Applied Mathematics Letters 20 (2007) 1189–1193

www.elsevier.com/locate/aml

A note on inversion of Toeplitz matricesI

Xiao-Guang Lv, Ting-Zhu Huang ∗
School of Applied Mathematics, University of Electronic Science and Technology of China, Chengdu, Sichuan, 610054, PR China

Received 18 October 2006; accepted 30 October 2006

Abstract

It is shown that the invertibility of a Toeplitz matrix can be determined through the solvability of two standard equations. The
inverse matrix can be denoted as a sum of products of circulant matrices and upper triangular Toeplitz matrices. The stability of
the inversion formula for a Toeplitz matrix is also considered.
c 2007 Elsevier Ltd. All rights reserved.

Keywords: Toeplitz matrix; Circulant matrix; Inversion; Algorithm

1. Introduction

Let T be an n-by-n Toeplitz matrix:

a a−1 a−2 · · · a1−n 
0
a1 a0 a−1 · · · a2−n 
a a a0 · · · a3−n  ,
 
T = 2 1
.. .. .. . . .. 
. . . . . 
an−1 an−2 an−3 · · · a0
where a−(n−1) , . . . , an−1 are complex numbers. We use the shorthand
T = (a p−q )np,q=1
for a Toeplitz matrix.
The inversion of a Toeplitz matrix is usually not a Toeplitz matrix. A very important step is to answer the question
of how to reconstruct the inversion of a Toeplitz matrix by a low number of its columns and the entries of the original
Toeplitz matrix. It was first observed by Trench [1] and rediscovered by Gohberg and Semencul [2] that T −1 can be
reconstructed from its first and last columns provided that the first component of the first column does not vanish.
Gohberg and Krupnik [3] observed that T −1 can be recovered from its first and second columns if the last component
of the first column does not vanish.

I Supported by NCET-04-0893 and applied basic research foundations of Sichuan province (05JY029-068-2).
∗ Corresponding author.
E-mail address: tzhuang@uestc.edu.cn (T.-Z. Huang).

c 2007 Elsevier Ltd. All rights reserved.

0893-9659/$ - see front matter
doi:10.1016/j.aml.2006.10.008
1190 X.-G. Lv, T.-Z. Huang / Applied Mathematics Letters 20 (2007) 1189–1193

In Heinig and Rost [4], an inversion formula was exhibited for every nonsingular Toeplitz matrix. The method
requires the solution of linear systems of equations (the so-called fundamental equations), where the right-hand side
of one of them is a shifted column of the Toeplitz matrix T . In [5], Ben-Artzi and Shalom proved that three columns
of the inverse of a Toeplitz matrix, when properly chosen, are always enough to reconstruct the inverse. Labahn and
Shalom [6], and Ng, Rost and Wen [7] presented modifications of this result. In [8], Georg Heinig discussed the
problem of the reconstruction of Toeplitz matrix inverses from columns.
In this work, we give a new Toeplitz matrix inversion formula. The inverse matrix T −1 can be denoted as a sum of
products of circulant matrices and upper triangular Toeplitz matrices. The results obtained show that this formula is
numerically forward stable.

2. Toeplitz inversion formula

Lemma 1. Let T = (a p−q )np,q=1 be a n × n Toeplitz matrix; then it satisfies the formula

K T − T K = f enT − e1 f T J,

where
 
0 1  
.. 1
.
 
1 .
, ,

K = J = ..
 
.. ..
. .
 
1
 
1 0
    0 
1 0
0  ..  an−1 − a−1 
.
e1 =  .  , en =  .  , .. .

f =
   
 ..  0  
a2 − a−n+2
 
0 1
a1 − a−n+1

Theorem 1. Let T = (a p−q )np,q=1 be a Toeplitz matrix. If each of the systems of equations T x = f, T y = e1 is
solvable, x = (x1 , x2 , . . . , xn )T , y = (y1 , y2 , . . . yn )T , then

(a) T is invertible;
(b) T −1 = T1 U1 + T2 U2 , where
   
y1 yn · · · y2 1 −xn · · · −x2
. . .. .. 
. . . 
  
 y2 y1
,
  1 ,
T1 = 
.. U =
. . 1
.

. .. .. y  
..
 
n
 −xn 
yn · · · y2 y1 1
   
x1 xn · · · x2 o yn · · · y2
. .
. . ..  .. .. 
. . 
  
x2 x1 o .

