STATISTICAL
REGULARIZATION OF SYSTEMS OF ALGEBRAIC EQUATIONS * E. L. ZHUKOVSKII Moscow (Received
THE QUESTION system
of the construction
AZ = u, is considered
experimental
realization
9 February 1971)
of the pseudo
when the vector
of it with a normally
solution
of the linear
u is not known exactly distributed
algebraic
but
error and covariance
matrix uoT are known. The method of solution used is the method of filtering of Markov processes. In the case where the error of the vector u is distributed arbitrarily, a filter
using
the regularization
of the solution
method of A. N. Tikhonov
[l], we present
of this problem.
1. Formulation of the problem Let A = lUijI be an m x I matrix, specifying the linear transformation m > 1, and z = {zr, . . . , z’} and ii = (3, . . . , P} be vectors
A : Rr +Rm,
the real spaces R, and R, respectively. In this paper we consider of the solution of the algebraic system of equations AZ = ii,
(1)
of
the question
iiERm,
when the vector ii is not known exactly, but its A-approximations z,, . . . . z’,... are known; a A-approximation ;);1 is a vector which is the sum of the two expectation, the vectors u”, = E + Au”,. Using M.O. to denote the mathematical expectation of the vectors of unknown vector ii = P.o.~“, is the mathematical the samples z,,; && is a vector, each coordinate of which is a normally distributed quantity with zero mathematical expectation and variance known in advance
(A&,h)2 = M.O.1ii,k - ii”1 2. ~~2 = 111.0.
The following measurements
problem
is solved:
of the coordinates
asymptotically unbiased “weighted” discrepancy
as a result
of the vector
of repeated ii E R,,,
estimate for the vector IIAZ; - u llB, deviating
*Zh. ujkchisl.Mat. mat. Fiz., 12, 1, 185-191. 1972.
230
experimental
find the consistent
z; E R,,,, minimizing the least from the given vector
and
Statistical
regularization
of systems
of algebmic
231
equations
m. E R, in the sense of the norm II- Ilc, which is defined by the scalar product (Cz, t), where B = (a~~)” and C is the matrix of weights, which will be defined in section 2, and oT is the transport matrix of cx Therefore,
we estimate
121. In this formulation
the pseudo
of the problem
solution
of problem
it is assumed
(1) in the sense
of
that the matrix A may have
rank r < I. The existence of some further information about the error AZ;, makes it possible to construct a stable statistical algorithm for the solution of problem (1) with an estimate
of the variance
of each coordinate
This information is naturally the assumption of the vector zn is normally distributed.
of the solution
found.
that the error AZ’ of the coordinates
The filter approach [3, 41 to the solution of problem (1) presented in this paper, based on the sequential Bayesian method of estimating the unknown vector Z, includes [5, 61, and enables In section
the method of statistical solution previously considered the pseudo solution of problem (1) to be constructed.
2 the formulation
problem
(1) is given
and the form of the
is indicated, in section 3 in Theorems 2 and 3 the main paper are formulated, in section 4 difference schemes for
regularizing algorithm results of the present solving
of problem
in
(1) are presented.
2.
Statistical
regularizing
algorithm
1. Let us have a sufficiently large number of realizations of the vector zn about which we suppose that E, = ii + w(n), where n is the number of the realization dimension
of the vector
m, and CJis an arbitrary
With these stochastic
chain
42)
w(n) E N(0, E)
ii,
assumptions
is a normal random vector
non-singular
about the solution
of
matrix. of problem
(1) the following
model may be used.
