See

discussions, stats, and author profiles for this publication at:

http://www.researchgate.net/publication/222738373

A theory of pseudoskeleton
approximations
ARTICLE in LINEAR ALGEBRA AND ITS APPLICATIONS AUGUST 1997
Impact Factor: 0.98 DOI: 10.1016/S0024-3795(96)00301-1

CITATIONS

DOWNLOADS

VIEWS

159

52

126

3 AUTHORS, INCLUDING:
E. E. Tyrtyshnikov
Russian Academy of Sciences
99 PUBLICATIONS 1,738
CITATIONS
SEE PROFILE

Nickolai Zamarashkin
Russian Academy of Sciences
24 PUBLICATIONS 385
CITATIONS
SEE PROFILE

Available from: E. E. Tyrtyshnikov

Retrieved on: 28 July 2015

A Theory of Pseudoskeleton Approximations

S. A. Goreinov,

E. E. Tyrtyshnikov,

Institute of Numerical
Russian Academy
Ul. Gubkina

and N. L. Zamarashkin

Mathematics

of Sciences

Moscow 117333,

Russia

Submitted by Richard A. Bru&

ABSTRACT
Let an m X n matrix A be approximated by a rank-r matrix with an accuracy E.
We prove that it is possible to choose r columns and r rows of A formin a so-called
pseudoskeleton component which approximates A with B<&<&
+ $ n )) accuracy
in the sense of the e-norm.

On the way to this estimate

we study the interconnection

between the volume (i.e., the determinant in the absolute value) and the minimal
singular value q of T x r submatrices of an n X r matrix with orthogonal columns.
We propose

a lower bound (better

by Hong and Pan) for the maximum

hypothesis

1.

on a tighter

bound.

than one given by Chandrasekaran

of o, over all these submatrices

0 Elsevier

Science

Inc.,

and Ipsen and

and formulate

1997

INTRODUCTION

Many applications
approximated

involve coefficient

by low-rank

tions can be computed,

matrices.

matrices whose blocks can be easily

Such block low-rank

for example,

by partial

SVD

matrix approxima-

algorithms

and may

underlie construction
of many efficient numerical methods. Partial SVD
algorithms typically require all entries of a block to construct a block low-rank
approximation.

However,

we may anticipate

dramatical

savings in memory

*E-mail: [email protected].

LINEAR ALGEBRA AND ITS APPLICATIONS 261: 1-21 (1997)

0 Elsevier Science Inc., 1997
655 Avenue of the Americas, New York, NY 10010

0024-3795/97/$17.00
PII SOO243795(96)00301-1

S. A. GOREINOV

and arithmetic

if it were possible

part of its entries.

construction
Let

to approximate

a block using only a small

In this paper we present a theory which lays the ground for

of such algorithms.

A E [WmXn, and assume that

rank A =T.

The;

ET AL.

there exists a nonsingular

(1.1)

r X r sybmatrix

A^ in A. If A^ lies in rows

i E Z = {iI,. . . , i,) and in columns j E J = (ji, . . . , j,},

we will write

A^=A(f,j).

(1.2)

A = CkR,

(14

It is easy to verif>l that

where

C=

A(Ljj,

z = {l,...,m},

=A(fJ),
J =

{l,...,n).

(1.4)

(1.3) is known as a skeleton decomposition of A.

Now let us suppose that rank A = r means that ranks A + E) = r, where
E = 0 in the sense of a prescribed matrix norm. The exact equality rank A = r

The decomposition

implies the exact equality A = CA-R,

and we wonder if the approximate
equality rank A = r may imply the approximate equality A = CA-iR.
Using a rather

standard

technique,

we can prove that for E sufficiently

small

IIA -

CA^+Rllz
= @(tIAII;IlA-%),

(1.5)

,.

provided that A is nonsingular (see A[4]). Consequently,

we cannot state in
general that A S= CA-R,
because A,may be ill conditioned or even singular.
We can be better off replacing A - with a more suitable square matrix
G, thus trying to approximate A by a matrix B = CGR of rank r or less. Any
matrix of the form B = CGR will be called a pseudoskeleton
component
of A.
The main purpose of this paper is to prove that whenever

A is approximated by a rank-r matrix with accuracy E, it can be as well a roximated by

its pseudoskeleton
component with accuracy @(J&(&K
+ F n >).

