ISSN 1064–5624, Doklady Mathematics, 2006, Vol. 74, No. 1, pp. 480–483. © Pleiades Publishing, Inc., 2006.

Original Russian Text © Yu.I. Vinogradov, 2006, published in Doklady Akademii Nauk, 2006, Vol. 409, No. 1, pp. 15–18.

MATHEMATICS

Solution Method for Linear Ordinary Differential Equations

Yu. I. Vinogradov
Presented by Academician V.A. Il’in January 13, 2006

Received January 19, 2006

DOI: 10.1134/S106456240604003X

It is well known that linear ordinary differential the second approximation

equations can be represented as a system of first-order x x x
differential equations and can be written in the matrix
form Y 2 ( x ) = E + A ( x ) dx + ∫ ∫ ∫
A ( x ) A ( x ) dx dx Y 0 ,
x0 x0 x0

Y ' ( x ) = A ( x )Y ( x ) + G ( x ). (1) and the third approximation

x x x
Methods for solving differential equations with con-
stant coefficients are well known. However, there is no
Y 3 ( x ) = E + A ( x ) dx + ∫ ∫ ∫
A ( x ) A ( x ) dx dx
general method for solving differential equations with x0 x0 x0

variable coefficients. x x x

To integrate a homogeneous differential equation,

+ ∫ A( x ) ∫ ∫
A ( x ) A ( x ) dx dx dx Y 0 .
we use Picard’s successive approximation method [1], x0 x0 x0

which is also known as the Picard iteration method.

Equation (1) with G(x) = 0 under the initial conditions Continuing the iteration yields
Y(x) = Y0 at x = x0 can be written as x x x

x
Y ( x ) = E + A ( x ) dx + ∫ ∫ ∫
A ( x ) A ( x ) dx dx

∫
x0 x0 x0
Y ( x ) = Y 0 + A ( x )Y ( x ) d x. (2)
x x x

∫ ∫ ∫
x0
+ A( x ) A ( x ) A ( x ) dx dx dx + … Y 0 .
x0 x0 x0
If the equation cannot be integrated analytically, the
following procedure is applied. To obtain the first By introducing the notation
approximation, we choose Y0, substitute it for Y(x), and x x x
obtain a new function Y1(x) (on the left) that is different
K x0 ( A ( x ) ) = E + A ( x ) dx + ∫ ∫ ∫
A ( x ) A ( x ) dx dx
x
from Y0. Next, Y(x) is substituted for Y1(x) in (2) to
obtain a new function Y2(x), i.e., the second approxima- x0 x0 x0
tion to Y(x), etc. x x x

Thus, we obtain the first approximation + ∫ A( x ) ∫ ∫

A ( x ) A ( x ) d x d x d x + …, (3)
x0 x0 x0
x
the solution to Eq. (2) can be written as
∫
Y 1 ( x ) = E + A ( x ) dx Y 0 ,
Y ( x ) = K x0 ( A ( x ) )Y 0 .
x
x0 (4)
To evaluate the integrals in solution (3) to the homo-
geneous differential equation, the interval [x0, x] is
divided by intermediate points xi (i = 1, 2, …, n) into sub-
Bauman State Technical University, intervals [xi – 1, xi] and the current argument is assigned
Vtoraya Baumanskaya ul. 5, Moscow, 107005 Russia an arbitrary value x = xn. The integral over [x0, xn] is

480
 SOLUTION METHOD FOR LINEAR ORDINARY DIFFERENTIAL EQUATIONS 481

determined as the sum of the integrals over [xi – 1, xi]. Integrating the terms in solution (3) gives
Then the second term in (3) can be rewritten as
( B∆x i ) ( B∆x i )
2 3
K x0 ( A ( x ) ) = E + B∆x i + ------------------
- + ------------------- + … (5)
x
x x1 x2
2! 3!
∫ A ( x ) dx = ∫ A ( x ) dx + ∫ A ( x ) dx …
Obviously, the right-hand side of this formula is a
x0 x0 x1
B∆ x i
xi xn matrix Taylor series for the matrix exponential e .
…+ ∫ A ( x ) dx + … + ∫ A ( x ) d x. As a result, the solution to the linear homogeneous
ordinary differential equation becomes
xi – 1 xn – 1
∞
( B∆x i )
m
Choosing the intervals [xi – 1, xi] (i = 1, 2, …, n) to be
K x0 ( A ( x ) ) = ∑
x
sufficiently short, averaging the functions of the matrix -------------------
-,
m!
A(x) over each of them, and using the same interval m=0

∆x n
∆x (6)
length ∆xi = xi – xi – 1, where ∆xi = ------ and ∆x = xn – x0,
n B = ∑ A ( τ ), i ∆x i = ------,
n
∆x = x n – x 0 ,
we find i=1

x
τ i ∈ [ x i – 1, x i ], i = 1, 2, …, n.

