Journal of Research of the National Bureau of Standards Vol. 49, No.

6, December 1952 Research Paper 2379

Methods of Conjugate Gradients for Solving

Linear Systems1
Magnus R. Hestenes 2 and Eduard Stiefel3
An iterative algorithm is given for solving a system Ax=k of n linear equations in n
unknowns. The solution is given in n steps. It is shown that this method is a special
case of a very general method which also includes Gaussian elimination. These general
algorithms are essentially algorithms for finding an n dimensional ellipsoid. Connections
are made with the theory of orthogonal polynomials and continued fractions.

1. Introduction method; (b) the conjugate gradient method presented

in the present monograph.
One of the major problems in machine computa- There are many variations of the elimination
tions is to find an effective method of solving a method, just as there are many variations of the
system of n simultaneous equations in n unknowns, conjugate gradient method here presented. In the
particularly if n is large. There is, of course, no present paper it will be shown that both methods
best method for all problems because the goodness are special cases of a method that we call the method
of a method depends to some extent upon the of conjugate directions. This enables one to com-
particular system to be solved. In judging the pare the two methods from a theoretical point of
goodness of a method for machine computations, one view.
should bear in mind that criteria for a good machine In our opinion, the conjugate gradient method is
method may be different from those for a hand superior to the elimination method as a machine
method. By a hand method, we shall mean one method. Our reasons can be stated as follows:
in which a desk calculator may be used. By a (a) Like the Gauss elimination method, the method
machine method, we shall mean one in which of conjugate gradients gives the solution in n steps if
sequence-controlled machines are used. no rounding-off error occurs.
A machine method should have the following (b) The conjugate gradient method is simpler to
properties: code and requires less storage space.
(1) The method should be simple, composed of a (c) The given matrix is unaltered during the proc-
repetition of elementary routines requiring a mini- ess, so that a maximum of the original data is used.
mum of storage space. The advantage of having many zeros in the matrix
(2) The method should insure rapid convergence is preserved. The method is, therefore, especially
if the number of steps required for the solution is suited to handle linear systems arising from difference
infinite. A method which—if no rounding-off errors equations approximating boundary value problems.
occur—will yield the solution in a finite number of
steps is to be preferred. (d) At each step an estimate of the solution is
(3) The procedure should be stable with respect given, which is an improvement over the one given in
to rounding-off errors. If needed, a subroutine the preceding step.
should be available to insure this stability. It (e) At any step one can start anew by a very
should be possible to diminish rounding-off errors simple device, keeping the estimate last obtained as
by a repetition of the same routine, starting with the initial estimate.
the previous result as the new estimate of the In the present paper, the conjugate gradient rou-
solution. tines are developed for the symmetric and non-
(4) Each step should give information about the symmetric cases. The principal results are described
solution and should yield a new and better estimate in section 3. For most of the theoretical considera-
than the previous one. tions, we restrict ourselves to the positive definite
(5) As many of the original data as possible should symmetric case. No generality is lost thereby. We
be used during each step of the routine. Special deal only with real matrices. The extension to
properties of the given linear system—such as having complex matrices is simple.
many vanishing coefficients—should be preserved. The method of conjugate gradients was developed
(For example, in the Gauss elimination special independently by E. Stiefel of the Institute of Applied
properties of this type may be destroyed.) Mathematics at Zurich and by M. R. Hestenes with
In our opinion there are two methods that best fit the cooperation of J. B. Rosser, G. Forsythe, and
these criteria, namely, (a) the Gauss elimination L. Paige of the Institute for Numerical Analysis,
1
This work was performed on a National Bureau of Standards contract with National Bureau of Standards. The present account
the University of California at Los Angeles, and was sponsored (in part) by the was prepared jointly by M. R. Hestenes and E.
Office of Naval Research, United States Navy.
23 National Bureau of Standards and University of California at Los Angeles.
University of California at Los Angeles, and Eidgenossische Technische
Stiefel during the latter's stay at the National Bureau
Hochschule, Zurich, Switzerland. of Standards. The first papers on this method were
409
 given by E. Stiefel4 and by M. R. Hestenes. 5
Reports all x, then A is said to be nonnegative. If A is sym-
on this 7method were given8 by E. Stiefel6 and J. B. metric, then two vectors x and y are said to be con-
Rosser at a Symposium on August 23—25, 1951. jugate or A-orthogonal if the relation (x,Ay) =
Recently, C. Lanczos 9 developed a closely related (Ax,y)=0 holds. This is an extension of the ortho-
routine based on his earlier paper on eigenvalue gonality relation (x,y)=0.
problem.10 Examples and numerical tests of the By an eigenvalue of a matrix A is meant a number
method have been by R. Hayes, U. Hochstrasser, X such that Ay=\y has a solution y^O, and y is
and M. Stein. called a corresponding eigenvector.
Unless otherwise expressly stated the matrix A,
2. Notations and Terminology with which we are concerned, will be assumed to be
symmetric and positive definite. Clearly no loss of
Throughout the following pages we shall be con- generality is caused thereby from a theoretical point
cerned with the problem of solving a system of linear of view, because the system Ax=k is equivalent to
equations the system Bx=l, where B=A*A, l=A*k. From a
numerical point of view, the two systems are differ-
ent, because of rounding-off errors that occur in
+a2nxn=k2 joining the product A*A. Our applications to the
(2:1) nonsymmetric case do not involve the computation
of A*A.
In the sequel we shall not have occasion to refer to
a particular coordinate of a vector. Accordingly
we may use subscripts to distinguish vectors instead
These equations will be written in the vector form of components. Thus x0 will denote the vector
Ax=k. Here A is the matrix of coefficients (atj), (a;oi, • • ., Xon) and xt the vector (xn, . . ., xin). In case
x and k are the vectors (xi,. . .,#„) and (Jcu. . .,&„).l a symbol is to be interpreted as a component, we shall
It is assumed that A is nonsingular. Its inverse A~ call attention to this fact unless the interpretation is
therefore exists. We denote the transpose of A by evident from the context.
A*. The solution oj the system Ax=k will be denoted by
Given two vectors x=(xi,. . .,#„) and y = h; that is, Ah—k. If x is an estimate of A, the differ-
(Viy - -yUn), their sum x-\-y is the vector ence r=k~Ax will be called the residual of x as an
Oi + 2/i,. . .,xn+yn), and ax is the vector (axu. . .,az»), estimate of h. The quantity \r\2 will be called the
where a is a scalar. The sum sauared residual. The vector h—x will be called the
error vector of x, as an estimate of h.
(x,y)=xiyi+x2y2+. . .+xnyn
3. Method of Conjugate Gradients (cg-
is their scalar product. The length of x will be denoted
Method)
by
M=G*M-. . .+£)*=(*,*)*. The present section will be devoted to a description
of a method of solving a system of linear equations
The Cauchy-Schwarz inequality states that for all Ax=k. This method will be called the conjugate
gradient method or, more briefly, the cg-method, for
(x,y)*<(x,x)(y,y) or \(x,y)\< \x\ \y\. (2:2)
reasons which will unfold from the theory developed
in later sections. For the moment, we shall limit
The matrix A and its transpose A* satisfy the ourselves to collecting in one place the basic formulas
relation upon which the method is based and to describing
briefly how these formulas are used.
The cg-method is an iterative method which
terminates in at most n steps if no rounding-off
If aij=aji) that is, if A=A*, then A is said to be errors are encountered. Starting with an initial
symmetric. A matrix A is said to be positive definite estimate x0 of the solution h, one determines succes-
in case (x,Ax)^>0 whenever z^O. If (x,Ax)^O for sively new estimates x0, Xi, x2, . . . of h, the estimate
Xi being closer to h than xw. At each step the
4
E. Stiefel, Uebereinige Methoden der Relaxationsrechnung, Z. angew. Math.
residual rt=k—Axt is computed. Normally this
Physik (3) (1952). vector can be used as a measure of the "goodness"
s M. R. Hestenes, Iterative methods for solving linear equations, N A M L of the estimate xt. However, this measure is not a
Report 52-9, National Bureau of Standards (1951).
6
E. Stiefel, Some special methods of relaxation techniques, to appear in the
Proceedings of the symposium (see footnote 8).
reliable one because, as will be seen in section 18,
7
J. B. Rosser, Rapidly converging iterative methods for solving linear equa- it is possible to construct cases in which the squared
tions, to appear in the Proceedings of the symposium (see footnote 8).
s Symposium on simultaneous linear equations and the determination of eigen- residual \rt\2 increases at each step (except for the
values, Institute for Numerical Analysis, National Bureau of Standards, held on
the campus of the University of California at Los Angeles (August 23-25,1951).
last) while the length of the error vector \h—xt\
9 C. Lanczos, Solution of systems of linear equations by minimized iterations, decreases monotonically. If no rounding-off error
N A M L Report 52-13, National Bureau of Standards (1951).
10 C. Lanczos, An iteration method for the solution of the eigenvalue problem of is encountered, one will reach an estimate xm(m^n)
linear differential and integral operators, J. Research NBS 45, 255 (1950) RP2133; at which r w =0. This estimate is the desired solu-
Proceedings of Second Symposium on Large-Scale Digital Calculating
Machinery, Cambridge, Mass., 1951, pages 164-206. tion h. Normally, m=n. However, since rounding-
410
off errors always occur except under very unusual Ax=k' . (3:4)
circumstances, the estimate xn in general will not be
the solution h but will be a good approximation of h. can be obtained by the formula
If the residual rn is too large, one may continue
with the iteration to obtain better estimates of h. _^i (puk')
Our experience indicates that frequently xn+i is X (3:5)
considerably better than xn. One should not con-
tinue too far beyond xn but should start anew It follows that, if we denote by ptj the jth component
with the last estimate obtained as the initial of pi} then
estimate, so as to diminish the effects of rounding-
off errors. As a matter of fact one can start anew
at any step one chooses. This flexibility is one of the y
principal advantages of the method. t=o (pi, Apt)
In case the matrix A is symmetric and positive is the element in the jth row and kth. column of the
definite, the following formulas are used in the con- inverse A~l of A,
jugate gradient method: There are two objections to the use of formula
(3:5). First, contrary to the procedure of the
Po=rQ—k—Axo (x0 arbitrary) (3:1a) general routine (3:1), this would require the storage
of the vectors p0, p\, . . . . This is impractical,
(3:1b) particularly in large systems. Second, the results
a t
- (puApt)' obtained by this method are much more influenced
by rounding-off errors than those obtained by the
(3:1c) step-by-step routine (3:1).
In the cg-method the error vector h—xis diminished
in length at each step. The quantity J(x) = (h—x,
A (h—x)), called the error function, is also diminished
at each step. But the squared residual \r\2=\k—Ax\2
normally oscillates and may even increase. There
is a modification of the cg-method where all three
quantities diminish at each step. This modification
is given in section 7. It has an advantage and a
In place of the formulas (3:1b) and (3-le) one may disadvantage. Its disadvantage is that the error
use vector in each step is longer than in the original
method. Moreover, the computation is complicated,
(Pt,rt) since it is a routine superimposed upon the original
(3:2a) one. However, in the special case where the given
linear equation system arises from a difference
approximation of a boundary-value problem, it can
(3:2b) be shown that the estimates are smoother in the
modified method than in the original. This may be
Although these formulas are slightly more compli- an advantage if the desired solution is to be differ-
cated than those given in (3:1), they have the ad- entiated afterwards.
vantage that scale factors (introduced to increase Concurrently with the solution of a given linear
accuracy) are more easily changed during the course system, characteristic roots of its matrix may be
of the computation. obtained: compute the values of the polynomials
The conjugate gradient method (cg-method) is RQ, Ri, . . . and Po, Pu . . . at X by the iteration
given by the following steps:
Initial step: Select an estimate x0 of h and com-
pute the residual r0 and the direction p0 by formulas
(3:1a).
General routine: Having determined the estimate
xt of h, the residual rt, and the direction pt, compute P —R A-b-P- (3*6)
xi+1, r^+i, and pi+x by formulas (3:1b), . . ., (3:If)
successively.
As will be seen in section 5, the residuals r0, ru The last polynomial Rm[X) is a factor of the charac-
. . . are mutually orthogonal, and the direction vec- teristic polynomial of A and coincides with it when
tors p0, pi, . . . are mutually conjugate, that is, m=n. The characteristic roots, which are the zeros
of i?m(\), can be found by Newton's methods without
(rt, r,)=0, (pi9APj)=0 (3:3) actually computing the polynomial Bm(\) itself.
One uses the formulas
These relations can be used as checks.
Once one has obtained the set of n mutually (3:7)
conjugate vectors p0, . . ., -pn-i the solution of
411
where Rm(\k), R'm(\n) are determined by the iteration Next select a direction pi+l conjugate to p0,- Pi,
(3:6) and, that is, such that
R'o=P'o=0
(4:2)
R'i+i=R'i-'KaiP'(-aiPt
In a sense the cd-method is not precise, in that no
formulas are given for the computation of the direc-
tions pOj pi, . . . . Various formulas can be given,
with \=\k. In this connection, it is of interest to each leading to a special method. The formula
observe that if m=n, the determinant of A is given (3: If) leads to the cg-method. It will be seen in
by the formula section 12 that the case in which the p's are obtained
by an A-orthogonalization of the basic vectors
(1,0, . . ., 0), (0,1,0, . . . ) , . . . leads essentially to
det 01) = the Gauss elimination method.
The basic properties of the cd-method are given by
The cg-method can be extended to the case in the following theorems.
which A is a general nonsymmetric and nonsingular Theorem 4 : 1 . The direction vectors p0, pi, • • • are
matrix. In this case one replaces eq (3:1) by the set mutually conjugate. The residual vector rt is orthogonal
to po, Pu - • •, Pi-i- The inner product of pt with each
ro=k—Axo, po=A*ro, of the residuals r0, ru • • - , / • < is the same. That is,

(4:3a)
(4:3b)

(3:8) (4:3c)
The scalar at can be given by the formula

(4:4)

pi+1=A*ri+1- in place of (4:1a).

