the NxN matrix defined in
A
example 3 , and l e t
ri 1. o^-N 0 1-N 01-Nv -,
R = diag (h 9 h , ... y 2 , 2 , 2 ) and
C = diag ( 1 , 2 , . . . , 2 N ~ 3 ^ 2 N ~ 2 , 1) .
-1 T -1
We observe that A 1 = (RA^C) , so k± and A±
are both well-conditioned matrices with elements no
larger than 1.
796
Let u = (0, 0, 0, ... , 0, 1 ) T and note
that
A. = A, + u u T .
Therefore, by a simple computation,
A 2 ~ X = A 1 " 1 - (1 + 2 X ~ N ) A ^ 1 u u T A ^ 1 .
Now recall that when row-pivoting alone is used with

Ap , the result can be an error tantamount to
replacing Ap by A.. , and if no other errors are
committed then one will compute A~ instead of
Ap" . Note that, when the usual Holder-norms are

used,
||A 1 -A 2 ||/||A 1 || = 0(1) and || A ^ - A , , " 1 1 | = 0(1) .
Next observe that either complete or partial

pivoting can be used with RApC ; the result is to
make an error which replaces RApC by RA-.C . But
now
IIRA1C - RA 2 CI1/HRA 1 C|| = 0(2" N ) and
|| ( R A ^ T 1 - (RA2C)-1||/||(RA2C)-1|| = 0 (2"N) .
In other words, despite some hocus-pocus

with shifted binary points, the error made in
applying Gaussian elimination to RApC is the same
as that made with A p , except for a change of
scale. But the former error looks negligible and
affects (R'A9C)~ negligibly, whereas the same
d
_1
errors look disastrous in A 2 and A 2 . And
nowhere is there any ill-conditioned matrix, nor do
any of the matrices look poorly scaled by the
usual criteria!
The moral of the story is that the choice of
R and C should reflect the norms by which the
errors are being appraised. But no one knows yet
precisely how to effect such a choice.
797
REFERENCES
D.W. Barron and H.P.F. Swinnerton-Dyer, Solution
of Simultaneous Linear Equations Using a Magnetic-
Tape Store. The Computer Journal, Vol. 3, (I960)
pp. 28-33.
F.L. Bauer, Optimally Scaled Matrices. Numerische
Math., Vol. 5, (1963) pp. 73-87.
G. Birkhoff, R.S. Varga, and D. Young, Alternating
Direction Implicit Methods. Advances in Computers,
Vol. 3, Academic Press (1962).
E. Bodewig, Matrix Calculus, 2nd ed. North Holland
(1959). (A catalogue of methods and tricks, with
historical asides.)
M.A. Cayless, Solution of Systems of Ordinary
and Partial Differential Equations by Quasi-
Diagonal Matrices. The Computer Journal, Vol. 4,
(1961) pp. 54-61.
M.M. Day, Normed Linear Spaces* Springer (1962).
J. Douglas and J.E. Gunn, A General Formulation
of Alternating Direction Methods, part I.
Numerische Math. Vol. 6, (1965) pp. 428-453.
Dunford and Schwartz, Linear Operators, part I:
General Theory. Interscience (1958).
M. Engeli et. al., Refined Iterative Methods for
Computation of the Solution and the Eigenvalues
of Self-Adjoint Boundary Value Problems.
Mitteilung Nr. 8 aus dem Inst, fur angew. Math,
an dur E.T.H., Zurich! Birkhauser (1959).
D.K. Faddeev and V.N. Faddeeva, Computational
Methods of Linear Algebra, translated from the
Russian by R.C. Williams. W.H. Freeman (1964).
(This text is a useful catalogue, but weak on
error-analysis. A new augmented Russian edition
has appeared.)
G.E. Forsythe and W.R. Wasow, Finite Difference
Methods for Partial Differential Equations.
Wiley (I960). (A detailed text.)
L. Fox, The Numerical Solution of Two-Point
Boundary Problems in Ordinary Differential
Equations. Oxford Univ. Press (1957).
798
L. Fox, Numerical Solution of Ordinary and
Partial Differential Equations. Pergamon
Press (1962) (Ed.). (Based on a Summer School
held in Oxford, Aug.-Sept. 1961.)
L. Fox, An Introduction to Numerical Linear

Algebra. Oxford Univ. Press (1964). (This
is an excellent introduction.)
C F . Gauss, Letter to Gerling, 26 Dec. 1823.

