c Pleiades Publishing, Inc., 2006.
Original Russian Text
c V.V. Amel’kin, 2006, published in Differentsial’nye Uravneniya, 2006, Vol. 42, No. 5, pp. 579–582.
DOI: 10.1134/S0012266106050016
Let π/2 and 3π/2 be the polar angles of two rays l1 and l2 issuing from the origin O(0, 0) of the
Cartesian coordinate system on the phase plane xOy of system (2).
Definition (cf. [1]). System (2) is strongly isochronous at the singular point O(0, 0) if all phase
points lying on the rays l1 and l2 at t = t0 and beginning to move along the trajectories of the
focus or center of system (2) at time t = t0 pass from one of the rays to the other in the same
time T = π.
In the mechanical interpretation, strong isochronicity means that the vibrations in the system
are isochronous on the time interval equal to the vibration period as well as on the time interval
equal to the half-period of vibrations around a stable equilibrium.
k−2
Bn Bk−n−1
A2 = 0, Ak = × , k = 3, 4, . . . (3)
n=1
n + 2 k − n + 1
Proof. By Theorem 1 in [2] (see also Theorem 8 in [3]), system (2) is strongly isochronous if
and only if some transformation of the form
∞
u = x, v=y+ βk,l xk y l (4)
k+l=2
If we rewrite system (5) in terms of the variables x and y taking into account the transforma-
tion (4) and system (2) and match the coefficients of xp y q (p + q = n) in the resulting relations,
615
616 AMEL’KIN
then for the coefficients of the transformation (4) and the coefficients of system (5), we obtain four
systems of algebraic equations, the first of which is
β2j−1,2m−2j+1 − γ2j−2,2m−2j+1 = 0, j = 1, . . . , m,
..., β1,2m−1 + γ0,2m−1 = 0,
(6)
(2m − 2l + 3)β2l−3,2m−2l+3 − (2l − 1)β2l−1,2m−2l+1 − γ2l−2,2m−2l+1 = 0, l = 2, . . . , m,
..., β2m−1,1 = μ2m,0 ,
β2j+1,2m−2j − γ2j,2m−2j = 0, j = 0, . . . , m,
..., β1,2m + γ0,2m = 0,
(2m − 2l + 4)β2l−3,2m−2l+4 (9)
− (2l − 1)β2l−1,2m−2l+2 − γ2l−2,2m−2l+2 = 0, l = 2, . . . , m,
..., 2β2m−1,2 − (2m + 1)β2m+1,0 − γ2m,0 = B2m .
DIFFERENTIAL EQUATIONS Vol. 42 No. 5 2006
STRONG ISOCHRONICITY OF THE LIENARD SYSTEM 617
Just as in the case of system (7), the analysis of system (9), which contains 2m + 2 equations
for 2m + 2 unknowns, shows that this system is consistent and has the unique solution
∞
Bn−1 n
u = x, v=y− x (10)
n=2
n+1
∞
Bn n ∞
Bn n
u̇ = −v − u u , v̇ = u − v u . (11)
n=1
n+2 n=1
n+2
Now we write out the entire system of conditions μn,0 = 0, n = 2, 3, . . . , in closed form. By rep-
resenting the second equation in system (11) in terms of the variables x and y with regard to the
transformation (2) and the change of variables (10) as
∞ ∞ ∞
Bk−1 k−1
x+ Ak x − y
k k
Bk x + y k x
k +1
k=2 k=1 k=2
∞ (12)
∞
Bk−1 k Bk k
=x− y− x x ,
k=2
k+1 k=1
k+2
we reduce this problem to establishing a relationship between the coefficients Ak and Bl in (12).
After some simplifications, relation (12) can be represented in the form
∞ ∞
Bk−1 k Bk k
∞
Ak xk = x x ,
k=2 k=2
k + 1 k=1 k + 2
∞ ∞ k−2
Bn Bk−n−1 k
Ak x = k
× x . (13)
k=2 k=3 n=1
n+2 k−n+1
Then, by matching the coefficients of like powers of x in (13), we obtain the desired conditions (3).
The proof of the theorem is complete.
The desired assertion readily follows from the easy-to-verify equivalence of conditions (3)
and (14).
The proof of Theorem 1 implies also the following assertion.
Theorem 3. System (2) is strongly isochronous if and only if there exists a unique transforma-
tion (10) taking system (2) to system (11).
Remark 1. If the Lineard system (2) is polynomial (respectively, nonpolynomial and holomor-
phic) system, then so is the transformation (10).
Theorem 4. System (2) has a strongly isochronous center at the origin O(0, 0) if and only if
k−1
B2m+1 B2k−2m−1
A2k = 0, A2k+1 = × , k = 1, 2, . . . (15)
m=0
2m + 3 2k − 2m + 1
Proof. Obviously, system (2) has a strongly isochronous center at the origin O(0, 0) if and only
if system (11) has a center at the point O(0, 0). But the latter condition takes place if and only if
the coefficients B2k , k = 1, 2, . . . , occurring in (3) are zero.
The necessity of the conditions B2k = 0, k = 1, 2,. . . , follows from the comparison of system (11)
∞
with a simplified system (11) such that the series n=1 (Bn /(n + 2)) un contains only odd powers
of u; the center curves of this system should be chosen as a topographic system of curves.
The sufficiency readily follows from the symmetry of the direction field of system (11) with
respect to the axis Ov. If the conditions B2k = 0, k = 1, 2, . . . , are satisfied, then conditions (15)
also hold.
Theorem 5. System (2) has a strongly isochronous center at the origin O(0, 0) if and only if
B(x) is an odd function and relation (14) holds.
Theorem 6. System (2) has a strongly isochronous center at the origin O(0, 0) if and only if
there exists a unique transformation
∞
B2n−1 2n−1
u = x, v =y−x x
n=1
2n +1
Remark 2. Theorem 5 was proved in [4, Corollary 1]. In the consideration of the isochronicity
of the center in [5] (the polynomial case), [6] (the analytic and locally Lipschitzian case), [7] (the
analytic case), in a sense, the sufficiency of the assumptions of Theorem 5 on the strong isochronicity
was proved.
Remark 3. By using the form of the Lienard polynomial system in the case of a strongly
isochronous center and the results in [8], we find that the domain of the strongly isochronous
polynomial center of the Lienard system, though unbounded, is not global, i.e., does not fill the
entire phase plane.
REFERENCES
1. Kukles, I.S. and Piskunov, N.S., Dokl. Akad. Nauk , 1937, vol. 17, no. 9, pp. 467–470.
2. Amel’kin, V.V. and Korsantiya, O.B., Mezhdunar. mat. konf. “Eruginskie chteniya–X”: Tez. dokl.
(Abstr. Int. Math. Conf. “Erugin Readings–X”), Mogilev, 2005, pp. 47–48.
3. Amel’kin, V.V. and Korsantiya, O.B., Differ. Uravn., 2006, vol. 42, no. 2, pp. 147–152.
4. Rudenok, A.E., Differ. Uravn., 1975, vol. 11, no. 5, pp. 811–819.
5. Algaba, A., Freire, E., and Gamero, E., Isochronicity via Normal Form, Preprint. 1998.
6. Sabatini, M., J. Differential Equations, 1999, vol. 152, pp. 467–487.
7. Christopher, C. and Devlin, J., J. Differential Equations, 2004, vol. 200, pp. 1–17.
8. Galeotti, M. and Villarini, M., Ann. Mat. Pura Appl. (4), 1992, vol. 161, no. 4, pp. 299–313.
DOI: 10.1134/S0012266106050028
The weakening of conditions imposed on the solution operators of the Cauchy problem for
abstract first- and second-order differential equations has led (see [1–3]) to the notion of integrated
semigroup and integrated cosine function.
In the present paper, we derive formulas relating the integrated cosine function to the solution
operator Yk (t) of the Cauchy problem
k
u (t) + u (t) = Au(t), t > 0, (1)
t
u(0) = u0 , u (0) = 0, (2)
for the Euler–Poisson–Darboux equation in a Banach space E. (Here k > 0 is a parameter.)
The operator function Yk (t) was introduced in [4] and named the operator Bessel function.
The set of operators A for which problem (1), (2) is uniformly well posed will be denoted by Gk .
Thus if A ∈ Gk , then problem (1), (2) has a unique solution, which continuously depends on the
initial data; moreover, u(t) = Yk (t)u0 , u0 ∈ D(A), and
Note that the condition for problem (1), (2) to be uniformly well posed and the properties of
the operator Bessel function Yk (t) were given in [4].
Next, recall the definition of integrated cosine function.
Definition 1. Let α > 0. A one-parameter family Cα (t), t ≥ 0, of bounded linear operators is
called an α-times integrated cosine function if the following conditions are satisfied:
1.
t+s s
2Γ(α)Cα (t)Cα (s) = (t + s − r) Cα (r)dr − (t + s − r)α−1 Cα (r)dr
α−1
t 0
t s
+ (r − t + s) α−1
Cα (r)dr + (r + t − s)α−1 Cα (r)dr, t > s > 0.
t−s 0
2. Cα (0) = 0.
3. Cα (t)x is a continuous function of t ≥ 0 for each x ∈ E.
4. There exist constants M > 0 and ω ≥ 0 such that
Cα (t) ≤ M eωt , t ≥ 0. (4)
619
620 GLUSHAK
The generator A of an integrated cosine function Cα (t) is defined as follows: the domain D(A)
is the set of elements x ∈ E such that there exists an element y ∈ E satisfying the relation
t
tα
Cα (t)x − x= (t − r)Cα (r)y dr, t ≥ 0, (5)
Γ(α + 1)
0
Theorem 1. Let α > 1, let an operator A be the generator of an α-times integrated cosine
function Cα (t), and let u0 ∈ D(A). Then problem (1), (2) is uniformly well posed (i.e., A ∈ Gk ),
and the corresponding operator Bessel function can be represented in the form
⎛ ⎞
t s
2α Γ(α + 1/2) ⎝ 1
Y2α (t)u0 = √ α Cα (t)u0 −
Pα−1 Cα (s)u0 ds⎠ , (6)
πt t t
0
Proof. Formula (6) can be obtained heuristically as follows. Consider an operator cosine func-
tion C(t) and set
t
IC(t) = C(τ )dτ.
0
t
2t1−k Γ(k/2 + 1/2)
m−1 dm (I m C(s)u0 )
Yk (t)u0 = √ t2 − s 2 ds (7)
π Γ(k/2) dsm
0
t s
2Pα−1 (1) + 2α 1 (α + 1)(α + 2)
− α+1
Cα (t)u0 + α Cα (t)u0 −
Pα−1 Cα (s)u0 ds
t t tα+3 t
0
⎞
t s t s
2α + 4 1
−
Pα−1 sCα (s)u0 ds −
Pα−1 s2 Cα (s)u0 ds⎠ .
tα+4 t tα+5 t
0 0
By [5, p. 206], the spherical Legendre function Pα−1 (s) is a solution of the equation
2α
Y2α (t)u0 + Y (t)u0 = AY2α (t)u0 .
t 2α
Therefore, the function Y2α (t)u0 is a solution of Eq. (1).
Note that the representation (6) (after the change of variables s = tτ in the integral), together
with (4), implies the estimate
Y2α (t) ≤ M1 eωt . (13)
To show that the function Y2α (t)u0 satisfies the initial condition (2), we use relation (5) and the
integral 2.17.1.4 in [6] and rewrite the expression (6) in the form
t s
2α Γ(α + 1/2)
Y2α (t)u0 = √ α Pα−1 Cα (s)u0 ds
πt t
0
⎛ ⎞
α t s α−1 s
2 Γ(α + 1/2) ⎝ s u0 +
= √ α Pα−1 Cα ()Au0 d⎠ ds
πt t Γ(α)
0 0
⎛ ⎞ (14)
α 1 t s s
2 Γ(α + 1/2) ⎝ 1
= √ α τ α−1 Pα−1 (τ )u0 dτ + Pα−1 ds Cα ()Au0 d⎠
πt Γ(α) t
0 0 0
1 tτ
2α Γ(α + 1/2)
= u0 + √ α−1 Pα−1 (τ )dτ Cα ()Au0 d.
πt
0 0
Now the desired assertion follows from (5), since the last term in (14) is of the order of t2 as t → 0.
We prove the uniqueness of the solution of problem (1), (2) by contradiction. Let u1 (t) and u2 (t)
be two solutions of problem (1), (2). Consider the function w(t, s) = f (Y2α (s) (u1 (t) − u2 (t))) of
two variables t, s ≥ 0, where f belongs to the adjoint space E ∗ . Obviously, it satisfies the equation
∂ 2 w 2α ∂w ∂ 2 w 2α ∂w
+ = + , t, s > 0,
∂t2 t ∂t ∂s2 s ∂s
and the conditions
∂w(0, s) ∂w(t, 0)
w(0, s) = = = 0.
∂t ∂s
DIFFERENTIAL EQUATIONS Vol. 42 No. 5 2006
ON THE RELATIONSHIP BETWEEN THE INTEGRATED COSINE FUNCTION 623
By using the change of variables t1 = (t + s)2 /4, s1 = (t − s)2 /4, one can reduce [7, Sec. 5, item 3]
the last problem to the problem whose uniqueness in the class of twice continuously differentiable
functions for t, s ≥ 0 was proved in [7, Sec. 5, item 2]. Moreover, the desired uniqueness is also
contained in Theorem 6.1 in [8], where even a more general equation was considered.
Consequently, w(t, s) ≡ 0, and since the functional f ∈ E ∗ is arbitrary, for s = 0, we obtain the
relation u1 (t) ≡ u2 (t), and the proof of the uniqueness is complete.
Thus the operator function Y2α (t) satisfies inequality (13), and the function Y2α (t)u0 is the unique
solution of problem (1), (2); consequently, problem (1), (2) is uniformly well posed. The proof of
the theorem is complete.
Remark 1. Theorem 1 remains valid for α = 1. In this case, the proof is much simpler, and
Y2 (t)u0 = (1/t)C1 (t)u0 , where C1 (t) can naturally be referred to as the operator sine function.
Theorem 2. Let A ∈ Gk , k > 0, and let Yk (t) be the corresponding operator Bessel function.
Then the operator A is the generator of an integrated cosine function Cn (t), where n is the least
positive integer such that 2n ≥ k.
Proof. First, note [4] that the operator Bessel function Y2n (t) can be expressed via the operator
Bessel function Yk (t) with the use of the formula for a parameter shift:
1
2
(m−k)/2−1
Ym (t) = 1 − s2 sk Yk (ts)ds, m > k,
B(k/2 + 1/2, m/2 − k/2)
0
Therefore, problem (16), (2) is exponentially uniformly n-well posed. It follows from Theorem 1.3
in [11] that the operator A is the generator of an integrated cosine function Cn (t). The proof of
the theorem is complete.
Remark 2. By virtue of the uniqueness of the solution, we have Cn (t)u0 = I n Y0 (t)u0 and
u0 ∈ D (An ), where Y0 (t) is given by (15). After integration by parts, we obtain
t
n
1 1 d 2n−1
n
Cn (t)u0 = I Y0 (t)u0 = (t − s)n−1 s s Y2n (s)u0 ds
(n − 1)! (2n − 1)!! s ds
0
n−2
t
−
(−1)n 1 d (t s)n−2
= s2n−1 Y2n (s)u0
(n − 2)! (2n − 1)!! s ds s
0
t
n−1
1 d (t − s)n−2
− s2n Y2n (s)u0 ds
s ds s
0
for n ≥ 2; moreover, the operator function I n Y0 (t), originally defined on the dense [4] set D (An ),
can be extended to the entire space E. For example,
t
t2 1
C1 (t) = tY2 (t), C2 (t) = Y4 (t) + sY4 (s)ds,
3 3
0
t
t3 t
C3 (t) = Y6 (t) + sY6 (s)ds,
15 5
0
In conclusion, we recall the definition of integrated semigroup and show how to use it so as to
weaken the conditions imposed on the operator A occurring in the problem [12]
k k
v (t) + v(t) = Av(t) + g, t > 0, (18)
kt
t
lim t v(t) = v0 . (19)
t→0
Theorem 3. Let k ∈ N,
let A be the generator of a k-times integrated semigroup Tk (t), and let
g ∈ D(A) and v0 ∈ D Ak+1 . Then the function
v(t) = t−k k! Tk (t)g + Tk(k) (t)v0
k
tj−k
v(t) ≤ M eωt g + Ak v0 + Aj v0 . (21)
j=0
j!
Proof. Note that the representation of the solution v(t) of problem (18), (19) was obtained
from Theorem 7 in [12], where it was assumed that A is the generator of a C0 -semigroup T (t), by
replacing the kth fractional integral of the semigroup T (t) by the integrated semigroup Tk (t).
By taking into account the relation (e.g., see [3])
(k)
k−1 j
t
Tk (t)v0 = Tk (t)Ak v0 + Aj v0 (22)
j=0
j!
and definition (20) of the generator of the integrated semigroup Tk (t), we compute v (t). We obtain
−k−1 (k) −k (k)
v (t) = −kt k! Tk (t)g + Tk (t)v0 + t k−1
k! ATk (t)g + kt g + ATk (t)v0 ;
therefore,
k k
v (t) + v(t) = Av(t) + g;
t t
i.e., the function v(t) satisfies Eq. (18).
It also follows from (20) and (22) that the initial condition is satisfied, since
(k)
lim tk v(t) = k! lim Tk (t)g + lim Tk (t)v0 = v0 .
t→0 t→0 t→0
The estimate (21) is a consequence of item 4 of Definition 2 and relations (20) and (22). Indeed,
k−1 j
t
v(t) ≤ t−k k! M1 tk eωt g + M1 tk eωt Ak v0 + Aj v0
j=0
j!
k
k
tj−k
= M e g + A v0 +
ωt Aj v0 .
j=0
j!
ACKNOWLEDGMENTS
The work was financially supported by the Russian Foundation for Basic Research (project
no. 04-01-00141).
REFERENCES
1. Arendt, W., Israel J. Math., 1987, vol. 59, pp. 327–352.
2. Arendt, W. and Kellermann, H., Pitman Res. Notes Math., 1989, vol. 190, pp. 21–51.
3. Mel’nikova, I.V. and Filinkov, A.I., Uspekhi Mat. Nauk , 1994, vol. 49, no. 6(300), pp. 111–150.
4. Glushak, A.V., Dokl. Akad. Nauk , 1997, vol. 352, no. 5, pp. 587–589.
5. Lebedev, N.N., Spetsial’nye funktsii i ikh prilozheniya (Special Functions and Their Applications),
Moscow: Gosud. Izdat. Fiz.-Mat. Lit., 1963.
6. Prudnikov, A.P., Brychkov, Yu.A., and Marichev, O.I., Integraly i ryady. Dopolnitel’nye glavy (Integrals
and Series. Additional Chapters), Moscow: Nauka, 1986.
7. Levitan, B.M., Uspekhi Mat. Nauk , 1951, vol. 1, no. 2(42), pp. 102–143.
8. Bragg, L.R., J. Math. Mech., 1969, vol. 18, pp. 607–616.
9. Glushak, A.V. and Shmulevich, S.D., Differ. Uravn., 1992, vol. 28, no. 5, pp. 831–838.
10. Tersenov, S.A., Vvedenie v teoriyu uravnenii, vyrozhdayushchikhsya na granitse (Introduction to Theory
of Equations Degenerating on Boundary), Novosibirsk: Novosibirsk Gos. Univ., 1973.