Equation (4:3a) follows from (4:2). Using (4: lc),
This system is discussed in section 10. we find that

4. Method of Conjugate Directions (cd- = (PJSJC)—ak(pj,Apk).

Method)1 If j=k we have, by (4:1a), (pk,rk+l) = 0. Moreover,
The cg-method can be considered as a special case and(4:3a)
by
(4:3c)
(p3,rk+1) = (pj,rk), (j'^k). Equations (4:3b)
follow from these relations. The formula
of a general method, which we shall call the method (4:4) follows from (4:3c) and (4:1a).
of conjugate directions or more briefly the cd-method. As a consequence of (4:4) the estimates Xi,x2, • • •
In this method, the vectors p0, pl7 . . . are selected of h can be computed without computing the resid-
to be mutually conjugate but have no further restric- uals r ,ri, • • •, provided that the choice of the
tions. It consists of the following routine: o
direction
Initial step. Select an estimate x0 of h (the solu- residuals. vectors p o ,pi, - • • is independent of these
tion), compute the residual ro=k—Axo, and choose a Theorem 4:2. The cd-method is an mstep method
direction p0. (m ^ n) in the sense that at the mth step the estimate xm
General routine. Having obtained the estimate is the desired solution h.
Xt of h, the residual rt=k—Axt and the direction pt, m be the first integer such that yo=h—xoris
compute the new estimate xi+1 and its residual ri+i in the let
For
subspace spanned by p0, • • •, pm-\. Clearly,
by the formulas m ^ n, since the vectors po,pi, • • • are linearly inde-
pendent. We may, accordingly, choose scalars
a1 = (4:1a) <*o> ' ' '•> am-i such t h a t
(
-CLm-lVm-\-
(4:1b)
Hence,
rl+1=ri—aiApi. (4:1c) h=xo+aopo+ '
1
This method was presented from a different point of view by Fox, Huskey, and Moreover,
Wilkinson on p. 149 of a paper entitled "Notes on the solution of algebraic linear
simultaneous equations," Quarterly Journal of Mechanics and Applied Mathe-
matics c. 2, 149-173 (1948). ro=k—Axo=A(h—xo) = (
412
Computing the inner product (Pi,r0) we find by (4:3a) Geometrically, the equation f(x) = const, defines
and (4:4) that at=aiy and hence that h=xm as was an ellipsoid of dimension n—1. The point at which
to be proved. f(x) has a minimum is the center of the ellipsoid and
The cd-method can be looked upon as a relaxation is the solution of Ax=k. The i-dimensional plane
method. In order to establish this result, we intro- Pi, described in the last theorem, cuts the ellipsoid
duce the function f(x)=f(x0) in an ellipsoid Et of dimension i—1,
unless Et is the point x0 itself. (In the cg-method,
,k). (4:5) Ei is never degenerate, unless xo=h.) The point xt is
the center of Et. Hence we have the corollary:
Clearly, f(x)^0 and f(x)=0 if, and only if, x=h. Corollary 1. The point xt is the center of the
The function f(x) can be used as a measure of the (i-l)-dimensional ellipsoid in which the i-dimensional
"goodness" of x as an estimate of h. Since it plays plane Pt intersects the (n— 1)-dimensional ellipsoid
an important role in our considerations, it will be f(x)=f(x0).
referred to as the error function. If p is a direction Although the function f{x) is the fundamental
vector, we have the useful relation error function which decreases at each step of the
relaxation, one is unable to compute f(x) without
(4:6) knowing the solution h we are seeking. In order to
obtain an estimate of the magnitude oif(x) we may
where r=k—Ax=A(h—x), as one readily verifies use the following:
by substitution. Considered as a function of a, Theorem 4:4. The error vector y=h—x, the residual
the function f(x-\-ap) has a minimum value at r=k—Ax, and the error function f(x) satisfy the
a=a, where relations
(4:11)

This minimum value differs from/(x) by the quantity where n(z) is the Rayleigh quotient
2
(v r)
J\X) j\X\~ap)==za \jp,^x.pj — ~/ ~i r* (^4 . o) (4:12)
KP>Ap)

Comparing (4:7) with (4:1a), we obtain the first two The Rayleigh quotient of the error vector y does not
sentences of the following result: exceed that of the residual r, that is,
Theorem 4:3. The point xt minimizes f(x) on the
line x=Xi-i + api_i. At the i-th step the error f(xt_i) (4:13)
is relaxed by the amount Moreover,
(4:14)
fixi-j-fixi)^^-1'^2,- (4:9)
The proof of this result is based on the Schwarzian
In fact, the point xt minimizes f(x) on the i-dimensional quotients { (Az,A2z)
plane Pt of points {
Lb)
(z,z) = (z,Az) =(Az,Az)' ^'
(4:10) The first of these follows from the inequality of
where a0, ..., at_i are parameters. This plane con-
Schwarz
tains the points xQ,X\,..., xt.
(4:16)
In view of this result the cd-method is a method by choosing p="Z, q=Az. The second is obtained
of relaxation of the error function/(x). An iteration by selecting p=Bz, q=Bzz, where B2=A.
of the routine may accordingly be referred to as In order to prove theorem 4:4 recall that if we
a relaxation. set y=h—x, then
In order to prove the third sentence of the theorem
observe that at the point (4:10) r=k—Ax=A(h—x)=Ay
f(x)=f(xo)-TJ tta&j^ f(x)=(y,Ay)
j=o
by (4:5). Using the inequalities (4:15) with z=y,
At the minimum point we have we see that
^(Ay,A*y)
=
(y,y) (y.Ay) -f(x)=(Ay,Ay)
and hence a.j=ah by (4:4). The minimum point is _(r,Ar)_
accordingly the point Xt, as was to be proved. (r,r) '
413
This yields (4:11) and (4:13). Using (4:16) with In view of theorem 4:6, it is seen that at each step
p=y and q=r we find that of the cd-routine the dimensionality of the space TT*
in which we seek the solution h is reduced by unity.
=(y,r)^\y\ \r\ Beginning with x0, we select an arbitrary chord Co of
Hence f(x) =f(xo) through x0 and find its center. The plane
TTI through Xi conjugate to Co contains the centers of
all chords parallel to Co. In the next step we re-
so that the second iuquality in (4:14) holds. The strict ourselves to TTI and select an arbitrary chord
first inequality is obtained from the relations Ci of f(x) =f(xi) through x1 and find its midpoint
x2 and the plane TT2 in 7rt conjugate to C\ (and hence
to Co). This process when repeated will yield the
answer in at most n steps. In the cg-method the
chord d of f(x) =f(xt) is chosen to be the normal
As is to be expected, any cd-method has 1within at x^
its routine a determination of the inverse A~ of A.
We have, in fact, the following: 5. Basic Relations in the cg-Method
Theorem 4:5. Let p0, . . ., pn-i be n mutually con-
jugate nonzero vectors and let p^ be the j-th component Recall that in the cg-method the following formulas
of pi. The element in the j-th row and k-th column of are used
A"1 is given by the sum Po=ro=k—Ax0 (5:1a)
n-l „1_ H 2 (5:1b)
(p,,APij
This result follows from the formula xm=xi+atpi
(5:Id)

b,=
for the solution h of Ax=k, obtained by selecting
xo=O.
We conclude this section with the following: (5: If)
Theorem 4:6. Let x* be the (n—i)-dimensional
plane through x{ conjugate to the vectors p0, plf . . ., One should verify that eq. (5:le) and (5: If) hold for
Pi-i. The plane TT* contains the points xt, xi+i, . . . ^=0,1,2,. . . if, and only if,
and intersects the (n—1)-dimensional ellipsoid fix) =
f(Xi) in an ellipsoid E[ of dimension (n—i—1).
The center of El is the solution h of Ax=k. The point (5:2)
xi+i is the midpoint of the chord d of E[ through xu
which is parallel to pt. In the eg-method the chord d
is normal to E[ at xt and hence is in the direction of The present section is devoted to consequences of
the gradient of f(x) at xf in Tt. these formulas. As a first result, we have
The last statement will be established at the end Theorem 5:1. The residuals r0, ru . . . and the
of section 6. The equations of the plane -KI is given direction vectors p0, pu . . . generated by (5:1) satisfy
by the system the relations

(Apj,x-xi) =0 0'=0,l,. . . , i - l ) . (5:3a)

Since Pi,pt+i,. . . are conjugate to pOj. . .,Pi-i, so also (pi,Apj)=0 (5:3b)
is
xk—xi=aipi+. . .+a*_i2?*-i (*>i). (5:3c)
The points Xi,xi+i,. . .,xm=h are accordingly in wi}
and h is the centerJ of E't. The chord d is defined by i), (rt,Apj)=0 (5:3d)
the equation x=xi rtaipi, where tis a parameter. As
is easily seen, The residuals r0, r1? . . . are mutually orthogonal and
the direction vectors Po, Pi, . . . are mutually conju-
t) =f(xi) - (2t- gate.
The proof of this result will be made by induction.
The second endpoint of the chord d is the point The vectors r0, p0 =r0 and r\ satisfy these relations
i
Xi~{-2aiPi at which / = 2 . The midpoint corresponds since
to t=l, and hence is the point xi+i as was to be
proved. = (po/0 = ko|2—<h(ro,Apo) —0
414
by (5:1b). Suppose that (5:3) holds for the vectors we accordingly suppose that the routine terminates
r0, . . ., rk and p0, . . ., pk-i. To show that pk can at the n-ih. step. Since the rt is orthogonal to
be adjoined to this set it is necessary to show that ri+i we have xt 9^xi+i and hence at 5^0. I t follows that
(Pift) ^ 0 by (4:1 a). We may accordingly suppose the
(5:4a) vectors^ihave been normalized so that (PiJri) = \ri\2.
In view of (4:3b) and (4:3c) eq (5:3c) holds. Select
(5:4b) numbers atj such that
(rk}Api) = (pk,Api) (i^kji^k—1) (5:4c)
Equation (5:4a) follows at once from (5:2) and
(5:3a). To prove (5:4b) we use (5: Id) and find that Taking the inner product of pt with r? it is seen by
(5:3c) that
(ri+1,pk) = (rupk) —at{Apupk).
By (5:4a) this becomes W! Ki)-
Consequently, (5:2) holds and the theorem is estab-
lished.
Since <x<>0, eq (5:4b) holds. In order to establish Theorem 5:3. The residual vectors r 0 , ru . . . and
(5:4c), we use (5:If) to obtain the direction vectors p0, pu . . . satisfy the further
relations
(pk,Ap%) = = (rk,Apt)
\rj\2\pi 2

(iS J) (5:6a)
\ri\2

It follows that (5:4c) holds and hence that (5:3) 12

holds for the vectors r0, T\ . . ., rk and p0, piy . . .,
P* (5:6b)
It remains to show that rk+l can be adjoined to this
set. This will be done by showing that >1 (5:6c)
(/•«,r*+i)=0 (i^k) (5:5a) -6?_i(Pi_i

(Api,rk+1)=0 (i<k) (5:5b) (5:6d)

