a note on systems of linear equations
TRANSCRIPT
A Note on Systems of Linear EquationsAuthor(s): David L. ElliottSource: SIAM Review, Vol. 3, No. 1 (Jan., 1961), pp. 66-69Published by: Society for Industrial and Applied MathematicsStable URL: http://www.jstor.org/stable/2027250 .
Accessed: 11/06/2014 11:15
Your use of the JSTOR archive indicates your acceptance of the Terms & Conditions of Use, available at .http://www.jstor.org/page/info/about/policies/terms.jsp
.JSTOR is a not-for-profit service that helps scholars, researchers, and students discover, use, and build upon a wide range ofcontent in a trusted digital archive. We use information technology and tools to increase productivity and facilitate new formsof scholarship. For more information about JSTOR, please contact [email protected].
.
Society for Industrial and Applied Mathematics is collaborating with JSTOR to digitize, preserve and extendaccess to SIAM Review.
http://www.jstor.org
This content downloaded from 62.122.72.94 on Wed, 11 Jun 2014 11:15:48 AMAll use subject to JSTOR Terms and Conditions
SIAM REVIBW Vol. 3, No. 1, January, 1961
A NOTE ON SYSTEMS OF LINEAR EQUATIONS'
DAVID L. ELLIOTT2
THIS NOTE IS A COMMENT on reference [1] and a generalization of the method there presented. We consider a system of m linear equations in n unknowns Xl, X X2 X* *XXnX
n (1) aijxj = ci i = 1,2,**,m,aij,c,real
j=1
or A *x = c in matrix notation. We distinguish three cases: (I) There is no finite vector x satisfying (1) (inconsistent case);
(II) There is a unique vector x satisfying (1); (III) There are an infinity of vectors x satisfying (1), such that their end-
points lie on some line, plane, or higher-dimensional linear manifold. Semarne considers Case II for m = n, but the presentation in reference [1]
requires modification, as will be shown below. Semarne's method of solving (1) combines the two classical ideas of the aug-
mented-matrix equation and Gram-Schmidt orthogonalization. (The method does not involve computing a matrix inverse to A.) As modified in this report, the method will handle any system (1) whether or not A has an inverse, furnish- ing a solution if one exists and informing us if one does not exist.
We shall use vectors (or points)
X = (X1, X2 X * Xn)
in En (euclidean n-space),
y = (yo,Yi, i *, Yn)
in E . We extend En by introducing, in addition to the xi, an artificial variable t. Multiplying (1) by t and transposing, we obtain:
n
( 1') -cit + , aijtxj = 0 i = 1, 2, * * m, j=1
or n
(2) E bijyj = O i = 1,2, * m, jaO
where
bio = -ci , bij = aij j)=1 2, * n,
and
(3) y =txj,1 <ijn,andyo=t.
1 Received by the editors August 10, 1959. 2 U. S. Naval Ordnance Test Station, Pasadena Annex.
66
This content downloaded from 62.122.72.94 on Wed, 11 Jun 2014 11:15:48 AMAll use subject to JSTOR Terms and Conditions
SYSTEMS OF LINEAR EQUATIONS 67
(This notation differs from that of reference [1].) If x is a solution of (1), then for arbitrary t (3) gives a solution of (2) lying on a line through the origin of Efn+l; (1, xi, X , * x n) is the intersection of this line with the hyperplane yo = 1. Conversely, if we find any point y* satisfying (2), then every point on the line y = sy*, - O0 < s < oc, also satisfies (2). If yo* # 0, let s* = l/yo*; for the
point y = st y, yo = 1. Then in (3) we have t = yo = 1,
(4) *Syj*= = j xjX 1 _ ji n,
and x satisfies (1 ) . If yo* = 0, the line y = sy* lies in the hyperplane yo = 0 and can have no
finite intersection with yo = 1, so no corresponding solution of (1) can be found. If yo = 0 for every y satisfying (2), we classify (1) as Case I, inconsistent. We shall give explicit forms for solutions of (2) and (1) below.
Equation (2),
B.y = 0,
where B is the m by n + 1 matrix (bij), can be interpreted geometrically in two ways. The first is that the vector y lies in the intersection of m hyperplanes (each passing through the origin). That is, the "solution" of (2) is a vector subspace, U, of E n+l: the origin itself, or some line, plane, etc., through the origin. De- fining the vector (bil , bi2 X , bin) = bt for any i = 1, 2, * * , m, each hyper- plane has equation
(2') b.y = O i = 1 2, **,m.