11. Zheng, Q., Int. J. Math. Math. Sci., 1996, vol. 19, no. 3, pp. 575–580.
12. Glushak, A.V., Differ. Uravn., 1998, vol. 34, no. 9, pp. 1284–1285.
DOI: 10.1134/S0012266106050041
1. INTRODUCTION
α
Let D0+,t u (x, t) be the partial fractional Riemann–Liouville derivative of order α > 0 of a
function u(x, t) with respect to the second variable [1, p. 342], given by
n t
α
∂ 1 u(x, τ )dτ
D0+,t u (x, t) =
∂t Γ(n − α) (t − τ )α−n+1 (1)
0
(α > 0, n = [α] + 1, x ∈ Rm , t > 0) ,
α
and let c D0+,t u (x, t) be the so-called partial fractional Caputo derivative of order α > 0, which
is defined via the derivative (1) by the formula
c α
n−1 k k
t ∂ u
D0+,t u (x, t) = D0+,t u(x, t) −
α
(x, 0) (x, t)
k=0
k! ∂tk (2)
(α > 0, n = −[−α], x ∈ Rm , t > 0) .
where λ > 0 and Δx is the Laplace operator with respect to the first variable
m
x = (x1 , x2 , . . . , xm ) ∈ Rm : Δx = ∂ 2 u/∂x2j .
j=1
If α = 1 and α = 2, then
c 1
∂u(x, t) c 2
∂ 2 u(x, t)
D0+,t u (x, t) = and D0+,t u (x, t) = ,
∂t ∂t2
respectively, and Eq. (3) coincides with the heat (diffusion) equation
∂u
= λ2 Δx u(x, t) (x ∈ Rm , t > 0) (4)
∂t
for α = 1 and with the wave equation
∂2u
= λ2 Δx u(x, t) (x ∈ Rm , t > 0) (5)
∂t2
638
THE CAUCHY PROBLEM FOR THE DIFFUSION-WAVE EQUATION 639
for α = 2. That is why Eq. (3) is called the diffusion-wave equation [2, p. 146]. In particular, for
m = 1, Eq. (3) acquires the form
c ∂ 2 u(x, t)
α
D0+,t u (x, t) = λ2 (x ∈ R, t > 0). (6)
∂x2
The interest in Eqs. (3) and (6) is due to their numerous applications to the solution of diffusion
problems in physics, mechanics, and other applied sciences [2, Secs. 4.2.1, 4.2.2; 3; 4]. (For historical
information and a survey of results, see [5, Chap. 7].)
In particular, the solution of Eq. (3) of order α ∈ (0, 1) with the initial condition u(x, 0) = f (x)
(x ∈ Rm ) and the fundamental solution [i.e., the solution for the case in which f (x) is the Dirac
delta function δ(x)] were found in [6]. Similar results for Eq. (6) of order α ∈ (0, 1) were obtained
in [7, 8].
The fundamental solution of Eq. (6) of order α ∈ (1, 2] with the initial conditions
∂u(x, 0)
u(x, 0) = δ(x), =0 (x ∈ R)
∂t
was constructed in [9]. Solutions of some boundary value problems in signal theory for Eq. (6) of
order α ∈ (0, 2] in the quadrant R+ = {(x, t) ∈ R2 : x > 0, t > 0} are given in [10].
In the present paper, we solve Eq. (3) of order α > 0 with the initial conditions
∂ku
(x, 0) = fk (x) (k = 0, . . . , n − 1, n = −[−α], x ∈ Rm ) . (7)
∂tk
If α = n ∈ N, then problem (3), (7) becomes the Cauchy problem for an equation with an nth
partial derivative:
∂ n u(x, t)
= λ2 Δx u(x, t) (x ∈ Rm , t > 0) , (8)
∂tn
∂ k u(x, 0)
= fk (x) (k = 0, . . . , n − 1, x ∈ Rm ) . (9)
∂tk
Therefore, by analogy, problem (3), (7) is also referred to as the Cauchy problem.
of the function u(x, t) with respect to t > 0 and x ∈ Rm , respectively, and the inverse transforms
γ+i∞
−1
1
Ls u (x, t) = est u(x, s)ds (x ∈ Rm , t > 0) , (12)
2πi
γ−i∞
1
Fσ −1 u (x, t) = u(σ, t)e−iσ·x dσ (x ∈ Rm , t > 0) (13)
(2π)m
Rm
inverse Laplace and Fourier transforms, e.g., see [11, Chap. II; 12, Chap. 16]. In particular, the
operators (12), (10) and (13), (11) are mutually inverse on sufficiently smooth functions u(x, t).
Suppose that there exist Fourier transforms (Fx fk ) (σ) of the functions fk (x) (k = 0, . . . , n − 1)
occurring in the initial conditions (7). By applying the Laplace transform (10) to both sides of
Eq. (3) and by using the initial conditions (7) and the formula
n−1
∂k u
Lt c D0+,t
α
u (x, s) = sα (Lt u) (x, s) − sα−k−1 (x, 0) (14)
k=0
∂tk
for the Laplace transform of the partial fractional Caputo derivative [2, Eq. (2.253)], we obtain
n−1
sα (Lt u) (x, s) − sα−k−1 fk (x) = λ2 (Δx Lt u) (x, s). (15)
k=0
By applying the Fourier transform (11) to this relation and by using the formula
(Fx Δx u) (σ, s) = −|σ|2 (Fx u) (σ, s)
for the Fourier transform of the operator Δx , we arrive at the relation
n−1
sα−k−1
(Fx Lt u) (σ, s) = (Fx fk ) (σ) (σ ∈ Rm , t > 0) , (16)
k=0
sα + λ2 |σ|2
m
where |σ|2 = i=1 σi2 .
By using the inverse Laplace and Fourier transforms (12) and (13), we find the solution u(x, t)
of the original problem (3), (7) in the form
n−1
sα−k−1
u(x, t) = Ls−1 Fσ −1 (Fx fk ) (σ) (x, t). (17)
k=0
sα + λ2 |σ|2
which is an entire function of z. By virtue of the formula for the Laplace transform of this function
(e.g., see [1, Eq. (1.93)]), we have
sα−k−1
Lt tk Eα,k+1 −λ2 σ 2 tα (s) =
sα + λ2 |σ|2 (19)
−2/α
k = 0, . . . , n − 1, s ∈ C, σ ∈ R , λ > 0, λ|σ||s|m
<1 .
By applying the inverse Laplace and Fourier transforms to both sides of relation (16) and by taking
into account (19), we obtain the solution of problem (3), (7) in the form
n−1
tk
u(x, t) = Eα,k+1 −λ2 |σ|2 tα (Fx fk ) (σ)e−ix·σ dσ. (20)
k=0
(2π)m
Rm
Theorem 1. Suppose that there exist Fourier transforms (Fx fk ) (σ) of the functions fk (x)
(k = 0, . . . , n − 1) and the integrals on the right-hand side in (20) are convergent. Then the
Cauchy problem (3), (7) is solvable, and the solution is given by (20). In particular, if m = 1, then
relation (20) defines a solution of problem (6), (7).
Corollary 1. The solution of the Cauchy problem (8), (9) is given by Eq. (20) with α = n
provided that the integrals on the right-hand side in (20) are convergent.
for the jth powers of the Laplace operator (Δ1x = Δx , Δjx = Δ1x Δj−1
x , j = 2, 3, . . .), we obtain the
representation of the solution (20) in the form
n−1
∞
(λ2 tα ) j
j
u(x, t) = tk Δx fk (x). (21)
k=0 j=0
Γ(αj + k + 1)
Theorem 2. Let fk (x) (k = 0, . . . , n−1) be infinitely differentiable functions in Rm , and let the
series on the right-hand side in (21) be convergent. Then the Cauchy problem (3), (7) is solvable,
and the solution is given by (21). In particular, if m = 1, then the solution of problem (6), (7) has
the form
n−1 ∞
(λ2 tα )
j
(2j)
k
u(x, t) = t fk (x). (22)
k=0 j=0
Γ(αj + k + 1)
Corollary 2. The solution of the Cauchy problem (8), (9) is given by Eq. (21) with α = n
provided that the series on the right-hand side in (21) are convergent.
Proof. We use the formula for the Fourier transform of a radial function in Rm (e.g., see
[1, Eq. (25.11)]):
∞
(2π)m/2
(Fx [ϕ(|x|)]) (σ) = ϕ()m/2 Jm/2−1 (|σ|)d. (24)
|σ|m/2−1
0
n−1 α(m+2)/4−k−1
s |x| α/2
(Lt u) (x, s) = |x|
1−m/2
Km/2−1 s ∗x fk (x) (x ∈ Rm , s > 0) (30)
k=0
λ(2πλ)m/2 λ
for x ∈ Rm and s ∈ C.
By applying the inverse Laplace transform (12) to Eq. (30), one can obtain an explicit solution
of problem (3), (7). To this end, one should find the inverse Laplace transform of the functions
|x| α/2
s α(m+2)/4−k−1
Km/2−1 s (k = 0, 1, . . . , n − 1).
λ
DIFFERENTIAL EQUATIONS Vol. 42 No. 5 2006
THE CAUCHY PROBLEM FOR THE DIFFUSION-WAVE EQUATION 643
Let us show that, for k = 0, 1, these functions can be represented via the Laplace transform of
the H-function
2,0 (a, 1/2), (b, α/2) 1 Γ(c + τ )Γ(d + τ /2) −τ
H2,2 z = z dτ. (31)
(c, 1), (d, 1/2) 2πi Γ(a + τ /2)Γ(b + ατ /2)
L
Here a, b, c, d ∈ R, and L is a special infinite contour such that the poles of the gamma functions
Γ(c + τ ) and Γ(d + τ /2) lie to the left of this contour. It follows from Theorem 1.1 in [15] that the
function (31) exists for 0 < α < 2, a, b, c, d ∈ R, and | arg z| < (2 − α)π/4, z = 0.
|x|
2,0 −α/2 (m/4, 1/2), (k + 1 − α(m + 2)/4, α/2)
Lt t k−α(m+2)/4
H2,2 t (s)
λ (m/2 − 1, 1), (1/2 − m/4, 1/2)
(32)
m/2 −1/2 α(m+2)/4−k−1 |x| α/2
=2 π s Km/2−1 s .
λ
2,0
Proof. We use the representation (31) for the H2,2 -function in (32) and choose a contour L such
that Re (τ ) > (m + 2)/2 − 2/α. Then k − α(m + 2)/4 + (α/2) Re (τ ) > −1 for 0 < α ≤ 1, k = 0
and 1 < α < 2, k = 0, 1, which provides the convergence of the improper integrals occurring below.
It follows from (10) and (31) that
−k+α(m+2)/4 2,0 |x| −α/2 (m/4, 1/2), (k + 1 − α(m + 2)/4, α/2)
Lt t H2,2 t (s)
λ (m/2 − 1, 1), (1/2 − m/4, 1/2)
−τ
1 Γ(m/2 − 1 + τ )Γ(1/2 − m/2 + τ /2) |x|
= dτ
2πi Γ(m/4 + τ /2)Γ(k + 1 − α(m + 2)/4 + ατ /2) λ
L
∞
× e−st tk−α(m+2)/4+ατ /2 dt (33)
0
−τ
−k−1+α(m+2)/4 Γ(m/2 − 1 + τ )Γ(1/2 − m/2 + τ /2) |x| α/2
1
=s s dτ
2πi Γ(m/4 + τ /2) λ
L
−k−1+α(m+2)/4 2,0 |x| α/2 (m/4, 1/2)
=s H1,2 s .
λ (m/2 − 1, 1), (1/2 − m/4, 1/2)
implies (32).
Let n = 1 if 0 < α ≤ 1 and n = 2 if 1 < α < 2. By using the expression (32), we rewrite
relation (30) in the form
n−1
2−m |x|1−m/2
(Lt u) (x, s) = 1+m/2 (m−1)/2 Lt tk−α(m+2)/4
λ π
k=0
|x|
(m/4, 1/2), (k + 1 − α(m + 2)/4, α/2)
× H2,2
2,0
t−α/2 ∗x fk (x) (s).
λ (m/2 − 1, 1), (1/2 − m/4, 1/2)
DIFFERENTIAL EQUATIONS Vol. 42 No. 5 2006
644 VOROSHILOV, KILBAS
By applying the inverse Laplace transform (12), we obtain the solution u(x, t) of problem (3), (7)
in the closed form
2−m |x|1−m/2 k−α(m+2)/4
n−1
u(x, t) = t
λ1+m/2 π (m−1)/2 k=0
(35)
|x|
−α/2 (m/4, 1/2), (k + 1 − α(m + 2)/4, α/2)
× H2,2
2,0
t ∗x fk (x).
λ (m/2 − 1, 1), (1/2 − m/4, 1/2)
2−m |x|1−m/2 −α(m+2)/4 2,0 |x| −α/2 (m/4, 1/2), (1 − α(m + 2)/4, α/2)
c
Gα0 (x, t) = 1+m/2 (m−1)/2 t H2,2 t (38)
λ π λ (m/2 − 1, 1), (1/2 − m/4, 1/2)
Theorem 4. If 1 < α < 2, m ∈ N, and λ > 0, then the Cauchy problem for Eq. (3) with the
initial conditions
∂u
u(x, 0) = f0 (x), (x, 0) = f1 (x) (x ∈ Rm ) (39)
∂t
is solvable, and its solution has the form
u(x, t) = [c Gα0 (x − τ, t)f0 (τ ) + c Gα1 (x − τ, t)f1 (τ )] dτ, (40)
Rm
|x| (41)
(m/4, 1/2), (2 − α(m + 2)/4, α/2)
× H2,2
2,0
t−α/2
λ (m/2 − 1, 1), (1/2 − m/4, 1/2)
by the formula
Corollary 4. If 1 < α < 2 and λ > 0, then the Cauchy problem for Eq. (3) with the initial
conditions (9) (with n = 2 and m = 1) is solvable, and its solution is given by Eq. (40) with m = 1,
where c Gα0 (x, t) is given by (44) and
1 1−α/2 α α |x| −α/2
c α
G1 (x, t) = t ϕ − ,2 − ;− t . (45)
2λ 2 2 λ
Remark 1. The solution of the Cauchy problem (36) with 0 < α < 1 and λ = 1 was ob-
tained in [6] in the form (37) with the kernel c Gα0 (x, t), different from (38). By using the properties
of H-functions [15, Chap. 2], one can readily show that this solution coincides with the solu-
tion (37), (38).
Remark 2. The solution of problem (6), (43) in Corollary 3 coincides with the known solution
obtained in [7].
Lemma 3. Let m ∈ N and 0 < α < 2, and let k = 0 for 0 < α ≤ 1 and k = 0, 1 for 1 < α < 2.
Then the asymptotic estimates
2,0 |x| −α/2 (m/4, 1/2), (k + 1 − α(m + 2)/4, α/2)
H2,2 t
λ (m/2 − 1, 1), (1/2 − m/4, 1/2)
α/(2−α)
(2/(2−α))[α(m+2)/4−k−1]
(2 − α) α|x| |x|
= Ak exp − (46)
2 2λtα/2 λtα/2
α/2 2/(2−α)
t
× 1+O (|x| → ∞, k = 0, 1),
|x|
are valid for each given t > 0, where the Ak (k = 0, 1) are some constants.
q,0
Proof. Formula (46) follows from the exponential behavior of the more general Hp,q -function
at infinity, which was obtained in Theorem 1.10 in [15].
From (46), we obtain asymptotic formulas for the functions c Gα0 (x, t) and c Gα1 (x, t) occurring
in (38) and (41):
α/(2−α)
(2 − α) α|x|
Gk (x, t) = Bk exp −
c α
|x|1−m/2 tk−α(m+2)/4
2 2λtα/2
(2/(2−α))[α(m+2)/4−k−1] α/2 2/(2−α)
|x| t (47)
× 1+O ,
λt α/2 |x|
2−m
Bk = Ak (k = 0, 1)
λ1+m/2 π (m−1)/2
for given t > 0 as |x| → ∞, where k = 0 for 0 < α ≤ 1 and k = 0, 1 for 1 < α < 2.
Since 0 < α < 2, it follows from (47) that c Gα0 (x, t) and c Gα1 (x, t) tend to zero for each t > 0
as |x| → ∞ :
lim c Gα0 (x, t) = lim c Gα1 (x, t) = 0 (x ∈ Rm ) . (48)
|x|→∞ |x|→∞
Consequently, the solutions (37) and (40) of the Cauchy problems (36) and (3), (9) (n = 2) vanish
at infinity for each t > 0 :
lim u(x, t) = 0. (49)
|x|→∞
This formula shows that the solutions (37) and (40) of the Cauchy problems for partial differential
equations of fractional orders 0 < α < 1 and 1 < α < 2 vanish with respect to x at infinity just as
the solutions of the Cauchy problem for the heat equation (4) and the wave equation (5).
Thus condition (49) could also be viewed as an initial condition for the Cauchy problem (36)
and (3) (with 1 < α < 2), (9) with n = 2.
6. EXAMPLES
Let us give examples of the solution of the Cauchy problem (3), (7) and illustrate the results
for u(x, 0) = (sin x)/x by graphs constructed with the use of Mathematica. Let us start from the
solution for a partial differential equation of order 1/2.
Example 1. According to (22), the Cauchy problem for Eq. (6) with α = 1/2 and u(x, 0) = f (x)
(x ∈ R, t > 0) has the solution
j
∞
λ2 t1/2
u(x, t) = f (2j) (x).
j=0
Γ(j/2 + 1)
The graph of the solution constructed for λ = 1 and f (x) = (sin x)/x with the use of Mathematica
is shown in Fig. 1.
Example 2. Consider the Cauchy problem for the first-order partial differential equation of
the form (4) with the initial condition u(x, 0) = f (x) (x ∈ Rm ).
By Theorem 3, this problem has a solution of the form (37) with α = 1.
Lemma 4. For α = 1, the function c Gα0 (x, t) in (38) has the form
1
√ m t−m/2 e−|x| /(4λ t) .
2 2
c
G10 (x, t) = (50)
(2λ π )
Fig. 1. Fig. 2.
Proof. By using properties 2.1, 2.2, and 2.5 of the H-function [15], we obtain
2,0 |x| −1/2 (m/4, 1/2), (1/2 − m/4, 1/2) 1,0 |x| −1/2 (m/4, 1/2)
H2,2 t = H1,1 t
λ (m/2 − 1, 1), (1/2 − m/4, 1/2) λ (m/2 − 1, 1)
m/2−1
|x| −1/2
1,0 |x| −1/2 (1/2, 1/2)
= t H1,1 t (51)
λ λ (0, 1)
m/2−1
|x| −1/2 1 1 |x| −1/2
= t ϕ − , ;− t .
λ 2 2 λ
By definition (42) of the Wright function, one can directly show that
1 1 1
ϕ − , ; z = √ e−z /4 .
2
(52)
2 2 π
2,0 |x| −1/2 (m/4, 1/2), (1/2 − m/4, 1/2)
H2,2 t
λ (m/2 − 1, 1), (1/2 − m/4, 1/2) (53)
−1/2 1/2−m/4 −|x|2 /(4λt)
= |x| m/2−1 1−m/2
λ π t e .
By substituting the expression (53) into Eq. (38) with α = 1, we obtain (50).
Therefore, for α = 1, Eq. (37) specifies the well-known solution of the Cauchy problem (4), (36)
(e.g., see [18, pp. 230, 482]):
1
√ m t−m/2 e−|x| /(4λ t) .
