mathematics of computation, volume 28, number 125, January, 1974

Triangular Factorization and Inversion by

Fast Matrix Multiplication*
By James R. Bunch and John E. Hopcroft

Abstract. The fast matrix multiplication algorithm by Strassen is used to obtain the
triangular factorization of a permutation of any nonsingular matrix of ordern in <Cxnlos'7
operations, and, hence, the inverse of any nonsingular matrix in <Cürtlog'7 operations.

1. Introduction. Strassen [3] has given an algorithm using noncommutative

multiplication which computes the product of two matrices of order 2 by 7 multi-
plications and 18 additions. Then the product of two matrices of order mlk could be
computed by m3lk multiplications and (5 + m)m2lk — 6(m2k)2 additions.
Let an operation be a multiplication, division, addition, or subtraction. Strassen
showed that the product of two matrices could be computed by < (4.7)«logl7 opera-
tions. The usual method for matrix multiplication requires 2n3 operations. Thus,
(4.7)n'°g*7 < 2«3 if and only if n3-'oga7 > 2.35, i.e. if n > (2.35)5 « 100.
Strassen uses block LDU factorization (Householder [2, p. 126]) recursively to
compute the inverse of a matrix of order «22*by mlk divisions, g (6/5)m37* — mlk
multiplications, and ;£ (6/5)(5 + m)m2lk — 7(m2k)2 additions. The inverse of a
matrix of order n could then be computed by ^ (5.64)«logi'7 arithmetic operations.
Let

A as Axx AX2 Axx 0 1 Axx AX2

_A2X A22_ _A2X Ax 0 A 0 /

where A = A22 — A2XAxx'1Al2, and

I Axx Ax2 An 0 / 0~
A~ =
_0 / 0 A"1 A%\ ^11 /_

Axx -- Axx Ax2 A A2XAX1 — Axx AX2 A

-A^AnAÏÎ A-1
if Axx and A are nonsingular.
Since the algorithm is applied recursively, it will fail whenever the inversion of
a singular principal submatrix in any of the reduced matrices is required.
For example, the block LDU factorization fails to exist for a matrix as simple as

Received February 26, 1973.

AMS (MOS) subject classifications(1970). Primary 15A06, 15A09; Secondary 65F05, 68A20.
Key words and phrases. LU decomposition, Strassen's method, fast matrix inversion, linear
equations, Gaussian elimination, computational complexity.
* This research was supported by ONR Grant N0O014-67-A-O077-O021.
Copyright © 1974, American Mathematical Society
231

License or copyright restrictions may apply to redistribution; see https://www.ams.org/journal-terms-of-use

 232 JAMES R. BUNCH AND JOHN E. HOPCROFT

0 0 0 1
0 0 10
0 10 0
_1 0 0 0_

Every principal submatrix in every reduced matrix is nonsingular if A is symmetric

positive definite, strictly diagonally dominant, or irreducibly diagonally dominant
(Varga [4, p. 23]). However, if A is only nonsingular, then we must, in general, pivot
(i.e., interchange rows or columns) in order to obtain a (point or block) LDU factori-
zation. If A is nonsingular, then there exist permutation matrices Px, P2, Qx, Q2
such that APX, QXA,Q2AP2have (point or block) LDU factorizations or LU factori-
zations (cf. Forsythe and Moler [1, p. 36]).
We shall ignore lower order terms. Then the usual method for inversion requires
2«3 operations [1, pp. 77-79]. Thus, (5.64)nlog!7 < 2n3 if n > (2.82)5 Srf 180.
However, one rarely needs the inverse of a matrix; rather, one usually wants to
solve systems of linear equations, and here Gaussian elimination (i.e. obtaining the
LU decomposition of a permutation of A) is more efficient [1, p. 79], since the LU
decomposition requires fn3 operations (and solving the two triangular systems
requires In2 operations).
Thus, a fairer comparison is

(5.64)n,og!7 < fn3 if n > 35(2.82)5 « 43,340.

If A is symmetric positive definite, then one should compare with symmetric

Gaussian elimination or Cholesky's method.
In Sections 2 and 3, we show that, by employing pivoting, we can use Strassen's
fast matrix multiplication algorithm to obtain the triangular factorization (LU de-
composition) of a permutation of any nonsingular matrix of order n = lk in <
(3.64)nlog*7 operations, and hence its inverse in < (ÎO.IS)«1"8'7 operations, where
an operation is defined to be a multiplication, division, addition, or subtraction.
In Section 4, we modify the algorithm so that we can find triangular factorizations
in < (2.04)nIogl 7 operations and inverses in < (5.70)nlogs 7 operations when n = 2k.
Then, for arbitrary n, we can find triangular factorizations in < (2.45)«logs 7 operations
and inverses in < (6.84)/ilog!!7operations.
In Section 5 we show that matrix multiplication, triangular factorization, and
inversion are equivalent in computational complexity.