T2 = 
.
 and U2 = 
.. .. .. ..
. . xn  .
  
 yn 
xn · · · x2 x1 o

Proof. From Lemma 1 and T x = f, T y = e1 , we have

K T = T K + f enT − e1 f T J
= T [K + xenT − y f T J ],
 X.-G. Lv, T.-Z. Huang / Applied Mathematics Letters 20 (2007) 1189–1193 1191

then,

K i T = K i−1 T [K + xenT − y f T J ]
= ···
= T [K + xenT − y f T J ]i .

Therefore,

K i e1 = K i T y = T [K + xenT − y f T J ]i y.

Let

ti = [K + xenT − y f T J ]i−1 y and T̂ = (t1 , t2 , . . . , tn ).

Then

T ti = T [K + xenT − y f T J ]i−1 y = K i−1 e1 = ei ,

T T̂ = T (t1 , t2 , . . . , tn ) = (e1 , e2 , . . . , en ) = In .

So the matrix T is invertible, and the inverse of T is the matrix T̂ = T −1 .

For (b): First of all, it is easy to see that

t1 = y, ti = [K + xenT − y f T J ]ti−1 (i = 1, 2, . . . , n),

ti = T −1
ei , J ei = en−i+1 ,
JT J = T , T
J J = I, J T = J.

Then, for i > 1

ti = K ti−1 + xenT ti−1 − y f T J ti−1

= K ti−1 + xenT T −1 ei−1 − y f T J T −1 ei−1
= K ti−1 + xenT J J T −1 J en−i+2 − y f T T −T J ei−1
= K ti−1 + xe1T T −T en−i+2 − y f T T −T en−i+2
= K ti−1 + x y T en−i+2 − yx T en−i+2
= K ti−1 + yn−i+2 x − xn−i+2 y.

So we have

t1 = y, t2 = K y + yn x − xn y,
...,
tn = K n−1 y + K n−2 x yn − K n−2 yxn + · · · + x y2 − yx2
T −1 = (t1 , t2 , . . . tn )
   
1 −xn ··· −x2 o yn ··· y2
.. ..  .. .. 
. .  . . 
 
1 o
= (y, K y, . . . , K n−1  + (x, K x, . . . , K n−1 x) 
 
y) 
.. ..

. .
   
 −xn   yn 
1 o
     
y1 yn ··· y2 1 −xn · · · −x2 x1 xn · · · x2 o yn ··· y2
.. . . ..  . .
. . ..   .. .. 
. . .  . . 
    
 y2 y1  1  + x2 x1
 o .
= 
..  .. 
.. .. .. . . ..
 
. . . . −xn  . . . . . xn   .
  
yn   yn 
yn ··· y2 y1 1 xn · · · x2 x1 o
1192 X.-G. Lv, T.-Z. Huang / Applied Mathematics Letters 20 (2007) 1189–1193

Remark. In Theorem 1, let the Toeplitz matrix T = (a p−q )np,q=1 be a circulant Toeplitz matrix. That is to say, the
elements of the matrix T = (a p−q )np,q=1 satisfy ai = ai−n for all i = 1, . . . , n − 1. It is easy to see that f = 0. Thus,
x = T −1 f = 0. From (b) of Theorem 1, we get
 
y1 yn · · · y2
 y2 y1 . . .
 
−1 .

T = .
. .

.. .. .. y 
n

yn · · · y2 y1
We conclude: The inverses of the circulant Toeplitz matrices are circulant Toeplitz matrices.

3. Stability analysis

In this section, we will show that the Toeplitz inversion formula presented in Section 2 is evaluation forward stable.
An algorithm is called forward stable if for all well conditioned problems, the computed solution x̃ is close to the true
solution x in the sense that the relative error kx − x̃k2 / kxk2 is small. In the matrix computation, round-off errors
occur. Let A, B ∈ C n,n and α ∈ C. If we neglect the O(ε2 ) terms, then for any floating-point arithmetic with machine
precision ε, then (cf. [9])
√
fl(α A) = α A + E, kEk F ≤ ε |α| kAk F ≤ ε n |α| kAk2 ,
√
fl(A + B) = A + B + E, kEk F ≤ ε kA + Bk F ≤ ε n kA + Bk2 ,
fl(AB) = AB + E, kEk F ≤ εn kAk F kBk F .
According to the floating-point arithmetic, we have the following bound.

