ERRORS IN LINEkR NONHOMOGENEOUS AILEBRAIC SYSTEI by ROBERT DONP.LD QEDER A TiE3IS submitted to OREGON STA TI COLLEGE in partial fulfillment of the requirements for the degree of EASTER OF SCIENCE June l9O APMtC Professor of Mathematics In Charge of Ihajor Head of Department of Mathematics Chairman of School Graduate Committee Deai of Graduate School Date thesis is presented Typed by helen C. Pfeifla May t, 19S0 ACKNWLEDGMENT To Doctor Arvid T. Lonseth, the author owes a debt of gratitude for his guidance, suggestions and constructive criticism in the preparation of this thesis. Also, for the encouragenent given to me, the writer wishes to expres s his gratitude to his w)íe, TABLE OF CONTENTS CHAPTERI. INTRODUCTION................ page i Statenientoftheproblem ..... ...... Method of Investigation An Inequality of K.O. Friedrichs Relatod Investigations . . . . CHAPTER Ii UPPER AND ILWR I , . . . . . . . . . . . . . . III. TriE ALGEBRAIC CASE ....... ....... I(k) -D(k)I ErrorinaDeterminant (k,y) Limitationof(xj-xjj 6 6 8 9 9 . Fori of solution of Algebraic Systems. Limitation of Limitation of S 5 .............. ............ ........... li DETENINANT BOUNDS FOR A Inequality of Hadamard Sufficient Condition that a Deterriinant be not zero. Positive Determinants .............. Inequality of Friedrichs . . CHA.PTiR 1 . . . .... ..... -D(k,y)I U . 12 . . 13 . . . . . . . 15 . . .16 CHAPTERIV.}UMRICALEX.pLß............,18 Example. . . . . Table of values LITuiATJt CITED. . . * . . . . * . . . ....... . . . . . . . . . . . . . . . . . . . * . . ..... 18 20 21 ERRORS IN LINEAR NONHOMOGENEOUS ALGEBRAIC SYSTEMS CHAPTER I INTRODUCTION STATENT OF THE PROBLßM. The purpose of this thesis is to consider a system of linear nonhomogeneous algebraic equations in n unknowns, x1, X2, - - - -, Xn, 31 where the n2 coefficients much the solutions X1, are subject to error, and to find by how - - - -, , of the approxinating equations j=1 x's. may differ frau the original rors in the coefficients frequently occur in the application of mathematics to natural plie- niiena, where they may appear as errors cf observation; they may also arise when decimal coefficients aro "rounded 0fft THOD OF INVESTIGATION . This problem has been solved before and partial or complete solutions may be found in papers by Etherington(3), Lonseth() and Moulton(6). The results of Lonseth and Moulton are compared with the results of this paper in a numerical example in Chapter four. The method of this paper is quite different above. fri those mentioned It follows Tricomi's solution(8) of the ar.1ogous problem with respect to a linear integral equation of Frociholm type and second kind. 2 Tricorni. - x(s) 1) considered the integral equation for x(s) K(s,t)x(t)cit = y(s), wilere K(s,t) and y(s) are known oils on the square O parameter A O 1, s functions. Also, K(s,t) is continu- i and y(s) is continuous on O s,t s 1 The is restricted to values such that the Fredholm f orrnula (9, p.211) for x(s) holds. If the kernel K(s,t) is replaced by (s,t), where (s,t) - , K(s,t)k E o, and "small," I on the square O 1, and if s,t is correctly restricted, the equation - i(s) 2) (s,t)(t)dt y(s), O s 1, Jo is solvable for 5(s) by Fredho].m's formula. Triconii derives an expros- sion which limits the maximum value of Ji(s) -x(s)I Suppose D( (s,t) and \) is , the Fred.holnì denontinator n/2 If and K Y s O xn for K(s,t), ( ) that * are upper bounds for K(s,t)J and y(s) , respectively, J and L J ?.. J(K + ), E being sorne "small" positive number, then Tricorni proves the following theorem: Tricorni' s Theorem., E and If for E is a positive nuirber such !