On the unique factorization of a non-singular matrix by William M Lowney A THESIS Submitted to the Graduate Faculty in partial fulfillment of the requirements for the degree of Master of Science in Applied Mathematics Montana State University © Copyright by William M Lowney (1954) Abstract: This paper describes a sequential method of factoring a square non-singular matrix into two triangular matrices, The purpose of this factorization is to enable the computation of the inverse of the original matrix to he more easily coded for machine computation, for the means of finding the inverse of a triangular matrix is well known. Herein we deal with a matrix whose principal minor determinants do not vanish/,- The method of attack is to make use of pivotal condensa-. tion of determinants and also the use of extensions identities to form the sequential method of evaluation. The results ob-tained indicate the possibility of this method being used for machine computation, OE THE HHIQ1 U E EACTORIZA T IOH OE A HOE-SIHGUIAR MATRIX ■ ty Jy W I L I I M M 0' IOWEEY A THESIS SuUmitted to the Graduate Eaculty in partial fulfillment of the requirements for the degree of Master of Science in Applied Mathematics at Montana State College Jj££adj Major Department Chairman Bozeman, Montana June, 1954 L -2* 4 ^p. table of contents A B S T R A C T ...................... .. 3 INTRODUCTION .................... .. 4 FUNDAMENTAL THEOREM............. .. 5 SEQUENTIAL COMPUTATION . . . . . . . . 16 EXAMPLE................................. 20 CONCLUSIONS.............................. 22 LITERATURE CONSULTED .................. , 24 110363 ' ABBTRAGT Tlais papej? deseriloes a sequential method o'f factoring a sq%a^"@' iaoti^aixiguTa^ patsi^ triahgiilaa? Wat-ridW-o -TWd ■pWpdde Cf this factGldidatioW--id' :td ehahid the 'Cdmputatiosat h e 'itive'rse df the drigihai matrix: to he mere easily' e'eSM “f w machine CCtopatatidB9, for the means of .finding t h e 'inverse of a triangular matrix id well known» Herein we' deal with ~a . matrix whose principal minor determinants.do not vanish. ' The method of attack is to make use of pivotal condensa*. tion Of determinants and also the use of extensional identities -to-, form ,the Seqnential method of' evaluation's' The results W * t a i n e d ’indicate the- possihility of this method "being used for m a c h i n e ..computation w - The P^Qhlem o f f i n d i n g the inverse of a matrix -is one of utmost impowtaaee in ,pfa^tioaiigr all Wahshes .of applied math#*ematlsm* 'The- theOfatleal'solution Wein#' t M 'adjoint method is almost wofthissS Sjfir aottihl .hWipEf'atiohi G thW methWs ii' *$#, task InOlnde partitioningapproximation^and statlstioal aetk* bds*.