(p,,r*+i)=0 (i^k). (5:5c) y than pt. The vector pt makes

an acute angle with pj
By (5:Id) we have The relations (5:6a) and (5:6b) follow readily from
(5:le), (5:lf), (5:2), and (5:3). Using (5:it) and
(5:3d), we see that
(r< A+i) = (rt,rk) —ak(ruApk).
If i<ik, the terms on the right are zero and (5:5a)
holds. If i=k, the right member is zero by (5:1b) Hence (5:6c) holds. Equation (5:6d) is a conse-
and (5:3d). Using (5: Id) again we have with i<Ck quence of (5: If) and (5:3b). The final statements
are interpretations of formula (5:6b) and (5:6a).
= {rk+ur%) —at(rk+uApt) =-• — at(rk+uAp%) . Theorem 5 : 4 . The d i r e c t i o n vectors pQ, p u . . .
satisfy the r e l a t i o n s
Hence (5:5b) holds. The equation (5:5c) follows
from (5:5a) and the formula (5:2) for pt. (5:7a)
As a consequence of (5: 3b) we have the first two
sentences of the following: (5:7b)
Theorem 5:2. The cg-method is a cd-method. It is
the special case of the cd-method in which the pt are Similarly, the residuals r0, ru . . . satisfy the relations
obtained by A-orthogonalization of the residual vectors
TV On the other hand, a cd-method in which the resid- r1=rQ—a0ArQ (5:8a)
uals r0, rit . . . are mutually orthogonal is essentially
a cg-method. (5:8b)
The term " essentially" is used to designate that
we disregard iterations that terminate in fewer than where
n steps, unless one adds the natural assumption that
the formula for pt in the routine depends continu- 6*-i=- (5:9)
ously on the initial estimate x0. T6 prove this result
415
 Equation (5:7b) is obtained by eliminating ri+x Theorem 6:1. The estimates xQ, xx, • • •, xm of h are
and Ti from the equations distinct. The point xt minimizes the error function
f(x) = (h—x,A(h—x)) on the i-dimensional plane Pt
passing through the points x0, xly • • •, xt. In the iih
step of the cg-method, f{x) is diminished by the amount
ri+l=ri —

where fx(z) is the Rayleigh quotient (4:12). Hence,

Equation (5:7a) follows similarly. In order to prove
(5:8b), eliminate Apt and Apt_i from the equations f^i)—J(xj)=ai\ri\2+ • • • +a,-i|r,_i| 2 <3)
(6:2)
The point xt is given by the formulas

(6:3)
j=o

This result is essentially a restatement of theorem

Equation (5:8a) holds since po=r0. 4:3. The formula (6:3) follows from (5:2) and
Theorem 5:5. The scalars at and b{ are given by the (6:2). The second equation in (6:1) is readily
several formulas verified.
Theorem 6:2. Let St be the convex closure of the
(5:10) estimates x0, Xi, • • •, xt. The point xt is the point in
Sf, whose error vector h—x is the shortest.
For a point x^Xi in St is expressible in the form
bt= (5:11)

The scalar at satisfies the relation where ai ^ 0 , a0 •Jrai=l.

, (5:12) We have accordingly

a0
Xi dj — OL§\d,i XojT
where fx(z) is the Rayleigh quotient (4:12). The recip-
rocal of at lies between the smallest and largest char-
acteristic roots of A.
The formula (5:10) follows from (5:1b) and (5:3c),
while (5:11) follows from (5:le), (5:If), (5:3b), and where the /3's are nonnegative. Inasmuch as all
(5:3d). Since ( ^ it follows that

by (5:6b) and (5:6d), we have Using the relation

(Pi, i9Art) xj—x\2=\Xj—xi\2Jr2(xj—xt, xt—x)-\-\xt—x\2 (6:4)

2
\Pi\ N we find that
The inequalities (5:12) accordingly hold. The last
statement is immediate, since n(z) lies between the
smallest and largest characteristic roots of A.
Setting j=m, we obtain theorem 6:2.
6. Properties of the Estimates x, of h in the Incidentally, we have established the
Corollary. The point xt is the point in St nearest to
cg-Method the point Xj (j^>i).
Theorem 6:3. At each step of the cg-algorithm the
Let now x0, X\, . . ., xm=h be the estimates of h error vector yi=h—Xi is reduced in length. In fact,
obtained by applying the cg-method. Let r0, rx,
• • ., rm=0 be the corresponding residuals and p0,
Pi, . . ., pm-i the direction vectors used. The pres- (6:5)
ent section will be devoted to the study of the prop-
erties of the points xQ, Xi, . . ., xm. As a first result
we have where JJL(Z) is the Rayleigh quotient (4:12).
416
 In order to establish (6:5) observe that, by (5:6a), is the point in Pi whose distance from the solution h is
the least. It lies on the line x^iXi beyond xt. Moreover,

(6.9)
and
(6.10)
In view of (6:2) and (5:1b) this becomes
In order to establish (6:9) and 6:10) we use the
( I)= (6:6) formula
^-*<- MS)-
Setting x=xi_l and j=m in (6:4), we obtain (6:5)
by the use of (6:6) and (6:1). which holds for all values of a in view of the fact that
This result establishes the cg-method as a method Xt minimizes f(x) on Pt. Setting «=/(#*)/|ri | 2
of successive approximations and justifies the pro- we have
cedure of stopping the algorithm before the final-
step is reached. If this is done, the estimate ob-
tained can be improved by using the results given
in the next two theorems.
Theorem 6:4. Let x$1: • • •, x£ be the projec- An algebraic reduction yields (6:9). Inasmuch as
tions of the points xi+1, • • • , xm=h in the i-
dimensional plane Pt passing through the points xQ,
# ! , • • • , xt. The points xt_u xu xQi, • • •, x^ lie |^ — %i\2— \h — X^\2= r—
on a straight 0line in the order given by their enumeration.
The point as* (k^>i) is given by the formulas
_i) — f(Xi) |r<_i
X fc — X i __ i " (6:7a)
we obtain (6:10) from (6:9) and (5:1b).
As a further result we have
x (i) _ _ x .
(6:7b) Theorem 6:6. Let x[, . . ., x'm_xbe the projections
of the points xx, . . ., x'm^x on the line joining the initial
point x0 to the solution xm=h. The points x0, x[, . . .,
In order to prove this result, it is sufficient to x'm-i, xm=h lie in the order of enumeration.
establish (6:7). To this end observe first that the Thus, it is seen that we proceed towards the solu-
vector tion without oscillation. To prove this fact we need
only observe that

\Xm X(),Xi Xi-i) = {Xm ~XQ,Q/i — i,Pi — i)

is orthogonal to each of the vectors p0, px, • • •,
Pi-\. This can be seen by forming the inner product m-1
with pi(l<Ci), and using (5:6a). The result is = at-i X) a>j(Pj,Pi-i)>0
j=o

\PhPj)— I
\ri-l\
12 \PhPi-V = by (5: 6a). A similar result holds for the line joining
The projection of the point Let -KI be the (n—i)-dimensional plane through xt
conjugate to p0, Pi, . . -, Pi-i- It consists of the set
of points x satisfying the equation
in P^ is accordingly
0"=0,l,
~«) ^ . a*_ 1 |r ? :_ 1 | 2 +.
Pi-i- This plane contains the points xi+i, .,xm and hence
the solution h.
Using (6:2), we obtain (6:7). The points lie in the Theorem 6:7. The gradient of the function fix) at xt
designated order, since f(xk) y>f(xk+1). in the plane Tr* is a scalar multiple of the vector pt.
Since f(xm) = 0, we have the first part of The gradient of/(as) at xt is the vector — rt. The
gradient qt oif(x) at xt in TTI is the orthogonal projec-
Theorem 6:5. T/i6 point tion of — Ti in the plane ?rz. Hence qt is of the form
(6:8) qi=—ri+aoAp{)+ . . . + ai_lApi-u
417
where a0, . . .,a*_i are chosen so that qt is orthogonal x
12 i

to ApQ, . . .jApi-i. Since *—T~SL 12 - (7:3b)

(7:3c)

rj+1=rJ—ajApj 1), The sum of the coefficients of x0, xu . . ., it in (7:3b)

(and hence of ro,ru . . ., rt in (7:3c)) is unity.
it is seen upon elimination of r0, /*i, . . ., /Vx succes-
sively that p* is also a linear combination of rt, Ap0, The relation (7:1) can be established by induction.
. . ,,Api_i. Inasmuch as pt is conjugate to p0, . . ., It holds for i = 0 . If it holds for i, then
Pi-u it is orthogonal to Ap0, . . ., Api_x. The vector
Pi accordingly is a scalar multiple of the gradient qf
off(x) at Xi in Tt, as was to be proved.
In view of the result obtained in theorem 6:7 it is
seen that the name "method of conjugate gradients" The formula (7:3a) follows from (7:2a), (5:le) and
is an appropriate name for the method given in sec- (5:6b). Formula (7:3b) is an easy consequence of
tion 3. In the first step the relaxation is made in the (7:2b). To prove (7:3c) one can use (5:2) or (7:3b),
direction p0 of the gradient of f(x) at x0, obtaining as one wishes. The final statement is a consequence
a minimum value of fix) at Xi. Since the solution h of (7:3a).
lies in TTI, it is sufficient to restrict x to the plane TTI. Theorem 7:2. The point xt given by (7:2) lies in
Accordingly, in the next step, we relax in the direc- the convex closure Si of the points x0, x1} • • •, xt. It is
tion pi of the gradient oifix) in xi at xu obtaining the the point x in the i-dimensional plane Pf through x0,
point x2 at which fix) is least. The problem is then xu - - -, xt at which the squared residual \k—Ax\2 has
reduced to relaxing fix) in the plane TT2, conjugate to its minimum value. This minimum value is given by
p0 and pi. At the next step the gradient in x2 in ir2 the formula
is used, and so on. The dimensionality of the space in
which the relaxation is to take place is reduced by
unity at each step, Accordingly, after at most n .= | 2 _KL 2 _ N 4 (7:4)
steps, the desired solution is attained.
The squared residuals |ro|2, \rt\2, • • • diminish mono-
7. Properties of the Estimates x, of h in the tonically during the cg-method. At the ith step the
cg-Method squared residual is reduced by the amount

In the cg-method there is a second set of estimates 19 *~' 9 (7:5)

#o— #o> #i> ^2? . . . of h that can be computed, and
that are of significance in application to linear
systems arising from difference equations approxi- The first statement follows from. (7:3b), since the
mating boundary-valve problems. In these applica- coefficients of x0, xu • • •, xt are positive and have
tions, the function defined by xt is smoother than unity as their sum. In order to show__ that the
that of xiy and from this point of view is a better squared residual has a minimum, on P at xit observe
approximation of the solution h. The point xt has that a point x in Pt differs from Xi byt a vector zt of
its residual proportional to the conjugate gradient pt. the form
The points Jo, xu x2, . . . can be computed by the
iteration (7:2) given in the following:
Theorem 7:1. The conjugate gradient pt is expressible
in the form
The residual r=k—Ax is^accordingly given by
pi=d{k-Axi)J (7:1)
r==rt—Azt
where d and Xi are defined by the recursion formulas

Co=l, c<+1= (7:2a)

Inasmuch as, by (7:3c), ri=pi/ci,'we have
(7:2b)