2 2
u(x, t) = G(x − τ, t)f (τ )dτ, G(x, t) =
(2λ π )
Rm
∂u ∂2u
(x, t) = λ2 2 (x, t), u(x, 0) = f (x) (x ∈ R, t > 0)
∂t ∂x
has the solution
1 −1/2 −|x|2 /(4λ2 t)
u(x, t) = G(x − τ, t)f (τ )dτ, G(x, t) = √ t e ;
2λ π
R
the graph of this solution constructed with the use of Mathematica for λ = 1 and f (x) = (sin x)/x
is shown in Fig. 2.
Example 3. By (22), the Cauchy problem for Eq. (6) with the initial conditions
∂u
u(x, 0) = f (x), (x, 0) = ϕ(x) (x ∈ R, t > 0) (54)
∂t
has the solution
2 3/2 j 2 3/2 j
∞
λ t ∞
λ t
(2j)
u(x, t) = f (x) + t ϕ(2j) (x). (55)
j=0
Γ(3j/2 + 1) j=0
Γ(3j/2 + 2)
The graph of this solution for λ = 1, f (x) = (sin x)/x, and ϕ(x) = 0 is shown in Fig. 3.
Fig. 3. Fig. 4.
Example 4. Consider the Cauchy problem for the wave equation with m = 1 :
∂2u 2
2∂ u
= λ (x ∈ R, t > 0) (56)
∂t2 ∂x2
with the initial conditions (54).
If f and ϕ are infinitely differentiable functions, then, by Theorem 2, the solution of this problem
can be represented in the form
∞
(λt)2j ∞
(λt)2j (2j)
u(x, t) = f (2j) (x) + t ϕ (x). (57)
j=0
(2j)! j=0
(2j + 1)!
Let Φ(x) be an antiderivative of ϕ(x). By using the Taylor expansion of f and ϕ in a neighborhood
of x, we obtain
∞
(λt)2j (2j) 1 (λt)2j+1 (2j+1)
∞
u(x, t) = f (x) + Φ (x)
j=0
(2j)! λ j=0 (2j + 1)!
∞
1 (λt)i (i) ∞
(−λt)i (i)
= f (x) + f (x)
2 i=0 i! i=0
i!
∞
1 (λt)i (i) ∞
(−λt)i (i)
+ Φ (x) − Φ (x) (58)
2λ i=0 i! i=0
i!
f (x + λt) + f (x − λt) Φ(x + λt) − Φ(x − λt)
= +
2 2λ
x+λt
f (x + λt) + f (x − λt) 1
= + ϕ(τ )dτ,
2 2λ
x−λt
which coincides with the well-known solution of the Cauchy problem (56), (54) (e.g., see [18, p. 56]).
The graph of the solution of this problem constructed for λ = 1, f (x) = (sin x)/x, and ϕ(x) = 0
with the use of Mathematica is shown in Fig. 4.
Example 5. By (22), the Cauchy problem for Eq. (6) with α = 5/2 and with the initial
conditions
∂u ∂2u
u(x, 0) = f (x), (x, 0) = ϕ(x), (x, 0) = ψ(x) (x ∈ R, t > 0)
∂t ∂t2
has the solution
j 2 5/2 j 2 5/2 j
∞
λ2 t5/2 ∞
λ t ∞
λ t
2j 2j 2
u(x, t) = f (x) + t ϕ (x) + t ψ 2j (x).
j=0
Γ(5j/2 + 1) j=0
Γ(5j/2 + 2) j=0
Γ(5j/2 + 3)
Fig. 5.
The graph of the surface of this function for λ = 1, f (x) = (sin x)/x, and ϕ(x) = ψ(x) = 0 is
shown in Fig. 5.
ACKNOWLEDGMENTS
The work was financially supported by the Belarus Republic Foundation for Basic Research
(project no. F05MS-050).
REFERENCES
1. Samko, S.G., Kilbas, A.A., and Marichev, O.I., Integraly i proizvodnye drobnogo poryadka i nekotorye
ikh prilozheniya (Fractional Integrals and Derivatives and Some of Their Applications), Minsk, 1987.
2. Podlubny, I., Fractional Differential Equations, San Diego, 1999.
3. Carpinteri, A. and Mainardi, F., Fractals and Fractional Calculus in Continuum Mechanics, Wien, 1997.
4. Hilter, R., Applications of Fractional Calculus in Physics, Singapore, 2000.
5. Kilbas, A.A. and Trujillo, J.J., Appl. Anal., 2002, vol. 81, no. 2, pp. 435–493.
6. Kochubei, A.N., Differ. Uravn., 1990, vol. 26, no. 4, pp. 660–670.
7. Mainardi, F., Waves and Stability in Continuous Media (Bologna, 1993 ). Ser. Adv. Math. Appl. Sci.,
1994, vol. 23, pp. 246–251.
8. Mainardi, F., Izv. Vyssh. Uchebn. Zaved. Radiofiz., 1995, no. 1–2, pp. 20–36.
9. Mainardi, F., Appl. Math. Lett., 1996, vol. 9, no. 6, pp. 23–28.
10. Gorenflo, R. and Mainardi, F., Matimyás Mat., 1998, vol. 21, pp. 109–118.
11. Ditkin, V.A. and Prudnikov, A.P., Integral’nye preobrazovaniya i operatsionnoe ischislenie (Integral
Transforms and Calculus of Operations), Moscow, 1974.
12. Nikol’skii, S.M., Kurs matematicheskogo analiza (Course of Mathematical Analysis), Moscow, 1967.
13. Erdélyi, A., Magnus, W., Oberhettinger, F., and Tricomi, F.G., Higher Transcendental Functions
(Bateman Manuscript Project), New York: McGraw-Hill, 1953. Translated under the title Vysshie
transtsendentnye funktsii, Moscow, 1967, vol. 3.
14. Prudnikov, A.P., Brychkov, Yu.A., and Marichev, O.I., Integraly i ryady. T. 3. Dopolnitel’nye glavy
(Integral and Series. Vol. 3. Additional Chapters), Moscow, 1986.
15. Kilbas, A.A. and Saigo, M., H-Transforms. Theory and Applications, Boca Raton, 2004.
16. Erdélyi, A., Magnus, W., Oberhettinger, F., and Tricomi, F.G., Higher Transcendental Functions
(Bateman Manuscript Project), New York: McGraw-Hill, 1953. Translated under the title Vysshie
transtsendentnye funktsii, Moscow, 1974, vol. 2.
17. Prudnikov, A.P., Brychkov, Yu.A., and Marichev, O.I., Integraly i ryady (Integral and Series), Moscow,
1986, vol. 2.
18. Tikhonov, A.N. and Samarskii, A.A., Uravneniya matematicheskoi fiziki (Equations of Mathematical
Physics), Moscow, 2004.
DOI: 10.1134/S0012266106050053
1. INTRODUCTION
In the present paper, we analyze the asymptotic stability of the Cauchy problem for the system
of partial differential equations with constant coefficients [1]
SG , A(1) , . . . , A(n) are real m×m matrices, I is the identity m×m matrix, and B is a large parameter.
There are two fundamentally different approaches to modeling nonequilibrium processes, micro-
scopic and phenomenological. The first method for the analysis of nonequilibrium processes uses
the description of system states via the distribution function introduced in statistical physics.
By establishing equations (kinetic equations) determining the variation of the velocity distribution
function of molecules or other microobjects (electrons, ions, etc.) in space or time, this approach
permits one to derive the laws governing certain systems from the analysis of their microstruc-
ture. The goal of kinetics [2, 3] is to establish relationships between micro- and macroscopic
variables, namely, between the distribution function and the fluxes of various thermodynamic vari-
ables. By defining thermodynamic variables (mass, velocity, energy, entropy, etc.) as moments
of the distribution function [2, 4], kinetics reduces the analysis of kinetic equations to an infi-
nite system of quasilinear partial differential equations with relaxation for the moment system
M = {m0 (x, t), . . . , mα (x, t), . . .}. (These equations are linear for the Fokker–Planck equation [1].)
There are numerous ways to close this system (truncate the infinite chain of equations), which
permit one to approximate it by a finite system of conservation laws with a twice divergent (in Go-
dunov’s terminology [5, p. 64]) form in the first N moments MN = {m0 (x, t), . . . , mN (x, t)}.
The simplest form (1.1) is given by the representation of the Fokker–Planck equation in the basis
of Hermite functions, that is, the Grad moment systems [1, 4].
For a linear system with constant coefficients, stability analysis of the Cauchy problem can be
reduced to the analysis of the dispersion equation
R (τ, ξ, B) := det Iτ + A(j) ξj − iBBG = 0, (1.2)
where BG = −SG . The polynomial pencil (1.2) is said to be stable if all of its roots τ = τj (ξ),
j = 1, . . . , N , lie in the open upper half-plane of the complex plane, i.e.,
R (τ, ξ, B) = 0, Im τ ≤ 0, |τ | + |ξ| > 0. (1.3)
The stability condition (1.3) is equivalent to the asymptotic stability of the Cauchy problem for
the original system.
650
ON THE PROPERTIES OF THE DISPERSION EQUATIONS 651
[N/2]
where Pj (τ, ξ) are homogeneous polynomials of order m − j with unit leading coefficients with
respect to τ .
The analysis of polynomial pencils (2.1) of dispersion equations of Grad moment systems for
the Boltzmann [7–9] and Fokker–Planck kinetic equations has shown that they have a very rigid
structure.
1. The constants γj , j = 1, . . . , N , satisfy the Routh–Hurwitz conditions [11, p. 66].
2. The homogeneous polynomials Pj (τ, ξ) are nonstrictly hyperbolic with respect to τ , and
the roots of successive polynomials nonstrictly separate each other; i.e., the Poisson brackets of
successive polynomials satisfy
Lemma 2.1. Suppose that the leading polynomial P0 in the pencil (2.1) has real roots of con-
n
stant multiplicity, the matrix j=1 ξj Aj of system (1.1) has elementary divisors only of degree 1
(the assumption of the Petrovskii theorem [12] on the reduction of a matrix to a canonical form),
and the relaxation matrix BN = −SN , where SN is the matrix representing the impact operator in
the basis of the first N Hermite functions, is a diagonal matrix with entries
where Nb is the number of basic equilibrium variables. Then the following assertions hold.
1. The polynomials P0 and P1 in the polynomial pencil (2.1) of the dispersion equation of the
Grad moment system form a proper hyperbolic pair of polynomials; i.e., they are hyperbolic and
have common multiple roots (the multiplicities of which are less by unity for P1 ), and their roots
nonstrictly separate each other.
2. If PN , N = m − Nb , is a strictly hyperbolic polynomial, then the polynomials PN −1 and PN
in the polynomial pencil (2.1) form a strictly hyperbolic pair of polynomials; i.e., they are strictly
hyperbolic, and their roots strictly separate each other.
3. If PN is an nonstrictly hyperbolic polynomial, then the polynomials PN −1 and PN form a
proper nonstrictly hyperbolic pair of polynomials, and all polynomials Pj , j = 0, . . . , N, in the chain
have a common real root.
Without loss of generality, it suffices to prove this assertion in the case Nb = 1, which corresponds
to the Grad moment system for the Fokker–Planck
n equation.
1. For the moment systems, the matrix j=1 ξj Aj has real eigenvalues and satisfies the assump-
tions of the Petrovskii theorem [12, pp. 100–104] on the reduction of a matrix to a canonical form.
Hence it follows that there exists a nondegenerate transformation C(ξ), det(C(ξ)) > 0, that reduces
this matrix to a canonical form,
n
ξj Aj = C(ξ) (−τj (ξ)δjk ) C −1 (ξ).
j=1
First, we show that the transposed factor-matrix of minors can be represented in the form
M 1
= C(ξ) δjk C −1 (ξ). (2.3)
P0 τ − τj (ξ)
d
τE + ξj Aj = C(ξ) ((τ − τj (ξ)) δjk ) C −1 (ξ);
j=1
i.e.,
−1
d
C(ξ) ((τ − τj (ξ)) δjk ) C −1 (ξ) τ E + ξj Aj = E.
j=1
that
1 −1
E − iC(ξ) δjk C (ξ)BN R = 0.
τ − τj (ξ)
DIFFERENTIAL EQUATIONS Vol. 42 No. 5 2006
ON THE PROPERTIES OF THE DISPERSION EQUATIONS 653
Hence we obtain
1 −1
det E − iC(ξ) δjk C (ξ)BN
τ − τj (ξ)
N
1 −1
=1−i C(ξ) δjk C (ξ)BN ej , ej + · · · ,
j=1
τ − τj (ξ)
1 −1
P1 (τ, ξ) = −P0 (τ, ξ) C(ξ) δjk C (ξ)SN ej , ej
τ − τj (ξ)
N
1 N
= −P0 (τ, ξ) (j − 1) C −1 (ξ)ej , es (ej , C(ξ)es ) .
s=1
τ − τs (ξ) j=1
Since det C > 0, it follows that the inner product (ej , C(ξ)es ) has the same sign as (C −1 (ξ)ej , es ).
2
Consequently, (C −1 (ξ)ej , es ) (ej , C(ξ)es ) has the same sign as (C −1 (ξ)ej , es ) > 0. Hence we obtain
N
μs (ξ) = (j − 1) C −1 (ξ)ej , es (ej , C(ξ)es ) > 0.
j=1
It follows from the Lagrange interpolation formula that the roots of the polynomial
N
μj
P1 = P0
j=1
τ − τj
are real, multiple roots coincide with those of the polynomial P0 (their multiplicities are less by
unity than the multiplicities of the corresponding roots of the polynomial P0 ), and the roots of P1
nonstrictly separate the roots of P0 .
2. Now we consider (for simplicity, for dimension d = 1) the extreme polynomials PN −2 and PN −1
in the pencil of the Grad moment system of the Fokker–Planck equation given by the determinant
det (τ E + ξA − iBN ) .
Here BN is the diagonal matrix with entries bjj = j − 1, j = 1, . . . , N , and a11 = 0 in the matrix A.
The lower polynomial PN −1 (τ, ξ) consists of all terms of the determinant containing (−i)N −1 . Hence
N
it follows that PN −1 (τ, ξ) = τ and γN −1 = j=2 (j−1) = (N −1)!. The polynomial PN −2 corresponds
to entries of the determinant with (−i)N −2 , i.e., is given by the sum of second-order minors
τ a1k ξ
det , k = 2, . . . , N,
ak1 ξ τ + akk ξ
N
with coefficient κk = 2=j=k (j − 1). Here ajkl are the entries of the matrix A. Therefore, we have
N
γN −2 PN −2 (τ, ξ) = κk τ 2 + akk ξτ − a1k ak1
k=2
N
1 2
= (N − 1)! τ + akk ξτ − a1k ak1 .
k=2
k−1
Hence for the Poisson brackets, we obtain
N
1 2
[γN −2 PN −2 , PN −1 ] = (N − 1)! τ + aik aki ξ 2
k=2
k−1
N
aik aki
= γN −2 τ 2 + (N − 1)! ξ 2 = γN −2 τ 2 + (N − 1)! A∗ B 2 Ae1 , e1
k=2
k−1
= γN −2 τ 2 + (N − 1)! B Ae1 , B Ae1
≥ 0 ∀(τ, ξ) ∈ R2 ,
DIFFERENTIAL EQUATIONS Vol. 42 No. 5 2006
654 ZAKHARCHENKO, RADKEVICH
N N
where γN −2 = k=2 2=j=k (j − 1), e1 = (1, 0, . . . , 0), and the entries of the diagonal N × N matrix
B have the form b11 = 0 and bjj = 1/ (j − 1), j = 2, . . . , N . Consequently, PN −2 is a strictly
hyperbolic polynomial, and its roots are strictly separated by the root τ = 0 of the polynomial PN −1 .
This simple proof gives hope to describe the structural properties of the representation matrices
of the transport operator and the impact operator in the Fokker–Planck equation, reproducing
properties 1–3 in Lemma 2.1 of moment approximations to the Fokker–Planck equation for an
arbitrary number N of equations in the Grad moment system.
Proposition 3.1. Let p(τ ) and q(τ ) be polynomials of degrees m and m − 1, respectively, in one
variable τ with real coefficients. Then the following assertions are equivalent.
(I) All roots of the complex polynomial r(τ ) := p(τ ) − iq(τ ) lie in the open half-plane Im τ > 0.
(II) The leading coefficients of the polynomials p and q have the same sign, and the zeros of p
and q are real and simple; moreover, the zeros of q strictly separate the zeros of p.
(III) All zeros of p are real and simple; if they are denoted by p1 , . . . , pm , then
q(τ ) μj
m
= ,
p(τ ) j=1
τ − pj
Following [7, 9], we say that homogeneous real polynomials P (τ, ξ) and Q(τ, ξ) of degrees m
and m − 1, respectively, form an nonstrictly hyperbolic pair if P and Q are solved in the highest
power of τ , the corresponding coefficients have the same sign, the roots pj (ξ) and qj (ξ) of these
polynomials are real, and the roots of the polynomial Q (nonstrictly) separate the roots of P .
The last assertion means that the inequalities
p1 (ξ) ≥ q1 (ξ) ≥ · · · ≥ qm−1 (ξ) ≥ pm (ξ) (3.2)
are valid for an appropriate numbering of the roots of these polynomials. A nonstrictly hyperbolic
pair P (τ, ξ) and Q(τ, ξ) is said to be strict if, for ξ = 0, all roots of P and Q are pairwise distinct
and inequalities (3.2) are strict.
For polynomials of many variables, we preserve notation (3.1) and set
[P, Q](σ, ξ) = ∂σ P (σ, ξ)Q(σ, ξ) − ∂σ Q(σ, ξ)P (σ, ξ). (3.3)
Let us return to the stability problem for pencils of the form (2.1). We represent a result in [9].
Theorem 3.1. Let the pencil (2.1) satisfy the stability condition R (τ, ξ) = 0, Im τ ≤ 0,
|τ | + |ξ| > 0. Then the following assertions are valid.
(I) The polynomial r(τ ) := R (τ, 0)/τ m−N of one variable is stable.
(II) The polynomials P0 (τ, ξ) and P1 (τ, ξ) form an nonstrictly hyperbolic pair.
(III) The polynomials PN −1 (τ, ξ) and PN (τ, ξ) form an nonstrictly hyperbolic pair.
(IV) [[P , Q ]](τ, ξ) > 0, Im τ ≤ 0, for all (σ, ξ) ∈ Rn+1 .
The study in [9] showed that necessary and sufficient stability conditions for polynomial pen-
cils (2.1) are quite complicated in the general case. The requirement of validity of the estimate in
assertion (IV) of Theorem 3.1 has a complicated form. This has forced us to return to directly com-
puted moment systems of the Fokker–Planck equation so as to single out their common properties,
which have permitted us to restrict the class of polynomial pencils considered.
with powers of τ of the same parity (even for even j and odd for odd j). Hence we obtain the
structure of these polynomials in the form
P (2k) (τ ) = gk τ 2 , Q (2k+1) (τ ) = τ fk τ 2 . (4.2)
Theorem 4.1 (Hermite’s theorem [11]). The following conditions are equivalent.
(I) The polynomial pencil P (τ ) − iQ (τ ) is stable.
m
(II) [P , Q ](τ ) = j=1 σj πj (τ )2 > 0 for all τ ∈ R.
(III) The constants σj > 0, j = 1, . . . , m, and the leading coefficients of the polynomials π2j (τ ) =
P (2j)
(τ ) and π2j+1 (τ ) = Q (2j) (τ ) of degree m − j have the same signs.
(IV) The polynomial pencils P (k) (τ ) − iQ (k+1 (τ ) and Q (k+1) (τ ) − iP (k+2) (τ ), k ≥ 0, are stable.