On the probability (z,,, &), n=l, Gl+i
=
space
&I+, =&a+
&a,
where both equations
are understood
of two random quantities. 0, we obtain
equation
(a, 3, 9) we consider
2, . . . . which controls
(l).)
aw(n + 1), in the sense
(It is obvious
the two-dimensional
the stochastic 20 =
difference 2,
a0
of the stochastic
that when the covariance
=
Markov system
u, equivalence matrix 00~ I
232
E. L- Zhukbvskii
We will interpret the sequence Z, as the *unobserved” component, and 2’ as the ‘observed” component. We denote by m,(o) = m,,(&(~), . . . . Z,(o)) the optimal
estimate
minimizing
,M. 0. ]]_zR- m, I$+ = s
(5 - m,,)T (z - m,) P (z,, E dz ) 5,
_ ), UI..... u,
Rl
where MO. llzn~lRlz < 00, for the sequence Z, which is controlled by the system (2), where M. 0. denotes the mathematical expectation subject to the conditional probability Q{zn E ds]9r;1, .... ;;, }, which will be calculated in Theorem variables
is the x-algebra of the wsets It is known [7], that
1, FZ,. ..$, & . . . . u,.
where rn is an estimate The calculation on the sequential connection
of the covariance
of these Bayesian
Theorem
generated
matrix of the vector
conditional
mathematical
method of estimating
by the random
z,.
expectations
the unknown
is based
vector Z. In this
1 holds.
Theorem 1 If z,, E N( mo, ro), and the sequence zn is controlled by the system estimates m, and rn exist, are normal and satisfy the system m,+i = m, + ~,,.AT(uu~ + AI’J*)-‘(~~++ (3)
(2), the
-Am,),
r,A=((JO= + ArnAT) -‘Ar,,
r ,,+1 = r, -
where mR is the vector
of conditional mathematical expectation: m, = r+~a.[z,]fi,, matrix of the unknown solution 2, with . ..) E*]; rh = jyiil is the covariance the elements vij = 111.0.[(zni- mni) (z,jm,,j) /u”,,, . . . , ai], i, j = 1,. . . , I; To = C’
i nitial
is the initial
approximation
Proof. normal
(4)
We consider
vector
covariance
approximation
~.o.(~n+*l&l+~) (s”+l cov(zn+I
-
P.O.
I f-L+,)
;h).
with mathematical
l3 = II”,
oo:II
By hypothesis expectation
a =
Zn)
-
E
( A > zn theorem
~u+t)[COV(Hn+r, fL+~)]-’
a,), cov(z,,
m, is the
it has a two-dimensional
_ By the normal correlation
= Bf.0. 2” + cov(zn+l,
=
matrix c
Z.
the pair (z,,
distribution
matrix
for the positive-definite
for the vector
C0V(Z,+1,U”n+1)
x
x
and [g],
Statistical
regularization
of systems
of algebmic
equations
233
{cov(Z*+i, Zn+i)l-‘[cov (Zn+i, u”n+i)P, which proves the existence r n = Cov(z,+il&+l). In order to calculate P (r,+,,
&+i I L+I).
m,+i = ~0. (z,+i) iL+l)
of the estimates
the quantities &+1
Vn+i =
Let
occurring
and
here we find the distribution
be a random variable;
we show that
( &+i 1
it is normally distributed. For this we write down the characteristic a two-dimensional random variable: f(t) =
rd. 0.
[nt.o.(exp
ar.o.[exp {iPa -
{it%+i} 1u’,, . . . (6, zn)l =
‘/#j3t} 1u”,,. . . , iii] =
116. o.(exp (itTaiz,} / E,, . . . , Qexp - i/#aiTrnait
exp (itaim, Thereby calculate
{-‘/zt’Br)
- ‘/#fit},
=
where
the normality of the distribution of the quantity the moments of v~+~. They are given by
cov(ii,,+i,
cov(z?l+i,
zn+1 I an) = ra,
cov(z,+L,
u”n+i1ii,) = I-,.@.
Note 1.
these
Solving
values
al = un+I is proved.
We
M.O.(u”,+~i E,) = Am,,,
1y.o.(z,+i I En) = mnr
Substituting
function
of the moments
the system
&+I 1ii,,) = AI’,A= + (J(T~‘,
in (4), we obtain
(3).