We have the relations

Consequently, (rt,Azl)=0 and
(7:3a)
W rr=
418
It follows that xt affords a proper minimum to \r\2 on that is,
Pi. Using (7:3c) and (7:3a) and the orthogonality
of r/s, it is seen that the minimum value of on Mt) -j{*i) = 6?-i (7:11)
Pi is given by (7:4). By (7:4) and (7:2a)
By (7:9) it is seen that this result can be put in the
2 form
/ • < - • .
I—
jf
k<l
This completes the proof of theorem 7:1. Since xo=xo this formula yields the desired relation
Theorem 7:3. The Rayleigh quotients of r0, rl, (7:7). Since 6i_i<l, it follows from (7:11) and
. . . and r0, ?i, . . . are connected by the formulas (7:7) that (7:8) holds. This proves the theorem.
Theorem 7:5. The error vector yi—h—Xi is shorter
_ix{r%) . jifo than the error vector yi=h—Xi. Moreover, yt is shorter
— ITT12 "I I- (7:6a) than i/i_i.
The first statement follows from (7:2). It also
follows from theorem 6:2, since xt is in St. By
(7:6b) (7:2) the point xt lies in the line segment joining
xt to ~x~i-\. The distance from h to xt exceeds the
distance from h to !ct. It follows that as we move
Rayleigh quotient of rt(i^>0) is smaller than that from Xi to xt-i the distance from h is increased, as
of r{, that is, M(P<) </*(*•«). was to be proved.
In order to prove this result we use (5:6d) and
obtain
8. Propagation of Rounding-Off Errors in
the cg-Method
In this section we take as basic relations between
Since |/'t| 4 =|^i| 2 ki| 2 &nd }x{p^ = ix(r^, this relation the vectors r0, rlf • • • and p0, ph • • • in the cg-
yields (7:6a). t h e eq (7:6b) follows from (7:6a). method the following:
The last statement follows from (5:12). ^o=^o, (8:1a)
In the applications to linear systems arising from
difference equations approximating boundary value |ri|2
problems, n(rt) can be taken as a measure of the a*=7 (8:1b)
smoothness oi_xt. The smaller /*(/•*) is, the smoother
Xi is. Hence xt is smoother than xt. (8:1c)
Theorem 7:4. At the point xt the error function
f(x) has the value
(8: Id)
(7:7)
and we have pt+i=r{+l+btpt.
(7:8)
As a consequence, we have the orthogonality
The sequence f(x0), f(xi), f(x2), • . . decreases mono- relations
tonically.
In order to prove this result, it is convenient to set
(rt,rk)=0, (APi,Pk)=0 (8:2)
2
ri Ci _ 1 ri
?\_ ! 2
(7:9) Because of rounding-ofT errors during a numerical
Ci
calculation (routine), these relations will not be
By (7:2) we have the relation satisfied exactly. As the difference \k—i\ increases,
the error in (8:2) may increase so rapidly that xn
(7:10) will not be as good an estimate of h as desired. This
error can be lessened in two ways: first, by intro-
Using the formula ducing a subsidiary calculation to reduce rounding-
off errors; and second, by repeating the iteration so
as to obtain a new estimate.. This section will be
concerned with the first of these and with a study of
which holds for any x in P f , we see that the propagation of the rounding-off errors. To this
end it is convenient to divide the section in four parts,
the first of which is the following:
419
 8.1. Basic propagation formulas by virtue of (8:1b) and (8:Id). Each of these propa-
gation formulas, and in particular the simple formu-
In this part we derive simple formulas showing las (8:9), can be used to check whether nonvanishing
how errors in scalar products of the type products (8:3) are due to normal rounding-off errors
or to errors of the computer. The formulas (8:10)
(8:3) have the following meaning. If we build the sym-
metric matrix P haying the elements (Apt,pk), the
are propagated during the next step of the computa- left side of (8:10b) is the ratio of two consecutive
tion. From (8:le) follows elements in the same line, one located in the main
diagonal and one on its right hand side. The
formula (8:10b) gives the change of this ratio as we
go down the main diagonal.
Inserting (8:1c) in both terms on the right yields
8.2. A Stability Condition
Even if the scalar products (8:2) are not all zero,
so that the vectors p§,P\, • • •, pn-i are not exactly
conjugate, we may use these vectors for solving
Applying (8:le) to the first and third terms gives Ax=k in the following way. The solution h may be
written in the form
(rt,ri+l) = (ri,rt)—at(pi,Apt) + bi-i<ii(Api-i,pi), (8:4)
h=x0+a'opQ+a'1p1 + ' • -+ai_!#„_!. (8:11)
which by (8:1b) becomes
Taking the scalar product with Apt, we obtain
(ri,rw)==bi-1ai(Api-l,pi). (8:5)
= (k,pt)
This is our first propagation formula.
Using (8:le) again,
or
(Api,pi+1)=Api,ri+1)+bi(Api,pi). (8:12)

Inserting (7:1c) in the first term, The system Ax=k may be replaced by this linear
system for a'o,- • -X_i. Therefore, because of
(Aptipi+1)=—— \ri+1\2+-~ (rt rounding-off errors we have certainly not solved
t^ a% the given system exactly, but we have reached a
(8:6) more modest goal, namely, we have transformed the
given system into the system (8:12), which has a
But in view of (8:1b) and (8:Id) dominating main diagonal if rounding-off errors have
not accumulated too fast. The cg-algorithm gives,
ri+1\2=aibi(Api,pi). (8:7) an approximate solution
Therefore,
h'=xo+aopo+- (8:13)
(8:8)
A comparison of (8:11) and (8:13) shows that the
number ak computed during the cg-process is an
This is our second propagation formula. approximate value of ak.
Putting (8:5) and (8:8) togetherjyields the third In order to have a dominating main diagonal in
and fourth propagation formulas the matrix of the system (8:12) the quotients

(Apt,pk)
(8:9a) (8:14)

(8:9b) must be small. In particular this must be true for

k=i+l. In this special case we learn from (8:10b)
which can be written in the alternate form that increasing numbers a0, aly- - • during the cg-
process lead to accumulation of rounding-off errors,
because then these quotients increase also. We
(8:10a) have accordingly the following stability condition.
The larger the ratios a*/a*_i, the more rapidly the
rounding-off errors accumulate.
(Api,pi+1)= at A more elaborate discussion of the general quotient
(8:10b)
(Api,pi) a<_! (8:14) gives essentially the same result.
420
 By theorem 5:5, the scalars at lie on the range 8.3. The End-Correction
The system (8:12) can be solved by ordinary re-
laxation processes. Introducing the numbers ak as
approximations of the solutions a'k, we get the resid-
where Xmin, Xmax are the least and the largest eigen- uals
values of A. Accordingly, the ratio p = Xmax>/Xmin is an (8:15)
upper bound of the critical ratio ai/ai-1, which deter-
mines the stability of the process. When p is near Inasmuch as ro'==ri+i-\-a0Ap0+. . , pi by (8:1c),
one, that is, when A is near a multiple of the identity, we have
the cg-method is relatively stable. It will be shown
in section 18 that examples can be constructed in i,pi). (8:16)
which the ratios &*/&*_i (i—1,- • -,n—l) are any
set of preassigned positive numbers. Thus the It follows that the residual (8:15) is"reduced to
stability may be low. However, this instability can
be compensated to a certain extent by starting the
cg-process with a vector x0 whose residual vector r0
is near to the eigenvector of A corresponding to Xmin. (8:17)
In this event aQ is near to the upper bound 1/Xmln of
the at. This result is brought out in the following This leads to the correction of at
theorem:
For a given symmetric and positive definite matrix A,
which has distinct eigenvalues, there exists always an
initial residual vector r0 such that («</«<_ i ) < l and hence
such that the algorithm is stable with respect to the (8:18)
propagation of rounding-off errors.
In order to prove this we introduce the eigenvalues A first approximation of a[ is accordingly

of A, and we take the corresponding normalized In order to discuss this result, suppose that the
eigenvectors as a coordinate system. Let a0, a\, ..., numbers at have been computed accordingly to
an_i be real numbers not equal to zero and e a small the formula
quantity. Then we start with a residual vector
(Pt,rt) (8:19)
r0z=(a:0,aie,a2e2, . . . ,an-ien-1). (8:14a)
at=-

Expanding everything in a power series, one finds (theorem 5:5). From (8:1c) it follows that
that iri+hpi)=0, and therefore this term drops out in
(8:18). In this case the correction Aaf depends
(8:14b) only on the products (Api,pk) with i<CJc. That is
to say, that this correction is influenced only by the
Hence rounding-off errors after the i-th step. If, for
instance, the rounding-off errors in the last 10 steps
of a cg-process are small enough to be neglected,
the last 10 values a* need not to be corrected. Hence,
if e is small enough. generally, the Aa* decrease rather rapidly.
As a by-product of such a choice of r0 we get by rounding-off From (8:18) we learn that in order to have a good
(8:14b) approximations of the eigenvalues of A. keep the products behavior, it is not only necessary to
Moreover, it turns out that in this case the successive satisfy (r pi) = 0 as (pk,Apt) (i^k) small, but also to
i+h
residual-vectors r0, rif ..., rn_i are approximations of it may be better to compute well as possible. Therefore,
the eigenvectors. the a* from the formulas
These results suggest the following rule: (8:19) rather than from (8:1b). We see this im-
mediately, if
The cg-process should start with a smooth residual (8:19) and (8:le) we have we compare (8:19) with (8:1b); by
distribution, that is, one for which /x(r0) is close to Xmin.
Ij needed, the first estimate can be smoothed by some
relaxation process.
Of course, we may use for this preparing relaxation
the cg-process itself, computing the estimates xf
given in section 7. A simpler method is to modify For ill-conditioned matrices, where at and bt may
the cg-process by setting bt=0 so that Pi=rt and become considerably larger than 1, the omitting
selecting at of the form at=a/ii(ri)i where a is a of the second summand may cause additional errors.
small constant in the range 0 < < l For the same reason, it is at least as important in
421
these cases to use formula (3:2b) rather than (8:Id) Introducing the correction factor
for determining bh since by (3:2b) and (8:1c)
1 (8:21)
{k
and taking into account the old value (8:1b) of aiy
Here the second summand is not directly made this can be written in the form
zero by any of the two sets of formulas for at and
hi. The only orthogonality relations, which are
directly fulfilled in the scope of exactitude of the 5<=gr (8:22)
numerical computation by the choice of at and bu
are the following: Continuing in the general routine of the (i+l)st step
we replace bt by a number 6* in such a way that
(Api,pt+i)=0. We use (8:6), which now must be
written in the form
Therefore, we have to represent (ri+i,r%) in terms
of these scalar products:
— |r f + 1 | 2 + (r
) = (ri+hpt) — b^
The term (r*,rf+i) vanishes by virtue of our choice
From this expression we see that for large 6?: and a* of at. Using (8:7), we see that aibi=aibi and from
the second and third terms may cause considerable (8:22)
rounding-off errors, which affect also the relation bi=bidi. (8:23)
(Pi+i,Api)=0, if we use formula (8:Id) for 6t. This
is confirmed by our numerical experiments (sec- Since rounding-off errors occur again, this sub-
tion 19). routine can be used in the same way to improve the
From a practical point of view, the following results in the (i+l)th step.
formula is more advantageous because it avoids the The corrections just described can be incorporated
computation of all the products (Apt,pk). From automatically in the general routine by replacing the
(8:1c) follows formulas (3:1) by the following refinement:
rn=ri+i— Po=ro=k—Ax0, do=l

and we have the result

(8:20) (8:24)

This formula gives corrections of the a* if, because

of rounding-off errors, the residual rn is not small
enough.
8.4. Refinement of the cg-algorithm
In order to diminish rounding-off errors in the
orthogonality of the residuals rt we refine our general
routine (8:1). After the ith step in the routine we Another quite obvious, but numerically more
compute (Api-i,pi), which should be small. Going laborious method of refinement goes along the
then to the (i+l)st step we replace at by a slightly following lines. After finishing the ith. step, compute
different quantity at chosen so that (ri}ri+i)=0. In a product of the type (Api,pk) with k<^i. Then
order to perform this, we may use (8:4), which now replace pt by
must be written
(8:25)

yielding The vector pt is exactly conjugate to pk. This

method may be used in place of (8:24) in case
k=i—l. It has the disadvantage that a vector must
be corrected and not just numbers, as in (8:24).
422
 9. Modifications of the cg-Process I. The vector p^ can be chosen to be the residual
vector rt described in section 7.
In the cg-method given by (3:1), the lengths of the In this event we select
vectors p0, Pi, . . . are at our disposal. In order to
preserve significant figures in computations, it is (9:3)
desirable to have all the pt of about the same magni-
tude. In order to aim at this goal, we normalized The formula (7:2b) for xi+1 becomes
the p's by the formulas
Po==^Qj Pi = r
'i~\~bi-ip»«—i
(9:4)
1 + bt
However other normalizations can be made. In II. The vector pt can be chosen so that the formula
order to see how this normalization appears, we
replace Pi by dtpt in eq (3:1), where df is a scalar
factor. This factor dt is not the same as that given
in section 8 but plays a similar role. The result is
holds.
In this event the basic formulas (9:1) take the
ro=k—AxQ, Po=-r simple form
a0
xi+i=Xi+aiPi
rQ=k—Ax0,

Pi
Pi+i= (9:1)
(9:5)
Apt
di=i

h \n+i\2dt
n+i
to>t,APt) Pt+i=Pi+
The connections between ai} biy dt are given by the
equation This result is obtained from (9:1) by choosing

In this case the formulas (9:5) are very simple and

are particularly adaptable to computation. It has
(i>0), (9:2) the disadvantage that the vectors p% may grow con-
siderably in length, as can be seen from the relations
where /*(r) is the Rayleigh quotient (4:12). In order I IO I In i J-
to establish these relations we use the fact that rt
and ri+1 are orthogonal. This yields
However, if "floating" operations are used, this
should present no difficulty.
I I I . The vector pt can be chosen to be the correction
to be added to xt in the (i-\-l)st relaxation.
by virtue of the formula ri+i=ri—aiApi. From the In this event, at—l and the formulas (9:1) take
connection between pt and rt) we find that the form
ro=k-- Ax0, pQ-
-- \ri\2=di(ri,Api) T
at

(9:6)
This yields (9:2) in case i > 0 . The formula, when Pi+l==
t-i+1
i = 0 , follows similarly.
In the formulas (9:1) the scalar factor dt is an
arbitrary positive number determining the length of
p t The case df=l is discussed in sections 3 and 5.
T he following cases are of interest.
227440—52- 423
 r
These relations are obtained from (9:1) and (9:2) by- If one does not wish to use any properties of the
setting di=l. cg-method in the computation of at and b t besides
IV. The vector pt can be chosen so that at is the recip- the defining relations, since they may be disturbed
rocal of the Eayleigh quotient oj rt. by rounding-off errors, one should use the formulas
The formulas for aiy bt and dt in (9:1) then become