PSEUDOSKELETON

APPROXIMATIONS

The choice of the proper pseudoskeleton

choice

of C,

submatrix
Section

R, and G) has much

with the best-bounded

2. Therein

component

(in other words, the

to do with the problem

inverse.

This

we improve the recently

problem

proposed

nontrivial

the 2-norm value of the inverse of such a submatrix [l,

absolutely trivial (and does not need the CS-decomposition
propose

also a hypothesis

discussion

on a tighter

of the interconnection

estimate.

between

of seeking

is considered
estimate

in
for

31. Our proof is

arguments). We

We close the section with a

the volume (i.e., the determinant

in the absolute value) and the minimal singular value uY of r X r submatrices of an 12 X + matrix with orthogonal columns.
In Section 3 we derive approximation estimates
decomposition
estimates

which do not involve the norms of A and of A^-. Also, these

do not use the assumption

these estimates
consider

and R, specifically

small. Therefore,

the approximation
We

consider

example which is chosen to illustrate

of pseudoskeleton

here

a matrix

conducting

may require

SVD techniques.

a numerical

capabilities

for a perfectly

approximation

theory. We

only from the entries of C

from the entries of A^.

In Section 4 we present

problem

that E is sufficiently

differ from the ones typical of small-perturbations

also the case when G may be constructed

application.

for the pseudoskeleton

components

arising

from

the

for a nontrivial
3D

scattering

sphere. We note that the pseudoskeleton

far less memory

and arithmetic

It allows one not to compute

than the partial

all the matrix entries,

working

instead only with small part of them.

In Section

2.

5, a brief discussion

SUBMATRICES

WITH

THE

of the results is given.

BEST-BOUNDED

From (1.5) we see that a pseudoskele;on

enough

approximation

component

understand
estimates

whether

provides an accurate

if the value of 11A- 11is not very large. An optimiza-

tion of this norm value by making a proper

rows can improve

INVERSES

choice

of the columns

the quality of the approximation.

an optimization

to

u priori

of the norm value of the submatrix with the best-bounded

inverse.

first a model (and apparently

A=

is feasible,

and the

in order

we need

Consider

such

However,

the simplest)

UT

case

(2.1)

0 1

where

UTU =

I,

U E

R,

r =g n.

(2.2)

S.A. COREINOV ET AL.

4
Denote by 4(U)

the set of all r X r submatrices in U, and set

t(r, n) =

min
u Psi,

%in()

(2.3)

where q,i,(P> denotes the minimal singular value of P.

In order to derive an upper estimate of t(r, n>, consider such a submatrix
P^E&(U) that has the maximal volume (i.e., the maximal absolute value of
the determinant). Due to the Binet-Cauchy formula we have
1 = det(UrU)

c
P4W)

(det P).

Since the number of summands equals CL = n!/[r!(n

that

- r)!], we conclude

(det @) > l/C,.

Let a,(i)

> **a > or( P^>be the singular values of P^.Then

(det P^) = a:(i)

. ..@(F)

< a,2(Pj[cri@)]r-1

Q a,($)

for c+,(P) < 1. Hence,

However (thanks to the results from [l, 3]), a much finer estimate for
a;(i) can be derived. We present below an even better estimate (with an
absolutely trivial proof).

LEMMA 2.1.

t(r, n) dejined by (2.3) satisfies

t(r,n)

~(72 -r)

+ 1.

(2.4)

Proof. Without loss of generality we may assume that the submatrix P^

of the greatest possible volume resides in the first r rows. It is then evident

PSEUDOSKELETON
that the submatrix

APPROXIMATIONS

of maximal volume in the matrix

I
[V1

fj=/g-l=

is located in the same r rows. If ICIjJ > 1, then by swapping the ith and the
jth rows we could obtain on the first r rows a submatrix whose volume was
greater than 1. This contradicts

the choice of P^.We thus have

and hence

t(r,n)
The equality in the estimate
cases: r = 1 and r = n -

<

r(n

of Lemma

-r)

?? (2.5)

+ 1.

2.1 is attained in the two extreme

1. We believe that for 1 < r < n -

1 this estimate

is not sharp. We do not know any matrix for which the inequality

is violated. However,
The orthogonal

so far (2.6) is still a hypothesis.

matrix

0
illustrates that the estimate (2.6) indeed is almost sharp. Moreover, there are
infinitely many values of r and 72, I < r < n - 1, for which (2.6) cannot be
improved.
LEMMA 2.2.