2. The Basic Algorithm. For simplicity, let M be of order « = 2*with det M^O-
Let M° = M. We shall construct a sequence P1, P2, ■■■ , P"'1 of permutation matrices
so that M = LUP, i.e., MP'1 = LU, where P = PlP2 ■■■ P""1 is a permutation
matrix, L= LiL2 ■■• Ln~' is unit lower triangular, U is upper triangular, and det M =
(det P) det U = ± ]I"-i «•<•Since (P*')_1= P' here,

p-1 = p"-1 ... p*p\

and

License or copyright restrictions may apply to redistribution; see https://www.ams.org/journal-terms-of-use

 triangular factorization and inversion 233

M'1 = P'1 V'L'1 = P" PV U-'iL"-1)-1 'L'rwr1,

where (LT1 = 11- V.
We define the algorithm sequentially for 1 g i ^ n — 1 as follows.
Let
Bi = \j : U = X, i = i^l"-1 + /t_22*"2 + • • • + Í.2' + /„20};
let

\t if S 5¿ Í,
t = maxjy : y*G 5,}, s = min{j : j £ #¿ ) and r =
i ;' [t - 1 if S = f.
Then
Mir1 m;2

0
Aft.
M^r1
where Mxx'~ is a nonsingular upper triangular matrix of order / — 1, MX2~ is
(i - 1) X (n - i + 1), 0 is the (2r+1 - / + 1) X (i - 1) zero matrix, Af«'"1 is
(« - 2r+1) X (i - 1), M22i_1 is (h - i + 1) X (n - i + 1), and M'"1 is nonsingular.
Since 2r+1 — / + 1 > 0 and M'_1 is nonsingular, there exists a nonzero element
in the first row of M22*-1. Hence, there exists a permutation matrix P' such that
N{ m M'"1?', nu* ?¿ 0, and A7*can be partitioned as

t/1

É \ F
N* = G' \ H'

X' r
where [P is (i - T) X (i - T), V*is (i - 2') X (n - i + 2'), F and G* are 2' X 2*,
F' and H* are 2' X (« - i). 0 is the (2r+1 - i + T) X (i - T) zero matrix, X¿ is
(n - 2r+1) X (i - T), and Y" is (« - i - T) X (n - i + V). Further, U* and F are
nonsingular upper triangular.
LetZ¿ = G'iET1 and

/.-

/■•
L'
Z'

/„-._

where /, is the identity matrix of order j.

License or copyright restrictions may apply to redistribution; see https://www.ams.org/journal-terms-of-use

 234 JAMES R. BUNCH AND JOHN E. HOPCROFT

Define M' ■ (L'y'N'. Then

if F'
£* F
0
M' - f where 7' = H{ - z'f'

X' r
At the last step, U = M" 1 is nonsingular and upper triangular.

3. Operation Count. Finding the permutation P' requires at most n — i com-

parisons, and, if Pu = 0, then the permutation involves n element interchanges.
Hence, at most n(n — l)/2 comparisons and at most n'n — 1) element interchanges
are required to obtain M = LU P. The computation of M"1 would require at most
an additional n'n — 1) element interchanges:
Let an operation be a multiplication, division, addition, or subtraction. Let
M'ri), MT(ri), and IT(ri) be the number of operations required to multiply two « X «
matrices, to multiply an n X « matrix by an upper triangular « X « matrix, and to
invert an « X « nonsingular upper triangular matrix (we shall ignore lower order
terms). Then M(l) = 1 and M(2*) = 7*+1for k ^ 1. Since Mr(2*) = 4MT(2k~1)+
2MÍ2*"1) + 2""1 and IAX) = 2/T(2*-1) + IM Al"'1),

Mri2k) = 2 Z 4,M(2*-'-1) < (yj/

and

ITi2k)= 2 Z 2'Mr(2*-'-1)< (11)7*.

Inverting all the U* for 1 t% i ^ n — 1 requires

operations. Forming all the multipliers Z* for 1 1%i á « — 1 requires

1
2"-1 Z t}j Mt'2¡)
i-a ¿
*<Ê>
operations. Forming all the reduced matrices J' for 1 ^ / g n — 1 requires
t-l r<n-2>')/2' +'
1)2'],^^MX)
z\ z I2k -Í21+
2''
22<t-u £ Af(2')

-a>"§(0'<-I^G)XD-
Hence, for n = 2*, triangular factorization requires < (91/25)7* = (3.64)nlog'7
operations.
Inverting U requires < (28/15)7* operations and U~1L~1 requires

License or copyright restrictions may apply to redistribution; see https://www.ams.org/journal-terms-of-use

 TRIANGULAR FACTORIZATION AND INVERSION 235

^§^-(i)^§(7)'<G2W7)'(D»(^
operations.
Hence, for n = 2", inversion requires (763/75)7* < (10.18)nlogï7 operations.
If M is a nonsingular matrix of order n, where 2* < n < 2*+1, then let 3TC=
M © I2k+,_„. We can find the triangular factorization of a permutation of 2(11,and
hence of a permutation of M, by < (91/25)7*+1 = (637/25)7* < (25.48)nlog*7opera-
tions, and the inverse of 9TC,and hence of M, by < (763/75)7*+1= (5341/75)7* <
(71.22)n'°g*7.