Example. In the two-dimensional case (d = 2), for the first six Grad moments, the approxima-
tion to the solution of the Fokker–Planck kinetic equation
1
f2 (x, t, c) = m0 (x, t)Ψ0 (c) + m10 (x, t)Ψ10 (c) + m01 (x, t)Ψ01 (c) + m20 (x, t)Ψ20 (c)
2 (5.3)
1
+ m11 (x, t)Ψ11 (c) + m02 (x, t)Ψ02 (c),
2
where Ψ(c)α are Hermite functions [1].
We obtain a system of six equations with constant coefficients,
T
(E∂t + ∂x1 A1 + ∂x2 A2 + B) M6 (x, t) = 0, M6 = (m0 , m10 , m01 , m20 , m11 , m02 ) , (5.4)
where γp,q (ξ) is a second-order homogeneous function, μp,q = p0 /q0 is the constant equal to the ratio
of the leading coefficients of the polynomials p and q, and the leading coefficient of the function
bp,q (τ, ξ) ∈ OPG2j−1 in τ is equal to unity.
A similar representation is valid in the case of an arbitrary homogeneous strictly hyperbolic
polynomial
p(τ, ξ) = f τ 2 , ξ (6.4)
of degree 2j + 1.
Lemma 6.1. An arbitrary homogeneous strictly hyperbolic polynomial (6.1) of degree m = 2j+1
[respectively, a polynomial (6.4) of degree m = 2(j +1)] and an arbitrary hyperbolic polynomial (6.2)
of degree m = 2j [respectively, a polynomial (6.5) of degree m = 2j + 1] whose roots nonstrictly
separate the roots of a polynomial p(τ, ξ) that is not a multiple of the polynomial q(τ, ξ) at any point
ξ ∈ Rd satisfy the inequality [p, q](τ, ξ) ≥ 0, (τ, ξ) = 0, and the function γp,q (ξ) in the expansion (6.3)
is a positive homogeneous function of degree 2.
p2 (ξ) = − 2
ck (ξ) , q2 (ξ) = − bk (ξ)2 ,
k=1 k=1
where ±ck , k = 1, . . . , j, and 0, ±bk , k = 1, . . . , j − 1, are the roots of the polynomials p and q,
respectively. Hence we obtain
p0
p(τ, ξ) = τ q(τ, ξ) − p(1) (τ, ξ), p(1) (τ, ξ) = (p2 (ξ) − q2 (ξ)) τ 2(j−1) + · · ·
q0
j−1
2
p2 (ξ) − q2 (ξ) = ck (ξ) − b2k (ξ) + c2j (ξ) ≥ 0,
k=1
and the corresponding strict inequality is valid if the polynomial p is not a multiple of q for any
ξ ∈ Rd . We set γp,q (ξ) = p2 (ξ) − q2 (ξ) > 0 for all ξ ∈ Rd . Then p(1) (τ, ξ) ≡ γp,q (ξ)bp,q (τ, ξ).
The second case can be considered in a similar way. Let us now justify the following assertion.
Lemma 6.2. Under the assumptions of the preceding lemma, the roots of the hyperbolic poly-
nomial bp,q (τ, ξ) ∈ OPGm−2 always nonstrictly separate the roots of the polynomial q(τ, ξ).
Consider the case in which m = 2 (m0 + 1). (The case in which m = 2m0 + 1 can be considered
in a similar way.) By the Lagrange formula,
m 0 +1
q(τ, ξ) = τ νj (ξ) τ 2 − c2k (ξ) . (6.6)
j=1 k=j
Hence we obtain
m 0 +1
νj (ξ)c2j (ξ) 2
bp,q (bs (ξ), ξ) = bs (ξ) − c2k (ξ)
j=1
γp,q (ξ) k=j
m
0 +1
b2s (ξ) − c2k (ξ) b2 (ξ)
m0 +1 m0 +1
=− μj (ξ) + s νj (ξ) b2s (ξ) − c2k (ξ)
k=1
γp,q (ξ) j=1
γp,q (ξ) j=1 k=j
m
0 +1
b2s (ξ) − c2k (ξ)
m0 +1
=− μj (ξ)
k=1
γp,q (ξ) j=1
for any root bs (ξ) = 0 of the polynomial q. Therefore, the roots of the polynomials bp,q (τ, ξ) and
q̂ = q/τ alternate, and consequently, the roots of bp,q (τ, ξ) strictly separate the roots of q.
Lemma 6.3. For any Grad pencil, there exists an analog of the Routh algorithm for constructing
a family of positive continuous functions
j (ξ), j = 1, . . . , m, and polynomials
(m−j)/2
k=0 akj (ξ)τ 2k if m − j is even
πj (τ, ξ) = (m−j−1)/2 j = 1, . . . , m,
τ k=0 akj (ξ)τ 2k if m − j is odd,
DIFFERENTIAL EQUATIONS Vol. 42 No. 5 2006
ON THE PROPERTIES OF THE DISPERSION EQUATIONS 659
We have γ2j,2j+1 (ξ) ≥ 0 for any ξ ∈ Rn . The leading coefficients of the polynomials P̂2j are
unity. Moreover, by the same lemma, we have
P2j+1 , P̂2(j+1) (τ, ξ) ≥ 0 ∀(τ, ξ) ∈ Rd+1 , j = 0, . . . , [m/2]. (6.11)
(1)
We set P2j (τ, ξ) = (γ2j − γ2j+1 γ0 /γ1 ) P2j (τ, ξ)+γ2(j−1),2j−1 (ξ)P̂2j (τ, ξ), j = 0, 1, . . . Hence it follows
that
γ τ
P (τ, ξ) = 0 Q (τ, ξ) − P (1) (τ, ξ),
γ1
γ0 2
[P , Q ](τ, ξ) = Q (τ, ξ) + Q , P (1) (τ, ξ),
γ1
P (1) (τ, ξ) = γ2(1) P2(1) (τ, ξ) − γ4(1) P4(1) (τ, ξ) + γ6(1) P6(1) (τ, ξ) + · · · ,
where
γ2j+1 γ0
(1)
γ2j (ξ) = γ2j − + γ2(j−1),2j−1 (ξ), j ≥ 1, N ≥ 2j.
γ1
Now we note that, by the definition of a Grad pencil, we have the inequality
P2j+1 , P2(j+1) (τ, ξ) ≥ 0 ∀(τ, ξ) ∈ Rd+1 , j ≥ 0.
(1)
Therefore, the polynomials P2j+1 and P2(j+1) in Q and P (1) , respectively, are related by formulas
like (6.10), which, together with Lemma 6.1, permit one to pass to the following step of the Routh
algorithm. Induction completes the proof of the lemma.
Now we note that the expansion (6.9) implies the stability of a connected pencil if the following
conditions are satisfied:
(1)
j (ξ) > 0 for all ξ ∈ Rn ,
(2) πj (τ, ξ), j = 1, . . . , m−1, are linearly independent polynomials, and πm (ξ) = 0 for all ξ ∈ Rn ,
ξ = 0.
Obviously, the second condition is satisfied for the chain of polynomials πj (τ, ξ). Their leading
coefficients, multiplying the powers τ m−j , are equal to unity.
CONCLUSION
We have singled out a class of stable polynomial pencils, referred to as Grad pencils, which,
as straightforward computations showed, are reproduced in the hierarchy of Grad moment systems
for the Boltzmann and Fokker–Planck kinetic equations.
We have shown that the second condition in the definition of a Grad pencil (the validity of
the Routh–Hurwitz conditions for the leading coefficients of the polynomials Pj in the pencil) is a
consequence of the dissipativity of the representation matrix of the impact operator in the basis of
Hermite functions.
As a second consequence of the dissipative property of the representation matrix of the impact
operator in the basis of Hermite functions, we have shown that the Poisson brackets of extreme
pairs of successive polynomials in the pencil of the dispersion equation are nonnegative.
The problem of finding out what causes the Poisson brackets of all pairs of successive polynomials
in a pencil to be nonnegative at each step of the construction of the approximation to a kinetic
equation by a Grad moment system remains open, which has been noted for all above-considered
examples.
ACKNOWLEDGMENTS
The work was financially supported by the Russian Foundation for Basic Research (project
no. 03-01-00189).
REFERENCES
1. Dreyer, W., Junk, M., and Kunik, M., Nonlinearity, 2001, vol. 14, pp. 881–906.
2. Levermore, C.D., J. Statist. Phys., 1996, vol. 83, pp. 1021–1065.
3. Chapman, S.C. and Cowling, T.C., The Mathematical Theory of Non-Uniform Gases, Cambridge:
Cambridge University, 1961.
4. Grad, H., Comm. Pure Appl. Math., 1949, vol. 2, no. 4, pp. 331–406.
5. Godunov, S.K., Elementy mekhaniki sploshnoi sredy (Elements of Continuum Mechanics), Moscow:
Nauka, 1978.
6. Müller, I. and Ruggeri, T., Extended Thermodynamics, Berlin: Springer, 1993.
7. Volevich, L.R. and Radkevich, E.V., Differ. Uravn., 2003, vol. 39, no. 4, pp. 485–499.
8. Radkevich, E.V., Contemp. Math., 2003, vol. 3, no. 3, pp. 5–32.
9. Volevich, L.R. and Radkevich, E.V., Tr. Mosk. Mat. Obs., 2004, vol. 65, pp. 69–113.
10. Hermite, Ch., Œuvres, Tome I , Paris: Gauthier-Villars, 1905.
11. Godunov, S.K., Obyknovennye differentsial’nye uravneniya s postoyannymi koeffitsientami (Ordinary
Differential Equations with Constant Coefficients), Novosibirsk: Izd. Novosibirsk. Univ., 1994.
12. Petrovskii, I.G., Izbr. tr. Sistemy uravnenii s chastnymi proizvodnymi. Algebraicheskaya geometriya
(Selected Works. Systems of Partial Differential Equations. Algebraic Geometry), Moscow: Nauka,
1986.
13. Chen, G.Q., Comm. Pure Appl. Math., 1993, vol. 46, pp. 755–781.
14. Whitham, G., Linear and Nonlinear Waves, New York: Wiley, 1974. Translated under the title Lineinye
i nelineinye volny, Moscow: Mir, 1977.
15. Volevich, L.R. and Gindikin, S.G., Smeshannaya zadacha dlya differentsial’nykh uravnenii v chastnykh
proizvodnykh s kvaziodnorodnoi starshei chast’yu (Mixed Problem for Partial Differential Equations with
Quasihomogeneous Leading Part), Moscow, 1999.
16. Zakharchenko, P.A. and Radkevich, E.V., Dokl. Akad. Nauk , 2004, vol. 395, no. 1, pp. 36–39.
DOI: 10.1134/S0012266106050065
∂U ∂I ∂I ∂U
= rI + L , = gU + C , (1)
∂x ∂t ∂x ∂t
where U (t, x) is the voltage between the wires, I(t, x) is the current in the line, and r, L, g, and
C are primary parameters of the line (r is the active resistance, g is the conductivity, L is the
inductance, and C is the capacitance), play an important role in the description of the dynamics
of a homogeneous line.
In√system (1), we
√ pass to dimensionless
√ variables by setting x → lx, where l is the length of line,
t → l LC t, U = L u1 , and I = C u2 , thus obtaining the system
∂u1 ∂u2 ∂u2 ∂u1
= − (a + μ)u1 , = − (a − μ)u2 . (2)
∂t ∂x ∂t ∂x
Here 0 ≤ x ≤ 1, a > 0, and |μ| < a, where
a = (l/2) r C/L + g L/C , μ = (l/2) g C/L − r L/C .
suggested in [2], where b < 1 is a positive parameter related to the passive resistance and k > 0 is
a parameter corresponding to energy amplification.
Our main goal is to find an algorithm that constructs solutions of the boundary value prob-
lem (2), (3) and takes into account its specific properties to the maximum possible extent.
2. AUXILIARY CONSTRUCTIONS
We need formulas for the solution of the Cauchy problem for system (2). We obtain them in a
form convenient for subsequent use.
By successively performing the changes of variables
u1 → u1 exp(−at), u2 → u2 exp(−at), u = u1 + u2 , v = u1 − u2 (4)
661
662 KOLESOV, KIRILLOV
Fig. 1.
∂u ∂u ∂v ∂v
= − μv, =− − μu, (5)
∂t ∂x ∂t ∂x
which we supplement with the initial conditions
where f and g are continuously differentiable functions defined on the entire real line.
By passing in system (5) to the independent variables ξ = μ(t + x) and η = μ(t − x), we succes-
sively obtain the equations
∂u 1 ∂v 1 ∂2u 1 ∂2v 1
= − v, = − u, − u = 0, − v = 0. (7)
∂η 2 ∂ξ 2 ∂ξ ∂η 4 ∂ξ ∂η 4
Therefore, the Riemann method [3, pp. 70–77] permits one to write out the solutions of problem (5),
(6) in closed form.
Let S be the boundary of the triangular domain QM P shown in Fig. 1. (Arrows show the sense
of the boundary.) It follows from Green’s formula that
∂R ∂u ∂R ∂u
u −R dξ − u −R dη = 0, (8)
∂ξ ∂ξ ∂η ∂η
S
where R = R(λ), λ = (ξ0 − ξ) (η0 − η), is the Riemann function, which in our case can be
represented [4, pp. 45–48] by the series
λ2 λ4 λ6
R=1+ + + + ··· (9)
22 22 × 42 22 × 42 × 62
Note that, by (9), R|ξ=ξ0 = R|η=η0 = 1 and ∂R/∂η|ξ=ξ0 = ∂R/∂ξ|η=η0 = 0. Therefore, it follows
from (8) that
Q
1 ξ0 η0 1 ∂u ∂R ∂u ∂R
u (ξ0 , η0 ) = f +f − + R −u dξ − R −u dη. (10)
2 μ μ 2 ∂ξ ∂ξ ∂η ∂η
P
ξ0
∂u du ∂R ∂R
2R +R −u + dσ,
∂η dσ ∂ξ ∂η
−η0
where dσ is the length differential on the segment over which the integration is performed. By using
the first relation in (7) and by transforming the middle term by integration by parts, we find that the
integral in question is equal to
μ0
∂R ξ0 η0
−vR − 2u dσ + f −f − . (11)
∂ξ μ μ
−η0
Passing to the variables t0 = (ξ0 + η0 )/(2μ), x0 = (ξ0 − η0 )/(2μ) and omitting the zero subscript,
after simple transformations from (10) and (11), we obtain the formula
t
μ
u(t, x) = f (x + t) − R(λ)[g(x + s) + g(x − s)]ds
2
0
(12)
2
t
μ R (λ)
+ [(t + s)f (x + s) + (t − s)f (x − s)]ds,
2 λ
0
√
where λ = μ t2 − s2 . By using the second relation in (7), we obtain
t
μ
v(t, x) = g(x − t) − R(λ)[f (x + s) + f (x − s)]ds
2
0
(13)
t
μ2 R (λ)
+ [(t − s)g(x + s) + (t + s)g(x − s)]ds.
2 λ
0
Lemma 1. Formulas (12) and (13) give the solution of problem (5), (6) for t > 0.
3. MAIN THEOREM
We supplement problem (2), (3) with the initial conditions
u1 |t=0 = f1 (x), u2 |t=0 = f2 (x), (16)
where f1 and f2 are smooth functions defined only on the interval 0 ≤ x ≤ 1 and related to the
initial conditions (6) on that interval via the second set of equations in (4) :
1 1
f1 (x) = [f (x) + g(x)], f2 (x) = [f (x) − g(x)]. (17)
2 2
It follows from (15) and (17) that, for −1 ≤ x ≤ 0, the function f1 should be defined as an even
function and f2 should be defined as an odd function. To preserve the smooth dependence on the
initial functions (16), one must additionally require that
f1 (0) = 0, f2 (0) = 0. (18)
Hence it suffices to define a single initial function f (x) on the interval −1 ≤ x ≤ 1; then, by (15)
and (17),
1 1
f1 (x) = [f (x) + f (−x)], f2 (x) = [f (x) − f (−x)].
2 2
This, together with (12) and (13), implies that the construction of solutions of problem (2), (3) for
0 < t ≤ 1 requires continuing the function f (x) to the intervals [−2, −1] and [1, 2].
By using identity (15), one can rewrite formulas (12) and (13) as
t
μ
u(t, x) = f (x + t) − R(λ)[f (−x − s) + f (−x + s)]ds
2
0
(19)
2
t
μ R (λ)
+ [(t + s)f (x + s) + (t − s)f (x − s)]ds,
2 λ
0
t
μ
v(t, x) = f (−x + t) − R(λ)[f (x + s) + f (x − s)]ds
2
0
(20)
2 t
μ R (λ)
+ [(t − s)f (−x − s) + (t + s)f (−x + s)]ds.
2 λ
0
Here the variable t on the right-hand side is positive by virtue of the above-performed normalization,
and the notation u− and v − indicates that we construct solutions u1 = u− −
1 and u2 = u2 related to
negative time.
By construction, u− − − − − −
1 = (v − u )/2 and v1 = (v + u )/2; therefore, from (4) and the second
boundary condition in (3), we obtain
(1 + b)v − (t, 1) + (1 − b)u− (t, 1) + k v − (t, 0) − u− (t, 0)
1 − 3 (26)
= v (t, 0) − u− (t, 0) exp(2at).
4
The general character of formula (26) is identic to that of (21). Their combination is a lin-
ear inhomogeneous Volterra system of integral equations for the unknown functions f (1 + t) and
f (−1 − t), which guarantees their existence and uniqueness on the interval 0 ≤ t ≤ 1.
Thus, by using (4) and relations (12) and (13) with 0 ≤ t ≤ 1, one can construct solutions
u1 (t, x) and u2 (t, x) of the nonlinear boundary value problem (2), (3). By setting f1 (x) = u1 (1, x)
and f2 (x) = u2 (1, x), we arrive at a situation similar to the original one; hence we have found an
iterative process that permits one to construct solutions on an arbitrary time interval.
Theorem. For smooth initial conditions (16) satisfying (18), the nonlinear boundary value
problem (2), (3) is nonlocally solvable in the classical sense.
Obviously, this assertion simultaneously contains a numerical integration algorithm. In what
follows, we consider details of its implementation.
The following assertion clarifies the possible dynamics of solutions of the boundary value prob-
lem (2), (3).
Lemma 3. If
k > k0 = cosh a + b sinh a, (27)
then the zero equilibrium of the linear approximation to the boundary value problem (2), (3) is
oscillatory unstable.
Proof. Let us make the first two changes of variables (4) and construct solutions of the linear
approximation in the Euler exponential form with exponent λ. The characteristic equation has the
form
μ μ2
cosh p + b sinh p − sinh p + k = 0, p = λ 1 + 2. (28)
p λ
It follows that
k = k0 , λs = a + i(2s − 1)π, s = 1, 2, . . . ,
for μ = 0. Hence for μ = 0, condition (27) necessarily provides exponential instability on higher
modes, i.e., eigenfunctions of the corresponding differential operator with large numbers.
For k = k0 and small |μ|, the coefficient γs of exponential growth (for μ > 0) or exponential
decay (for μ < 0) on the sth mode satisfies
a sinh a
γs = μ + O μ2 , s = 1, 2, . . .
(a2 + (2s − 1)2 π 2 ) (sinh a + b cosh a)
u1 = f (t + x) exp(ax) + f (t − x) exp(−ax),
(29)
u2 = f (t + x) exp(ax) − f (t − x) exp(−ax),
DIFFERENTIAL EQUATIONS Vol. 42 No. 5 2006
666 KOLESOV, KIRILLOV
where f is a smooth function defined on the interval [−1, 1]. By virtue of the second boundary
condition in (3) and formula (29), its values on the interval [1, 2] are determined by the relation
f (t + 1) = −p1 f (t) − p2 f (t − 1) + p3 f 3 (t), (30)
where
0 ≤ t ≤ 1, p1 = 2k(1 + b)−1 exp(−a),
p2 = (1 − b)(1 + b)−1 exp(−2a), p3 = 8(1 + b)−1 exp(−a).