, AA*ri+1)

btai+1
do=l, In this case the error function f(x) is the function
f(x)=\k—Ax\2, and hence is the squared residual.
This is sufficient to indicate the variety of choices It is a simple matter to interpret the results given
that can be made for the scalar factor d0, d1} . . . . above for this new system.
For purposes of computation the choice di = 1 appears It should be emphasized that, even though the use
to be the simplest, all things considered. of the system (10:2) is equivalent from a theoretical
point of view to applying the cg-algorithm to the
10. Extensions of the cg-Method system (10:1), the two methods are not equivalent
In the preceding pages we have assumed that the from a numerical point of view. This follows because
matrix A is a positive definite symmetric matrix. rounding-off errors in the two methods are not the
The algorithm (3:1) still holds when A is nonnegative same. The system (10:2) is the better of the two,
and symmetric. The routine will terminate when because at all times one uses the original matrix A
one of the following situations is met: instead of the computed matrix A*A, which will
(1) The residual rm is zero. In this event xm is contain rounding-off errors.
a solution of Ax~k, and the problem is solved. There is a slight generalization of the system (10:2)
(2) The residual rm is different from zero but that is worthy of note. This generalization consists
(Apm,pm)=Q, and hence Apm=0. Since Pi=Ctri, of selecting a matrix B such that BA is positive defi-
it follows that Arm=0, where rm is the residual of the nite and symmetric. The matrix B is necessarily of
vector xm defined in section 7. The point xm is the form A*H, where H is positive definite and
accordingly a point at which \k—Ax\2 attains its symmetric. We can apply the cg-algorithm to the
minimum. In other words, xm is a least-square system
solution. One should observe that Pm^O (and hence BAx=Bk. (10:3)
rm5*0). Otherwise, we would have rm=—bm-ipm-i,
contrary to the fact that rm is orthogonal to pm-\. In place of (10:2) one obtains the algorithm
The point xm fails to minimize the function
rQ=k—Ax0, pQ=Br0,
g(x)=(x,Ax)-2(k,x),
for in this event „ _ \Br<\*
at
2
-(pt,BAPt)'
g(xm+tpm)=g(xm)—2t\rm\ .
In fact, g(x) fails to have a minimum value.
It remains to consider the case when A is a general (10:4)
nonsingular matrix. In this event we observe that
the matrix A*A is symmetric and that the system
Ax=k is equivalent to the system
A*Ax=A*k. (10:1)
Applying the eq (3:1) to this last system, we obtain
the following iteration,
Again the formulas for at and bi} which are given
r o = k — Ax0, po= A*r0, directly by the defining relations, are

ai
~ \APi\* ai
(pitBAPi)
xi+1=xi+aipi
(10:2) h _ (Bri+1,BAPi)
(piyBAPi)

When B=A*, this system reduces to (10:2). If A

is symmetric and positive definite, the choice B=I
gives the original cg-algorithm.
424
 There is a generalization of the cd-algorithm con- given. In the cg-method the choice of the vector
cerning which a few remarks should be made. In Pi depended on the result obtained in the previous
this method we select vectors p0, . . ., pn-\ and step. The vectors p0, pXj . . . are accordingly deter-
</o, • . •, qn-i such that mined by the starting point x0 and vary witli the
point x0.
Assume again that A is a positive definite, sym-
(10:5) metric matrix. In a cd-method the vectors p0,
Pi, . . . can be chosen to be independent of the
starting point. This can be done, for example, by
The solution can be obtained by the recursion for- starting with a set of n linearly independent vectors
mulas u0, ui, . . ., un-\ and constructing conjugate vectors
ro=k—Ax0, by a successive ^L-orthogeonalization process. For
example, we may use the formulas
(qj,ro) ^
~(qifApt)'
(10:6) Pi=U1—alopo,

The problem is then reduced to finding the vectors

Pi, qt such that (10:5) holds. We shall show in a
moment that qt is of the form
q.=B*Pi, (10:7)
where B has the property that BA is symmetric and
positive definite. The algorithm (10:6) is accordingly
equivalent to applying the cd-algorithm to (10:3). The coefficient aij(i^>j) is to be chosen so that pt is
To see that qt is of the form (10:7), let P be the conjugate to pj. The formula for atj is evidently
matrix whose column vectors are p0, . . ., pn-\ and Q
be the matrix whose column vectors are q0, . . ., qn-\-
The condition (10:5) is equivalent to the statement (11:2)
that the matrix D= Q*AP is a diagonal matrix whose Observe that
diagonal terms are positive. Select B so that (pi}Auj)=0
Q=B*P. Then D=P*BAP from which we conclude
that BA is a positive definite symmetric matrix, as
was to be proved. (11:3)
In view of the results just obtained, we see that Using (11:3) we see that alternately
the algorithm (10:4) is the most general cg-algorithm
jor any linear system. Similarly, the most general (Aui,pj)
=
cd-algorithm is obtained: by (i) selecting a matrix aij (11:4)
B such that BA is symmetric and positive definite, (Auj,p3)
(ii) selecting nonzero vectors p0, . . ., pn_x such that
As described in section 4, the successive estimates of
(Pi,BApj)=0, (i^j) the solution are given by the recursion formula
and (iii), using the recursion formulas (11:5)
where
ro=k—Ax0
di = 7 (11:6)
„ _ (pi9Brt) _ (Pi,Br0)
There is a second method of computing the
xi+1=xi + a vectors p0, pu . . ., pn_h given by the recursion
formulas
ri+1=ri — ulo)=Ui (11:7a)
11. Construction of Mutually Conjugate PJ=U}°, (11:7b)
Systems
As was remarked in section 4 the cd-method is not
complete until a method of constructing a set of
iJ
mutually conjugate vectors p0, pi} . . . has been (phAuj)
425
We have the relations (11:3) and, ing in (12:Id) and (12: If) can be computed. A
systematic scheme for carrying out the computations
(11:8a) will now be given. The scheme is that commonly
used in elimination. In the presentation that we
(11:8b) now give, extraneous entries will be kept so as to
give the reader a clear picture of the results obtained.
i>j) (11:8c) We begin by writing the matrices A, I and the
vector A: as a single matrix
(pi)Apj)=O
(III #12 #13 <hn 1 0 0 0
The eq (11:8a) hold when k=j+l by virtue of
(11:7c) and (11:7(1). That they hold for other #21 #22 #23 #2n 0 1 0 0
values of j<Ck follows by induction. Equation #31 #32 #33 <hn. 0 0 1 0
(11:8c) follows from (11:7c).
If one selects successively uQ=r^ Ui=ru . . ., (12:2)
un-i=rn-i, the procedure just described is equivalent
to the cg-method described in section 3, in the sense
that the same estimates x0, Xi, . . . and the same 0
direction vectors p0, pu . . . are obtained. If #«3 • • #71 1
one selects uQ=k, Ui=Ak, . . ., un-i=An~1k, one
again obtains the same estimates x0, Xi, . . . as in The vector px is the vector (1,0, . . .,0), and ax =
the cg-method with a ^ O . However in this event ki/an is defined by (12:lf). Hence,
the vectors p0, pu . . . are multiplied b y nonzero
scalar factors. On the other hand if one selects
^ = ( 1 , 0 , . . .,0), Ui = (0,l, . . .,0), . . ., ttn_i = (0,
. . .,0,1) the cd-method is equivalent to the Gauss
elimination method. This case will be discussed in is our first estimate. Observe also that
the next section.

12. Connections With the12Gauss Elimina- #n

tion Method Multiplying the first row by an and subtracting the
In the present section it will be convenient to result from the ith. row (i=2, • • • ,ri), we obtain the
use the range 1, . . ., n in place of 0, 1, . . ., n— 1 new matrix
used heretofore, except for the notations x0, xiy . . .,
xn describing the successive estimates of the solution. #11 #12 #13 # i « Pll Vm
Let ei, . . ., en be the unit vectors (1,0, . . ..0), 0 U/22 U'23 •• #2n P21
(0,1,0,. . .,0), . . ., (0,. . .,0,1). These vectors (2)
will play the role of the vectors u0, . . ., un-x of 0 Us 32 a #3n kf
section 11. The eq (11:7), together with (11:4) and
(11:5), yield the recursion formulas
u^=ei (<=!,. . .,n) (12:1a)
up (12:1b) (12:3)
(} (
(i=7-[-l r . ,)U) (12:1c) One should observe that (0,i | ,. . .,/: ?) is the residual
of xx. By the procedure just described the ith. row
( i > l ) of the identity matrix has been replaced by
uj2), the second row yielding the vector p2=u{f. Ob-
serve also that
xQ=0, a}f=(Auf,et) {i=2,...,

a1=
(Apt,et)
Hence,
These formulas generate mutually conjugate vectors
Pu • • •> Vn and corresponding estimates Xi, . . ., xn
of the solution of Ax=k. In particular xn is the
desired solution. The advantage of this method is the next estimate of the solution. Moreover,
lies in the ease with which the inner products appear-
12
4§
cf. Fox, Huskey, and Wilkinson, loc. cit. #22

426
Next multiply the 2nd row of (12:3) by ai2 and sub vectors are pu p2, • • •, pnj then the matrix (12:4)
tract the result from the iih row (i—2>y—,n). We ob is the matrix
tain \\P*A P * P*Jfc||.
#13 Pn Vm The matrices P*A and P are triangular matrices
(2) n(2)
with zeros below the diagonal. The matrix D=P*AP
0 a #23 P21 Vin is the diagonal matrix whose diagonal elements are
0 0 n (3)
#33
n(3)
#3n Vzn kf #11, a™, . . ., a<nnn- The determinant of Pis unity and
4,(3)
the determinant of A is the product
0 0
n n(2) <n)
#11#22.
As was seen in section 4, if we let
0 0 ,(3) (3)
u.'nn Jfe*. f(x) = (h—x,A(h—x)),
( ( }
The vector (0,0,& l\. . .,& n ) is the residual of x2. the sequence
The elements u}?,. . .,u$ form the vector uf (i = 3)
and Vz—U (3} • We have
a^) = (-4u?), e*) (i=3,. . .,n) decreases monotonically. No general statement
can be made for the sequence