4. A Modified Algorithm. We can modify the algorithm in Section 2 so that the

coefficient of nlogî7 is smaller in the operation counts of Section 3. In particular, we
find m and k such that the number of operations is minimized subject to the constraint
n g mlk.
First, let n = T = ml". Then m = T~k = 1', and M(2r) < (5 + 2w)w27* =
/(s)7r, where f(s) = (5 + 2m)m27"' = (5 + 2*+I)22*7_*.Since minos.s, f(s) = /(3) =
192/49, we take m = 8, k = r — 3, and use regular multiplication and inversion
for submatrices of order ^8. Then, M(2r) ^ (192/49)7r for r ^ 0 (rather than
M(2r) ^ (7)7r in Section 3). Hence, each coefficient in Section 3 is multiplied by
(1/7X192/49).
Triangular factorization requires

< (l)(I)7' < <2o4>"'-'

operations, and inversion requires

<(fXI)7'<*■"*-■
Now let n be arbitrary. Taking k = [log2 n — 4] and m = [w2~*]+ 1 (cf. [3]),
we have n ^ w2* and (5 + 2m)m2T < (4.7)«logl7.
Now,

MTim2k) < 2(5 + 2m)m2lk-x Z (j) < \ (5 + 2m)m2lk

and

ITim2k) ^ ^ (5 + 2m)m27*-1 Z (f)' < ^ (5 + 2m)m2lk.

Triangular factorization thus requires < (39/75X5 + 2m)m2T < (2.45)«"""7

operations, and inversion requires < (109/75)(5 + 2m)m2lk < (6.84)nlogï7 operations.

5. Remarks. As seen above, the coefficient of nlogl7 is very sensitive to the

implementation of the algorithm. Another modification of the algorithm might
reduce the coefficient. Further, the bounds we have given on the coefficient are
pessimistic.
The algorithm as stated in Sections 2 and 4 may not be numerically stable since

License or copyright restrictions may apply to redistribution; see https://www.ams.org/journal-terms-of-use

 236 JAMES R. BUNCH AND JOHN E. HOPCROFT

we cannot guarantee that the elements in the reduced matrices are bounded. However,
there may be a modification of our algorithm which guarantees stability; this question
deserves further investigation.
If a fast matrix multiplication algorithm were given for multiplying two matrices
of order u in v multiplications where log„ v > 2, then algorithms similar to those in
Sections 2 and 4 could find the triangular factorization of a permutation of any
nonsingular matrix, and hence the inverse of any nonsingular matrix, in < cnlog°°
operations. The algorithms would be expressed in terms of the expansions of integers
modulo u and hold over any ring.
Similarly, given an algorithm for multiplying two n X n matrices in 0(n") opera-
tions, then algorithms similar to those in Sections 2 and 4 can find the triangular
factorization of a permutation of any n X n nonsingular matrix in 0(n") operations,
and hence the inverse of any n X n nonsingular matrix in 0(n") operations, where
2 < a g 3. Conversely, given an algorithm for calculating the inverse of an n X n
nonsingular matrix in ^cn" operations, the product of two n X n matrices A, B
can be computed in ^(3ac)na operations, since
/ A ol"1 (~7 -A AB
0 I B =0 I -B
.o o i\ Lo o /.
(Winograd [5]).
Thus, matrix multiplication, triangular factorization, and inversion are equivalent
in computational complexity.

Addendum. A. Schönhage has proposed another method for forming the inverse
of any nonsingular matrix (over the reals or complexes), based on the generalized
inverse identity M"1 = (M*M)~1M*. Since M*M is symmetric positive definite,
Strassen's inversion algorithm will not fail (but the condition number [1, pp. 20-26]
of the problem will be squared). Strassen's matrix multiplication algorithm is used
to form M*M and to multiply (M*M)~1 by M*. This gives the inverse of any non-
singular matrix in ^cnios'7 operations (but the constant c is larger than the ones
given above).

Department of Computer Science

Cornell University
Ithaca, New York 14850

1. G. E. Forsythe & C. B. Moler, Computer Solution of Linear Algebraic Equations,

Prentice-Hall, Englewood Cliffs,N.I., 1967. MR 36 #2306.
2. A. S. Householder, The Theory of Matrices in Numerical Analysis, Blaisdell, New
York, 1964. MR 30 #5475.
3. V. Strassen, "Gaussian elimination is not optimal," Numer. Math., v. 13, 1969, pp.
354-356. MR 40 #2223.
4. R. S. Varga, Matrix Iterative Analysis, Prentice-Hall, Englewood Cliffs, N.J., 1962.
MR 28 #1725.
5. S. Winograd, Private communication.

License or copyright restrictions may apply to redistribution; see https://www.ams.org/journal-terms-of-use