∫ A ( x ) dx = [ A ( τ ) + A ( τ ) + … + A ( τ ) 1 2 i … If the differential equation has constant coefficients

A(x) = A = const, then A(τi) = A (i = 1, 2, …, n),
x0
n
… + A ( τ n ) ]∆x i ,
τ i ∈ [ x i – 1, x i ], i = 1, 2, …, n.
∑ A ( τ ) = nA, B∆xi = nA∆xi = A∆x, and its integral is
i
i=1
determined by the formula
By introducing the notation B = A(τ1) + A(τ2) + … +
∞
A(τi) + … + A(τn), the last expression can be rewritten ( A∆x )
m
K x0 ( A ) ∑ ∆x = x n – x 0 .
x
in the shortened form = ------------------, (7)
m!
m=0
x

∫ A ( x ) dx = B∆x i , ∆x i = x i – x i – 1 , When x0 = 0, solutions (6) and (7) become

m
⎛ B --x-⎞
x0
∞
∆x ⎝ n⎠ n
∆x i = ------, ∆x = x n – x 0 , i = 1, 2, …, n. K x0 ( A ( x ) ) ∑ ∑ A ( τ ),
x
n = ---------------, B = i (8)
m!
m=0 i=1
The third term in solution (3) is rewritten by using the ∞
( Ax )
expression obtained for the second term: m
K x0 ( A ) = ∑
x
--------------. (9)
x x x m!
m=0

∫ ∫
A ( x ) A ( x ) ( dx ) dx = ∫ A ( x )B∆x i d x.
Thus, for the first time, we obtained solution (8) in
x0 x0 x0 the form of a converging (as proved) matrix power
series (matrix Taylor series) for a matrix linear homo-
Putting the constant matrix B in the integrand and geneous differential equation. This solution is valid if
repeating the previous procedure, which involves the the elements of A(x) are functions that are continuous in
averaging of the functions of A(x) over the intervals ∆xi , some range [x0, xn] of x.
we obtain
The resulting solutions (8) and (9) to the homoge-
x x neous differential equation have a remarkable property.
( B∆x i )
2

∫ ∫
A ( x ) A ( x ) d x d x = ------------------
2!
-. Specifically, they become an identity matrix E at x = x0
and, hence, satisfy the arbitrary initial conditions Y(x) =
x0 x0
Y0 in the Cauchy problem. By using the Cauchy
method, functions with these properties were first
Similarly, we find
derived by A.N. Krylov for the computation of beams
x x x lying on an elastic foundation. The functions given by
( B∆x i )
3
formulas (8) and (9) will be referred to as Cauchy–Kry-
∫ A( x ) ∫ ∫
A ( x ) A ( x ) d x d x d x = ------------------
3!
-. lov functions; and their values, as the values of
x0 x0 x0 Cauchy–Krylov functions for differential equation (1).

DOKLADY MATHEMATICS Vol. 74 No. 1 2006

482 VINOGRADOV

It is well known that a particular solution Y*(x) to G( τ *i ) (which is independent of x) out of the integral,
Eq. (1) can be determined via the matrix of the Cauchy– we obtain
x
Krylov functions K x0 (A(x)) (see [1]):
x1 x2
⎛ x x ⎞ ⎛ x x ⎞
x
⎝x∫ ⎠ ⎝x ⎠ ∫
Y*( x n) = ⎜ K xn1 K τ11 dτ 1⎟ G(τ *1 ) + ⎜ K xn2 K τ22 dτ 2⎟ G(τ *2 ) …
∫K
τ –1
Y *( x ) = x 0 ( A ( x ) ) [ K x 0 ( A ( x ) ) ] G ( τ ) dτ,
x
(10) 0 1

x0
xi
⎛ ⎞
∫
x x
where K(x, τ) = K x0 (A(x))[ K x0 (A(x))]–1 is the Cauchy
x τ … + ⎜ K xni K τii dτ i⎟ G ( τ *i ) + …,
⎝x ⎠
matrix. i–1

As before, the integration interval [x0, x = xn] is i = 1, 2, …, n.

divided by intermediate points xi into n subintervals
x
[xi – 1, xi], i = 1, 2, …, n. Particular solution (10) is The matrices of the functions K xni (i = 1, 2, …, n) are
rewritten as independent of τi on the corresponding intervals [xi – 1, xi].
n xi Then

∑ ∫ [K
x τi –1
Y *( x n) = K xn0( A( x)) i ) dτ i .
x 0( A(τ i)) ] G(τ * (11) x1 x2
x ⎛ ⎞ x ⎛ ⎞
⎝∫ ⎝∫
x x2
i = 1 xi – 1 Y * ( xn ) = K xn1 ⎜ K τ11 dτ 1⎟ G 1 + K xn2 ⎜ K τ 2 dτ 2⎟ G 2 …
x0
⎠ x1
⎠
x
The matrix K x0 (A(x)) is put outside the integral, xi xn
since the former is independent of τi, where τi ∈ [xi – 1, x ⎛ ⎞ ⎛ ⎞
⎝ ∫ ∫
x x
…+ K xni ⎜ K τii dτ i⎟ G i +…+⎜ K τnn dτ n⎟ G n ,
1 ⎠ ⎝x ⎠
xi] and τ *i = --- (xi – 1 + xi) is an intermediate point in the xi – 1 n–1 (15)
2
interval [xi – 1, xi]. where Gi = Gi ( τ *i ) for i = 1, 2, …, n.