We have accordingly \yo\,\yi\, . • -,\yn-i\,\yn

of lengths of the error vectors y1=h—Xi. In fact,
„ (3) we shall show that this sequence can increase mono-
#33
and tonically, except for the last step. A situation of
this type cannot arise when the cg-process is used.
If A is nonsymmetric, the interpretation given
above must be modified somewhat. An analysis of
Proceeding in this manner, we finally obtain a matrix the method will show that one finds implicitly two
of the form triangular matrices P and Q such that Q*AP is a
diagonal matrix. To carry out this process, it may
#11 #12 #13 ... dXn Vn Pin be necessary to interchange rows of A. By virtue
of the remarks in section 10, the matrix Q* is of the
0 a (2)
• •• #3» k[2) form J5*P. The general procedure is therefore equiv-
#23
0 0 #33} • ' - #3n
alent to application of the above process to the sys-
tem (10:3).
13. An Example
0 0 Pnl In the cg-method the estimates xo,x1}... of the solu-
(12:4) tion h of Ax=k have the property that the error
vectors yo=h—#0, yi=h—x1} . . . are decreased in
The elements p , • • •, pin define a vector pt. The length at each step. This property is not enjoyed
vectors ^x, • •-, pn are the mutually conjugate by every cd-method. In this section we construct
vectors defined by the iteration (12:1). At each an example such that, for the estimates #o=O,#i, . . .,
stage of the elimination method,
\h-Xi-iK\h-Xi\ —1).
If the order of elimination is changed, this property
may not be preserved.
The example we shall give is geometrical instead
Moreover the estimate of the solution A is given of numerical. Start with an (n—1)-dimensional
by the formula ellipsoid En with center xn=h and with axes of un-
equal length. Draw a chord Cn through xni which
is not orthogonal to an axis of En. Select a point
xn^\9^xn on this chord inside En, and pass a hyper-
plane Pn-\ through xn-\ conjugate to Cn, that is,
The vector 0, • • •, 0, a/j}, • • •, a/^ defined by the parallel to the plane determined by the midpoints
first 71 elements in the ith row of (12:4) is the vector of the chords of En parallel to Cn. Let en be a unit
Apt. If we denote by P the matrix whose column vector normal to Pw_i. It is clear that en is not
427
parallel to Cn. The plane Pn-\ can be shown to cut tinued fraction expansions. To develop this, we
En in an (n—2)-dimensional ellipsoid En-\ with cen- first study connections between orthogonal poly-
ter at xn-i and with axes of unequal length. nomials and ^-dimensional geometry.
Next draw a chord Cn-\ of En_x throughx n -\ which Let m(X) be a nonnegative and nondecreasing
is not orthogonal to an axis of i?w_i, and which is not function on the interval 0 < \ < Z . The (Riemann)
perpendicular to h—ajn_i. One can then select a Stieltjes integral
point xn-2 on Cn-i which is nearer to Athana;n_i. Let
Pn-2 be the hyperplane through xn-2 conjugate to
Cn-i. It intersects 2?w_i in an in—3)-dimensional
ellipsoid En-2 with center at xn-2- The axes ofEn~2
can be shown to be of unequal lengths. Let en-\ be then exists for any
r /(X)dm(X)
Jo continuous function /(X) on
a unit vector in Pn-\ perpendicular to Pn-2- 0<X<Z. We call m(X) a mass distribution on the
We now repeat the construction made in the last positive X-axis. The following two cases must be
paragraph. Select a chord Cn~2 of En-2 through distinguished.
#w_2 that is not orthogonal to an axis of En-2 and that (a) The function m(X) has infinitely many points
is not perpendicular to h—xn-2- Select xn-Z on Cn-2 of increase on 0<X<Z.
nearer to h than #w_2, and let Pn-% be a plane through (b) There are only a finite number n of points of
#rc_3 conjugate to Cn-2. It cuts En_2 in an (n—4)- increase. In both cases we may construct by
dimensional ellipsoid En_z with center at #w_3 with orthogonalization of the successive powers 1, X, X2,
axes of unequal lengths. Let en-2 be a unit vector in • • •, \n a set of w+1 orthogonal polynomials 13
Pn-i and Pn-2 perpendicular to Pn-z- Clearly, en,
eu-ly en-2 are mutually perpendicular. 2?o(X),B1(X), (14:1)
Proceeding in this manner, we can construct with respect to the mass distribution. One has
(1) Chords Cn, Cn-u • • •> @u which are mutually
conjugate.
(2) Planes Pn-i, . . ., Pi such that Pk is conjugate fi2i(X)BA;(X)dm(X)=0 (i^ky (14:2)
to Ck+1. The chords Ci, . . ., Ck lie in P*. Jo
(3) The intersection of the planes P w -i, . . ., Pk, The polynomial Rt(\) is of degree i. In case (b),
which cuts En in a (k—1)-dimensional ellipsoid Ek Rn(X) is a polynomial of degree n having its zeros at
with center xk. the n points of increase of m(X). In both cases the
(4) The point xiy which is closer to h than zeros of each of the polynomials (14:1) are real and
iC
i C n l . distinct and located inside the interval (0,Z). Hence
(5) The unit vectors ew, . . ., e2, e\ (with 6X in the we may normalize the polynomials so that
direction of Ci), which are mutually orthogonal.
Let x0 be an arbitrary point on Ci that is nearer to 5,(0) = 1 ( i = l , • . •, n). (14:3)
h than xx. Select a coordinate system with x0 as the
origin and with ely . . ., en as the axes. In this co- The polynomials (14:1) are then uniquely determined
ordinate system the elimination method described by the mass distribution.
in the last section will yield as successive estimates During the following investigations we use the
the points xly . . ., xn described above. These esti- Gauss mechanical quadrature as a basic tool. It can
mates have the property that xt is closer to xn=h be described as follows: If Xi, • • *,XW denote the
than xi+i if i < n — 1 . zeros of Rn(\), there exist positive weight coefficients
As a consequence of the construction just made we m1,m2, • • -,mn such that,
see that, given a set of mutually conjugate vectors
Piy . . ., pn and a starting point x0, one can always
choose a coordinate system such that the elimination . . . + mnB(Xn)
method will generate the vectors pi, . . ., pn (apart (14:4)
from scalar factors) and will generate the same esti-
mates Xi, . . ., xn of h as the cd-method determined by whenever R(\) is a polynomial of degree at most
these data. One needs only to select the origin at 2n—l. In the special case b) the \k are the abscissas
x0, the vector ex parallel to pXj the vector e2 in the where m(X) jumps and the mk the corresponding
plane of px and p2 and perpendicular to ei, the vector jump.
ez in the plane of pu p2) p3 and perpendicular to ex In order to establish the duality mentioned in the
and e2, and so on. This result may have no practi- title of this section, we construct a positive definite
cale value, but it does serve to clarify the relation- matrix A having Xb X2, • • •, Xw as eigenvalues (for
ship between the elimination method and the cd- instance, the diagonal matrix having Xb • • •, Xn in
method, and also the relationship between the the main diagonal and zeros elsewhere). Further-
elimination method and the cg-method. more, if 6i, e2y - - ', en are the normalized eigen-
vectors of A, we introduce the vector
14. A Duality Between Orthogonal Poly-
nomials and n-Dimensional Geometry -anen, (14:5)
l
The method of conjugate gradients is related to » The various properties of orthogonal polynomials used in this chapter may be
found in G. Szego, Orthogonal Polynomials, American Mathematical Society
the theory of orthogonal polynomials and to con- Colloquium Publications 33 (1939).

428
where Again we may restrict ourselves to the powers
•,»)•
(14:6) 1, X, . . . ,Xn-1. That is, we must show that

We then have
+an\n*en (14:7)
r
Jo
X'+1X*dm (X) =
(14:10)
for k=0,l, • • •, n—1. The vectors rOrAro, • • -, If j<Cn—lj this has already been verified, The
An~lr§ are linearly independent and will be used as a remaining case
coordinate system. Indeed their determinant is up
to the factor a^a2 • • • an Van der Monde's determi-
nant of Xi, • • -,Xn. By the correspondence
Jo
f <n-l) (14:11)

\k->Akr0 (k=0,l, • • -,n — l) (14:8) follows in the same manner from Gauss' integration
formula, since n+k<2n—l.
every polynomial of maximal degree T? — 1 is mapped Theorem 14:3. Let A be a positive definite sym-
onto a vector of the ^-dimensional space and a one- metric matrix with distinct eigenvalues and let r be a
0
one correspondence between these polynomials and vector that is not perpendicular to an eigenvector of A.
vectors is established. The correspondence has the There is a mass distribution m(X) related to A as
following properties: described above.
Theorem 14:1. Let the space of polynomials 22(X) In order to prove this result let eh . . ., en be
of degree <n—l be metrized by the norm the normalized eigenvectors of A and let Xx, . . .,
\n be the corresponding (positive) eigenvalues. The
i vector r0 is expressible in the form (14:5). According
to our assumption no ak vanishes. The desired mass
distribution can be constructed as a step function
Then the correspondence described above is isometric, that is constant on each of the intervals 0<Xi<X 2 <
that is, • • • <XW<^, and having a jump at \k of the amount

r
Jo7
mk^aiy>0, the number I being any number greater
than \n.
We want to emphasize the following property of
our correspondence. If A and r0 are given, we are
where 22(X), 22 (X) are the polynomials corresponding able to establish the corredence without computing
to r and rf. eigenvalues of A. This follows immediately from the
It is sufficient to prove this for the powers 1, X, basic relation (14:8). Moreover, we are able to com-
2 1 j k
X , . . ., X"" . Let \ , \ be two of these powers. pute integrals of the type
From Gauss' formula (14:4) follows
[lR (X) R' (X) dm (X), VR (X) R' (X) \dm (X),
l k j+k
f V\ dm (X) = (\ dm (X) Jo Jo
Jo Jo (14:12)

where R, B/ are polynomials of maximal degree

n—1 without constructing the mass distribution.
But (14:5), (14:6), and (14:7) show that this is Indeed, the integrals are equal to the corresponding
exactly the scalar product (Ajr0,Akr0) of the cor- scalar products (r,r'), {Ar,rf) of the corresponding
responding vectors. vectors, by virtue of theorems 14:1 and 14:2. Finally,
Theorem 14:2. Let the space of polynomials 22(X) the same is true for the construction of the orthog-
of degree <n—l be metrized by the norm onal polynomials RQ(\), RI(\), . . ., 22W(X) because
the construction only involves the computation of
integrals of the type (14:12). The corresponding
f r*22(X) vectors r0, rX) . . ., rn-i build an orthogonal basis in
the Euclidian n-space.

Then for polynomials 22(X),22'(X) corresponding to 15. An Algorithm for Orthogonalization

r,rf one has
In order to obtain the orthogonalization of poly-
l nomials, the following method can be used. For
( R(h)R' (14:9) any three consecutive orthogonal polynomials the
Jo recurrence relation holds:
that is, the correspondence is isometric with respect to
the weight function \dm(X) and the metric, determined
by the matrix A. (15:1)
429
where au ct, dt are real numbers and at^0. Taking Applying (15:3) to the first term yields
into account the normalization (14:3), we have
\=di-ci. (15:2) &i-\ Jo l Jo
Hence Combining this result with (15:7), we obtain

This relation can be written

-ft (15:8)
Ri+i—Rj __ T> | Cj Rj—R
Jo
"~ <Lt X
The formulas (15:5), (15:7), (15:8), together with
From this equation it is seen by induction that i?o=l, 6_i=0, completely determine the polynomials
R R R
(15:3) 16. A New Approach to the cg-Method,
Eigenvalues
are polynomials of degree i. Introducing the num-
bers In order to solve the system Ax=k, we compute the
residual vector ro=k—Ax0 of an initial estimate x0 of
&*-! = 6_i = 0 (15:4) the solution h and establish the correspondence based
at on A} r0 described in Theorem 14:3. Without com-
we have puting the mass distribution, the orthogonalization
i-iP<-i(X) (15:5a) process of the last section may be carried out by
(15:5), (15:7) and (15:8) with 5 0 = l , &-i = 0. The
vectors riy pi corresponding to the polynominals
,XP<(X). (15:5b) Rt, Ft are therefore determined by the recurrence
relations
Beginning with RQ=1, we are able to compute by
(15:5) successively the polynomials PQ=R0=1, Rh Pt=ri~{-bi-.iPi-i, ri+i=ri—d.iApi. (16:1)
Pi, R2, Pi, • . ., provided that we know the numbers
at, bi. In order to compute them, observe first the Multiplication by X in the domain of polynominals
relation is mapped by our correspondence into applying A
in the vector space according to (14:11). In fact,
f P,(X)P*(X)Xdm(X)=O
Jo
d^k). (15:6)
Pt=Pt(A)r0, r^ (i=0,
Indeed this integral is up to a constant factor The numbers ah bt are computed by (15:7) and (15:8).

r
For k<i% this is zero, because the second factor is of
Using the isometric properties described in theorems
14:1 and 14:2, we find that

&«-,=
\r*\'
degree k^i
Using (15:5a) and (15:6), we obtain The vectors rt are orthogonal, and the pt are con-
jugate; the latter result follows from (15:6). Hence
the basic formulas and properties of the cg-method
f Ri(X)Pi(X)\dm(\)= fZ listed in sections 3 and 5 are established. It remains
J° Jo to prove that the method gives the exact solution
after n steps. If we set xi+1=Xi+aiPi, the corre-
Combining this result with the orthogonality of sponding
Ri+i and Rt, we see, by (15:5b), that residual is ri+i as follows by induction:

Rt(\)*dm(\)
____ Jo (15:7) For the last residual rw we have (i=0, 1, . . . ,,n—l)
at=
Jo (rn,rt) = (rn-iVi) —dn-^Apn-ifi)

Using (15:6) and (15:5a), = I Rn-Jtidm — dn-i I Pn-_Ri\dm

0= f = (RnRidm = 0.
Jo Jo
430
 Our basic method reestablishes also the methods This gives the following result, let A be a symmetric
of C. Lanczos for computing the characteristic poly- matrix having the roots of the Legendre polynomial
nomial of a given matrix A. Indeed the polynomials Rn(K) as eigenvalues, and let
Ri, computed by the recurrence relation (15:5), lead
finally to the polynomial 2?W(X), which, by the basic
definition of the correspondence in section 14, is the
characteristic polynomial of A, provided that r0 satis- where eu . . ., en are the normalized eigenvectors of
fies the conditions given in theorem 14:3. It may A, and mi=ai, ra2=a|, . . ., mn=<x2n are the
be remembered that orthogonal polynomials build a weight-coefficients for the Gauss' mechanical quad-
Sturmian sequence. Therefore, the polynomials rature with respect to Rn. The cg-algorithm applied
Ro, i?i, . . ., Rn build a Sturmian sequence for the to AirQ yields the numbers at, bi given by (17:1).
eigenvalues of the given matrix A. Moreover,
Our correspondence allows us to translate every
method or result in the vector-space into an anal-
ogous method or result for polynomials and vice
versa. Let us take as an example the smoothing
process in section 7. It is easy to show that the
vector rt introduced in that section corresponds to a Hence the residuals decrease during the alogrithm.
polynomial__i?i(X) characterized by the following It may be worth noting that the Rayleigh quotient
property: Rt(\) is the polynomial of degree i with of rt is
22i(O) = l having the least-square integral on (0,1)- (r^Ari) 1 , &t_i 1
In other words, if r0 is given by (14:5), then

afRi(\1)2+afR2(\2)2+ All residual vectors have the same Rayleigh quotient.