Werke Vol. 9, (1823) pp. 278-281. A translation
by G.E. Forsythe appears in MTAC Vol. 5 (1950),
pp. 255-258.
C F . Gauss, Supplementum ... . Werke, Gôttingen,
Vol. 4, (1826) pp. 55-93.
G. Golub and W. Kahan, Calculating the Singular
Values and Pseudo-Inverse of a Matrix. J. SIAM
Numer. Anal. (B), Vol. 2, (1965) pp. 205-224.
J.E. Gunn, The Solution of Elliptic Difference
Equations by Semi-Explicit Iterative Techniques.
J. SIAM Numer. Anal. Ser. B, Vol. 2, (1965)
pp. 24-45.
A.S. Householder, The Theory of Matrices in
Numerical Analysis. Blaisdell (1964). (An
elegant but terse treatment, including material
on matrix norms which is otherwise hard to
find in Numerical Analysis texts.)
IFIP: Proceedings of the Congress of the
International Federation for Information
Processing, held in New York City, May 24-29,
1965. Spartan Books (1965).
C.G.J. Jacobi, Uber eine neue Auflftsumgsart ... .
Astr. Nachr.Vol. 22, No. 523, (1845) pp. 297-306.
(Reprinted in his Werke Vol. 3, p. 467.)
L.V. Kantorovich and G.P. Akilov, Functional
Analysis in Normed Spaces, translated from
the Russian by D.E. Brown. Pergamon (1964).
R.B. Kellog and J. Spanier, On Optimal
Alternating Direction Parameters for Singular
Matrices. Math, of Comp., Vol. 19, (1965)
pp. 448-451.
799
V.V. Klyuyev and N.I. Kokovkin-Shcherbak, On
the Minimization of the Number of Arithmetic
Operations for the Solution of Linear Algebraic
Systems of Equations. Journal of Computational
Math, and Math. Phys., Vol. 5, (1965) pp. 21-33
(Russian). A translation, by G.J. Tee, is
available as Tech. Repft CS24 from the Computer
Sci. Dep't of Stanford University. (My copy has
mistakes in it which I have not yet sorted out.)
J. Liouville, Sur le développement des fonctions
en series ... . II, J. Math, pures appl. (1),
Vol. 2, (1837) pp. 16-37.
D.W. Martin and G.J. Tee, Iterative Methods for
Linear Equations with Symmetric Positive
Definite Matrix. The Computer Journal Vol. 4,
(1961) pp. 242-254. (An excellent survey.)
W.A. Murray and M.S. Lynn, A Computer-Oriented

Description of the Peaceman-Rachford ADI Method.
The Computer Journal, Vol. 8, (1965) pp. 166-175.
J. von Neumann and H.H. Goldstine, Numerical
Inverting of Matrices of High Order. Bull.
Amer. Math. S o c , Vol. 53, (1947) pp. 1021-1099.
J. von Neumann and H.H. Goldstine, ".... part II"
Proc. Amer. Math. S o c , Vol. 2, (1951) pp. 188-
202.
L.B. Rail, Error in-Digital Computation. Two
volumes (1965) Wiley. (Contains a valuable
bibliography.)
G.D. Smith, Numerical Solution of Partial
Differential Equations. Oxford Univ. Press
(1965). (This is an introductory text.)
E.L. Stiefel, Some Special Methods of Relaxation
Technique appearing in Simultaneous Linear
Equations and the Determination of Eigenvalues.
National Bureau of Standards Applied Math.
Series No. 29 (1953). (A subsequent article
by J.B. Rosser in this same book contains more
details about conjugate gradient methods.)
E.L. Stiefel, Kernel Polynomials in Linear