- ^or p complete OmtTine Of ‘methods- m b - refer to Eors^the ts article (Paige and Tanseky9 6}0 All such methods^ however^, ^ ■fOOhire IahoriOn-S,Oo%pnt#ion and oonsidefahle- riaohin# tiaao* _ On the other Iiand5- if the matrix id- triangnlar5 ,the OOding for machine OompntatSon has "been Set np and is W l l known. -- , This paper treats of the factorIsation of & non^-singnlar matrix into conditions > two triangular ■ . matrices* e ! ', -ThS . ^ 1. . - • . imposed .i » on the Original matrix; are.rslat-iwlgr 1ESah-?.. ,Wo dsrslop a -method of sequential computation of' the elements Cf the- trian­ gular matrices such that it may he coded for electronic com­ puters. We- Shall show that the elements of Our triangular matrixoss are formed "by simple operations' on the two "by two detcrmi% rants compiled in evaluating’the determinant of the matrix by successive pivotal condensations^ Alw G Then If A f. SG we have ^ and if B and G are triangular our inverse ,Is ,read!# Iy Computed#. -5- THE E u h d a i -IEHTAL t h e o r e m ¥e take A = /a-ij) as our given non-singular matrix with the added condition that none of the first principal minor deter­ minants of any order vanishes. Ve shall factor A into two triangular matrices, B and C, such that A = BC, where B and C are of the form exhibited; 0 0 0 bjj bjj. 0 0 0 ^3} O 0 b„ 0 "bji (I) I "k/i*."bnI "bflv b nn I G/z Gli 0/y • . • • • Olfj 0 I Czl Ca y • • • • • c-t/z 0 0 1 C1Y . . • e • C Jz1 0 0 0 0 (2) 6 Then h„ C,j • e • • • • • ^ IZ ^iZ C /i- ^0 U f\z c Ztf ^iz hj, Cjl^h3-L ^3Iiz ^3Ziz cUtHii ^Z C< Om • • • • ji cZj^^3Ji Pti+^aj • • 1iz C"7+ '3Ji c-»k+ ^3Sj 9?»v BC - H i cIifH i c^ tHi * * 5T °n<" ^int^3Z 6V However, if the two matrices are equal, we know that the corresponding elements must he equal. y f / h /lr h “ KJ That is w3 ere p is the minimum of (i.j), for hy our method of construction it is obvious that b ih = O for K k , that Clfj = O for k> j , and C ic = I. We shall prove the following theorem: Theorems (4) I l« a # # « • a ir-z I^z a^tr IaZZ aU e • • • a K-Iz Ir-I I I^k 7 (5) I it # • ta^aMfrlfluf7)./ Jnll Srjl* • j?k • a frk I where we use the notation (Aitken, l); a (< . . « . • • • , a,j a//f a z/ IaV * • • • • • • ♦ Jj a Zk 1Xl * * * a Jj a ii(l 1 aJ' ajj &i* a i( a y a Lk To begin our proof, we note that b„-a„ and, in fact, that the first column of B equals the first column of A, thus b£,-atl. IIow a .j -- b„ C/j c,j = aJu' • au The solution for b £1 is as follows* a ix- C12 + b£i b,z = a £l - b £| cu = - I^ i J?j£x L aU ^ a^x - a£, a.x -8- Solving for c 2j Il • *1 a C2j ’ cU + a!j” a 2, H j J a /i I aK IJ111 a,, J Iai, Bill . I a„ B 11I To prove the theorem we shall use the first law of induc­ tion# noting that the column of V s is to he computed first. Assume the theorem true for b (j w h e r e where j - I < k. I ? j - I and C j Then we have h£ - Ia1, . . . . |a„ • • • • • • Cj., i a/fj-,/ . . .. i-i.a Iall e . • • .'’j-.,,'-! a I * I Uow consider (6) Q — {a,. . . . . »a ;-1 , I a tj( Ia„ . . . * • • a I . B y use of our extensional identities (Aitkenf I) we get the above equal to 9 K • • ia„ • Ia 11 • Ia ,, e e • a J-lJ 1 a j-',j/ eaJfyj-I / • eaJ -V z a ^ » Ia 11 • « e8lJfyj-I I /a., e a a J-Ij-JL a t j / # • • »a Z-'y j-2 I "Dividing the first row of the numerator by and the second row by |a „ . . ,Ayiii-J Ia,, # . we have: 0J-',) 1 t)£ j~( lOfIi # # # |a u . k&ii, ,f t Ia K • • . j-i .B j-Iyj-I I _ .9- lij J—.. " • a /-iyj-i I • But this becomes, upon another application of the extensional r, I ii # # # ' I'l-j-S i^j-2 J I a ll • • • • Bj-J1J-Z. I • • ay-Jyj-J a JIyjl I n • • • /a,, . 