(3), we obtain n
mn = [E +
45)
ro_4=((J@)-' AR]-’
mo + I’oAT(wT)-’ El a=,
In the particular
case
where u is commutative _
UUTO-’ + ATA
(‘3)
-
a.
with the matrix A, we obtain
ugTro-*rno + AT-
a=, r, = &{ro-i
(au*) + ATAn]-‘.
As will be shown below, the estimate
the solution
m,, Tn for the pseudo
solution
of the difference
system
(3) gives
for
E. L. Zhukovskii
234
2; = lim m, 78400 and its covariance Note 2.
matrix r.
The statistical
regularizing
algorithm
of the solution
of problem
(1) is constructed similarly when we require that not only z~+~ t z,, but also the k-th difference of the solution sn with respect to n remain constant from realization of the input vector zn. Then the Bayesian filtering problem is written in the form ii n-l-l
(2’)
4, + uw(n + I),
=
Gl+i
=
(1) Gl+i
=
2,
2, +
(i)+ 2 ? . . . . . . . . . . . &a
z’“:,--n z’R’ 11
with the initial
conditions
z(r))
=
zp, . . . , z’A)(o) = 0.
z(i) (0) =
zo,
Then the estimate mk(n) of the regularized solution z of problem (l), the k-th difference of which equals zero from experiment to experiment, will be the solution
of the system Amo(n) = roo(n)AT(aa')-'(u"n+l-AAm~(n)), -Am,(n)), Ami = m,(n) + I’~(n)A~(~d-‘(~n+i . * * . . . . . . . . . . . . . . . * * * AmA
(3’)
AI’(n) =
-[(E+
= mA_,(n)
+ rA,(n)AT(~UT)-'(~7t+i -Am~(n))v
ar(n)
d(n) + rbw -
a)r(n)aT]r~~*+Ar(n)A7-1[(E+4rwm
with the initial conditions rno (0)
=
no, rni(0) = mi,
...q
mA(0) =
0;
r(o)
=
rO*
Here 0 E 0 ...O OOE...O a= . . . . . , OO...E oo...o
are matrices vector,
of dimension
whose elements
kl x kl, and A = (A, 0, . . . . 0) is a (12+ l)-component are matrices
of dimension
m x 1. The regularizing
Stutisticcd regularization of systems of algebraic equations
algorithms
(3) and (3’) of the solution
controlled
by the system
2.
We consider
the random parametric
T”[z, G] = Il.42- k2
(7)
a > 0 is a parameter functional
of problem
(2) and (2’) are thereby
(1) for the case where it is constructed.
functional
+ alb- mb2,
which consists
is the square
235
of the sum of two functionals:
of the deviation
the first
of the vector
I/Z - m,ljc From AZ in the norm of the matrix B = (00. T)*‘; the second functional is the square of the deviation of the unknown solution z from m, in the norm \I - \I,, where C = r,‘It is easy to prove that there exists a vector, unique with probability ma = hi. G.z*, realizing the minimum of the smoothing functional (7) and identical with (5) for a = a, = l/n.
1,
This can easily be verified if Euler’s equation for the functional T’lIz,^ul is used and the system (4) is summed. It is also obvious that if we assume that the matrices A and 0 are commutative and in (5) put P = (oaT) r,“, we obtain a Tikhonov regularization for problem (1) with a = a, = l/n. This
shows
that the sequential
Bayesian
method of estimating
the unknown
vector zU in problem (1) or the problem of the linear unsteadxfiltering of the ‘unobserved” component zn from the uobserved” component U, of the Markov by the system (3), in the case where the sequence (2% 2 n) t which is controlled input vector U, has a normally distributed error, defines a Tikhonov regularizing algorithm with a = l/n. The solution Bayesian
sequential
of problem solution
(1) realized
by equation
(5) will be called
the
below.