(s,t) -K(s,t)< E that 3 then (f y Ji(s) -x(s)J ïhe coerricieflt3 XL(L)i1'(L) + La'2(L) +í1(L)c1n(L)] -II'(L) E] in the algebraic systern can always be written aj.z in the £oxa - ajj = Ii ir gij = s i { ?hen tze i. s are real mribcrs, tim the ïstems Xj-EkjXjYj 3) 1) where ijj = E j "ii ± two integral eçation ana1oes "sìil,' are algebraic E , i) arxï 2 ) , with ¿1 of the 1. it is supposed that the detei,ninants D,(k) and (k) of the two systems are not zero se that both systcs can be solved uniquely. stead of the transcendental integral functiri (j) (x) and fl (z) , In- two polyniials are used, where Q n s) ( ) = (Ïmn t ) and x((t)2() n 6) ¿p n (x) When !kjjJ<s paper rw i IEijI< be stated as e.nd I I IYil< Y, the rrincipal reu1t of this Theorem 1. (k)I 6< and is a positive number such that If il' [<E - , then (Xj - J(k)I < Coroflary 1. (a) - [IUnOc)I + E)] e If the coefficients in equation L) satisfy I (b) )(K)cpt(K+e)] + xj for i , - 1, 2, -, n; r (1 - ku)ÌT11(k) 6 where TT(k) i-1 rrki - ) - ) J then i-iYE - Xj I xii (1-i11)T7(k) Corollary 2. (a) [K "(K +) + 9)(JÇ) q'(K [1_iiirk - E + + )] 12 the coefficients in equation Li.) satisfy 1-1jl1JIjjI,fori1,2, e (b) E (K) '(K+ e(K) = (1-K)(1-2K)...(1-(n-1)K), where () ' (n-1)K then - xi - xj< ' 1; Jj(K)çt(K +) +q(K) q'(K e(K) %(K) [ -c'(K A table of values for 9)3(x), be found at the end of Chapter IV. i +e)J + L)] q"3(x), p3(x) and '3(X) villi 5 LN INEÇULLITY O? K.O. FR1Ltt1CHS.1 of an inoquality of K.O. bound for t Fridrchs certain typo or ie given in this paper It is the belief of the uality has never been published and ry L&T1) ThUSIOATIOiS. inc1uìes a proof hieh specifies a positIve lower detriant. writer that a proof of this in the This tbesi be new. Adazs(1) considers the hcxogeneou8 tra. caso, :ì() = O, of equa tions 1) nd 2) and derives bounds for the errors in the characteristic values aìd characteristic functions associated With the prcblen. Lonseth derives the inequality -x< 7) where th + fAjjI l - is the nonvanishing deterrriinant of the eyste ¿ ajjxj = i = 1, 2, , , n, . i satisíiee the systen n : (ajj + jj)xj = yj + li' Lij is the co.factor f &jj in L i , = I, 2, , n, and 4t ijl;IiiI< I il 1. 5uested :pAii j]. by Dr. A. T. Lonseth, ?rofeasor or Oregon State College. athentics, CHAPTER II UPPER AND LVER BOUNDS FOR A DETErth1INANT In order to prove Theorem i essential use is made of an inequality of Hadarnard which specifies an upper bound for a determinant. A proof is presented below. Also shown in this chapter is a sufficient condition for a determinant to be not zero and a sufficient condition for a determinant to be positive or zero. This enables the writer to prove Friedrichs inequality which cives a positive lower bound for a certain type of determinant. INEQUALITY OF HADARD. whose elements are a.1, i,j D2 TT i=l Proof ;t Theorem 2. 1, 2, If D is the determinant , n, then a k1 (, cxjx Suppose that p.3l): where Cii , caj, i , 3=1 is a positive definite quadratic forni1 arrì let whose elements are c1 cC21 . C1 ¿i . c . C22- ... . . C2 =0 ... C,_ is a polynomial of degree n in 1. be the determinant Then the determinant . c ¿ , ,P( It can be shown (2, p.171) Positive for all real values of the variables x xl = X2 = . . = Xrl O. . except for 7 P( ?) that E Ecu = -, has n positive roots, and . , and that Laking use of the fact that the geometric mean of n positive nunibers is less than or equal to the arithmetic mean, it foliovfs that n 8) ) If cil O for all i, then the form ) n n cii ii ') L__ - i,j=l is also positive definite. L L_ i,j=l Applying 3) to this form it follows that B where B 2. the determinant consisting of elements i,j = 1, 2, ¿1 --, n. Now L1l, B = c11c22 ... therefore ¿X . 