0. L j ] e and following the same procedure as above this becomes: * Ia lt • eaJ-IyJ-J./ O C • • o' p. • I Ia1, 3.0«* I 0J-SJ IfnI1 Si s s 9/i-3 ji-l cX/j I I'it • • • aJ-JyJ-*I ljSi-*- IB h e • • IaB • • • ai-j i-i a / • aJ'SJ-J1 la„ • . 'aJ-Sj-JI |a„ • • •aJ-SH - / K-I * 0J-fZj "b ^zJ-K °J-'<,j IaIZ • » 'a Z-IJ--V a J-SJ I a Lj-jI Ia1l • • 'a J-Sj-J I Ia,, • • •Bj-VfJ-V a Lj I - I \ is r ~ J-KzJ Ia II • * •a J-J-V I c J-SJ SJ-J 1Xj-, Z- -I Iaii— s— s_«a I-VjJ-V a.j/ /a„ • • • av Sj-YI la„ . . «aj-v.j-v &•ifI IaIf • • • • I - Z H SJ-K- V/'fr CJ-Kj tJ-AfyJ A=, or in general IB-a # » «aj-ifj-i a ,J I I8 It • • • • aJ-/,j-,1 I 'it > « »Rj-BjJ-H R/j I Ib ii • • • • a I — y TDtK /TM c J-KzJ 11 If we let n = j - 2 we have Q — / -I Ii i/1) /a„ a^i la„ aJ Cj'-k, j Ar2z Iall aXJi Iall a l a « a t;/ Ia,, azll a„ I 0IJ /a,, a ^1 -1 a„ la„ a,jI Qftl = - V * Z- A-=; by -k 0J-",) 2 K'l 0J-^j j-i - ati Z _ h Vj'k 0J kyJ - bt# 0U " K uJ-KyJ Uow let j - k = I and we have a LJ - 2 1 b t-i Cfj "but this is a finite sum —12— Pa ij - Z J Jt-I % s 0 Ki - i k-i * Finally replace ^ by k but this is the value of b tjil for from our original matrices a <-j = £7, 0Hj f N ' and so we see that Q c. Ia // » , . - a , __ HujL I3"!/ • • • • • aJ-Iyj-II = How we shall examine the C matrix. Istif » . Ia U • • for If we take . a IjW lJ - / Qiftcl ' a JOj-' a JjI and multiply by bjj , which we now know, we have R — la,. » # aJkjI Jalt . • •ajWyJ-' av’jI — « faIi « . , a ,11 I8*/ • • • aJ-Iyj-If I ,, . . I6 !! • ' • ' Qjfl • I Again using our identities, we have la,, . . •aJH,I IaII • • 'aZ-JyJ * aJ-IykI Ia1, • • •OfijJ-I aJyj-II la n • • la(( . • 'aJ oJ-; I . » , R j - I jj. 2 a Jkl 13- Ivlding the first row by by Ial, . • . I and the second row (a,, . . .aHjJ , we have I c which equals (&U # » #Clj-T.i2.Q*IIrI |a„ e • • • I la*i-S— S— ' (&„ Ia11 . • ' a J - L iJ - ! • • • I J-W-Z RilrL la„ • • Jatt . . *a J-J7j ) aJ7Jjf “ 0 J-I7K J-L1J4-L I • 'a J-J1J-J aJ-Z7ZrI laH • • • -J-J7J-J a j> I 0 Z llM Ia Il • • * a Jl7J 11 Ia Il • • 0Z-I1Jc » ' a J-J,/-! I and following the same procedure we have I c/-V< I' - if # # #c ^ i I b u • • S • B j-^/-51 J//-2. %_ Ia-I1 « # ia n • • • or in general, I-JlL a Jtc/ • aJ-J1/-j/ # # la,, e e • • aJ'J-'I bj,r* 0H k -14- )a,, # # #& *-/)^-/7 StipJ lau • • • • aJ-P1J-*)! jJ//-* - f k * -I;/ If we let n - j - 2, we have j-3 -Z ^j/j-Q- c>-*,k /a,, a2x/ I--I Iall H1J )a„ aik/ /a Il a JxI Ia Zf a >/ bZJ--* 0Z-Z-* 1 =1 a i/ /u A v - I I I# /a« a we/ a.i aii«/ Z/Jg Hr-* cJ-AAr 0/-*,/< 1-1 j-i aj k — a j I a < a„ ' a ;K “ aA -JI c'k ~ " b AJ i - X CH > JL-i bJjM 0Z-^zAr cJ-A- 1-/ which hy a similar method as applied to the previous case re- 15 duces to but from the matrix product j I a 11 * « »a i/1 # i a i, Iav • • • or R Ip .. . . .a , /-I a /S/i • * • . a jjl * « I-' 0 ,1 i d /Bi, • • • * ayj I c as desired. Hence, if the theorem is true for the (j-l)th column of B and the (j-l)th row of C, It vi3I be true for the jth column of B and the jth row of C. But we have seen it is true for the first and second columns of B and the first and second rows of C, and our induction is complete. 