Zf r- is a positive integer and n is divisible by r + 1, then

S. A. GOREINOV

Proof.
that

Let Q = [A b] E

ET AL.

Rr+ I) cr+ ) be any orthogonal matrix such

We construct the CS decomposition of Q which corresponds to the diagonal

blocks of orders r and 1, and we can easily see that if P E RrXr resides in
the first r rows of A then

Repeating the same arguments with different permutations of rows of Q, we

conclude that the above equality holds true for any P EJ( A). Now, let

The required inequality follows from the observation that

We close the section by noting that the choice of the maximal-volume

submatrix ensures only that its minimal singular value is sufficiently large.
This minimal singular value is not guaranteed to be the greatest possible one
among all submatrices.
For example, consider an n X 2 matrix U of the form
T

...
[0
E

ZE

_c2

E2

Assume that
n > 3.

-82

E2

...

-82

I.

PSEUDOSKELETON

APPROXIMATIONS

and set
3

l/3

( i

,y=

n-3

Then the two columns

of U are orthogonal.
possible volume of all 2 X 2 submatrices

For large n the greatest

equal to .s and there are exactly three such submatrices.

These

of U is

submatrices

are located in the first three rows.

After normalizing
orthogonal

matrix

the columns

C in which

of U by the scalars d, and d,, we get the

the same

three

submatrices

still have the

maximal volume. It can be easily seen that their minimal singular values differ
in the general case.
For instance,

For any

the 2 X 2 submatrices

d, > 0 the minimal singular value of A, is greater than that of A,.

Obviously,
preserve

consider

we can construct

small perturbations

the norms and the orthogonality

of a special form which

of the two columns

of C, do not

change the entries of A,, and decrease the volume of A,. Sufficiently small
perturbations will preserve as well the inequality between minimal singular
values of submatrices

A, and

A, while the three

above still contain the submatrix

3.

MAIN

submatrices

considered

of maximal volume.

RESULTS

We now realize that a matrix

A of the form

has an (r - 1) X (r - 1) submatrix with a sufficiently well-bounded

In the general case the situation may get much worse; for example,
the matrix

A=

VT
0

...

0,

v?cr

[WnX(r-l)

inverse.
consider

S. A. GOREINOV

The minimal singular value of the unique nonsingular

not greater

than

components

with r columns

obviously

come

approximation

E. We want to estimate
to nothing

properties

submatrix of order r is

the accuracy

of pseudoskeleton

and r rows for an c-perturbation

here

when

using

of pseudoskeleton

ET AL.

(1.5).

of A. We

Nevertheless,

components

certain

hold true even in

this case.

Amme that A, F E [WmXn,rank(A

THEOREM 3.1.

- F) < r, and jlF[lg

< E for some E > 0. Then there exist r columns and r rows which determine
a pseudoskeleton

component

CGR such that

IA - CGRllz<
Consider

Proof.

~(1

+ [dm

where Z = diag(cr,,...,u,),

ye

rzrl >

= UCV,

(3.2)

**a z ur > 0, UTU = WT

fi, G E Rrx r of U, V, respectively,

submatrices,

Ifi-II2 G t( r, m),

(3.3)

ll~-ll2 Q t(r, n).

(3.4)

rows and columns.

by the choice of .!? and

and R and FE denote

m X r submatrices,

of A and o,f F, respectively,

the intersections
doskeleton

denote

which correspond

component

rows and columns

which occupy

in A and F. Then

CGR, G E [w r, can be presented

r X n

to the selected

Let A and F denote the r X r submatrices

of these

= I, and sub-

such that

now select the r rows and r columns determined

V, respectively.
Let C and F,

(3-l)

the decomposition

A -F

matrices

+ \lt(r,m)]).

any pseu-

in the following

form:

CGR = (Uzc

+ Fc)G(&V

= UcTiG&V

E = (Us?

+ E,

+ Fc)GF,

= Uri-(&,TjG)F,

= U+(@G)F,

+ FR)

(3.5)
+ FcG(tiI;V

+ FE) - FcGF,

+ F,(Gfi~~)~.-V

+ F,(G@)+V

+ FcGFR

+ F,GF,,

(3.6)

PSEUDOSKELETON

APPROXIMATIONS

where
(3.7)
We also rewrite (3.2) and (3.5) in terms of @:

ulj-@+\~

(3.8)

CGR = Ufi-'((DG@)+V
Now consider

@=

the singular-value

tis,

decomposition

f: = diag(G1,...,

Let 7 > 0 be a threshold

+ E.

a,),

(3.9)

of @:
$6

= $TTj

value which will be specified

= 1.