Algebra and their Applications. in Further
Contributions ..., National Bureau of
Standards Applied Math., Series No. 49 (1958).
800
33. A.M. Turing, Rounding-off Errors in Matrix
Processes. Quart. J. Mech. Appl. Math. 1,
(1948) pp. 287-308.
34. R.S. Varga, Matrix Iterative Analysis. Prentice
Hall (1962). (An important treatise on those
iterative methods most widely used to solve
large boundary-value problems.)
35. J.H. Wilkinson, Rounding Errors in Algebraic

Processes, in Information Processing, (i960)
pp. 44-53. Proceedings of a UNESCO conference
held in Paris in 1959.
36. J.H. Wilkinson, Error Analysis of Direct
Methods of Matrix Inversion. J. Assoc.
Computing Machinery, Vol. 8 (1961) pp. 281-330.
37. J.H. Wilkinson, Rounding Errors in Algebraic
Processes. National Physical Lab. Note on
Applied Science No. 32(1963), Her Majesty's
Stationery Office.
University of Toronto
801

Numerical Linear Algebra

Uploaded by

Copyright:

Available Formats

Numerical Linear Algebra

Uploaded by

Document Information

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Numerical Linear Algebra

Uploaded by

Copyright:

Available Formats

NUMERICAL LINEAR ALGEBRA

Ax = b> (i.e., E.a..x = b.) ,

and the solution of the eigenvalue problem

For the next five years it will be economically

Select an element from the matrix,

Let us count up those arithmetic operations.

This simple process is repeated until there

2. The Time Needed to Solve a Linear System

Although N equals the number of lattice points in

Type of Storage required Time required for

Incidentally, there is no need to find space

3. Iterative Methods for Solving Linear

Hence, the error is bounded by

The practicality of this scheme depends upon

The last consideration is quite important in

One generalization begins with the iteration

is z, t s "residual" in the equation Ax. = b_ .

where P, (w) is a k degree polynomial defined

To simplify the exposition now I shall assume

(CA) n = z XnE(x) for all n

whence comes the following theorem:

This theorem is not applied directly in

Given R and k , one might seek a polynomial

max |P, (w) | over w in R

max |P k (w)| _> r/Lk(0) for w in ^ ( r ) .

Proof. Apply the maximum modulus theorem on

to conclude that this quotient has magnitude at

which shows that, if A = I-B and we know only

Having chosen a polynomial *V( W ) > the

Proof. If |P k (z)| <_ l/Tk(L(0)) on

Now, the zeros \."" of T,(L(z)) include

which can be implemented conveniently and is

z k = x + Tk(L(CA))(zQ - x)/ T k ( L (°))

for all k >_ 0 . One great convenience of this

is the "best" such polynomial for each k , so that

More generally, given a norm for the error

is smaller than for z, . In Gauss-Seidel

Another approach to iterative methods is

4. Errors in the Solution of Linear Systems.

no other vector z_ specified to 4 dec. can have

Linear systems with this kind of pathological

A common vector norm is the Holder norm

||x||p = (Zjlxjl?) 1 ^ for 1 I P < - .

T h i s norm i s e a s i l y shown t o have t h e p r o p e r t i e s

||A|| = max ||Ax|| / | | x | L with s u i t a b l y chosen

oo^oo = max llAxllyHxJI^ = ma X l | ^ | a i J | (max r o w - s u n )

IIAIL,I = m a x iML/yii =max ij i a iji

Ul\E L M2$2 > WE/^rank (A) . )

Bz.T||a max lz.Tï.|/||x|l over

And the Hahn-Banach theorem guarantees that to each

K(A) -= S2L M/l over a l l x ï 0_ and £ ï jO

(fl£ - xJl/fl^l)/(|IlJ|/IlbJ|) * (l/2)/(l(T 8 ) < K(A) .

Let us return to the example again. If the

The pathological behaviour of ill-conditioned

A more nearly modern error-analysis was

WilkinsonTs main result about Gaussian

||AA||< c g N NP 3 " S ||A||

The constants c and p depend upon the details of

In short, JJAAjl is comparable to rounding

So, a small residual