16 SEQU UTIAL COMPUTATION Ve have now proved. It la possible to factor a non-singular matrix whose first principal minor determinants do not vanish into two triangular matrices. Let us now consider the deter­ minant A from the aspect of pivotal condensation (Aitken, l). a i/7 a,? • ' B 1T (7) /Al - ax,i° S • • # e $ # # where # a „n • B y pivotal condensation we have P-u a^| I IalT Pu (a) H-. J_ (a,rrL Iaw0 aU I • S • /aC a W I• # e |a,° aJ/ a WiI • # # s # # e # • # e # • • e # • Bpil Pu aAij • # * Ia,? |a,r Let (9) - K then (8) takes form < a,ie • aVl aj. * * S/H , • • I • V * * • • 3j F t - ' , n - t a ,;/ iJ e 17 Hence, we see that (10) . a*k •••• • • e# a# *#a #-** •*, ,• 1 • • • It-KjM-U M -- JLW a* a,* . . . • s s e where ISi, k-f „ k*', t I (H) k^l a,7-1 S • r In this notation we shall see that B takes the form a,7 0 aj a,r / aJ, a* a„° • • V 2j l u a,,1 Osee . . 0 0 .. . a^e . . a,,' # # * # • a/I-I1X1e e • • • Q and that C takes the form I 0 C 0 • S • 0 0 0 17'/ «'*• /I • S l U- I 18- Hence, we see that j-i (12) "b U ^ S,L-i ,I a,rz t"-i (13) cij = a- ^ ; ^ 7 ^ Il To prove the latter true we refer to (ll) a k ij k-« I a ;>/. I<-t I 7-x a and following the extensional identities of Aitken we have k-i I a w/> I' I Ia K • • * Kkl la« • • •ak-Z,k-1 a i+Ky3+kl | - |a1( . hut (14) • *a IHjK-I I K I * • eaKk a,V = Iali * . •n Ir-Iy k a ij ' /a /, • • •aKk a LtkyJtkI • • •a Jrfc a CtkykfI I a if •= Ki a CtkyJtk) m O Iar # How from (4) l0Ukftl -- I B 11 . . •a a:K a Lt-Ic^k I-/1 )a„ . • •a Kk/ K = -a JJa Il or the same form as (12). — 19— Prom 04) we see that a Ij K 1 • • * Kk v I f A j J t * ) ~ hut from (5) ^ . l, ..i. ^ I. # # # /r Ar & /c»/ . IaIf » • • • 8- a ' or (13) as desired. a,A Hence, we see that the elements of B and C ma y he compu­ ted sequentially as a product of 2*2 determinants hy computing the Jth column of B and then the Jth row of C. We see that the kth column of B is found hy dividing the first column of the (k-l)th pivotal condensation of A hy the first prlnci al minor of the (k-2)th pivotal condensation. Likewise, the kth row of C is found hy dividing the first row of the (k-l)th pivotal condensation hy its first principal minor. 20- AH rXAMFXE Consider the matrix A S 2 3 1 0 1 2 0 0 1 3 0 1 1 2 1 1 2 1 2 1 How we have* hy pivotal condensation 1 - 1 4 |A|=b 1 2 2 2 4 9 -I 4 -I 1 1 2 1 21 2 O O O O I 3 2 I 2 O I 2 I I O O O O I -I 4 I 2 BC ? O I 2 O O O O I -3 I 2 3 -9 2 -5 25 2 O O O O I 13 25 I -I 2 O 3 -M 25 O O O O I G0Q8&B8IOBB Xn this paperg the respite 'derived, show a simple sequent tial method of factorization of a non»singiular matrix whose '' - first principal minors do not vanish=. ' i . ■■ ■ A method for computing the 'IpverSes .of the resulting triangular matrices Iihs.."been . oodea for machine computation (##a&er* Punean