(3.10)

later. Introducing

the

notation

2, = diag(

C$),

a;, =

a,

if

otherwise,

Oi

if

otherwise,

-I

f;:

= diag( 6;:

Q7 =
q,+ =
7

),

tiS,G,
cTz+ijT

aT: =

ai>7,

a,>7,

(3.11)

(3.12)

(3.13)

(3.14)

we see that
(3.15)
and moreover,
(3.16)
It we set
(3.17)

S. A. GOREINOV

10
then (3.6) and (3.15)

ET AL.

imply that

IlkII2 + II~-lllz +

IJEll
Using this inequality

(3.18)

in conjunction

with (3.81, (3.9),

and (3.15),

we get the

estimate

((A - CGRI12 6

~~~d-~~~~~~-
+ ;t~~+ ellfi-ll~ + #%.
(3.19)

Now setting

r = E/ diiFKii%,

we complete

the proof of the theow

rem.
Note

that in this proof

necessary;

the singular-value

decomposition

of @ is not

we can get the same estimate by applying the filtering

(3.10)-(3.16)

directly to the decomposition

@ = UX$.

However,

technique
in that case

we must take

7 = EJm
Ilu- IlaW 112.
COROLLARY 3.1. Under the hypotheses
pseudoskeleton component such that
])A - CGR(]s

This bound

immediately

of Theorem

< ~(1 + 2&

follows from (3.1)

3.1 there

+ 2fi).

exists a

(3.20)

and the above estimates

for

t(r, n) (see Section 2).

REMARK 3.1.

We would like to emphasize

that Theorem

3.1 is somewhat

different from theorems of small-perturbations

theory, where E must be
sufficiently small (for a rather standard result of this type, see [4, Theorem
3.11). However, in the most interesting and important cases E may depend on
m and n, and usually we are interested in construction of pseudoskeleton
approximations for a family of matrices generated by one application.

PSEUDOSKELETON
We believe

11

APPROXIMATIONS

that the estimate

(3.1) is sharp up to a numerical

The following example provides the illustration

constant.

for the case r = 1, n = m.

EXAMPLE. Let an n x n matrix A be given by its singular-value

decom-

position

where

are orthogonal

n x n matrices,

ur E RnX1,

TV]E RIXn,

and E is a positive

number.
Let Or = [l/ G,.

. . , l/

(Ci,Jl) = (Aei,ul) =

Jrt]. Then for any column ci of A,

(%'ei,e,)

~r(e,,V~er)

= &(ej,uT)

= E.

and similarly

(ci>uk) = E(vk)i>
where uk and ok denote
By orthogonal

columns

invariance

l<k<n,

(3.21)

of U and rows of V, respectively.

of the 2-norm

for any pseudoskeleton

compo-

nent cigri we have the inequality

(3.22)

The equality (3.21)

implies that

Il2TciIIZ
= E2 l-i

1
n i

S. A. GOREINOV

12
ht-hand

Notice

that

minimized

f(i)

side of the inequality

is a quadratic

in the

can be estimated

function

from below

g^ and can be

readily

with the bound

Il]6u, Suppose

(3.22)

ET AL.

cigrjl12

now that (n -

>

(3.23)

1)~ = 1. Then

using

(3.23)

and the

triangle

inequality

IIu,Gv, - cigrjll2 < Il~,h,

- AlI2 + IIA - CigrjII2,

we get the estimate

IIA - CigrjII2b E

for any pseudoskeleton

component

-E=

E(l

to Theorem

(3.24)

ci grj.

Note that this example can be easily extended

Thanks

- fi)

to the case T > 1.

3.1, we know that accurate

enough

pseudoskeleton

approximations do exist. We also note that our proof is almost constructive

and involves two stages:
(1) the choice of appropriate
(2) the choice of G.

C and R;

Both stages of the proof make use of the explicit knowledge of F. The first
stage consists indeed in looking in A - F for an T X r submatrix of the
maximal volume. The second stage is something like the pseudoinversion
procedure applied to A^- F^ [see (3.7) and the remainder of the proof].

PSEUDOSKELETON

APPROXIMATIONS

1:3

We
n now want to exploit another choice of G with no explicit information
on F.
THEOREM3.2.
Assume that C and R are chosen in the same way as in