1122 Cirn Now suppose that the form n n :: = i,j=l where a11 ... a ... J(ax1 + a2x + ... + ax)2, [.1 L!] a1 so that Cj1 + a2 + ... + a. Also, if D O, D = is the determinant of a positive definite form and n D2 c11c22 .., c = n Ea k1 n ,IIa2k k1 k1 or ft D2 k1 i=1 SUFFICIENT CONDITION THAT A DEThRJOENANT B NOT ZFJtO. Lengia 1. If D is the determinant whose elements are ajj, i,j = 1, laul then D II)a 7 n, and A O. Proof (7, p.672): assume . + ax a11x1 + = + ... + , i ax1 + = that D O, then the system of equations O O x, has a nontrivial solution lxii , 1, 2, , x2, ---, x1.. Let be the n, and consider the rth equation in the system: . . . + arrxr + . . . + ax = O or arrxr axk - whence n I3rrIIcr( But since XrI> jaIx.1 : , )Ar\\X. this contradicts the hypothesis. Therefore D O. POSITIVE DETIHttNTS. = eleluent$ are ajj, i,j ajj a Lemma 2 - 1, 2, If D is a determinant whose n, and , 1 i then D - O. Proof (7, p.6Th): i j. The lemma is obviously true if ajj = O for Since D is a continuous function of n2 variables, D . O by lemma 1. INEQUALITY OF FRIEDRICHS. whose elements are ajj, i,j Ja Theorem 3. 1, If D is a deterrilnant n, and , , then n-1 a11(a Proof: - 1a211) ..... a11 a12 O . D . By lemma Therefore . 2 . Ia,l) k1 Consider the determinant alI...aTh a -E (a a13...a a22-I2i) a23...a2 & a21 la21! O ... O a32 a33...a3fl + a31 a32 . . a1 a2 &3 a a .aun a the second determinant in the sum i Ian, a1 -- _i_J.. O a22-a21) a, ... a1,j - _&._i a23 ... a2n . aun T . a a13...aTh = a31 .aun -Jn a12 a1 . a3 . . 3 n2 a33...a3 n3 ...a nfl positive or zero1 lo By repeating the step outlined above to the remaining n-2 rows, the point is reached, after a finite number of steps, where a11 O >1 a12 a2-aj a13 ... a3 a23 ... O O o o O O O O a33-1a3Jj--Ia32L a2 ... n n-1 ej.A ..U. . .. a-E k1 Ea II CHAPTER III TE ALGEBRAIC CASE FORE OF SOLUTION OF ALGEBRAIC SYSTE&S. D(k) O and By Cramer' s Rule, if O, then x1-Ekjx=y 3) j=1 and = Yj, i - ) i = 1, 2, - J=l have unique solutions of the form yD :l and Xj = D(k) where l-k11 l2 l-k .. -k ... i-1 -k - andD(k) D(k) = n2 and Uj D :in - ''nn are the cofactors of - kj in l2 2l l'422 2n Cnl 4n2 "nn D(k) and 1k (k) respectively. To repeat, the purpose of this oaper is to find a bound for \ 9) X. -Xj.- (k,y) - D(k,y) _____ D(k) (k) where D1( k ,y) = EyDj and ( k ,y) kik 12 Inequality may be written as 9) I(k,1) lo) n(k)-1)n(k)I k kOc) JD(k,y)-U(k,y) (k) (k)Il To find a bound for the J maximum - value of to determine bounds for the expressions and 1D(k,y) - (k,y). bounds and cOEnpiete (k) - (k)II lL) it will be necessary ID(k,y)J, JU(k) -D(k)),JD)I The following nìteria1 will develop these the proof of Theorem 1, Corollary 1, and Corollary 2. ERROR Th A DTERThANT. whose elements are ajj + E ; IaiiV if a. and . Theorem 1. n (a) If Da(a) is a determinant is a determinant whose elements are further K, then flW2[(K + -D(a)J )fl Proof: U(a) = E a12+ E21 a22+ E22 l2 aTh+ Ein a2+E2 a11 a12+ E12 . . . aTh+ a21 a22+22 ... a2+ E2 + . . E nl an2+ E n2 a+ Eun . a1 an2n2 ayml Eun By continuing the decomposition started above, the point is reached where 13 E a12 621 a22 ... a11 ... a1,1C . + ... + Da(a) + + . s .. a 6ii + 12 e s . t Ep That is, a13 ari1 .. a t t.. n2 a113 t. a . a,_i E1 .. E C ... . E nl (a) is equal to Da(a) plus n determinants in each of which just one column of a's has been replaced by the corresponding £-coluxnn, plus n(n-1)/2 determinants in which two co1unns of a's have been replaced by Consequently, with the aid of 's, etc. Theorem 2, ¡(a)-D(a)J <. (a)-D(a)J n(n-l)..2 nfh2[n:fl n2[(k: + LI1ITATION OF 1D(k) D(k) may be expanded 2 E+...