Theorem 2. Let T = (a p−q )np,q=1 be a nonsingular Toeplitz matrix and be well conditioned; then the formula in
Theorem 1 is forward stable.

Proof. Assume that we have computed the solutions x̃, ỹ in Theorem 1 which are perturbed by the normwise relative
errors bounded by ε̃,
kx̃k2 ≤ kxk2 (1 + ε̃), k ỹk2 ≤ kyk2 (1 + ε̃).
Therefore, we have
√ √
kT1 k F = n kyk2 , kT2 k F = n kxk2 ,
√ q √
kU1 k F ≤ n 1 + kxk22 , kU2 k F ≤ n kyk2 .

Using the perturbed solutions x̃, ỹ, the inversion formula in Theorem 1 can be expressed as
 
T̃ −1 = f l T̃1 Ũ1 + T̃2 Ũ2
= f l ((T1 + ∆T1 )(U1 + ∆U1 ) + (T2 + ∆T2 )(U2 + ∆U2 ))
= T −1 + ∆T1 U1 + T1 ∆U1 + ∆T2 U2 + T2 ∆U2 + E + F.
Here, E is the matrix containing the error which results from computing the matrix products, and F contains the error
from subtracting the matrices. For the error matrices ∆T1 , ∆U1 , ∆T2 and ∆U2 , we have
√
k∆T1 k F ≤ ε̃ kT1 k F = ε̃ n kyk2 ,
√
k∆T2 k F ≤ ε̃ kT2 k F = ε̃ n kxk2 ,
√ q
k∆U1 k F ≤ ε̃ kU1 k F ≤ ε̃ n 1 + kxk22 ,
√
k∆U2 k F ≤ ε̃ kU2 k F ≤ ε̃ n kyk2 .
 X.-G. Lv, T.-Z. Huang / Applied Mathematics Letters 20 (2007) 1189–1193 1193

It follows that
kEk2 ≤ kEk F
≤ εn(kT1 k F kU1 k F + kT2 k F kU2 k F )
q 
≤ εn 2 kyk2 1 + kxk22 + kxk2
√
≤ εn 2 kyk2 (1 + 2 kxk2 ), kFk2 ≤ ε n T −1 .

2
Adding all these error bounds, we have
√
T̃ − T −1 ≤ n(2ε̃ + nε) kyk2 (1 + 2 kxk2 ) + ε n T −1 .
−1
2 2
Note that T y = e1 and T x = f ; then

kyk2 ≤ T −1 and kxk2 ≤ T −1 k f k2 .

2 2
Thus, the relative error is

−1
T̃ − T −1

  √
2 ≤ n(2ε̃ + nε) 1 + 2 −1
k f k 2 + ε n.
T −1 T
2
2

As T is well conditioned, thus, T −1 2 is finite. Obviously, k f k2 is finite. Therefore, the formula presented in
Theorem 1 is forward stable. 

References

[1] W.F. Trench, An algorithm for the inversion of finite Toeplitz matrix, J. SIAM 13 (3) (1964) 515–522.
[2] I. Gohberg, A. Semencul, On the inversion of finite Toeplitz matrices and their continuous analogues, Mat. Issled. 7 (1972) 201–223 (in
Russian).
[3] I. Gohberg, N. Krupnik, A formula for the inversion of finite Toeplitz matrices, Mat. Issled. 7 (12) (1972) 272–283 (in Russian).
[4] G. Heinig, K. Rost, Algebraic methods for Toeplitz-like matrices and operators, in: Operator Theory: Advances and Applications, vol. 13,
Birkhäuser, Basel, 1984.
[5] A. Ben-Artzi, T. Shalom, On inversion of Toeplitz and close to Toeplitz matrices, Linear Algebra Appl. 75 (1986) 173–192.
[6] G. Labahn, T. Shalom, Inversion of Toeplitz matrices with only two standard equations, Linear Algebra Appl. 175 (1992) 143–158.
[7] M.K. Ng, K. Rost, Y.-W. Wen, On inversion of Toeplitz matrices, Linear Algebra Appl. 348 (2002) 145–151.
[8] G. Heinig, On the reconstruction of Toeplitz matrix inverses from columns, Linear Algebra Appl. 350 (2002) 199–212.
[9] S. Zohar, The solution of a Toeplitz set of linear equations, J. ACM 21 (1974) 272–276.