This shows that, unlike many other relaxation
This result may be used to estimate a single eigen- methods, the cg-process does not necessarily have
value of A. In order to compute, for instance, the the tendency of smoothing residuals.
lowest eigenvalue Xi, we select r0 near to the corres- The Chebyshev polynomials yield an example
ponding eigenvector. The first term in the expan- where aix hi are constant for i > 0 .
sion being dominant, the smallest root of JR*(X) will
be a good approximation of Xi, provided that i is not 18. Continued Fractions
too small. Hence the last residual vanishes, being
orthogonal to r0, rlf . . ., r n _ t . It follows that xn is
the desired solution. Suppose that we have given a mass distribution of
type (b) as described in section 14. The function
m(X) is a step function with jumps at 0<Xi<X2<C • •
17. Example, Legendre Polynomials <X n <7, the values of the jumps being mu m2, . . .,
Any known set of orthogonal polynomials yields mn, respectively. It is well known u that the orthog-
an example of a cg-algorithm. Take, for instance, onal polynomials i20(X), -Ri(X), . . ., J?rt(X), corre-
the Legendre polynomials. Adapted to the interval sponding to this mass distribution, can be constructed
(0,1), they satisfy the recurrence relation by expanding the rational function

mn (18:1)
t+1 X-Xi

From (15:1) and (15:4) in a continued fraction. The polynominal Rt(\) is

the denominator of the iih convergent. For our
purposes it is convenient to write the continued
at=— ct=7 fraction in the form
i+r

-1

a 2— ctoX —
(18:2)

* H. S. Wall, Analytic Theory of Continued fractions, Van Norstrand (1948).

431
The denominators of the convergents are given by puted by expanding into a continued fraction the quo-
the recursion formulas tient built by the characteristic polynomial of A with
respect to r0 and the ordinary characteristic polynomial
o=O. (18:3) of A.
This is the simplest form of the relation between a
This coincides with (15:1). However, in order to matrix A, a vector r0 and the numbers at, bf of the
satisfy (14:3), the expansion must be carried out corresponding cg-process. The theorem may be used
so that di=Ci + l, by virtue of (15:2). The numbers to investigate the behavior of the af, bt if the eigen-
bt are then given by (15:4). It is clear that values of A and those with respect to r0 are given.
The following special case is worth recording. If
.Qn-l(X) m]=m2= . . . = m n = l , the rational function is the
logarithmic derivative of the characteristic poly-
where nomial. From theorem (18:1) follows
Theorem 18:2. If the vector r0 of a cg-process is the
sum of the normalized eigenvectors of A, the numbers
di, bf may be computed by expanding the logarithmic
derivative of the characteristic polynomial of A into a
Let us translate these results into the n-dimensional continued fraction.
space given by our correspondence. As before, we Finally, we are able to prove
construct a positive definite symmetric matrix A Theorem 18:3. There is no restriction whatever on
with eigenvalues Xi, . . ., \n. Let eu . . ., en be the positive constants at, bt in the cg-process, that is,
corresponding eigenvectors of unit length and choose, given two sequences of positive numbers a0, aly . . .,
as before, an_i and b0, bi, . . ., 5w_i, there is a symmetric positive
definite matrix A and a vector r0 such that the cg-
anen, algorithm applied to A, r0 yield the given numbers.
The demonstration goes along the following lines:
The eigenvalues are the reciprocals of the squares of From (15:2) and (15:4), we compute the numbers ciy
the semiaxis of the (n—1)-dimensional ellipsoid dt, the d being again positive. Then we use the con-
(x,Ax) = l. The hyperplane, (ro,x) = O, cuts this tinued fraction (18:2) to compute F(\) which we
ellipsoid in an (n—2)-dimensional ellipsoid, En_2, the decompose into partial fractions to obtain (18:1).
squares of whose semiaxis are given by the reciprocals We show next that the numbers Xi, mt appearing in
of the zeros of the numerator Qn-i(\) of F(\). (18:1) are positive. After this has been established,
This follows from the fact that if Xo is a number our correspondence finishes the proof.
such that there is a vector xQ?£0 orthogonal to r0 In order to prove that Xj>0, m t >0we observe that
having the property that (Axo,x) = \o(xo,x) whenever the ratio RwIRt is a decreasing function of X, as can
(ro,a0 = O, then Xo is the square of the reciprocal of be seen from (18:3) by induction. Using this result,
the semiaxis of En-2 whose direction is given by x0. it is not too difficult to show that the polynomials
If the coordinate system is chosen so that the axes i?0(X),i?i(X), . . ., i?»(X) build a Sturmian sequence
are given by ex, . . ., en, respectively, then X=X0 in the following sense. The number of zeros of Rn(\)
satisfies the equation in any interval a< X< b is equal to the increase of the
number of variations in sign in going from a to b.
0 0 0 At the point Xo there are no variations in sign since
Bt(0) = 1 for every i. At X= + °°, there are exactly n
0 X2—X 0 0 variations because the coefficient of the highest power
of X in i?i(X) is (— 1)%0&I • • • <t>i-i- Therefore, the
0 roots Xi, X2, . . ., Xw of i?TC(X) are real and positive.
That the function F(\) is itself a decreasing func-
tion of X follows directly from (18:2). Therefore, its
residues mx, m2, . . ., mn are positive.
In view of theorem. 18:3 the numbers at in a cg-
0 0 ... K —X an process can increase as fast as desired. This result
was used in section 8.2. Furthermore, the formula
ai a2 ... an 0
as was to be proved.
Let us call the zeros of Qn-i (X) the eigenvalues of A
with respect to r0 and the polynomial Q»-i(X) the shows that there is no restriction at all on the be-
characteristic polynomial of A with respect to r0. The havior of the length of the residual vector during the
rational function F(X) is accordingly the quotient of cg-process. Hence, there are certainly examples
this polynomial and the characteristic polynomial of where the residual vectors increase in length during
A. Hence we have, the computation, as was stated earlier. This holds
Theorem 18:1. The numbers at, bt connected with in spite of the fact that the error vector h—x decreases
the cg-process of a matrix A and a vector x0 can be com- in length at each step.

432
 19. Numerical Illustrations A sufficient number of experiments have not been
carried out as yet so as to determine the '"best"
A number of numerical experiments have been formulas to be used. Our experiments do indicate
made with the processes described in the preceding that floating operations should be used whenever
sections. A preliminary report on these experiments possible. We have also observed that the results
will be given in this section. In carrying out these in the (r& + l)st and (r&+2)nd iterations are normally
experiments, no attempt was made to select those far superior to those obtained in the nth. iteration.
which favored the method. Normally, we selected Example 1. This example was selected to illus-
those which might lead to difficulties. trate the method of conjugate gradients in case
In carrying out these experiments three sets of there are no rounding off errors. The matrix A
formulas for a^ bt were used in the symmetric case, was chosen to be the matrix
namely,
1 2 -1
b K }
2 5 0 2
\ri\ -1 0 6 0
di = 7 (19:2)
\rt
1 2 0 3
a
1
= If we select k to be the vector (0,2, —1,1), the
( computation is simple. The results at each step
(19:3) are given in table 1.
Normally, the computation is not as simple as in-
In the nonsymmetric case, we have used only the dicated in the preceding case. For example, if one
formulas selects the solution h to be the vector (1,1,1,1), then
k is the vector (3,9,5,6). The results with (0,0,0,0)
(19:4) as the initial estimate is given by table 2.
TABLE 1.
Our experience thus far indicates that the best
results are obtained by the use of (19:1). Formulas
(19:2) were about as good as (19:1) except for very Components of the vector
ill conditioned matrices. Most of our experiments Step Vector en bi-i
1 2 3 4
were carried out with the use of (19:2) because they
are somewhat simpler than (19:1). Formulas (19:3) Xo 1 0 0 0
were designed to improve the relations
ro -1 0 0 0
0
Po -1 0 0 0
(pt,Api+1) = 0, (19:5)
Apo -1 -2 1 i
1

which they did. Unfortunately, they disturbed the

0 0 0 0
first of the relations Xi

r\ 0 2 -1 1 6
1
(19:6) Pi -6 2 -1 1
(pt,r i+i)=0,
Api 0 0 0 1 6 .
A reflection of the geometrical interpretation of the
method will convince one that one should strive to X2 -36 +12 -6 6
satisfy the relations (19:6) rather than (19:5). It is T2 0 2 -1 -5 5
2
for this reason that (19:1) appears to be considerably P2 -30 12 -6 0
superior to (19:3). In place of (19:2), one can use AP2 0 0 -6 -6 5/6
the formulas
Xz -61 22 -11 6
\rt\ n 0 2 4 0 2/3
at==-/ 3
\ri\ P3 -20 10 0 0
Apz 0 10 20 0 1/5
to correct rounding off errors. A preliminary
experiment indicates that this choice is better than 4 Xi -65 24 -11 6
(19:2) and is perhaps as good as (19:1).

433
 TABLE 2. The system just described is particularly well
suited for elimination. In case k is the vector
a. times components of vector (3, 9, 5, 6) the procedure described in section 12 yields
Step Vector a the results given in table 3. In this table, we start
1 2 3 4 with the matrices A and / . These matrices are
transformed into the matrices P*A and P* given at
Xo 0 0 0 0 1 the bottom of the table.
0
ro 3 9 5 6 1 It is of interest to compare the error vectors
po 3 9 5 6 1 yi=h—Xi obtained by the two methods just described
Apo 22 63 27 39 1
with Jc=(3, 9, 5, 6). The error \yt\ is given in the
following table.
zi 453 1359 755 906 ft
ri -316 -495 933 123 ft \Vi\ cg-method Elimination method
1
Pi -1935 -2799 6461 1140 ft7l
Api -12854 -15585 40701 -4113 ft7l M 2.0 2.00
toil 0. 7 2.65
X2 131702 419553 298277' 304149 ft
1689 -34360 -27345 73483 ft M .67 4.69
2
Pi -116022 -1684085 -381080 3066641 ft72 \v*\ .65 6.48
Ap2 -66471 -2579187 -2140458 5685731 ft72 M .0 0.00

X3 27589274 84526651 62344884 73103513 ft

n 542343 -188185 92550 -66019 ft In the cg-method \yt\ decreases monotonically, while
3
P3 41725242 -15212135 6969632 -3788997 ft73
in the elimination method \yt\ increases except for
the last step.
AP3 542343 -188185 92550 -66019 ft73 Example 2. In this case the matrix A was chosen
to be the matrix
Xi 1 1 1 1 1
4 . 263879 T 014799 .016836 .079773 -.020052 .011463
0 0 0 0 1
-. 014799^ . 249379 .028764 .057757 -. 056648 -.134493
ft = 1002, (82=326123, ft = 69314516, . 016836 .028764 .263734 —.033628 -.012128 .084932
7i=ft/151, 72=02/8149, 73=ft/899615 .079773 .057757 -.033628 .215331 .090696 -.037489
ao=l/7i> «i=71/72, 02=72/73, 03=73 —.020052 -.056648 -.012128 .090696 .324486 -.022484
6o=8149//jJ, 61 = 899615ft 71/$, 62=380689ft72/$- .011463 -.134493 .084932 -.037489 -.022484 .339271

TABLE 3. This matrix is a well-conditioned matrix, its eigen-

values lying on the range Xi = .6035 ^X^X 6 =4.7357.
k Xi Ti The computations were carried out on an IBM
card programmed calculator with about seven sig-
1 2 -1 1 1 0 0 0 3 3 0 nificant figures. The results for the case in which
2 5 0 2 0 1 00 9 0 3 x0 is the origin and h the vector (1,1,1,1,1,1) are
-1 0 6 0 0 0 1 0 5 0 8
given in table 4.
1 2 0 3
Example 3. A good illustration of the effects of
0 0 0 1 6 0 3
rounding can be obtained by study of an ill-con-
ditioned system of three equations with three un-
1 2 -1 1 1 0 00 3 -3 0
knowns, namely, the system
0 1 20 -2 1 0 0 3 3 0
0 2 5 1 1 0 10 8 0 2
0 0 1 2 -1 0 0 1 3 0 3
I3x+29y-38z=2
1 2 -1 1 1 0 00 3 7 0 -l7x-38y+50z=-3, >
0 1 20 -2 1 0 0 3 -1 0
0 0 1 1 5 -2 1 0 2 2 0
whose solution is x=l, y=—3, z=—2. The system
was constructed by E. Stiefel. The eigenvalues of
0 0 1 2 -1 0 0 1 3 0 1
A are given by the set
1 2 -1 1 1 0 0 0 3 1 0
X1=.O588, X2=.2OO7, X3=84.7405.
0 1 20 -2 1 0 0 3 1 0
0 0 1 1 5 -2 1 0 2 1 0
The ratio of the largest to the smallest eigenvalue
00 0 1 -6 2 -1 1 1 1 0 is very large: X3/\i = 1441. The formulas (19:1),
(19:2), and (19:3) were used to compute the solution,
434
 TABLE 4. TABLE 5