The identity of the Tikhonov regularized solution and the Bayesian sequential solution in the case of a Gaussian error makes it possible to establish some statistical properties of the solution obtained by the Tikhonov regularization
method (asymptotic
unbiasedness
and independence
as b- + 0); and
conversely, since the Bayesian sequential solution is a solution of Euler’s equation for the functional T’?z, ~1,it will be shown in Theorem 2 that for a = l/n + 0 it is a pseudo solution of problem (1) and is calculated as the limit
236
E. L. Zhukouskii
zu = lim 7nn. *-cQ 3.
As is obvious
from the statement
of section
2, the parameter
in the
statistical regularization is the number of realizations of the input vector zh. It is known that one of the fundamental criteria of the accuracy of the approximations
pR (As, G), 1 which in our case may be the first term of the functional Ta[z, u^l, that is, (A~ - &)~(oo~)-i(~z - ^u), which has a x2 distribution with (I - 1) degrees of freedom. statistical
of z obtained
is the magnitude
And therefore, the ‘discrepancy check of the hypotheses. 3.
Properties
principle”
of the Bayesian
Let To and uoT be the matrices holds.
of the discrepancy
defined
in this case
sequential previously.
will be a
solution The following
lemma
Lemma. The condition
~‘~A~(ocr~)-l~x = 0
is equivalent
to the condition
A=AZ =
0, x E RI.
We denote by L the subspace in R, of vectors such that L = ix E R[ : Gx = 0) = ker G, where G = iToAT(uu~) -*A, and N=(~ER~:~IL}~=I~G,
and
L CBN = R1.
Theorem2 The sequence m, = [E + I',,AT(uuT)--l An]-’
mo + I’oAT(uuT)-’ [
Cl S=i u”,
with probability 1 in R, to the vector iii = mO'+ fi",wheremo’ = np,mo, z; = E = lim m, = mp’ + ro%(u-tAro’lz)+u-‘~ is a pseudo and iii”= np.,+i n solution of problem (1).
converges
We calculate -t + roAT(uuT)-LA
x I
Statistical
regularization
of systems
ro-’ + AT(UCP) -I A
of algebmic
237
equations
--I 1ro-‘mo + 1 1 n
n
--i
r,,-’ + rl~(cw)-’
A
a=,
Because (~a’)-r
of the positive =
definiteness
of the matrices
r0 and B we can write
Q2, I’,,-’ = ‘P, Q, 0 > 0.
Writing D=
V = QA,
VW’
and
f=-
ii,,
we obtain lim m, =
n-rm
lim CP-~ $
n-m
which is satisfied inverse
i
E +DT D]-'D,Q~ =
TI-WCI 1
J n( +
r;JZ (o-lAr~~2)+a-1ii
with probability
1.
= zlb’
The symbol ( )+ denotes
the pseudo-
matrix.
Lemma
1 of [2] was used in the calculation
Note 3. and m,lker defines
E + DTD]-~~ @,,'ro-lmo+ lim@-i
[
In particular, ATA,
if problem (1) has a solution
the theorem proved implies
an estimate
of the second
m, converging
limit.
in the ordinary
that the iterative
to the normal solution
scheme
sense (3)
in the mean-square
sense. We return to the question of the estimate
of the asymptotic
of the Bayesian
m,=~.o.(z,/n,
sequential
,...,
unbiasedness
and independence
solution
Ei).
Theorem3 The Bayesian independent
sequential
estimate
The asymptotic
solution
for the exact unbiasedness
m, is an asymptotically
solution
unbiased
and
of problem (1).
of the estimate
m, follows from the proof of
238
E. L. Zhukouekii
Theorem
2.
We prove the independence of the estimate m,. For this we show that cov(m,-zz;)+O_We have cov(m,-zz;;) =~.o.(m,-i&,)(m,,-a,,)~+ (FE,,-zz;;:) (iii” - 2;;)‘. where 1 = mn
n
z&o.-
ii ,. n c 8-l
Using
of Theorem 2, we obtain
the notation
then ~.o.(m,--Eiii,)(m,-iiiiii,)T=-Q)-i
It follows tends
1 n
from-the proof used in Theorem
Ln f
1
-1
DTD
DTD@-‘.