We set z1 (t) = f (t) and z2 = f (t − 1), which permits one to interpret relation (30) as the
mapping
z1 → −p1 z1 − p2 z2 + p3 z13 , z2 → z1 (31)
of R2 defining the values of variable at time t + 1 on the basis of their values at time t.
In vector notation, the mapping (31) can be written concisely as
T
z → Π(z), z = (z1 , z2 ) . (32)
It is almost obvious that if inequality (27) is valid, √ then the mapping
√ (32) has the cycle z 0 , −z 0 of
period 2, where the vector z has the components Δ/2 and − Δ/2, Δ = k − k0 .
0
1 1
Lemma 4. Each of the nonsymmetric cycles z± and z∓ is exponentially stable if Q(−1) > 0
and dichotomous if Q(−1) < 0.
This lemma just summarizes results; all of its assertions are obvious. However, they show a way
for solving the problem.
It follows from (35) that the condition Q(−1) < 0 is equivalent to the inequality
2
2 p2 + p21 > 1 + p22 + 9 (1 + p2 ) ,
1
k2 > 5k02 − 1 − b2 . (39)
2
1
f (t + 1) = −f (t), f 2 (t) = Δ. (40)
4
For example, this set contains the 2-periodic function
√
f (t) = √Δ/2 if 0 < t < 1 (41)
− Δ/2 if −1 < t < 0.
In particular, it follows from (29) and (41) that the normalized current u2 (t, x) satisfies the
formula
u2 (0, x) = √
0 if x = 0
Δ cosh ax if 0 < x ≤ 1,
√ (42)
u2 (t, x) = √Δ sinh ax if 0 ≤ x < t
Δ cosh ax if t < x ≤ 1,
DIFFERENTIAL EQUATIONS Vol. 42 No. 5 2006
668 KOLESOV, KIRILLOV
t t
μ2 R (λ) μ
(1 + b)z1 + (1 + b) (t + s)z1 ds − (1 + b) R(λ)z2 ds
2 λ 2
0 0
(47)
t 2 t
μ μ R (λ)
− (1 − b) R(λ)z1 ds + (1 − b) (t − s)z2 ds = (1 + b)φ1 (t),
2 2 λ
0 0
t t
μ2 R (λ) μ
−(1 − b)z2 + (1 + b) (t − s)z1 ds + (1 + b) R(λ)z2 ds
2 λ 2
0 0
(48)
t 2
t
μ μ R (λ)
+ (1 − b) R(λ)z1 ds − (1 − b) (t + s)z2 ds = −(1 − b)φ2 (t).
2 2 λ
0 0
Recall that φ1 (t) and φ2 (t) are determined by the initial conditions.
We introduce the matrix
⎛ ⎞
1−bμ μ2 R (λ) μ 1 − b μ2 R (λ)
− R(λ) + (t + s) − R(λ) + (t − s)
⎜ 1+b2 2 λ 2 1+b 2 λ ⎟
K(t, s) = ⎜
⎝
⎟.
⎠
1 + b μ2 R (λ) μ 1+bμ μ2 R (λ)
− (t − s) − R(λ) − R(λ) + (t + s)
1−b 2 λ 2 1−b2 2 λ
By virtue of the relations R(0) = 1 and R (λ)/λ|λ=0 = 1/2 and the notation
1−b 1+b
α= − μt, β= − μt,
1+b 1−b
we have
μ −α 1
K(t, t) = − .
2 1 β
T T
If we now set w(t) = (z1 (t), z2 (t)) and Φ(t) = (φ1 (t), φ2 (t)) , then, obviously, relations (47)
and (48) can be reduced to the form
t
w + Aw = Φ(t), Aw = K(t, s)w(s)ds. (49)
0
Formulas (49) describe the essence of the main theorem. They are also used for numerical experi-
ments.
We divide the interval 0 ≤ t ≤ 1 into n equal parts by setting n = 50 and introduce the notation
wp = w(ph) and Φp = Φ(ph), p = 1, . . . , n, where h = 1/n.
Fig. 2. Fig. 3.
We approximate the integral operator given by the second relation in (49) by a finite-dimensional
operator by introducing the matrix
⎛ ⎞
0 0 0 ... ... 0
⎜ ⎟
⎜ K(h, 0)/2 K(h, h)/2 0 ... ... 0 ⎟
⎜ ⎟
Ah = h ⎜
⎜ ... ... ... ... ... ...⎟⎟; (50)
⎜ ⎟
⎝ K(ph, 0)/2 K(ph, h) . . . K(ph, ph)/2 . . . 0 ⎠
... ... ... ... ... ...
where Fp = Φp − (h/2)K(ph, 0)Φ0 . In (50) and (51), we have used the relation w0 = Φ0 .
In the construction of the explicit finite-difference scheme (51), the formulas
−1
h μh μ2 h2 2 3 α 1
I + K(ph, ph) =I+ B(μt) + B (μt) + O h , B=
2 4 16 1 β
To perform numerical experiments for small μ, for the initial conditions u1 (0, x) and u2 (0, x),
we take smoothed initial conditions of various 2-periodic solutions for a distortion-free line.
One of the main conclusions is the following: if μ = 0, then self-induced oscillations are most
likely to be smooth with respect to t and x within the accuracy of computations.
Other important specific properties of dynamics related to changes of Δ and μ are shown in
Fig. 2, where we have used the following notation: A is the domain of parameter values for which
there is a unique attractor given by a cycle (its structure is to be specified below), B is the zone
of chaos, where waves are relatively weakly deformed in a random way with the preservation of
the above-mentioned physical meaning of specific properties of wave motion in the periodic case,
C is the zone in which oscillations grow infinitely, which, from the physical viewpoint, implies the
overloading of the self-excited oscillator.
Note that, for an appropriate change of Δ and μ, we have a rapid passage from periodic oscil-
lations to chaotic ones. Therefore, in Fig. 2 the domain related to this passage is included in B.
We point out that self-oscillations have a complicated structure: Fig. 3 represents two specific
examples related to periodic [Δ = 1, μ = 0.5 (a)] and chaotic [Δ = 4, μ = 0.5 (b)] changes of the
normalized current. The time moments at which chaotic oscillations are also structurally similar
to periodic ones are highlighted in each of these figures. In both cases, in the course of time, the
peak of oscillations diminishes; in addition, the wave does not move; then to the right from its peak,
there appears a domain of negative values of the current, then it fully becomes negative, and finally,
at the right-end point, there appears a domain of positive values, after which the picture coincides
with the original one neglecting the sign. In particular, the wave half-period is equal to unit with
high accuracy.
To complete the exposition, we return to the possibility of the appearance of the buffer property
as a general phenomenon. Our study showed that for small |μ| and for initial conditions adjusted,
for example, to one and a half waves fitting in the interval 0 ≤ x ≤ 1, in a relatively long time,
we return to the basic case of a half-wave. (For example, if μ = ±0.1, then the passage takes
approximately 102 units of normalized time.)
ACKNOWLEDGMENTS
The work was financially supported by the Russian Foundation for Basic Research (project
no. 03-01-00456).
REFERENCES
1. Atabekov, G.I., Teoreticheskie osnovy elektrotekhniki (Foundations of Electrical Engineering), Moscow,
1996.
2. Kolesov, Yu.S., Differentsial’nye uravneniya i ikh primenenie (Differential Equations and Their Appli-
cation), Vilnyus, 1971.
3. Sobolev, S.L., Uravneniya matematicheskoi fiziki (Equations of Mathematical Physics), Moscow, 1966.
4. Mikhlig, S.G. et al., Lineinye uravneniya matematicheskoi fiziki (Linear Equations of Mathematical
Physics), Moscow, 1964.
5. Kolesov, Yu.S., Mat. Zametki, 1994, vol. 56, no. 1, pp. 41–49.
6. Kolesov, Yu.S., Mat. Zametki, 1997, vol. 62, no. 5, pp. 744–750.
7. Kolesov, Yu.S. and Kulikov, A.N., Mat. Zametki, 1999, vol. 64, no. 6, pp. 948–951.
DOI: 10.1134/S0012266106050090
When studying numerous real phenomena involving nonuniform passages between different phys-
ical characteristics, one has to analyze singularly perturbed problems. Such problems attract the
attention of numerous mathematicians (e.g., see [1–16]). The studies in the field were intensified
after the papers [6, 7], which are a substantial contribution to the theory of singularly perturbed
problems. However, most of the problems studied were related to classical equations in a do-
main with a single viscous boundary or two disjoint boundaries. Note that a viscous boundary is
understood as a boundary in a neighborhood of which there is a boundary layer.
If two or more viscous boundaries meet, then the boundary-layer functions overlap and, as
a result, the boundary conditions of the original problem are not satisfied at the corners. Such
problems for the classical equations were considered in the paper [11–13]. There are only few papers
dealing with the asymptotics of solutions of boundary value problems for nonclassical equations for
the case in which two or more viscous boundaries meet. Here we note the papers [14–16].
In the present paper, we consider the equation
∂ 2m+1 U 4
2∂ U ∂U ∂2U
Lε U ≡ (−1)m ε2m + ε + − + aU = f (t, x), (1)
∂t2m+1 ∂x4 ∂t ∂x2
where m is an arbitrary positive integer, ε > 0 is a small parameter, a > 0 is a constant, and f (t, x)
is a given sufficiently smooth function. We divide the paper into two parts, where two different
boundary value problems are considered for Eq. (1) in the rectangle
D = {(t, x) | 0 < t < T, 0 < x < 1}.
I. In the first part, the boundary conditions for Eq. (1) are posed as follows:
∂U ∂ m U
U |t=0 = = ··· = = 0,
∂t t=0 ∂tm t=0
(2)
∂ m+1 U ∂ m+2 U ∂ 2m U
= = ··· = = 0,
∂tm+1 t=T ∂tm+2 t=T ∂t2m t=T
In this part, we construct the complete asymptotics of the solution of problem (1)–(4) with
respect to the small parameter. To construct the asymptotics, we use iterative processes.
In the first iterative process, we construct an approximate solution of Eq. (1) in the form
W = W0 + ε2 W1 + · · · + ε2n Wn , (5)
698
ON BOUNDARY VALUE PROBLEMS FOR A CLASS OF SINGULARLY PERTURBED . . . 699
where the functions Wi are found from the following recursive chain of problems:
∂W0 ∂ 2 W0
− + aW0 = f0 (t, x); W0 |t=0 = 0, W0 |x=0 = W0 |x=1 = 0,
∂t ∂x2
∂Wi ∂ 2 Wi
− + aWi = fi ; Wi |t=0 = 0, Wi |x=0 = Wi |x=1 = 0,
∂t ∂x2
where fi (W0 , W1 , . . . , Wn−1 ) are known functions.
If f (t, x) is a sufficiently smooth function all of whose even x-derivatives vanish for x = 0 and
x = 1, then all functions Wi occurring in the expansion (5) are uniquely determined and smooth
and satisfy the conditions
∂ 2 Wi ∂ 2 Wi
= = 0, i = 0, 1, . . . , n.
∂x2 x=0 ∂x2 x=1
Therefore, in this part of the paper, we need not construct boundary-layer functions near the
boundaries x = 0 and x = 1.
The function W satisfies only the first condition in (2) and does not necessarily satisfy the re-
maining boundary conditions for t = 0. Therefore, we use another iterative process and supplement
the function W with the boundary-layer function
V = ε V0 + εV1 + · · · + ε2n+m−1 V2n+m−1 (6)
near the boundary t = 0 so as to ensure that the sum W + V satisfies the remaining m conditions
for t = 0 :
∂ ∂2 ∂m
(W + V ) = 2 (W + V ) = · · · = m (W + V ) = 0. (7)
∂t t=0 ∂t t=0 ∂t t=0
(i) (i)
Vs = as0 (x) + as1 (x)τ + · · · + a(i)
ss (x)τ s
eλi τ ,
i=1
Vj |t=0 = 0, j = 0, 1, . . . , 2n + m − 1,
we adjust them as follows:
m
∂W0 (0, x) λi τ
V0 = − a0i e − 1 ,
∂t i=1
s
m
m
Vs = a(i)
s0 (x) e
λi τ
−1 + a(i)
sj (x)τ
j
eλi τ , s = 1, 2, . . . , 2n + m − 1.
i=1 i=1 j=1
Let us multiply all functions Vj by a smoothing function and denote the resulting functions
again by Vj , j = 0, 1, . . . , 2n + m − 1, again. The sum W + V does not necessarily satisfy
the boundary condition (2) for t = T . Therefore, to the sum W + V , we add the function
η = εm+1 (η0 + εη1 + · · · + ε2n+m−1 η2n+m−1 ) of boundary-layer type near the boundary t = T so as
to ensure that the sum W + V + η satisfies the boundary conditions
∂ m+1 ∂ m+2 ∂ 2m
(W + V + η) = m+2 (W + V + η) = · · · = 2m (W + V + η) = 0.
∂t m+1 ∂t ∂t
t=T t=T t=T
W0 (T, x)
m+1 m
m∂
η0 = (−1) b0i eλi y ,
∂tm+1 i=1
s
m
ηs = b(i)
sj (x)y
j
eλi y , s = 1, 2, . . . , 2n + m − 1,
i=1 i=1
where y = (T − t)/ε, boi are known numbers, and b(i)sj (x) are known functions; moreover,
d2k (i) d2k (i)
b (x) = b (x) = 0.
dx2k sj x=0 dx2k sj x=1
We multiply the functions ηj by smoothing functions and retain the notation ηj for the resulting
functions.
Therefore, for the solution of problem (1)–(4), we obtain the asymptotic representation
n
2n+m−1
2n+m−1
2i 1+j
U= ε Wi + ε Vj + ε1+m+j ηj + ε2n+2 z, (11)
i=0 j=0 j=0
where ε2n+2 z is the remainder; here z satisfies the homogeneous boundary conditions corresponding
to (2)–(4).
By taking into account the equations obtained from the iterative processes and by performing
some transformations, we obtain
Lε z = F, (12)
where F (t, x, ε) is a bounded function in D for any ε ∈ [0, ε0 ].
Theorem 1. Let f (t, x) and all of its even x-derivatives vanish for x = 0 and x = 1. Then
the solution of problem (1)–(4) admits the asymptotic representation (11), where Wi are functions
found by the first iterative process, Vj and ηj are boundary-layer functions near the boundary t = 0
and t = T, respectively, which are found from the corresponding iterative processes, and ε2n+2 z is
the remainder in which the function z admits the estimate (13).
II. In the second part of the paper, we preserve the boundary conditions (2) and (3) and
replace (4) by the condition
∂U ∂U
= = 0; (14)
∂x x=0 ∂x x=1
i.e., we consider the boundary value problem (1)–(3), (14).
Suppose that the function W0 has been constructed just as in the first part of the present paper.
Since the first derivative of W0 with respect to x does not necessarily satisfy the homogeneous
boundary conditions corresponding to (14) for x = 0 and x = 1, we see that one should construct
boundary-layer functions near the boundaries x = 0 and x = 1 to compensate for the lost boundary
conditions on these boundaries. It suffices to restrict considerations to the construction of the first
terms in the asymptotics of the solution of problem (1)–(3), (14).
First, let us construct boundary-layer functions near the boundaries t = 0 and t = T . The func-
tion V0 (of the boundary-layer type near the boundary t = 0) is sought in the form V = εV0 ,
where V0 is some solution of Eq. (8). Here the boundary conditions for Eq. (8) are found from the
condition
∂k
(W + V ) = 0, k = 1, 2, . . . , m.
∂tk
0
t=0
Hence we find that Eq. (8) should be solved with the boundary conditions
∂V0 ∂W0 ∂ 2 V0 ∂ 2 W0
=− , = −ε , ...,
∂τ τ =0 ∂t t=0 ∂τ 2 τ =0 ∂t2 t=0
(15)
∂ m V0 m
m−1 ∂ V0
= −ε .
m
∂τ τ =0 ∂t t=0
m
where cij are known numbers. To satisfy the condition V |t=0 = 0, we define the function V by the
formula
m j λi τ
j ∂ W0
V = ε cij e − 1 . (16)
i,j=1
∂tj t=0
We multiply the function V by a smoothing function and denote the resulting function by the
same symbol V .
The function η (which is a boundary-layer function near the boundary t = T ) will be sought in
the form η = εm+1 η0 ; moreover, η0 is a boundary-layer solution of the problem
∂ 2m+1 η0 ∂η0
(−1)m 2m+1 + = 0, (19)
∂y ∂y
∂ m+1 η0 m ∂
m+1
W0 ∂ m+2 η0 m+1 ∂
m+2
W0
= (−1) , = (−1) ε , ...,
∂y m+1 y=0 ∂t2m+1 t=T ∂y m+2 y=0 ∂tm+2 t=T
∂ 2m η0 2m−1 m−1 ∂
2m
W0
= (−1) ε , (20)
∂y 2m
y=0 ∂t2m t=T
where y = (T − t)/ε. Note that the boundary conditions (20) follow from the condition
∂ m+k
m+k
(W0 + V + η) = 0, k = 1, 2, . . . , m. (21)
∂t t=T
By writing out the boundary-layer solution of problem (19), (20), we find that η is given by the
formula
m
∂ m+j W0
η= εm+j m+j
dij eλi y , (22)
i,j=1
∂t t=T
The first derivative of the sum W0 + V + η with respect to x does not necessarily satisfy the
homogeneous condition (14) for x = 0 and x = 1. Therefore, we construct boundary-layer functions
near the boundaries x = 0 and x = 1.
Let us write out a new decomposition of the operator Lε near the boundary x = 0; to this end,
we make the change of variables t = t, x = εξ. In the variables (t, ξ), the operator Lε has the form
4
−2 ∂ ∂2 ∂ m 2m+2 ∂
2m+1
Lε,3 ≡ ε − +ε 2
+ a + (−1) ε .
∂ξ 4 ∂ξ 2 ∂t ∂t2m+1
The function ϕ (of boundary-layer type near the boundary x = 0) will be sought as an approx-
imate solution of the equation Lε,3 ϕ = 0 in the form
The characteristic equation corresponding to the ordinary differential equations (26) has the
unique negative root λ = −1. Consequently, problem (1)–(3), (14) is regularly degenerate for x = 0.
The boundary conditions for Eqs. (26) can be found from the condition
∂
(W + V + η + ϕ) = 0. (27)
∂x x=0
The boundary-layer solutions of Eq. (26) with the boundary conditions (28) and (29) have the
form
∂W0 ∂
ϕ0 =
−ξ
e , ϕ1 = (V0 + ε η0 )
m
e−ξ .
∂x x=0 ∂x x=0
For the function ϕ given by (25) to vanish for x = 0, we modify ϕ0 and ϕ1 as follows:
∂W0 −ξ
ϕ0 = e −1 , (30)
∂x x=0
∂ −ξ
ϕ1 = (V0 + ε η0 )
m
e −1 . (31)
∂x x=0
Obviously, the functions ϕ0 and ϕ1 given by (30) and (31) are also solutions of Eq. (26) and
satisfy the corresponding boundary conditions (28) and (29). We multiply the functions ϕ0 and ϕ1
by smoothing functions.