We introduce the notation K x0 (A(x)) = K x0 and use

x x The lengths of [xi – 1, xi] are chosen so that the ele-
ments of A(x) in Eq. (1) can be averaged over these
the following property of the matrix of Cauchy–Krylov x
functions: intervals. Then the Cauchy–Krylov functions K τii in the
x x x x
integrands are determined by the matrix Taylor series
K xn0 = K xni K xii – 1 K xi0– 1 , (12)
xi

∫K
xi
x x x x x τ i ( A ( τ* ) ) dτ
where K xni = K xnn – 1 K xnn –– 12 … K xii + 1 and K xi0– 1 =
xi – 1
x x x
K xii –– 12 K xii –– 23 … K x10 . xi
A i ∆τ i ( A i ∆τ i )
2
⎛ E + ------------
- + -------------------- + …⎞ dτ i .
In the interval [xi – 1, xi] (i = 1, 2, …, n) with an arbi-
= ∫ ⎝ 1! 2! ⎠
x x τ xi – 1
trary point τi ∈ [xi – 1, xi], we have = K xii – 1 K τii K xii – 1 .
Then matrix product (12) can be rewritten as Integrating and substituting the results into (15), we
obtain a formula for a particular solution to the inhomo-
x x x τ x geneous differential equation (1):
K xn0 = K xni K τii K xii – 1 K xi0– 1 , i = 1, 2, …, n. (13)
n

∑K
xn
τ τ x Y * ( xn ) = x i ( A ( x ) )T i G i ∆x i , (16)
Using the notation K xi0 = K xii – 1 K xi0– 1 and the matrix i=1
inversion rule, we have
where
τ –1 x –1 τ –1
[ K xi0 ] = [ K xi0– 1 ] [ K xii – 1 ] . (14) A ( τ *i )∆x i ( A ( τ *i )∆x i )
2
- + ------------------------------ + …,
T i = E + ----------------------
x
2! 3!
Putting K xn0 given by (13) and the inverse matrix
∆x
τ
[ K xi0 ]
–1
given by (14) inside the integral and taking G i = G i(τ *i ), ∆x i = ------, ∆x = x n – x 0 , i = 1, 2, …, n.
n

DOKLADY MATHEMATICS Vol. 74 No. 1 2006

 SOLUTION METHOD FOR LINEAR ORDINARY DIFFERENTIAL EQUATIONS 483

By using a multiplicative representation of solu- Thus, we have obtained the general multiplicative
tion (12) to the homogeneous differential equation, solution to the inhomogeneous linear ordinary differen-
formula (16) can be rewritten as tial equation (1):
Y * ( x n ) = K n K n – 1 …K 2 F 1 + K n K n – 1 …K 3 F 2 … Y ( x ) = K x0 ( A ( x ) )Y 0 + Y * ( x ).
x
(19)
… + K n K n – 1 …K k + 1 F k + … + K n F n – 1 + F n , (17) Since the solution is multiplicative, the resulting
formulas can be used to evaluate the integral in the
where Fi = TiGi∆xi . accepted positive direction of variation of x. When the
Formulas (16) and (17) for a multiplicative particu- direction of integral evaluation reverses, the formulas
lar solution to Eq. (1) have no permutation property in change.
the product of Cauchy–Krylov matrix functions. The remarkable properties of Cauchy–Krylov func-
For Eq. (1) with constant coefficients A(x) = A = tions were used in [2] to construct efficient priority
x
algorithms for solving boundary value problems for
const, formula (17) simplifies, since K xii – 1 = Ki = K and stiff linear ordinary differential equations and to solve
Ti = T are identical for all ∆xi = xi – xi – 1 = const, where many applied problems in the mechanics of shells and
thin-walled structures. For example, the problems were
∆x
∆xi = ------ and ∆x = xn – x0: solved of deforming isotropic, orthotropic, and layered
n shells loaded on the surfaces contoured by circles or
ellipses whose principal axes coincide with the direc-
Y * ( xn ) = [ K T G2 …
n–1 n–2
T G1 + K tions of the principal curvatures of the shells.
… + KT G n – 1 + T G n ]∆x i , (18)
REFERENCES
where the superscript of K denotes an exponent.
1. F. R. Gantmacher, The Theory of Matrices (Chelsea,
In formulas (16)–(18), the respective matrices New York, 1959; Fizmatgiz, Moscow, 1988).
x 2. A. Yu. Vinogradov and Yu. I. Vinogradov, Dokl. Phys.
K xii – 1 (A(x)), Ki, and K of Cauchy–Krylov functions are 48, 392–394 (2000) [Dokl. Akad. Nauk 373, 474–476
given by (6)–(9). (2000)].

DOKLADY MATHEMATICS Vol. 74 No. 1 2006