and collar* 10881* ' : Por computational purposes j, We note that we are evaluate ihg two. by two determinants to solve for the jth column Of B i51 and using these results to solve for the jth row of 8* wherein again we have only to evaluate two by two- determinants* Ihe significant advantage Of this method*, the# is the simple ee* quentlal method of attack=. We also, note that by weakening the conditions of our original matrix further work can be done= BOf example* change ing the non-~vanishing condition of the first principal minors would lead u&* in effect* to results obtained from interchange Ing rows by elementary operations* 1 • » The possibility Of computing the inverses of the triangu-* V Iar matrices in a similar sequential method could lead to an* other line of research* This method may have an advantage in enabling the computation of A inverse directly through proper codification* The next, condition to be eliminated could be the non* «*23** Singularity of our original matr!%* T h i s method probably could he applied to singular matrices.- ' In this case it would seem that the B matrix would have n-r hull column vectors* where r denotes the rank of the matrix. *8# LITlRAfTO CONSULTED it 1946&/ M f M I E A B T B W mmiG@8^ ' ' 2.43 p^a j,:,lnterseieiice Publishers fada g B w Y o r k b:, ■ •, 8* 1 ^ ‘ -* •> I 't ‘ t I „ n • '* » " ! * i C i I . ‘ I D&m8d8&e&,Bid.GiHd Be*dba*#t'i* R*. &98&*, THB OEARACfiR* ' ,IBfIG ROOTS Ol OERfAm-KAfRIGlS# 'BaiIlet 1 # Re8em?d'h ' LaLo^atorIee Report 814* Aberdeen ProTing Grounds M d ;» , , 3.6 * 4 ^ : * Ji ‘ IragSa?'^ R.* A»» Dmaeaiij,. W* B i m m f ARY.m t m i T O ; ^ » - y ‘ I • 4 I i i •» ' ‘ V * and G o I W % " A 0. 1938* - './ " / - ........... 416 ppff WSrerslty Prel^.f Casibrldge0. " • 1S 1 4* $ O a M s t W f H# ■* < -* * * * S ! I ‘ *\ ■ - > g and VoA H e w a n n * J*. mvmfim ', • 1J '1 a , 1#81# V , mmmiGAL Hion O m m - ..........- PrdOti A m e r 0 Mathb Soc0; £§288*BQO0 5&. M a d B u f W r O* G* 1946b T m I M O B Y W MATRICES* 410 pptir GBelbea Ptiblishlng COyfiWew Ydrky6b PaWr Jb V (WtOrA)* aa* \' 1953*. , 8I#L4 TASEOUS LIBEAR EqUATIQSS ASD THE DETERMHATIOW OE EIGEB- , VALUES$ .Batldnal Btirdati of Standards Applied' Mathematics series- 29 c • ..126 ppo-j, UtiS0 Dept* of CdSJmerddi. Ifeehti- D*., C 0r, Vo P e r l # * 8* 1968% THEORY OE MATRICES# , . ' ^ ,., ;- S37 p p b # A d d i s o h ^ e e l e y Press IntiUV Cambrlige^ Massy"8* Petrief G 0' WV '1.931*. MATRIX: IBVERSIOB ABD SOLUfIOB OB SlMULfABEOUS LIBEAR ALGEBRAIC EQUATIOBS IfIfH T O IBM fYEE 604 ELEOfROBIG OALOULAfIBG PUBCHf .: . . -, ' Pr Odeed Ings-O emputat IOn Seminar9 105-111 *' IBM Corp**; Bexi York* 9ti -• „ Sohreier* Dti/'and Sperher^ Ok MATRIX THEORY*. ' . 1931^ MODERB ALGEBRA ABD 371 ppb* Chelsea Publishing Cd06 Bew 'Ydfk* ,, "i'-'hiMiv, !" 1 110363 MONTANA STATF im r v c B c , ™ _________ 3 1762 10014733 7 - cop.2 110363 Lowney, William M On the unique factorization oi a non-singular matrix. Uamk a n B . ^ / f LU'AfWA, A B B e M e J. I V I' — /tf'— ^ /S'/p 110363 N318 L =1540 ec'.. 4.