+ n(n-l) . (ri-m+1) mt ]. It can be shown (9, p.211i) -D(k)l. in the following marner. -k11 1 - . )fl .. k1k1 n D(k) = . + kj _1 Let I i . . Dn(k)1 < K , then by Theorem 2, i + Since n(n-1)2K2 nl( + (21)2 { mm/2 + e n(n-l) ... (n+1)m/2K (,)2 i we may define o0/2 to be 1, and write that iL n ' \ or, by U) fl (mt)2(n_m)) < ) V)n(c) < (-j)Oc). (k) to In the same manner, expand the determinant LE41 + + . i1 k+ E and -JJ(k)I < E < n k1 in -k1 ... n(n-l) + 2 -k .23 [(KE)2 + (21)2 )m (n_m+1)mm/2[(K - km] (mt)2 With the aid of the mean value theorem of E k+ , by Theorem + (K + )m = E m( K + E. )XTh.1 o differential calculus, write < < i (K+E.)m_Kl< Em(K+6)m to get (k) I 12) jD(k) kj kj jj - - n(n-l) jj k+ -k11 ' -k- nfn1 K E .. -k11-E11 Let )k k1+ n -D(k) et -D(k)JK n i ntmm/2 \ I E((mI)2 (n-ni) 1/)ii(K m=O\ +)m_1 is LflITATION OF (9, expanded -D(k,y). I(k,y) The determinants Dik may be p.2114) as n kjkkjj 3=1 to a finite number ji jk 3--- terris. ol' 33/2 . k41 kk k1 h- K, by Theorem 2, + (n-m+1) (m+1) (,)2 + kik k K3 (202 + k1 IkI< Since 1) K + 2nJ(2 n( n-1) 1 kik (m+1 )/2m+1 n jDikt<KE((nit)2(n_m)t J or, by 6) 13) (DikI £(k,y) Since iL') <4(K). yD if lyj< , Y, then YP(x). ID(k,y))< In a nanner analogous to that of deriving inequality 12), t ik - Dik < i Elk / n k+ Eik kj+ j=1 kjk+Ejk k3+E L to a finite number of terms. Theorem J ;.k Now ¡k ) kik k1 kik k4 * + ': K, ¿e E , so by ) !, D1K + + 2n n(n-1) ((K)2 y2}+ . . i (nm+1)(m+1)(1)/2 (t)2 )m+1 .3a±1] s . )l E mt_(i)(11»'2 (rnt)(n) )+ 16 n - Again, with the aid of the mean value theorem of write ( £)" (m+].)(K+6E )m, - Km e o K j di'ferential calculus, < i (m+l)(Ke)m Therefore n (r+1)/2\ E mo( TSik - I) (m')2(n-m)1 (m+1)(K+6)m jDik_Dikl< £j'(KE). If jjÎ< then Y, 1) J(k,y) -Lj(k,y)J< n!EP'(K+). LThITATION OF inequalities which will establish n' 11) ID(k)I 12) I(k) -D(k)I< lL) tD(k,y)I4 i) IU(k,y) -D(k,y) l) The i are, the proof of Theorem < xj. - t(K+E), nYp(K) <. and ny q'(K i- E). If < ,- substitution of 16) Ii - x e1 U), 12), iLi) and i) 6[(K)ys(i+) 1Un( [I(k)I _ in 10) results in + (K) '(K '(+ 17 which is the inequality of Theorem 1. U(k) To prove Corollary 1, assume that is a determinant that satisfies the inequality of Friedrichs, i.e. 17) (1 - k11) lrì(k) TT(k) = 77(k) , where i - k11) and that (1 - k11) < TT(k) 9'(k+e) (k) in 16) by the right hand side of 17) then replace (i) + - 18) J (1-k11) < ( k)[.)( k) to get - G ' ( ' which is the inequality of Corollary 1. (k) is a determinant that To prove Corollary 2, assume that satisfies the inequality of Friedrichs, kj< K, (n-l)K < 1, and ê( K) E <t(y + e) where 8(K) K)(l - 2K) ... (1 - (n-1)K), (1 ' then 19) (k) Now replace - ri(k) 2O)Ix_xjI< nY in 16) by the right hand side of 19) to get + %(K)[6(K)- Eq)'(K+ which is the inequality of Corollary 2. N1J1ERIG&L EXMAPL1 EXU1PI. Consider the system of equations, 3.99x1 - 2.01x2 + 2. 