It follows from (25), (30), and (31) that ϕ is given by the formula
∂ −ξ
ϕ=ε (W0 + V + η) e −1 . (32)
∂x x=0
Following (21), (23), (24), and (32), we find that the sum W0 + V + η + ϕ satisfies (27) as well
as the conditions
∂
(W0 + V + η + ϕ)|t=0 = (W0 + V + η + ϕ) = ···
∂t
t=0
∂m
= m (W0 + V + η + ϕ) = 0,
∂t
t=0
∂ m+1 ∂ m+2
(W0 + V + η + ϕ) = m+2 (W0 + V + η + ϕ) = ···
∂t m+1 ∂t
t=T
t=T
∂ 2m
= 2m (W0 + V + η + ϕ) = 0,
∂t t=T
(W0 + V + η + ϕ)|x=0 = (W0 + V + η + ϕ)|x=1 = 0.
One can readily show that the function ψ (which is a boundary-layer function near the bound-
ary x = 1), providing the validity of the condition
∂
ψ|x=1 = (W + V + η + ψ) =0
∂x x=1
where εz is a remainder.
The results obtained in the second part of the present paper can be stated as the following
assertion.
Theorem 2. Let the function f (t, x) have continuous derivatives of order ≤ 2m with respect to
t and of order ≤ 4 with respect to x. Let the function f (t, x) and its second and fourth derivatives
with respect to x vanish for x = 0 and x = 1. Then the solution of problem (1)–(3), (14) admits the
asymptotic expansion (33), where W0 is a solution of the degenerate problem, εV0 is a boundary-
layer function near t = 0, εm+1 η0 is a boundary-layer function near t = T, ε (ϕ0 + εϕ1 ) is a
boundary-layer function near x = 0, ε (ψ0 + εψ1 ) is a boundary-layer function near x = 1, and εz
is a remainder in which the function z satisfies the estimate (13).
ACKNOWLEDGMENTS
The authors are grateful to V.A. Il’in for attention to the research.
REFERENCES
1. Tikhonov, A.N., Mat. Sb., 1948, vol. 22, no. 2, pp. 193–204.
2. Vasil’eva, A.B., Mat. Sb., 1952, vol. 31, no. 3, pp. 587–644.
3. Mishchenko, E.F. and Pontryagin, L.S., Izv. Akad. Nauk. Ser. Mat., 1959, vol. 23, no. 5, pp. 643–660.
4. Oleinik, O.A., Mat. Sb., 1952, vol. 31, no. 1, pp. 104–117.
5. Ladyzhenskaya, O.A., Vestnik Leningrad. Univ., 1957, vol. 2, no. 7, pp. 104–120.
6. Vishik, M.I. and Lyusternik, L.A., Uspekhi Mat. Nauk , 1957, vol. 12, no. 5, pp. 3–122.
7. Vishik, M.I. and Lyusternik, L.A., Uspekhi Mat. Nauk , 1960, vol. 15, no. 5, pp. 3–80.
8. Trenogin, V.A., Uspekhi Mat. Nauk , 1960, vol. 16, no. 1, pp. 163–169.
9. Dzhavadov, M.G., Dokl. Akad. Nauk , 1963, vol. 152, no. 4, pp. 790–793.
10. Dzhavadov, M.G., Izv. Akad. Nauk Azerb. Fiz.-Mat. Tekhn. Nauk , 1963, no. 6, pp. 3–10.
11. Butuzov, V.F., Mat. Sb., 1977, vol. 104, no. 3, pp. 460–485.
12. Butuzov, V.F., Vestnik Moskov. Univ. Ser. XV Vychisl. Mat. Kibernet., 1978, no. 2, pp. 49–56.
13. Butuzov, V.F. and Udodev, Yu.P., Zh. Vychil. Mat. Mat. Fiz., 1981, vol. 21, no. 3, pp. 665–677.
14. Dzhavadov, M.G. and Sabzaliev, M.M., Dokl. Akad. Nauk , 1979, vol. 247, no. 5, pp. 1041–1046.
15. Sabzaliev, M.M., Uspekhi Mat. Nauk , 1979, vol. 34, no. 4, p. 172.
16. Salimov, Ya.Sh. and Sabzalieva, I.M., Izv. Akad. Nauk Azerb. Fiz.-Mat. Tekhn. Nauk , 1998, no. 5,
pp. 11–14.
DOI: 10.1134/S0012266106050107
with rapidly varying kernels one of which has an unstable spectral value (μ1 (t) ≡ 0 for all
t ∈ S ⊂ [0, T ]). A similar problem with a single spectral value μ0 (t) was considered in [1], where
it was assumed that μ0 (t) = 0 for all t ∈ [0, T ] and instability is caused by the vanishing of an
eigenvalue a(t) of the differential operator of problem (1). Even in this case, the algorithm of the
Lomov regularization method needs essential modifications, since the regularization with respect
to the spectrum [2] does not take into account the singularities induced in the solutions of an
integro-differential equation by the instability of the eigenvalue a(t). Therefore, the regulariza-
tion in [1] was performed with the use of normal forms; this technique was developed in detail
for differential systems in a number of papers (e.g., see [3, 4]). In the case of integro-differential
equations, this method needs essential modifications due to a nonstandard regularization of the
integral operator, since if an eigenvalue is unstable, then such a regularization cannot be performed
with the use of classical integration by parts alone. In this connection, it was suggested in [1] to
use a procedure that permits one to derive expansions of integrals in asymptotic series (in powers
of ε) only after their preliminary “normalization” with the use of the corresponding equation of a
normal form. Owing to such a “normalization,” one can use the well-known operation of integration
by parts for integral operators and perform their complete regularization. However, for integral
operators with an unstable spectral value μ1 (t) of the kernel such an operation does not provide
the
t desired result, since when using the normal form, one cannot eliminate an integral of the form
0
Eε (t, s; μ1 ) y0 (s)ds, to which integration by parts cannot be applied by virtue of the identity
μ1 (t) ≡ 0 for all t ∈ S. In the general case, this problem remains open; therefore, in the present
paper, we consider its special version, namely, the case of a degenerate kernel with an unstable
spectral value μ1 (t).
1. REDUCTION OF EQ. (1) TO AN EQUIVALENT
INTEGRO-DIFFERENTIAL SYSTEM
We introduce the function
t
z= Eε (t, s; μ1 ) q(s)y(s, ε)ds.
0
705
706 BOBODZHANOV, SAFONOV
By differentiating it with respect to t, we obtain the equation εż = μ1 (t)z + εq(t)y and reduce the
original problem (1) to the equivalent system
t
dy
ε = a(t)y + p(t)z + Eε (t, s, μ0 ) K(t, s)y(s, ε)ds + h(t), y(0, ε) = y 0 ,
dt
0 (2)
dz
ε = μ1 (t)z + εq(t)y, z(0, ε) = 0.
dt
t
dw
ε = A0 (t)w + εA1 (t)w + Eε (t, s; μ0 ) G(t, s)w(s, ε)ds + H(t),
dt (3)
0
w(0, ε) = w0 ,
where w = {y, z} and the 2 × 2 matrix A0 (t) has a simple spectrum σ (A0 ) = {a(t), μ1 (t)}. (We as-
sume that, in the original problem (1), the functions a(t), μ0 (t), and μ1 (t) do not coincide for any
t ∈ [0, T ].) In system (2), the matrices A0 (t), A1 (t), and G(t, s) and the vector function H(t) have
the form
a(t) p(t) 0 0
A0 (t) = , A1 (t) = ,
0 μ1 (t) q(t) 0 (4)
G(t, s) = diag(K(t, s), 0), H(t) = (h(t), 0)T .
du
l+1
ε = Λ(t)u + εr gr (t)e2 , u(0, ε) = {1, 1, 1}, (5)
dt r=1
where Λ(t) = diag (λ1 (t), λ2 (t), λ3 (t)), e2 = {0, 1, 0}, and the scalar functions gr (t) are unknown
for now and will be specified when constructing the asymptotic solution of problem (2). For a
vector function ỹ = ỹ(t, u, ε) such that ỹ(t, u(t, ε), ε) ≡ y(t, ε) [where u = u(t, ε) are solutions of the
normal form (5) and y(t, ε) is an exact solution of system (3)], it is natural to pose the following
problem:
∂ w̃ ∂ w̃
l+1
ε + Λ(t)u + εr gr (t)e2 − A0 (t)w̃ − εA1 (t)w̃
∂t ∂u r=1
t (6)
− Eε (t, s; λ3 ) G(t, s)w̃(s, u(s, ε), ε)ds = H(t),
0
w̃ (0, 1̄, ε) = w0 ,
where u(t, ε) is a solution of the regularizing normal form. Although system (6) is an extension of
system (3), the regularization of the integral term
t
J w̃(t, u(t, ε), ε) = Eε (t, s; λ3 ) G(t, s)w̃(s, u(s, ε), ε)ds
0
has not been performed in it; therefore, this extension is practically useless in the construction
of an asymptotic solution of problem (3). To regularize J w̃, one should introduce a class Mε
asymptotically invariant under the operator J (see [2, p. 62]).
Definition 1. We say that a vector function w(t, u) = {y, z} belongs to the space U if it can
be represented by a sum
3
w(t, u) = wj (t)uj + w0 (t), (7)
j=1
asymptotically convergent (as ε → +0) uniformly with respect to t ∈ [0, T ]. Indeed, on func-
tions (7), the image Jw has the form
t t
Jw(t, u) = Eε (t, s; λ3 ) G(t, s)w2 (s)u2 (s, ε)ds + Eε (t, s; λ3 ) + Eε (s, 0; λ1 ) G(t, s)w1 (s)ds
0 0
(9)
t t
+ Eε (t, s; λ3 ) G(t, s)w0 (s)ds + Eε (t, 0; λ3 ) G(t, s)w3 (s)ds.
0 0
[Here we have used the relation uj (t, ε) = Eε (t, 0; λj ), j = 1, 3.] By denoting the second and
third integrals by J1 (t, ε) and J0 (t, ε), respectively, and by integrating by parts in these integrals,
we obtain
∞
J0 (t, ε) = (−1)m εm+1 [(I0m (G(t, s)w0 (s)))s=t − (I0m (G(t, s)w0 (s)))s=0 Eε (t, 0; λ3 )] ,
m=0
∞
(10)
J1 (t, ε) = (−1)m εm+1 [(I1m (G(t, s)w1 (s)))s=t Eε (t, 0; λ1 )
m=0
− (I1m (G(t, s)w1 (s)))s=0 Eε (t, 0; λ3 )] ,
DIFFERENTIAL EQUATIONS Vol. 42 No. 5 2006
708 BOBODZHANOV, SAFONOV
1 1 ∂ m−1
I00 = − , I0m = − I (m ≥ 1),
λ3 (s) λ3 (s) ∂s 0
1 1 ∂ m−1
I10 = , I1m = I (m ≥ 1).
λ1 (s) − λ3 (s) λ1 (s) − λ3 (s) ∂s 1
One can readily show (see [5]) that the series (10) converge asymptotically to the corresponding
integrals as ε → +0. To regularize the first integral in (10) [we denote it by J2 (t, ε)], we use the
following assertion.
Lemma 1. Let b(t), c(t) ∈ C[0, T ] be functions such that b(t) = c(t) for all t ∈ [0, T ], and let
g(t, ε) be a continuous function of t ∈ [0, T ] for each ε > 0. If a function v = v(t, ε) is a solution
of the equation
εv̇ = b(t)v + g(t, ε),
then
ε d g(t, ε)
v(t, ε) = Eε (t, 0; c) (Eε (0, t; c)v(t, ε)) − (11)
b(t) − c(t) dt b(t) − c(t)
for arbitrary t ∈ [0, T ] and ε > 0.
l+1
By using Lemma 1 with c(t) ≡ λ3 (t), b(t) ≡ λ2 (t), v ≡ u2 (t, ε), and g(t, ε) ≡ r=1 εr gr (t),
we obtain
t
ε d
J2 (t, ε) = Eε (t, s; λ3 ) G(t, s)w2 (s) Eε (s, 0; λ3 ) (Eε (0, s; λ3 ) u2 (s, ε)) ds
λ2 (s) − λ3 (s) ds
0
t l+1
l+1 t
− εr Eε (t, s; λ3 ) I20 (G(t, s)w2 (s)gr (s)) ds,
r=1 0
t
∂ 0
− I20 (G(t, s)w2 (s)) s=0 u2 (0, ε) − Eε (0, s; λ3 ) I (G(t, s)w2 (s)) u2 (s, ε)ds
∂s 2
0
0 0
= I2 (G(t, s)w2 (s)) s=t u2 (t, ε) − I2 (G(t, s)w2 (s)) s=0 Eε (t, 0; λ3 )
t
∂ 0
− Eε (t, s; λ3 ) I2 (G(t, s)w2 (s)) u2 (s, ε)ds.
∂s
0
Here the last integral is of the type J2 (t, ε) where G(t, s)w2 (s) is replaced by the vector function
∂ 0
I (G(t, s)w2 (s)) ; therefore, one can again use Lemma 1. Next, we have [see (10)]
∂s 2
t
∞
0
Eε (t, s; λ3 ) I2 (G(t, s)w2 (s)gr (s)) ds = (−1)m εm+1 I0m I20 (G(t, s)w2 (s)gr (s)) s=t
0 m=0
− I0m I20 (G(t, s)w2 (s)gr (s)) s=0 Eε (t, 0; λ3 ) .
Therefore, by using Lemma 1 and integration by parts many times, one can rewrite J2 (t, ε) in
(1)
the form of the series (8), where vj (t) ≡ 0. One can readily show that it converges to J2 (t, ε)
asymptotically as ε → +0 (uniformly with respect to t ∈ [0, T ]); therefore, the image Jw(t, u) can
be represented in the form of the series (8) asymptotically converging to Jw uniformly with respect
to t ∈ [0, T ]. It follows that the class Mε = U |u=u(t,ε) is asymptotically invariant under the integral
operator J.
By grouping the coefficients of like powers of ε in Jw(t, u), we obtain
∞
Jw(t, u) = R0 w(t, u) + εm+1 Rm+1 w(t, u),
m=0
where u = u(t, ε) is the solution of the normal form (5) and the Rm : U → U are the opera-
tors (the order operators in ε) such that the image Rm w(t, u) is the sum of all coefficients of εm
in Jw(t, u). The explicit expressions for these operators for m = 0, 1, 2 are the following:
t
R0 w(t, u) = u3 G(t, s)w3 (s)ds,
0
R1 w(t, u) = (G(t, s)w0 (s)) s=t − I00 (G(t, s)w0 (s)) s=0 u3
I00
2
0
+ Ij (G(t, s)wj (s)) s=t uj − Ij0 (G(t, s)wj (s)) s=0 u3 ,
(12)
j=1
R2 w(t, u) = − I01 (G(t, s)w0 (s)) s=t − I01 (G(t, s)w0 (s)) s=0 u3
2
1
− Ij (G(t, s)wj (s)) s=t uj − Ij1 (G(t, s)wj (s)) s=0 u3
j=1
+ I00 I20 (G(t, s)w2 (s)g1 (s)) s=t − I00 I20 (G(t, s)w2 (s)g1 (s)) s=t u3 .
The explicit expressions for Rm with m ≥ 3 are quite cumbersome, and we omit them. If
∞
w̃(t, u, ε) = εk wk (t, u) (13)
k=0
is a series with coefficients wk (t, u) ∈ U , then the formal expansion J˜ of the operator J on such
series is constructed by the well-known rule [5]:
∞
def
∞
ν
J˜w̃(t, u, ε) ≡ J˜ k
ε wk (t, u) = εν Rν−s ws (t, u).
k=0 ν=0 s=0, ν−s≥0
Moreover, the regularized [with respect to (3)] problem has the form
∂ w̃ ∂ w̃
l+1
ε + Λ(t)u + ε gr (t)e2 − A0 (t)w̃ − εA1 (t)w̃ − J˜w̃ = H(t),
r
∂t ∂u r=1
(14)
w̃ (0, 1̄, ε) = w0 .
This problem is meaningful in the class of functions w̃(t, u, ε) that can be represented by se-
ries (13) convergent asymptotically (as ε → +0) uniformly with respect to (t, u) ∈ [0, T ] × Π. (Here
Π = {u : |uj | < 1 + δ, j = 1, 2, 3} and δ > 0 is a small constant.)
By substituting the series (13) into (14) and by matching the coefficients of like powers of ε,
we obtain the iterative problems
∂w0
L0 w0 (t, u) ≡ Λ(t)u − A0 (t)w0 − R0 w0 = H(t), w0 (0, 1̄) = w0 , (140 )
∂u
∂w0 ∂w0
L0 w1 (t, u) = − − A1 (t)w0 − g1 (t)e2 + R1 w0 , w1 (0, 1̄) = 0, (141 )
∂t ∂u
∂w1 ∂w0 ∂w1
L0 w2 (t, u) = − − A1 (t)w1 − g2 (t)e2 − g1 (t)e2 + R2 w0 + R1 w1 ,
∂t ∂u ∂u (142 )
w2 (0, 1̄) = 0, ...,
∂wk−1
∂wj k−1
k−1
L0 wk (t, u) = − − A1 (t)wk−1 − gk−j (t)e2 + Rk−j wj ,
∂t j=1
∂u j=0 (14k )
wk (0, 1̄) = 0, k ≥ 2,
where gj (t) ≡ 0 for j > l + 1 and uj are treated as independent variables.
The solutions of iterative problems (14k ) are sought in the subspace V of the space U given by
the formula
3
V = w(t, u) ∈ U : w = wj (t)uj + w0 (t), (w0 (t), χ2 (t)) ≡ 0, t ∈ [0, T ] .
j=1
The right-hand sides of Eqs. (14k ) do not necessarily belong to the space V . Their embedding in
the space V is performed with the use of the functions gj (t) occurring in the normal form (5). Let
us illustrate this using the iterative equation (141 ) as an example. Since it contains the solution
w0 (t, u) of the first iterative problem (140 ), we start from the computation of this solution.
3
(0) (0)
w0 (t, u) = wj (t)uj + w0 (t). (15)
j=1
3 t
(0) (0) (0) (0)
λj (t)wj (t)uj − A0 (t)wj (t)uj − A0 (t)w0 (t) − u3 G(t, s)w3 (s)ds = H(t).
j=1 j=1 0
Here, by separately matching the coefficients of uj and the free terms, we obtain
System (16) has the solution wj (t) = αj (t)ϕj (t), where αj (t) ∈ C∞ ([0, T ], C1 ) is an arbitrary
(0) (0)
(0)
function, j = 1, 2. The integral system (18) has the unique solution w3 (t) ≡ 0 (for all t ∈ [0, T ]),
since λ3 (t) ∈ σ(A(t)). To compute the solution of system (17), we make the transformation
(0)
w0 (t) = Φ(t)ξ in this system, where Φ(t) = (ϕ1 (t), ϕ2 (t)) and ξ = {ξ1 , ξ2 }. Then
−A0 (t)Φ(t)ξ = H(t) ⇔ − Φ−1 (t)A0 (t)Φ(t) ξ = Φ−1 (t)H(t)
−λ1 (t)ξ1 = (H(t), χ1 (t))
⇔ −Λ0 (t)ξ = Φ−1 (t)H(t) ⇔
−λ2 (t)ξ2 = (H(t), χ2 (t)) .
ξ1 (t) = −λ−1
1 (t) (H(t), χ1 (t)) .