2 that (iii, - zi) (iii, - 2;)~.
as n + m
to 0 in the norm 11- llRl. It is obvious from this that the matrix c& (m, also proves the independence of the estimate m,, for the pseudo
2;) + 0, and this
solution
2;.
.4. Numerical method of solution The approximate method of solving problem (1) in the case of a Gaussian error reduces to the solution of equation (S), requiring only one inversion of the when n realizations of the vector are known. matrix [E + l?oAT(aaT)-lAn]-i, The solution z’, can also be obtained by minimizing functional (7) by direct methods of minimizing quadratic
the smoothing functionals.
We consider problem (1) with the error of the vector z,, distributed arbitrarily. In this case a filter of the solution of problem (1) can be constructed by means of Tikhonov’s regularization method. Let the vector zrr have an arbitrarily distributed error with a non-singular covariance matrix uuT. Let equation
za = (c& + AT(oar) -‘A) -~A=(ooT) -l^u of problem (1) for I0 = E, where UC-
calculated
be the Tikhonov
regularized
1 +& P/l *=1
for some realizations
of the vector
z?, and a > 0 is the parameter
Statistical
defined, Then
for example,
the following
regularization
of systems
as the solution iteration
of algebraic
of the equation
algorithm
is proposed
It is obvious
that as the initial
za, and at subsequent Numerical
steps
calculations
approximation
we will improve performed
ljdz=-^ulls (Tikhonov
(.:X1
z,,+! = zn +(a,!? + AT(~uT)-iA)-‘AT(uuT)-i
equations
239
= lla, _ ills.
filter):
u”,+i -dAz,
. 1
*=*
Z, = 0, at the first
step we obtain
it.
at the Moscow University
Computer
on model problems have demonstrated the efficiency and good agreement Bayesian sequential solution with the given theoretical solution. In conclusion I wish to thank A. N. Tikhonov for supervising A. N. Shiryaev and R. Sh. Liptser for their interest and assistance paper. The author thanks V. A. Morozov for reading making a number of valuable comments.
Centre of the
the paper, with the
the paper in manuscript
and
and
Translated by J. Berry REFERENCES 1.
TIKHONOV, A. N. On incorrect problems of linear algebra and a stable method of solving
them.
Do&l. Akad. Nauk SSSR,
163. 3, 591-595,
1965.
2.
MOROZOV, V. A. On pseudo solutions. 1387-1391.1969.
3.
LIPTSER, R. SH. and SHIRYAEV. A. N. Non-linear filtering of diffusion processes. Tr. Mat. in-ta Akad. Nauk SSSR, 104, 135-180, 1968.
4.
GLONTI. 0. A. Sequential Markov chain. Litovskii
5.
PETROV, A. P. Estimates of linear functionals for the solution problems. Zh. uychisl. Mat. mat. Fiz., 7, 3, 648-654.1967.
6.
TURCHIN, V. F. Solution of a Fredholm equation of the first kind in a statistical ensemble of smooth functions. Zh. uychisl. Mat. mat. Fiz., 7, 6,1270-1284, 1967.
7.
GIKHMAN, I. I. and SKOROKHOD. Processes (Vvedenie v teoriyu
8.
ANDERSON, T. W. An Introduction to Multiuariate v mogomernyi statisticheskii analiz). Fizmatgiz,
Zh. vychisl.
Mat. mat. Fiz.,
filtering and the interpolation mat. sb., 2, 263-279, 1969.
9, 6,
of components
Markov
of a
of some inverse
A. V. Introduction to the Theory of Random sluchainykh protsessov), ‘Naukam, Moscow 1965.
Statistical Moscow.
Analysis 1963.
(Bvedenie