the proof of Theorem 3.1. Then G can he chosen using only A, and this choice
guarantees the accuracy

11A + CGR((2 < &[2 + 2t( r, n) + 2t( r, m) + 5t( r, n)t( r, m)].

Consider

Proof.

the singular-value

(3.25)

decomposition

and define Z,, %:, &, A^: by f ormulas similar to (3.1 lH3.14).
We essentially use the relationships
(3.6)-(3.91,
which hold true for any G, and
consider

them in the following form:

E = Ul?-(

A^c)F,

- Ufi-+GF,
A

(Jfi-Aij-V

CGR = U+(

kGi)$-

+ Fc(GA^)Tj-V
- FcG%-IV.
_ Ufi-l@lV,
V + E

- Ufi-(i+GA^)+V

Uri-(

+ Uri-($G$)Tj-%.
If we set

then relations

similar to (3.15)-(3.16)

+ FcGF,

are valid:

A^GF)Tj-V

14

S. A. GOREINOV

ET AL.

Therefore,

E2 c2
+ .&112 + - + -+-1112
7

+ fllB~llz

+ 2Ell~-111~ll~-l12 + ~l,0,l~13l12.
Taking the threshold

value r = E, we derive the estimate (3.25).

Note that unlike the proof of Theorem

tion of Q, is essential
REMARK 3.2.

to the statement

The estimate

(3.25)

IIA - CGRll 2 G dm{l

where

p = min(m,

THE

Here

as follows:

+ dt(r,m)]),

(3.26)

n).

NUMERICAL

we present

capabilities
are making
purpose of
doskeleton

decomposi-

3.2.

can be improved

+ [dm

The proof of the inequality

4.

3.1, the singular-value

of Theorem

??

(3.26)

can be found in [4].

EXAMPLE

a numerical

example

to illustrate

the approximation

of pseudoskeleton
components
for a nontrivial application. We
no attempt here to devise really efficient algorithms. The only
the experiments was to demonstrate
the existence of an pseucomponent with approximation quality comparable with that of

the SVD.
All the constructions

were carried

out as close to the paths described

in

the proofs as possible, and all the estimates were obtained in the most
conservative way (for example, if we present below an estimate of )IAlIs, it
arises from the SVD run).
We have picked a 1048 X 1040 block of a complex matrix of order 20,808
originating from the 3D scattering problem for a perfectly conducting sphere

PSEUDOSKELETON

APPROXIMATIONS

S. Namely, we solve the so-called

electric

1.5
field integral equation:

{(k:
+graddv)LG(r)j'(x)

dx)=3,
7

where

G(r)
T means the tangential

= -

component,

field (we used a plane-wave

current

on the surface

r = Ix - yl,

I? is proportional

excitation),

and T(x)

We have adopted

the Gale&n

by Rao, Wilton,

are presented
corresponding
the chosen

the wavenumber

discretization

technique

and Glisson [5]. Fragments

k,, of the

in the version

of the triangular

mesh

in Figure 1. Each edge matches exactly one unknown in the

linear algebraic system. Shown are the edges corresponding to

matrix block, which had block indices (20, 1) in a 20 X 20 block

partition with blocks of approximately

had 20,808
The

electric

S.

whereas
The radius of the sphere was 1 111,
incident wave was 6.8 m-.
proposed

to the incident electric

is the unknown

equal sizes. The linear system involved

complex unknowns.

construction

of the pseudoskeleton

component

was based

on the

SVD of the chosen block A = U CV * = Zr= 1uiciuT,

computed by the
standard LAPACK routine. For a given rank r we defined the matrix F as

ui aivF. To choose

F = C;=,+,
ces

and

submatrices

R,

we

applied

r rows and columns

an algorithm

[l] to the m X r and rr X r orthonormal

and [vl,. . . , v,]. The approximation error matrices

explicitly.
Figures 2 and 3 display the matrix approximation
Frobenius

which constitute

searching

norms for SVD and pseudoskeleton

matri-

for well-conditioned
matrices

A - CGR

[ul, . . . , ur]
were formed

error in the spectral and

components