Since λ2 (t) ≡ μ1 (t) ≡ 0 for t ∈ S and λ2 (t) = 0 for t ∈ [0, T ]/S, it follows that the second equation
in the system has the solution
0 for t ∈ [0, T ]/S
ξ2 (t) =
γ2 (t) for t ∈ S,
where γ2 (t) is an arbitrary smooth continuation of zero from the set [0, T ]/S on the set S. In this
case, system (5) has the solution
(H(t), χ1 (t))
w0 (t, u) = − ϕ1 (t) + ξ2 (t)ϕ2 (t). (19)
λ1 (t)
However, it is required that w0 (t, u) ∈ V ; therefore, the function (19) should satisfy the condition
(0)
w0 (t), χ2 (t) ≡ 0, whence we obtain ξ2 (t) ≡ 0 (for all t ∈ [0, T ]); i.e., γ2 (t) ≡ 0 for all t ∈ S.
The solution (15) of system (140 ) is thereby found in the form
2
(H(t), χ1 (t))
(0) 2
w0 (t, u) = α(0)
j (t)ϕj (t)uj − ϕ1 (t) ≡ αj (t)ϕj (t)uj + w0(0) (t), (20)
j=1
λ1 (t) j=1
where αj (t), j = 1, 2, are scalar functions so far unknown. By subjecting the function (20) to the
initial condition w0 (0, 1̄) = w0 , we obtain
2
(0) (H(0), χ1 (0))
αj (0)ϕj (0) = w0 + ϕ1 (0).
j=1
λ1 (0)
The inner product of this relation by χj (0) provides the initial values for the functions αj (0) :
(H(0), χ1 (0))
α1 (0) = w0 , χ1 (0) + , α2 (0) = w0 , χ2 (0) . (21)
λ1 (0)
2
2
(0) (0) (0) (0)
L0 w1 (t, u) = − αj (t)ϕj (t) uj − ẇ0 (t) − αj (t)A1 (t)ϕj (t)uj − A1 (t)w0 (t)
j=1 j=1
3
(1) (1)
w1 (t, u) = wj (t)uj + w0 (t), (23)
j=1
[λj (t)I − A0 (t)] wj(1) (t) = − α(0)
j (t)ϕj (t) − A1 (t)ϕj (t)α(0)
j (t)
(24)
G(t, t)ϕj (t) (0)
+ α (t), j = 1, 2,
λj (t) − λ3 (t) j
t
G(t, 0)w0 (0)
G(t, 0)ϕj (0)αj (0)
(0) 2 (0)
(1) (1)
[λ3 (t)I − A0 (t)] w3 (t) − G(t, s)w3 (s)ds = − , (25)
λ3 (0) j=1
λj (0) − λ3 (0)
0
(0)
(1) (0) (0) (0) G(t, t)w0 (t)
− A0 (t)w0 (t) = −ẇ0 (t) − A1 (t)w0 (t) − α2 (t)ϕ2 (t)g1 (t) − . (26)
λ3 (t)
For system (24) to be solvable in the space C ∞ ([0, T ], C2 ), it is necessary and sufficient to satisfy
the orthogonality condition (e.g., see [2])
(0) G(t, t)ϕj (t) (0)
− αj (t)ϕj (t) − A1 (t)ϕj (t) + α (t), χj (t) ≡ 0, j = 1, 2, ∀t ∈ [0, T ],
λj (t) − λ3 (t) j
or
(0) G(t, t) (0)
α̇j (t) = − A1 (t) ϕj (t) − ϕ̇j (t), χj (t) αj (t) ∀t ∈ [0, T ], j = 1, 2. (27)
λj (t) − λ3 (t)
DIFFERENTIAL EQUATIONS Vol. 42 No. 5 2006
EQUATIONS WITH AN UNSTABLE SPECTRAL VALUE OF THE KERNEL 713
(0)
By taking into account (21), we find the functions αj (t) :
(0) 0 (H(0), χ1 (0))
α1 (t) = w , χ1 (0) +
λ1 (0)
⎧ t ⎫
⎨ ⎬
G(θ, θ)
× exp − A1 (θ) ϕ1 (θ) − ϕ̇1 (θ), χ1 (θ) dθ ,
⎩ λ1 (θ) − λ3 (θ) ⎭
(28)
0
⎧ t ⎫
⎨ ⎬
(0) G(θ, θ)
α2 (t) = w0 , χ2 (0) exp − A1 (θ) ϕ2 (θ) − ϕ̇j (θ), χj (θ) dθ ,
⎩ λ2 (θ) − λ3 (θ) ⎭
0
j = 1, 2.
Therefore, the solution (20) of the iterative problem (140 ) is uniquely determined in the space V .
The system of integral equations (24) is also uniquely solvable in the space C∞ ([0, T ], C2 ), since
λ3 (t) ∈ σ (A0 (t)). One can readily see that system (26) is uniquely solvable in the class
W = C∞ [0, T ], C2 ∪ {(w0 (t), χ2 (t)) ≡ 0 ∀t ∈ [0, T ]}
if and only if its right-hand side belongs to W , i.e.,
(0)
G(t, t)w (t)
−ẇ0(0) (t) − A1 (t)w0(0) (t) − α(0)
2 (t)ϕ2 (t)g1 (t) −
0
, χ2 (t) ≡ 0.
λ3 (t)
Note that, in the solution of the first two iterative problems (140 ) and (141 ), we construct a
regularizing normal form (5) of order l + 1 = 1. It has the form
du
ε= Λ(t)u + εg1 (t)e2 , u(0, ε) = 1̄,
dt
where g1 (t) is the function (29).
Theorem 2. Let conditions (1◦ )–(3◦ ) and (5◦ ) be satisfied, let (w0 , χ1 (0)) = 0, and let the
3
right-hand side P (t, u) = j=1 Pj (t)uj + P0 (t) of system (30) belong to the space U . System (30)
is solvable in the space V if and only if
P (t, u) ∈ V ⇐⇒ (P0 (t), χ2 (t)) ≡ 0 (∀t ∈ [0, T ]), (32)
P (t, u), χj (t)uj ≡ 0, j = 1, 2 (∀t ∈ [0, T ]). (33)
Note that if conditions (32) and (33) are satisfied, then in the space V , system (30) has the
solution
2
2
(Pj (t), χs (t))
w(t, u) = αj (t)ϕj (t) + ϕs (t) uj + w3 (t)u3
j=1
λ (t) − λs (t)
s=j, s=1 j (34)
(P0 (t), χ1 (t))
− ϕ1 (t),
λ1 (t)
where αj (t) ∈ C ∞ ([0, T ], C1 ) are arbitrary functions and w3 (t) is a vector function satisfying the
integral equation
t
[λ3 (t)I − A0 (t)] w3 (t) − G(t, s)w3 (s)ds = P3 (t). (35)
0
∗
The initial condition w (0, 1̄) = w leads to the system
2
2
2
(Pj (0), χs (0)) −1
αj (0)ϕj (0) = w∗ − ϕs (0) − [λ3 (0)I − A0 (0)] P3 (0)
j=1 j=1 s=j, s=1
λj (0) − λs (0)
(P0 (0), χ1 (0))
+ ϕ1 (0).
λ1 (0)
By taking the inner product of this relation by χk (0) and by using the biorthonormality property
of the systems {ϕj (t)} and {χk (t)}, we obtain
(Pj (0), χk (0))
2
−1
αk (0) = (w∗ , χk (0)) − − [λ3 (0)I − A0 (0)] P3 (0), χk (0)
j=1, j=k
λj (0) − λk (0)
(36)
(P1 (0), χ1 (0))
+ δ1k , k = 1, 2.
λ1 (0)
For the final computation of the functions αj (t), one should apply Theorem 2 to system (31) [first
having computed its right-hand side in view of the function (34)]. In this case, for the functions
αj (t), we obtain the linear differential equations
G(t, t)
α̇j (t) = − A1 (t) ϕj (t) − ϕ̇j (t), χj (t) αj (t) + lj (t), j = 1, 2,
λj (t) − λ3 (t)
DIFFERENTIAL EQUATIONS Vol. 42 No. 5 2006
EQUATIONS WITH AN UNSTABLE SPECTRAL VALUE OF THE KERNEL 715
where lj (t) ∈ C ∞ ([0, T ], C1 ) is a known function. By supplementing these equations with the initial
conditions (36), we obtain unique functions αj (t), j = 1, 2. The condition (32) of the embedding
of the right-hand side of system (31) in the space V permits one to find the function g(t) also
uniquely. Let us state the corresponding assertion.
Theorem 3. Let Q(t, u) ∈ U, let conditions (1◦ )–(3◦ ) and (5◦ ) be satisfied, let (w0 , χ2 (0)) = 0,
3
and let the right-hand side P (t, u) = j=1 Pj (t)uj +P0 (t) ∈ U of system (30) satisfy conditions (32)
and (33). Then under the additional conditions
! "
∂w
− − A1 (t)w + R1 y + Q(t, u), χj (t) ≡ 0, j = 1, 2 (∀t ∈ [0, T ]),
∂t
problem (30) is uniquely solvable in the space V . Moreover, there exists a unique function
g(t) ∈ C∞ ([0, T ], C1 ) such that the right-hand side of system (31) belongs to the space V .
Theorem 4. Let (w0 , χ2 (0)) = 0, and let conditions (1◦ )–(5◦ ) be satisfied. Then problem (3) is
uniquely solvable in the class C1 ([0, T ], C2 ) , and its solution w(t, ε) admits the estimate
where CN > 0 is a constant independent of ε for ε ∈ (0, ε0 ] (ε0 > 0 is sufficiently small) and
wεN (t) = SN t, u(N ) (t, ε), ε is the above-constructed formal solution of order N .
Let us now study the passage to the limit in problem (3) under the following conditions, more
restrictive than (4◦ ) :
(6◦ ) Re a(t) < 0 and Re μ0 (t) < 0 for all t ∈ [0, T ]; Re μ1 (t) < 0 for all t ∈ [0, T ]/S.
Let Q = [α, β] be an arbitrary closed interval lying in the half-closed interval (0, T ] and disjoint
with the set S̄ [S̄ is the closure of S]. Then the interval Qν = [α − ν, β] has the same property for
sufficiently small ν > 0.
Lemma 2. Let (w0 , χ2 (0)) = 0, and let conditions (1◦ )–(3◦ ), (5◦ ), and (6◦ ) be satisfied. Then
for all t ∈ Q and for sufficiently small ε > 0 (0 < ε ≤ ε0 ), the components u0j (t, ε) of the solution
u(0) (t, ε) of the normal form
du
ε = Λ(t)u + εg1 (t)e2 , u(0, ε) = 1̄, (38)
dt
where g1 (t) is the function given by (29), satisfy the inequalities
' '
' (0) ' ε
'uj (t, ε)' ≤ c0 e−χν/ε + δj2 g1 , j = 1, 2, 3, (39)
χ
where δjk is the Kronecker delta, c0 > 0 is a constant independent of ε for ε ∈ (0, ε0 ] , g1 =
maxt∈[0,T ] |g1 (t)| , and χ = minj=1,2,3 mint∈Qν (− Re λj (t)).
Proof. The solution of the normal form (38) for t ∈ [0, T ] has the form
t
(0)
uj (t, ε) = Eε (t, 0; λj ) + δj2 Eε (t, s; λ2 ) g1 (s)ds, j = 1, 2, 3,
0
are valid for all t ∈ [0, T ] and ε > 0. Let us now write out a solution of the normal form (38) for
t ∈ Qν = [α − ν, β] :
t
(0)
uj (t, ε) = Eε (t, α − ν; λj ) uj (α − ν, ε) + δj2 Eε (t, s; λ2 ) g1 (s)ds, j = 1, 2, 3.
α−ν
χν t
χ(t − s)
≤ c0 exp − + g1 exp − ds
ε ε
α−ν
χν 's=t
ε χ(t − s) ''
≤ c0 exp − + g1 exp − '
ε χ ε
χν s=α−ν
ε χ(t − α + ν)
= c0 exp − + g1 1 − exp −
ε χ ε
χν g
1
≤ c0 exp − + ε
ε χ
for t ∈ [α, β] = Q. The proof of the lemma is complete.
We use the resulting estimates to analyze the passage to the limit in problem (3) on the inter-
val Q. By Theorem 1 and the estimate (37), we have
( (
(
2 (
( (0) (0) (0) (
C1 ε ≥ (w(t, ε) − w0 (t) − αj (t)ϕj (t)uj (t, ε)(
( j=1
(
C(Q)
( (
2 ( (
( (0) ( ( (0) (
≥ (w(t, ε) − w0 (t)( − qj (uj (t, ε)(
C(Q) C(Q)
j=1
or
( (
2 ( (
( (0) ( ( (0) (
(w(t, ε) − w0 (t)( ≤ C1 ε + qj (uj (t, ε)( ,
C(Q) C(Q)
j=1
( ( χν χν ε
( (0) (
(w(t, ε) − w0 (t)( ≤ C1 ε + q1 c0 exp − + q2 c0 exp − + g1 → 0
C(Q) ε ε χ
Theorem 5. Let (w0 , χ2 (0)) = 0, and let conditions (1◦ )–(3◦ ), (5◦ ), and (6◦ ) be satisfied. Then
(0)
the limit system 0 = A0 (t)w̄ + H(t) has a unique smooth solution w̄(t) = w0 (t) ∈ V, and
( (
( (0) (
(w(t, ε) − w0 (t)( →0 (ε → +0),
C(Q)
Let us analyze the passage to the limit on the set S. For simplicity, we assume that S is
the segment [t1 , t2 ] ⊂ (0, T ). Since Re λ2 (t) is a continuous function on the segment [0, T ] and
Re λ2 (t) ≡ 0 for t ∈ [t1 , t2 ], we have Re λ2 (t) → 0 as t → t1 − 0. However, this convergence is not
necessarily monotone; therefore, we impose one more requirement.
(7◦ ) The function Re λ2 (t) is strictly increasing in some left neighborhood of the point t = t1 .
[That is, there exists a δ0 > 0 such that Re λ2 (t) is strictly increasing for t ∈ [t1 − δ0 , t1 ).]
Following the proof of Lemma 1, one can show that the estimate
' '
' (0) '
'uj (t, ε)' ≤ c0 exp (−χ0 ν/ε) , j = 1, 3, (40)
is valid for t ∈ [t1 , t2 ], where χ0 = minj=1,3 mint∈[t1 −ν,t2 ] (− Re λj (t)) and ν > 0 is a number such
that [t1 − ν, t2 ] ⊂ (0, T ).
From (37) and (40), we obtain the estimates
or ( (
( (0) (0) (0) (
(w(t, ε) − w0 (t) + α2 (t)ϕ2 (t)u2 (t, ε) ( ≤ C1 ε + q1 c0 exp (−χ0 ν/ε) . (41)
C(S)
(0)
The regularizing function u2 (t1 , ε) on the set S = [t1 , t2 ] can be represented in the form
t
(0) (0)
u2 (t, ε) = u2 (t1 , ε) + g1 (s)ds
t1
⎛ ⎞
t1 t
≡ ⎝Eε (t1 , 0; λ2 ) + Eε (t1 , s; λ2 ) g1 (s)ds⎠ + g1 (s)ds.
0 t1
Here, by condition (6◦ ), the exponential Eε (t1 , 0; λ2 ) tends to zero as ε → +0. By using the ideas
of [1], one can show that the second term on the right-hand side in the last relation tends to zero
and consequently,
( ⎡ ⎤(
( t (
( (
(w(t, ε) − ⎣w0 (t) + α2 (t)ϕ2 (t) g1 (s)ds⎦(
(0) (0)
→0 (ε → +0). (42)
( (
( (
t1 C(S)
Theorem 6. Let (w0 , χ2 (0)) = 0, and let conditions (1◦ )–(3◦ ) and (5◦ )–(7◦ ) be satisfied. Then
the passage to the limit (42) is valid on the set S = [t1 , t2 ] ⊂ (0, T ), where w(t, ε) is an exact
solution of problem (3), w0 (t) = −λ−1
(0) (0)
1 (t) (H(t), χ1 (t)) ϕ1 (t), and α1 (t) and g1 (t) are the functions
computed with the use of formulas (29) and (20).
The exact solution w(t, ε) has two layers, a boundary layer in a neighborhood of the point t = 0
and an internal transition layer (a contrast structure; see also [6]) in a neighborhood of the set S.
By taking into account (29) and (20), we rewrite w̄ ¯ (t) for t ∈ S in the form
⎧ t ⎫
t ⎨ G(θ, θ)ϕ (θ) ⎬
2
¯ (t) = w0(0) (t) − ϕ2 (t) exp
w̄ − A1 (θ) ϕ2 (θ) − ϕ̇2 (θ), χ2 (θ) dθ
⎩ λ2 (θ) − λ3 (θ) ⎭
t1 s (43)
(0)
(0) G(s, s)w0 (s)
× ẇ0 (s) + A1 (s)w0 (s) + , χ2 (s) ds,
λ3 (s)
to embed the right-hand sides of systems (14k ) in the space V with the use of the functions gj (t).
These functions are absent [gj (t) ≡ 0, j = 1, . . . , l + 1], and the normal form (5) itself acquires
the form εu̇ = Λ(t)u, u(0, ε) = 1̄; i.e., the regularization of problem (3) is performed with respect
to the spectrum {λj (t)} of the matrix A0 (t) and the spectral value λ3 (t) = μ1 (t) of the kernel of
the integral operator. This completely coincides with the ideas in the paper [2], where an integro-
differential system with a stable spectrum was considered. Therefore, the above-developed method
of normal forms should be treated as a generalization of the Lomov regularization method [2] to
integro-differential problems with unstable spectrum.
REFERENCES
1. Bobodzhanov, A.A., Kalimbetov, B.T., and Safonov, V.F., Vestn. Moskov. Energ. Inst., 2002, no. 6,
pp. 15–27.
2. Lomov, S.A., Vvedenie v obshchuyu teoriyu singulyarnykh vozmushchenii (Introduction to General Sin-
gular Perturbation Theory), Moscow, 1981.
3. Kalimbetov, B.T. and Safonov, V.F., Uzbek. Mat. Zh., 2002, no. 1, pp. 36–42.
4. Rumyantseva, M.A. and Safonov, V.F., Vestn. Moskov. Energ. Inst., 1995, no. 6, pp. 91–108.
5. Bobodzhanov, A.A. and Safonov, V.F., Mat. Sb., 2001, vol. 192, no. 8, pp. 53–78.
6. Vasil’eva, A.B., Butuzov, V.F., and Nefedov, N.N., Fundam. Prikl. Mat., 1998, vol. 4, no. 3, pp. 799–851.
DOI: 10.1134/S0012266106050132
where n is the inward normal on ∂Ω, since this choice guarantees that the zero-order boundary
layer is lacking in the asymptotics.
We require that the following conditions be satisfied. (The remaining conditions will be stated
when constructing and justifying the asymptotics.)
I. The function g is sufficiently smooth. [In the construction of the asymptotics with remainder
of the order O (εn+1 ), it suffices to require that g is n + 2 times continuously differentiable.]
II. We assume that the degenerate equation
L(u, u, x, 0) = 0, x ∈ Ω̄, (1.3)
has discontinuous solutions of the form
ϕ1 (x, C) for x ∈ Ω̄\Ω̄i
ϕ(x, C) =
ϕ3 (x, C) for x ∈ Ωi ,
739
740 NEFEDOV, NIKITIN
where C ≡ ∂Ωi is an arbitrary simple smooth curve lying in the domain Ω, ϕ1 (x− , C) < ϕ3 (x+ , C)
for x− = x+ = x ∈ C [here and throughout the following, ϕ1 (x− , C) and ϕ3 (x+ , C) stand for the
limit values of ϕ1 (x, C) and ϕ3 (x, C) as x → C from the exterior domain Ω̄\Ω̄i and the interior
domain Ωi , respectively], and
We also assume that Eq. (1.3) has a solution ϕ2 (x) continuous in the entire domain Ω̄ such that
ϕ1 (x, C) < ϕ2 (x) < ϕ3 (x, C) for x ∈ C and Lu (ϕ2 , ϕ2 , x, 0) < 0 for x ∈ Ω̄. To be definite,
we assume that ϕ2 ≡ 0.