Then the vector bi is perpendicular to the ith hyperplane. Our second interpreta- tion of (2) is "any solution-vector y is perpendicular to the m vectors bl, * b* , and a "solution" is the subspace U which can be described as the orthogonal complement in En+l of the subspace V spanned by bl, * , bn. (That is, any point in V can be written as a linear combination of the bt.) Then we need to find a set of vectors which span U, in order to define the solution of (2) com- pletely.
From the viewpoint of the second interpretation, above, a good solution method is evident, namely, the Gram-Schmidt orthogonalization procedure. We note these facts:
a. One complete basis of orthogonal vectors for En+l is the set
I = (1, 0,...,0,0),
= (0, 1, ... ,0, 0),
{u} = {U ....................
u n= ((0, 0, ... 0, 1).
b. We can form an orthogonal basis for V, consisting of M mutually orthogo- nal vectors (M < m)f',f2, * M , from the (not necessarily independent) vectors bl, * , bm by using the Gram-Schmidt process.
This content downloaded from 62.122.72.94 on Wed, 11 Jun 2014 11:15:48 AMAll use subject to JSTOR Terms and Conditions
68 DAVID L. ELLIOTT
c. If we continue the Gram-Schmidt process by adjoining vectors-say, from {ul-which do not belong to V, we eventually will obtain a new complete orthogonal basis for En+l,
fl f2 I fM, fM+1 ... ffn+l,
and the vectors (if any) fM+1 **, fn +l will be a complete orthogonal basis for U, hence a "solution". If M = n + 1, U contains only the origin.
As Semarne points out, the beauty of this approach is that the Gram-Schmidt computation is quite simple; the basis vectors are:
f1 = bl,
f = 2 I(f ) f X
*................
1 bm \ {2 bmX - mn-1 bm
fm = - ( J) fi - J f2 f - . b fm-1 f
= bm ~~~fl .f 1 f f2 .f 2 f,;n-1 ff-
Now M ( m) of these vectors are non-zero; we may omit the others and re- number the non-zero ones as fl, * * *, fM; they span V. Now we continue com- puting until we obtain a total of n + 1 non-zero vectors:
fm - (f *) - (fM fM
= 0- (f/l) f' - (4o) fM
(5') =M+2 Ul - ( fi )+ - - I f+l /fTM1+
fl = - f - f) . - _( )
where again we have dropped all f's which are zero. The "solution" of (2) is then the space U consisting of all y such that:
(6) y= +lfM+l + ... + an+if8f, for any real a? .
We can obtain the solution of (1) as the intersection (if any) of U with the hyperplane yO = 1; that is, points of U having yO = 1 satisfy ( 1 ). We can now, from the vectors of (5'), decide which case (of the first paragraph) we have:
(I) IffoR = 0 for all thefR in (5'), U lies in the hyperplane yo = 0, and no finite solution of ( 1 ) exists.
(II) If it happens that M = n and that fo= l0+l , (6) becomes y =a we take St = 1/fol+l and write:
(I)xIf solves (1) uniqey. = (S t bl. i * * * = a)
x solves (1) uniquely. (See note below.)
This content downloaded from 62.122.72.94 on Wed, 11 Jun 2014 11:15:48 AMAll use subject to JSTOR Terms and Conditions
SYSTEMS OF LINEAR EQUATIONS 69
(III) If M < n and some foK # 0, K > AM, (2) is satisfied, from (6), by n+1
E aR f n+l (7) Y n=+l whenever E aK fo # 0.
21 aKfo K-M+1
It is evident that go = 1, so y = (1, xl, * , xn) and x solves (1) if the con- ditions of (7) are satisfied. Thus, (1) is completely solved.
NOTE. Reference [1] considers only Case II, for m = n, but the author forms fn+l not from u?, but from his vector:
en+1 (1, ,**,1 )
This vector is unsatisfactory for the following reason. In reference [1], the matrix A of (1) is assumed n-by-n non-singular; then no combination of its row vectors at can satisfy:
n ~~~~n ,8?as = 0, E8 fl O.
Then no combination of the bt of the matrix B can equal u0, so fn+l $ 0 and we obtain our solution. But if we replace u? with efn+l in the orthogonalization it may occur that, for a non-singular A, some combination of the rows of B will equal e8+'; then we get a zero solution of (2), and the solution of (1) is inde- terminate, as in the example:
2x, + X2 = -2, (8)
X1 + 0 = -1.
In such a case, reference [1] givesf3 = (0, 0, 0), whereas ourf3 = ( T7, 0), giving the solution x1 =-1, X2 = 0.
REFERENCE
1. H. MANVEL SEMARNE, New direct method of solution of a system of simultaneous linear equations, SIAM Review, 1, 1(1959), pp. 53-54.
This content downloaded from 62.122.72.94 on Wed, 11 Jun 2014 11:15:48 AMAll use subject to JSTOR Terms and Conditions