as a function

of

their rank r.
The curve labeled PSA, G = f( A - F) shows the approximation
of the pseudoskeleton

component

error

with

G = argmin]]A

- CGR](r.

The Frobenius

norm is preferable,

since it allows for a complete

solution of

the minimization problem (see [4]). However, this choice of G is apparently

of no practical value, since it uses all elements of the block.

S. A. GOREINOV

16

ET AL.

FIG. 1. The mesh fragments corresponding to the chosen block. Highlighted are
the edges constituting the well-conditioned 60 X 60 submatrix found as in the proof
of Theorem 3.1.

The curves labeled

spond to the choices

PSA,

G = f( A^- I$ and PSA, G = f( A^) corre-

of G described

in the proofs of Theorems

3.1 and 3.2,

respectively.
Figures

4 and 5 show the upper bounds of the matrix approximation

error

provided by Theorems 3.1 and 3.2. The corresponding experimental data are
repeated for comparison purposes.
Note that some of the edges depicted in Figure 1 are highlighted. These
are edges corresponding
to the 60 X 60 submatrix obtained by a search
algorithm of [I] applied to the orthonormal matrices containing 60 senior left
and right singular vectors of the chosen block.
In Tables 1 and 2 we present the numerical values of approximation
errors displayed in Figures 2-5.
Assume now that we want to approximate the considered matrix block to
a lop5 accuracy. From Table 1 we can see the difference between the two
following approaches:

PSEUDOSKELETON

APPROXIMATIONS

17

PSA, G = f (A - F)

____--__

SVD
IO&o IIA -

FIG. 2.
nents.

(1) We

CGRll2

40

20

Approximation error in 2-norm

compute

approximately

60

for SVD and pseudoskeleton compo-

10 matrix

entries

and find 20 senior

singular triplets of a 1048 X 1040 matrix.

(2) We compute the entries of 30 rows and columns (which is only 6% of

the total number
Both

approaches

of entries)
provide

and find the SVD of 30 X 30 matrix.

the same approximation

quality,

but the second

approach may require far less memory and arithmetic. Note that it allows us
not to compute all the matrix entries, working instead only with a small part
of them.

5.

DISCUSSION

The principal conclusion

of the above theory can be formulated
as
follows: the existence of a rank-r approximation generally means that such an
approximation can be found only from some r columns and r rows of the
matrix to be approximated. If the rank-r approximation has the accuracy .s,
then the same (asymptotic) accuracy can be preserved by the pseudoskeleton
approximation. Restrictions on the information available for the construction

S. A. GOREINOV

18

..................

PSA, G = f(A)

____-

PSA, G = f(A - P)

log,,

ET AL.

IIA- CGRIIF

-2
-3
-4
-5
-6
-7
-8

0
FIG. 3.
nents.

20

40

60

Approximation error in F-norm for SVD and pseudoskeleton compo-

of G degrade

the accuracy

by a factor of U(G).

Apparently

this addresses

the worst case. If the original matrix has an r- X r submatrix with sufficiently
(for given E) well-bounded inverse, then it is possible to preserve the same
accuracy E. Moreover,

we can expect the same result in the case of an order-r

submatrix with well-separated

small singular values.

Our concern
in this paper has been the estimation
of the
accuracy which can be obtained, in principle, via pseudoskeleton

possible
compo-

nents. Another important matter is how to construct such components

in
practice, and in particular, how to select appropriate columns and rows. We
believe that results of [l] will promote a suitable algorithm.
Some proposals