00x1 + 3. O1x .99x3 = i - 3. Oli3 = i + 2.00x2 + 3.99x3i Rewrite this systeni as Ii.. oc1 - 2. OO2 + 1. OQx3 = i 2.OQ1 + 3.00x2 - 3.oaE3 i 1.O1+2,OQx+1.OÇ.bc3-1 Then,n3,K3.99,Yz1,and=102. The solutions, by Craners rule, to four places are; = .3272 = .1882 X3 - .3263 - X3 - .0736 .189t and I - xii = .0009 j. Ix2 " x2J .0012 k3 .0002 Lonsetht - X31 s inequality gives, Iii-xii i; < .0052 - x2J < .OlU k33I < .0075 For n = 3, and first approximation, Ii-xiH [l+ (ouiton writes, IAiiIJ .O731 where E is the largest error in the system, vanishing determinant of the system and A is the non- is the cofactor of a1 inn. For this example, Ik-xiIi' .0052 2H .0067 Ix2 - Ix3 .00LS. X31 By Theorem 1, i e. inequality 20), of this paper, and the table on the following page, I±_xiI< .6)366,i=l, 2, 3 The example indicates that inequality 20) of this paper does not give as close a bound for the maximum value of - xii as the one derived by Lonseth, nor the approximation of 1oulton. table of values for the polynomial functions i1(x), However, if a P(x) and their derivatives were made available, the calculation of 16) would be quite easily done. The methods of Lonseth and Moulton require the calculation of the determinant of the coefficients in the approximating system in addition to n2 cofactors, in of this determinant. the maximum value of computation. Longet, and n in Moulton's, Also, the method of this paper gives a bound for j - xjj, for all i 1, 2, - -, n, in a single TABLE OF VALUES )(x) (x) X (ni (m+1)(m+1'2) =o(mt) (n-ni)t X 4)3(X) (f)13(X) /23(X) 00 1 0000 3 . 0000 o; O 0000 3.3065 3.6260 0657 .16% 2.298 3.985 .3039 Li.3039 I8S8 3.3621 !.I2O6 .3!; 1.1576 1.3309 1.2Olj 1.7269 I.9S31 2.2L01 2.15h6 .140 2.7351k . .10 .15 .20 2 .30 .145 3.O36i .50 .55 .60 .65 .70 .75 .80 .85 .90 .95 3.2603 3.7016 1.00 2.00 3.00 14.00 5.00 Li.0671 L..)3%3 L.867O 5.3029 5.76314 6.21493 6.7613 7.3000 7.8660 25.5282 59.0325 113.2253 193.0025 L.6621 .O338 1 0000 .7166 .6281 6.9921k 11.014147 1.0018 1.3h78 1.7608 2.6979 2.8162 3.L735 .228O 5.0880 6.0626 7.1609 8.3927 9.7681 11.2976 12.9921 11.5980 25.3923 114.8628 142.61482 147.01493 110.2359 203.86140 1413.2961 1115.22014 21471.1855 1105.7896 2978.9003 S.La83 5.8157 6.2261 6.6195 7.0860 7.5353 7.9977 8.L73O 8.96]J4 9.14627 9.9771 10.501414 1414.3826 68.5691 97.9518 3.5220 1O.221j3 12.1070 1L.179O 16.1479 18.9218 21.6085 2Lt.162 27.6527 31.0262 314.614147 38.5160 5314.1414140 21 LITERATURE CITED 1. Adame, Rachel 13. On the approxinate solution of Fredholru's homogeneous integral equations. American Journal of Mathematics 51: 139 - ]J.8. 1929. 2. Bocher, Maxime. Introduction to higher algebra. Macmillan, 1938. 3lSp. 3. Etherington, I.M.H. On errors in determinants. Proceedings of the Edinburgh Lathentìca1 Society 3: 107 - 117. New York, 1932. Li. Hardy, G.H., J.E. Littlewood and G. Polya. New York, Macmillan, l931. 299p. Inequalities. !. Lonseth, A.T. Systeìs of linear equations with coefficients subject to error. Annals of i.theiiatical Statistics 13: 332 - 337. 19h2. 6. Moulton, F.R. On the solutions of linear equations having sria11 determinants American Mathematical Monthly 20: 2L2 - 2L9. 1913. 7. Taussky, Olga. A recurring theorem on determinants. Mathematical Monthly S6: 672 - 675. 191.9. 8. Tricorni, Francesco. 9. Whittaker, E.T. and G.N. Watson. Macmillan, l9Li8. 59Op. American Sulla risoluzione numerica deflo equazioni integrali di Fredholm. Atti della Reale Accademia Nazionale dei Lincei, Rendiconti, Classe di Scienze Fisiche, Matematiche, e Naturali 33: 10 sexa. L83 - L86, 2° sem. 26 - 30. 1921.j. Modern analysis. New York,