Starting vector k = (3.371, 1.2996, 3.4851, 3.7244, 3.0387, 2.412) Xo= (1,0,0)

1 2 3
Step Xi r» Pi at, bi Case. Formula (19:2) F o r m u l a (19:3)
1 with 10 digits
Formula (19:1)

0 3.37100 3.37100 5 5 5 5
0 1.29960 1. 29960 ao=3.092387 po 11 11 11 11
-14 -14 -14 -14
0 3.48510 3.48510

0 3.72440 3.72440 6o=O. 02360156 CO .011804 .011804 .011804 . 01180409347

0 3.03870 3.03870 .94098 .94098 . 94098 . 9409795326

XI -.12984 -.12984 - . 12984 - . 1298450282
0 2.41200 2.41200
.16526 .16526 .16526 . 1652573086

1.042444 -0.3176047 -0.02380454 . 14856 .14856 .14856 . 1485175838

.4018866 1.011922 .1042594 ai=3.487517

n . 18754 . 18754 .18754 .1874503815
.20021 . 20021 .20021 . 2003244444
1. 077728 . 2194351 . 03016873

1.151729 -.02954774 . 005835219 6i=0.1411714

n« .097325 . 097325 . 097325 .09732500125

. 9396836 -.3199108 -.02481941 60 .00028458 . 00028458 .00027639 . 0002845760270

. 7458837 .03016107 . 008708692 .14998 .14998 . 14994 .1499404639

Pi .19067 .19067 .19058 . 1905807178
.9594250 - 0 . 009951160 -0.1331168
.7654931 .004267497 .1898594 02=5.448597 .19623 .19623 . 19634 .1963403800
2 1.1829418 - . 01781102 -.1355206
1.1720790 -.009187803 - . 08364038 62=0.3997728 a\ 7.0058 7.0393 7.0059 7.006740263
.8531255 .01514192 . 1163813
.7762554 . 03244676 . 3367617
-.10975 -.11477 -.10948 -.1096143529
r'
.8868953 . 1476560 . 009443967 XI -1.46564 -1.47202 -1.46502 —1.4651946170
.8689395 .1042268 .01801273 a 3 =4. 580482
- 1 . 20949 -1.21606 -1.21028 -1.2104487372
3 1.1091023 -.1643885 -.02185659
1.1265069 -.0091902 - . 004262730 63=0.3769l45
.9165367 .0633472 .01098733 -.15045 -.15188 -.12747 - . 1275876043
.9597427 -.0907231 .004390484
f n . 030400
. 085455
.029648
.084906
.081611
.018197
. 0814215368
.0184025802
.930153 .08593514 .1215308
.951447 .00050406 . 06839666 tti = 5.464933
-1.008989 .05108757 -.O3129307 r22 . 030682 .031156 . 023240 . 02324671838
4 -1.106982 -.12107954 —.1371464
.966864 -.03445640 . 006956427 64=0. 2541540
. 979853 .03607941 . 05262778 61 . 31710 .31860 . 23870 . 2388565947

. 996569 -.002365634 . 007231114 - . 10289 -.10387 - . 091679 - . 0917733357

. 988825 . 000616167 . 02354492 as=4.742589
. 991887 .002508661 .01713337
p% .090861 .090685 .12710 . 1269429981

5 1.032032 .003267702 - . 06753326 . 14768 .14772 .065039 . 0652997748

.970666 .006006834 . 06183634 65=0
1.008614 .003155791 -.01818237 a2 • .047688 .047713 12.039 12.09069098

. 999998 -.00000252
- . 10484 -.05079 . 99424 .9999886893
. 999991 -.00000084
6 1.000013 -.00002271 x% -1.46997 —1.34651 -2.99518 -3.000023179
1.000004 .00000645
- 1 . 21653 -1.38837 -1.99328 -1.999968135
. 999992 .00001636
. 999991 . 00000825
-.057616 -.058572 -.086092 -.0009108898
r% .23615 .23643 - . 19036 -.0020300857
-.18543 -.18733 .25063 .0026663150
keeping five significant figures at all times. For . 093471 . 094422 .10646 . 000012060204
comparison, the computation was carried out also
with 10 digits, using (19:2). The results are given 62 3.0287 3.0306 4.5804 . 000518791676

in table 5. In the third iteration, formula (19:1) -.36924 -.37336 -.50602 -.0009585010
gave the better result. In the fourth iteration, Vi .51134 .51126 .39181 -.0019642287
formulas (19:1) and (19:2) were equally good, and . 26185 .26035 .54853 .0027001920

superior to (19:3). The solution was also carried out as 2.9923 2.9762 . 011854 .0118007358
by the elimination method using only five significant
figures. The results are 1.00004 1.06040 1.00024 1.0000000003
Xi -3.00005 - 2 . 86812 -2.99982 -2.9999999997
-2.00006 —2.16322 -1.99978 -1.9999999993

.00064408 .00014843 .00005181 0

cg-method (19:1) Elimination
n .0014340 - . 00035647 .0000152 .00000000008
- . 0018823 .00094441 .0000364 .00000000002

X . 99424 1. 00603
y -2.99518 -3.00506
z — 1.99328 -2.00180

435
In this case the results by the cg-method and elimin- Several symmetric systems, some involving as
ation method appear to be equally effective. The many as twelve unknowns, have been solved on the
cg-method has the advantage that an improvement IBM card programed calculator. In one case,
can be made by taking one additional step. where the ratio of the largest to the smallest eigen-
This example is also a good illustration for the value was 4.9, a satisfactory solution has been
fact that the size of the residuals is not a reliable obtained already in the third step; in another case,
criterion for how close one is to the solution. In where this ratio was 100, one had to carry out fifteen
step 3 the residuals in case 1 are smaller than those steps in order to get an estimate with six correct
of case 3, although the estimate in case 1 is very far digits. In these computations floating operations
from the right solution, whereas in case 3 we are were not used. At all times an attempt was made
close to it. to keep six or seven significant figures.
Further examples. The largest system that has The cg-method has also been applied to the
been solved by the cg-method is a linear, symmetric solution of small nonsymmetric systems on the
system of 106 difference equations. The computa- SWAC. The results indicate that the method is
tion, was done on the Zuse relay-computer at the very suitable for high speed machines.
Institute for Applied Mathematics in Zurich. The A report on these experiments is being prepared
estimate obtained in the 90th step was of sufficient at the National Bureau of Standards, Los Angeles.
accuracy to be acceptable. 15
is See U. Hochstrasser,"Die Anwendung der Methode der konjugierten Gradi-
enten und ihrer Mqdifikationen auf die Losung linearer Randwertprobleme,"
Thesis E. T. H., Zurich, Switzerland, in manuscript. Los ANGELES, May 8,1952.

INDEX TO VOLUME 49

A Page
Page Compressibility of natural high polymers, effect ol moisture on, RP2349 135
Absorbency and scattering of light in sugar solutions, RP2373 365 Computing techniques, RP2348 133
Absorption spectrum of water vapor, RP2347 91 Configurations, low even, of first spectrum of molybdenum (Mo i), RP2378. _. 397
Acoustic materials, long-tube method for field determination of sound- Constitution diagram for magnesium-zirconium alloys, RP2352 155
absorption coefficients, RP2339 17 Copper, high-purity, tensil properties, RP2354 167
Acrylic plastic glazing, tensile and crazing properties, RP2369 61 Copper-nickel alloys, crystal orientation and polarized-light extinctions
Acquista, Nicolo, Earle K. Plyler, Infrared properties of cesium bromide of, RP2351 149
prisms, RP2343 61 Corrosion of galvanized steel in soils, RP2366 299
—, —, Calibrating wavelengths in the region 0.6 to 2.6 microns, RP2338 13 Corrosion, soil, of low-alloy irons and steels:
Aliphatic nitrocompounds, diphenylamine test for, RP2353 163 RP2366 299
Alloy, copper-nickel, crystal orientation and polarized light extinctions of, RP2367 315
RP2351 149 Creep, influence on tensile properties of copper, RP2354 167
Annealed and cold-drawn copper, structures and properties of, RP2354 167
Annealing optical glass, effect of temperature gradients, RP2340 21
Arc and spark spectra of rhenium, RP2355 187 D
Argon, spectrum of, RP2345 73
Attenuation index of turbid solutions, RP2373 365 Dam-break functions, hydraulic resistance effect upon, RP2356 217
Axilrod, B. M., M. A. Sherman, V. Cohen, I. Wolock, Effects of moderate Deitz, V. R., N. L. Pennington, H. L. Hoffman, Jr., Transmittancy of com-
biaxial stretch-forming on tensile and crazing properties of acrylic plastic mercial sugar liquors: Dependence on concentration of total solids, RP2373. 365
glazing, RP2369 331 Denison, Irving A., Melvin Romanoff, Corrosion of galvanized steel in soils,
RP2366 299
B —, —, Corrosion of low-alloy irons and steels in soils, RP2367 315
Deuterium and hydrogen electrode characteiistics of lithia-silica glasses,
Badger, Florence T., Leroy W. Tilton, Fred W. Fosberry, Refractive uni- RP2363 267
formity of a borosilicate glass after different annealing treatments, RP2340-_ 21 4,4'-Diaminobenzophenone, dissociation constants from spectral-absorb-
Beer's law, deviations in sugar solutions, RP2373 365 ancy measurements, RP2337 7
Benedict, W. S., H. H. Claassen, J. H. Shaw, Absorption spectrum of water Dibeler, Vernon H., Fred L. Mohler, R. M. Reese, Mass spectra of fluoro-
vapor between 4.5 and 13 microns, RP2347 91 carbons, RP2370 343
—, Earle K. Plyler, Fine structure in some infrared bands of methylene —, Mass spectra of the tetramethyl compounds of carbon, silicon, ger-
halides, RP2336 1 manium, tin, and lead, RP2358-.. 235
Brauer, G. M., 1. C. Schoonover, W. T. Sweeney, Effect of water on the induc- Digges, Thomas G., William D. Jenkins, Influence of prior strain history
tion period of the polymerization of methyl methacrylate RP2372 359 on the tensile properties and structures of high-purity copper, RP2354 167
Breckenridge, F. G., W. F. Hosier, Titanium dioxide rectifiers, RP2344 65 Diphenylamine test for aliphatic nitrocompounds, RP2353 163
Burnett, H. C, J. H. Schaum, Magnesium-rich side of the magnesium- Dissociation constants of 4,4'-diaminobenzophenone, RP2337 7
zirconium constitution diagram, RP2352 155 Dressier, Robert F., Hydraulic resistance effect upon the dam-break func-
tions, RP2356 217
c Duncan, Blanton C, Lawrence M. Kushner, James 1. Hoffman, A visco-
metric study of the micelles of sodium dodecyl sulfate in dilute solutions,
Calorimetric properties of Teflon, RP2364 273 RP2346 85
Carson, F. T., Vernon Worthington, Stiffness of paper, RP2376 385
Cements, magnesium oxychloride, heat of hardening, RP2375 377
Cements, portland-pozzolan, heats of hydration and pozzolan content, E
RP2342 - 55
Cesium bromide prisms, infrared properties, RP2343 61 Edelman, Seymour, Earle Jones, Albert London, Long-tube method for field
Claassen, H. H., W. S. Benedict, J. H. Shaw, Absorption spectrum cf water determination of sound-absorption coefficients, RP2339 17
vapor between 4.5 and 13 microns, RP2347 91 Eigenvalue problems, numerical computation, RP2341 33
Cleek, Given W., Donald Hubbard, Deuterium and hydrogen electrode charac- Electrode, lithia-silica glass, pH and pD response, RP2363 267
teristics of lithia-silica glasses, RP2363 267 Emerson, Walter B., Determination of planeness and bending of optical
Cohen, V., B. M. Axilrod, M. A. Sherman, I. Wolock, Effects of moderate flats, RP2359 241
biaxial stretch-forming on tensile and crazing properties of acrylic plastic Energy levels, rotation-vibration, of water vapor, RP2347 91
glazing, RP2369 331 Evans, William H., Donald D. Wagman, Thermodynamics of some simple
Combustors, gas turbine, analytical and experimental studies, RP2365 279 sulfur-containing molecules, RP2350... _._ 141

436