can be derived right away from our estimates.

Since the

deterioration
factor incorporates
(Theorem
3.1) the original matrix size, it
seems reasonable to partition the matrix into a few blocks, find pseudoskeleton approximations
for these blocks, and then try to compress the whole
mosaic-skeleton
approximation
by the standard Lanczos-like
techniques.
This strategy evidently leads to a highly concurrent algorithm. The details are
left for another paper.
The function t(r, n) defined by (2.3) and its upper estimates play a key
role in our proofs. We have presented here a new estimate (Lemma .%I>,

PSEUDOSKELETON

APPROXIMATIONS
Bound of Thm 3.2

________

PSA, G = j(a)

I
-8

FIG. 4.
components.

Upper

_--~___~___!__1__L__I

bounds

20

40

for approximation

_______.

error

60
in 2-norm

for pseudoskeleton

Bound of Thm 3.1

PSA, G = f(A - P)

A log,, IIA- CGRll2

-2
-3
-4
-5
-6
-7
-8

_-_I__ _I_ - J_ _
P

FIG. 5.
components.

Upper

bounds

20
for approximation

40

60

error in F-norm

for pseudoskeleton

20

S. A. GOREINOV

ET AL.

TABLE 1
THEMATRIXAPPROAIMATION

ERRORFORSVDANDPSEUDOSKELETON
MEASURED
G =f(A

SVD

0
5
10
15
20
25
30
35
40
45
50
55
60

2.647E-02
5.035E-03
3.958E-04
7.112E-05
6.609E-06
1.714E-06
4.476E-07
9.6013-08
4.850E-08
2.703E-08
1.556E-08
4.411E-09
1.958E-09

which improves
readers

COMPONENTS

IN SPECTRALNORM

- F)

2.647E-02
7.086E-03
8.588E-04
8.923E-05
1.569E-05
4.889E-06
1.790E-06
3.2573-07
9.954E-08
7.8833-08
3.486E-08
1.273E-08
6.584E-09

=f(Al- 6)

2.6473-02
2.6473-02
1.227E-03
l.l31E-04
6.242E-05
5.351E-06
5.892E-06
3.457E-07
2.8643-07
1.944E-07
1.383E-07
2.307E-08
9.814E-09

2.647E-02
2.647E-02
2.647E-02
1.1653-03
l.l37E-04
5.7903-05
l.l06E-05
3.521E-06
8.058E-07
7.507E-07
1.964E-07
1.543E-07
1.307E-07

=f(i)

the previous one [l, 31. Finally, we would like to draw the

attention

to our hypothesis

(2.6),

since we still have no proof of it.

We would like to thank Gene Golub for good advice about seeking the
best-conditioned
submatrices.
Our special thanks go to Skivkumar Chandrasekaran for kindly sending us the report 111.

TABLE 2
THE MATRIX APPROXIMATION

ERROR FORSVDANDPSEUDOSKELETON

MEASUREDIN
r

SVD

0
5
10
15
20
25
30
35
40
45
50
55
60

3.883D-02
5.276D-03
5.788D-04
9.038D-05
1.283D-05
3.409D-06
8.667D-07
2.081D-07
l.O85D-07
5.943D-08
2.226D-08
8.558D-09
5.076D-09

G =f(A

FROBENIUS

- F)

3.883D-02
8.8933-03
1.434E-03
1.485E-04
3.012E-05
8.654E-06
2.8973-06
5.5423-07
2.425E-07
1.716E-07
6.915E-08
2.609E-08
1.521E-08

COMPONENTS

NORM

G =f(/i

- 6)

3.883D-02
3.8833-02
2.287E-03
2.246E-04
6.914E-05
9.177E-06
6.423E-06
5.842E-07
4.326E-07
3.229E-07
1.570E-07
3.5853-08
1.785E-08

G =f(A)
3.883D-02
3.883D-02
3.883D-02
1.999E-03
2.164E-04
l.O06E-04
1.8083-05
6.716E-06
1.338E-06
l.O37E-06
4.194E-07
2.926E-07
1.842E-07

PSEUDOSKELETON

21

APPROXIMATIONS

REFERENCES
1 S. Chandrasekaran

and I. Ipsen,

YALEU/DCS/RR-880,
2 6. Golub
Berkeley,

and C. F. Van Loan,

Matrix

Dec.

QR Factorizations,

Res. Rep.

1991.

Computations.

2nd ed., Johns

Hopkins

U.P.,

and singular

value

1989.

3 Y. P. Hong

and C.-T.

decomposition,

Math.

4 S. A. Goreinov,
approximations

E.

Pan,
Camp.
E.

shape,

IEEE

Rank-revealing

QR

58(197):213-232

Tytyshinikov,

of matrices,

5 S. M. Rao, D. R. Wilton,
of arbitrary

On Rank-Revealing

Yale Univ.,

and

factorizations

(1992).
N.

L. Zamarashkin,

Rep. Russ. Acad. Sci. 343(2):151-152

and A. W. Glisson,
Trans. Antennus

Electromagnetic
Propagat.

scattering

AP-30:40$-418

Aeceiwi 25 April 1994; f ~na1 manuscript accepted 19 March 1996

Pseudo-skeleton
(1995).
by surfaces

(1982).