Remark. The degenerate equation (1.3) has the form of the nonlinear integral equation of the
first kind
g(u(x), u(s), x, s)ds = 0.
Ω
In the preceding (Condition II), we require that this equation has a solution of a specific form.
It follows from Condition IV below that the linearization of the degenerate equation on this solution
is a linear integral equation of the second kind; consequently, this solution is stable under small
perturbations.
In neighborhoods of the curves C and ∂Ω, we introduce local orthogonal coordinate systems in
a standard way. We seek the asymptotic expansion of the solution of problem (1.1), (1.2) with the
use of the boundary function algorithm [4] in the form
∞
u(x, ε) = εn (ūn (x) + Pn u(η, l) + Πn u(τ, m)) ≡ ū(x, ε) + P u(η, l, ε) + Πu(τ, m, ε), (1.5)
n=0
where ūi (x) are regular terms of the asymptotics, Pi u(η, l) are internal transition layer functions,
and Πi u(τ, m) are boundary layer functions. Here the arguments η and l refer to a local coordinate
system (r, l) introduced in a neighborhood of the curve C∗ given by the condition
u|x∈C∗ = 0, (1.6)
and the arguments τ and m refer to a local coordinate system (, m) introduced in a neighborhood
of the boundary ∂Ω in a standard way (see [4]). Consider the local coordinate system (r, l) in
more detail. As will be shown below, the curve C∗ undergoes small changes in a neighborhood
of some curve C0 independent of ε as the parameter ε varies. In a neighborhood of the unknown
(for now) curve C0 , whose equation will be specified below, we introduce a local coordinate sys-
tem (r, l), where |r| is the distance from a point M (x) to the curve C0 along the normal to the curve
through M and l is the coordinate of the foot of the normal on C0 . The coordinate r is positive
if the point M lies in the domain Ωi0 bounded by the curve C0 and negative if the point lies in
the domain Ω\Ωi0 . If C0 is sufficiently smooth and the neighborhood is sufficiently small, then the
normals issuing from different points of the curve do not meet in this neighborhood; therefore,
there exists a one-to-one correspondence between the coordinates (x1 , x2 ) and (r, l) of the point M .
In the local coordinate system, the curve C0 is given by the equation r = 0, and we seek the
equation of the curve C∗ in this coordinate system in the form
r = f (l, ε) = εr1 (l) + ε2 r2 (l) + · · · , (1.7)
where ri (l) are sufficiently smooth periodic functions. The variable η is the 1/ε-dilated distance
from the curve C∗ , η = (r − f (l, ε))/ε.
In a similar way, in a neighborhood of the boundary ∂Ω, we introduce a local coordinate system
(, m), where is the distance from a point M ∈ Ω to the boundary ∂Ω along the inward normal
to the boundary passing through M and m is the coordinate of the foot of the normal on the
boundary. The variable τ = /ε is the 1/ε-elongated distance from the boundary of Ω.
In Section 2, we construct a formal asymptotics of the solution of problem (1.1), (1.2). In Sec-
tion 3, we justify the asymptotics by the method of asymptotic differential inequalities [1, 3] mod-
ified for the new type of solutions and prove the local uniqueness of the solution of problem (1.1),
(1.2) and the Lyapunov asymptotic stability of the resulting solution treated as a stationary solution
of the corresponding integro-parabolic boundary value problem
∂u
= ε2 Δu − L(u, u, x, ε), x = (x1 , x2 ) ∈ Ω, (1.8)
∂t
with the homogeneous boundary conditions (1.2) and with initial conditions lying in a neighborhood
of the asymptotic solution constructed.
in the standard way. To this end, we represent the function g in the form
g ≡ ḡ + Γg + Ψg + T g + Qg,
where
ḡ = g (ū(x, ε), ū(s, ε), x, s, ε) ,
Γg + T g = g(u(x, ε), u(s, ε), x, s, ε) − g (ū(x, ε), u(s, ε), x, s, ε) ,
Ψg + Qg = g (ū(x, ε), u(s, ε), x, s, ε) − ḡ.
In turn, we rewrite the functions Γg and T g as
Γg + T g ≡ P g + Φg + Πg + Θg,
where P g + Πg = g (u(x, ε), ū(s, ε), x, s, ε) − g (ū(x, ε), ū(s, ε), x, s, ε).
We expand these functions in powers of the small parameter:
∞
∞
∞
n n
ḡ = ε ḡn (x, s), Pg = ε Pn g(η, l, s), Πg = εn Πn g(τ, m, s),
n=0 n=0 n=0
∞
∞
Φg = εn Φn g(η, l, σ, t), Θg = εn Θn g(τ, m, ξ, p),
n=0 n=0
∞
∞
Ψg = εn Ψn (x, σ, t), Qg = εn Qn g(x, ξ, t),
n=0 n=0
where the coordinates (σ, t) and (ξ, p) of the point M (s) correspond to the coordinates (η, l) and
(τ, m) of the point M (x). Note that, after the substitution of these expansions into the operator L,
the functions Pn g and Πn g occur in the equations for the transition layer terms and the nth-order
boundary-layer terms near the curve C and the boundary of Ω, respectively, and the functions Φn g
and Θn g occur in equations for the (n + 1)st-order functions depending on the fast variables. This
is due to the fact that, as will be shown below, the boundary functions are exponentially decreasing
with respect to the boundary layer variables; therefore, integration with respect to the slow variable
s increases the order of smallness of the terms in the expansion by unity. After the integration,
the functions Ψn g and Qn g no longer depend on the boundary layer variables and hence occur
in the equations for the regular terms of the (n + 1)st-order asymptotics.
In the terms in which the operator Δ acts on the P - and Π-functions, we pass to the variables
(η, l) and (τ, m). Let us consider the passage to local coordinates in a neighborhood of the curve
C0 in detail. In the local coordinates (r, l), the operator Δ has the form
∂2 ∂2 ∂ ∂
Δ= 2
+ a(r, l) 2
+ b(r, l) + c(r, l) ,
∂r ∂l ∂r ∂l
DIFFERENTIAL EQUATIONS Vol. 42 No. 5 2006
742 NEFEDOV, NIKITIN
where
2 2
∂l ∂l ∂2r ∂2r ∂2l ∂2l
a(r, l) = + , b(r, l) = 2
+ 2, c(r, l) = 2
+ 2,
∂x1 ∂x2 ∂x1 ∂x2 ∂x1 ∂x2
and r = r (x1 , x2 ) and l = l (x1 , x2 ) are formulas describing the transition from the coordinates
(x1 , x2 ) to the coordinates (r, l). Note that b(0, l) is the curvature of the curve C0 at the point (0, l).
In the variables (η, l), the operator ε2 Δ has the form
∂2
1 + ε2 + εA1 + ε2 A2 + · · · ,
∂η 2
where = (εη, l, ε) is a known function and A1 , A2 , . . . are linear differential operators containing
the derivatives ∂/∂η, ∂/∂l, and ∂ 2 /∂l2 ; in particular, A1 = b(0, l)∂/∂η. Following the method of
boundary functions, we obtain a sequence of problems for the coefficients in the expansion of u(x, ε).
As an additional condition, which permits one to find the coefficients ri (l) in the expansion (1.7)
for the curve C∗ , we use the condition of continuity of the r-derivative of u on C∗ . For convenience,
we write out this relation with the factor ε :
∂u x(+) , ε ∂u x(−) , ε
ε =ε for x ∈ C∗ . (2.1)
∂r ∂r
First, let us find the zero terms of the asymptotics. For the zero-order term of the regular part
of the asymptotics, we obtain the nonlinear integral equation
For the solution of this equation, we take ū0 (x) = ϕ (x, C∗ ). Note that the curve C∗ is not de-
fined yet.
The zero-order boundary layer term is determined from the nonlinear integro-differential equa-
tion
∂ 2 Π0 u
= Π0 g ds, τ > 0, (2.3)
∂τ 2
Ω
where
Obviously,
Π0 u ≡ 0 (2.4)
is a solution of this problem. We take this solution for the zero-order boundary layer term.
The zero-order term of the transition layer is determined from the equation
∂ 2 P0 u
= P0 g ds, η < 0, η > 0,
∂η 2
Ω
where
±
P0 g = g ϕi x±
0 , C0 + P0 u(η, l), ϕ (s, C0 ) , x0 , s, 0 − g ϕi x0 , C0 , ϕ (s, C0 ) , x0 , s, 0
(i = 1 for η < 0, i = 3 for η > 0). From now on, x0 stands for x ∈ C0 , which corresponds to
coordinates (0, l) in the local coordinate system (r, l).
In view of (2.2), the last equation acquires the form
∂ 2 P0 u ±
→ 0 as η → ±∞,
P0 u (2.6)
−
P0 u(−0, l) + ϕ1 = P0 u(+0, l) + ϕ3 x+
x0 , C0 0 , C0 = 0, (2.7)
∂P0 u ∂P0 u
(−0, l) = (+0, l). (2.8)
∂η ∂η
Condition (2.6) is the usual condition for a transition layer function, condition (2.7) follows from
(1.6) in the zero approximation, and condition (2.8) follows from (2.1) in the zero approximation.
We introduce the function
ϕ3
(x,C0 )
⎨ ϕ1 x0 , C0 + P0 u(η, l) if η < 0
ũ(η, l) = 0 if η = 0
⎩ ϕ x+ , C
+ P u(η, l) if η > 0,
3 0 0 0
which is defined in a unified way for arbitrary values of η. The problem for ũ has the form
∂ 2 ũ
= g (ũ, ϕ (s, C0 ) , x0 , s, 0) ds, −∞ < η < +∞, (2.9)
∂η 2
Ω
ũ(−∞, l) = ϕ1 (x0 , C0 ) , ũ(0, l) = ϕ2 (x0 ) ≡ 0, ũ(+∞, l) = ϕ3 (x0 , C0 ) . (2.10)
For problem (2.9), (2.10), Condition III implies that the phase plane of Eq. (2.9) contains a
cell consisting of two saddles ϕ1 (x, C0 ) and ϕ3 (x, C0 ) (x ∈ C0 ) joined by separatrices. Owing
to the presence of the cell, there exists a solution of problem (2.6)–(2.8) describing the motion
along the upper separatrix and tending to ϕ1 (x, C0 ) as η → −∞ and to ϕ3 (x, C0 ) as η → +∞.
Note that ∂ ũ/∂η > 0. Consequently, there exists a solution of Eq. (2.5) satisfying all additional
conditions (2.6)–(2.8). We have thereby constructed the zero-order asymptotics.
Let us proceed to the construction of terms of the first-order asymptotics. The regular term of
the expansion can be found from the linear integral equation
Lu (ϕ, ϕ, x, 0)ū1 (x) + gv (ϕ(x), ϕ(s), x, s, 0)ū1 (s)ds
Ω
m0 +∞ l0 +∞ (2.11)
= −Lε (ϕ, ϕ, x, 0) − Q0 g(x, p, ξ)dp dξ − Ψ0 g(x, σ, t)dt dσ,
0 0 0 −∞
where
Q0 g = [g (ϕ (x, C0 ) , ϕ1 (s, C0 ) + Π0 u(ξ, p), x, s, 0) − g (ϕ (x, C0 ) , ϕ1 (s, x0 ) , x, s, 0)]|s∈∂Ω ≡ 0
by virtue of (2.4) and
Ψ0 g = [g (ϕ (x, C0 ) , ũ(σ, t), x, s, 0) − g (ϕ (x, C0 ) , ϕi (s, C0 ) , x, s, 0)]|s∈C0
By taking into account (2.4), we obtain π1 (τ ) ≡ 0; therefore, the solution of Eq. (2.12) has the
form
∂ exp −L1/2
u (ϕ1 (x, C0 ) , ϕ, x, 0) τ
Π1 u(τ, m) = (ϕ1 (x, C0 )) .
∂ 1/2
Lu (ϕ1 (x, C0 ) , ϕ, x, 0)
x∈∂Ω
P1 u(±∞, l) = 0. (2.16)
Conditions (2.14) and (2.16) define P1 u as a function of the variable η and the parameter
∂ ũ
r1 (l), and condition (2.15) specifies the value of r1 (l). Note that the function (η, l) satisfies the
∂η
homogeneous equation corresponding to (2.13) and the boundary conditions (2.16) and vanishes
at the point η = 0. This permits one to write out P1 u in closed form with the use of the function
∂ ũ ∂ ũ
z(η, l) = (η, l) (0, l) :
∂η ∂η
τ ±∞
∂ϕ
P1 u(η, l) = − ū1 x±
0 + r1 (l) −2
z(η, l) − z(η, l) z (χ, l) z(ζ, l)p1 (ζ, l)dχ dζ,
∂r (x± ,C0 )
0
0 χ
(2.17)
where the plus sign corresponds to η > 0 and the minus sign corresponds to η < 0.
∂ ũ
The exponential estimates as η → ±∞ follow from the exponential estimates for (η, l)
∂η
and p1 (η, l). By substituting the expression (2.17) for P1 u into (2.15), for r1 (l), we obtain a linear
equation, where the coefficient of the unknown r1 (l) is equal to ∂J (x0 , C0 )/∂r0 . By Condition III,
the unknown r1 (l) is uniquely determined. Higher-order transition-layer terms can be found from
linear equations similar to (2.13), and the terms of the expansion (1.7) can be found from equations
similar to the equation for r1 (l). We have thereby constructed an asymptotics with arbitrary order
of accuracy in powers of the small parameter ε.
In the sequel, we use the following theorem on differential inequalities [5, Th. 7.2].
Theorem 3.1. Suppose that there exists a pair of lower and upper solutions α(x, ε) and β(x, ε)
of problem (1.1), (1.2) and Condition V is satisfied. Then there exists a solution u(x, ε) of prob-
lem (1.1), (1.2) such that α(x, ε) ≤ u(x, ε) ≤ β(x, ε).
The proof of the existence of a contrast structure is based on a constructive method for finding
lower and upper solutions. The functions β(x, ε) and α(x, ε) are represented by the expressions
∗
β(x, ε) = ū0 (x) + εū1β (x) + P0 u (ηβ , l) + εP1β u (ηβ , l) + ε2 P1β u (ηβ , l, ε)
+ εΠ1 u(τ, m) + ε2 exp(−kτ )D(),
∗
(3.4)
α(x, ε) = ū0 (x) + εū1α (x) + P0 u (ηα ) + εP1α u (ηα , l) + ε2 P1α u (ηα , l, ε)
2
+ εΠ1 u(τ, m) + ε exp(−kτ )D(),
where
ηβ = (r − εrβ (l))/ε, ηα = (r − εrα (l))/ε, rα (l) = r1 (l) + δ,
rβ (l) = r1 (l) − δ, ū1α (x) = ū1 (x) − A, ū1β (x) = ū1 (x) + A.
Here ū1 (x) is the first-order regular term in the expansion (1.5) constructed above, where the
function f (r, l) is replaced by rα in the first case and by rβ in the second case, and A > 0, k > 0,
and δ > 0 are some constants. The functions P0 u (ηβ , l) and P0 u (ηα , l) are found from problems
of the form (2.5)–(2.8), where η is replaced by ηβ and ηα , respectively. Likewise, the functions
P1β u (ηβ , l) and P1α u (ηα , l) are found from problems of the form (2.13)–(2.16), where η, r1 are
replaced by ηβ , rβ and ηα , rα , respectively. The functions P1β u (ηβ , l) and P1α u (ηα , l), as well as
solutions of problem (2.13)–(2.16), can be represented in closed form and admit an exponential
∗ ∗
estimate. In turn, the functions P1β u (ηβ , l, ε) and P1α u (ηα , l, ε) are chosen from the condition
that β(x, ε) and α(x, ε) are continuous on the curves Cβ and Cα , respectively (here Cβ and Cα
are the curves whose equations differ from the equation of C by the replacement of the function f
∗ ∗
by rβ and rα , respectively); moreover, P1β u and P1α u are solutions of the homogeneous equations
for P1β u and P1α u, and D() is a cutoff function. To construct a smooth continuation of the
boundary and transition layer functions to the entire domain Ω, they should also be multiplied
by the corresponding cut off functions. We assume that this standard procedure, which preserves
the order and properties of asymptotics, has been performed [4]. Note that the functions β(x, ε)
and α(x, ε) are continuous in the entire domain Ω̄ but are not smooth (their normal derivatives
experience a jump) on the curves Cβ and Cα , respectively.
Lemma. There exist constants A, δ, and k such that the functions α(x, ε) and β(x, ε) are a
pair of lower and upper solutions of problem (1.1), (1.2).
Proof. The proof is by a straightforward verification of the definition of upper and lower so-
lutions. The verification of the condition α(x, ε) ≤ β(x, ε) reproduces the proof of a similar fact
in [3] nearly word for word. Let us verify condition (3.1) for β(x, ε). We represent L by the sum
By using the explicit representations (3.4) and by representing L(β, β, x, ε) in the form of four
terms like (1.5), we obtain
where o(ε) stands for exponentially small terms. Since, by the construction of the asymptotics, the
sum of the first two terms on the right-hand side in this relation is zero, we have
∂β
<0
∂n x∈∂Ω
DIFFERENTIAL EQUATIONS Vol. 42 No. 5 2006
METHOD OF DIFFERENTIAL INEQUALITIES FOR STEP-LIKE CONTRAST . . . 747
for a sufficiently small k. In a similar way, one can show that the choice of a sufficiently large k
provides the validity of the corresponding inequality for the lower solution α on the boundary.
Let us show that a condition of the form (3.2) for the jump of the derivative holds on the line
Cβ on which β(x, ε) has a jump of the derivative. For the difference of the inward and outward
normal derivatives of the function β(x, ε) on the line Cβ , one can obtain
−1
∂ ∂ ∂ ũ ∂J
β − − β = (0, l) δ − RA ,
∂r + x∈Cβ ∂r x∈Cβ ∂η ∂r x∈C0
where
+∞
By Condition III,
∂J
< 0.
∂r x∈C0
∂ ũ
Therefore, by taking into account the inequality (0, l) > 0, one can choose δ large enough to
∂η
ensure that δJ (x0 ) − RA < 0, and hence β(x, ε) satisfies inequality (3.2) for sufficiently small ε.
In a similar way, one can verify the corresponding inequality for the inward and outward normal
derivatives of the lower solution α(x, ε) on the line Cα .
By Un (x, ε) we denote the nth partial sum of the asymptotic series (1.5) in which the right-hand
n+1
side of Eq. (1.7) for the curve C is replaced by the partial sum i=1 εi ri (l).
Theorem 3.2. Let Conditions I–VI be satisfied. Then for sufficiently small ε, there exists a
solution of problem (1.1), (1.2) satisfying the inequality
The proof of the theorem is similar to that of the lemma but uses a modification of terms of
order n + 1 in the asymptotics.
The proof of the Lyapunov asymptotic stability of the constructed solution of problem (1.1),
(1.2) lying in a small neighborhood of Un (x, ε), viewed as a stationary solution of the integro-
parabolic problem (1.8), (1.2), is performed by the method of asymptotic differential inequalities.
For the solution of the integro-parabolic problem (1.8), (1.2) with initial conditions lying between
the upper and lower solutions of the original problem (1.1), (1.2), we suggest the barrier functions
where us is a stationary solution of problem (1.8), (1.2), which exists by virtue of Theorem 3.2 and
satisfies inequality (3.5), and αn and βn are the lower and upper solutions of problem (1.1), (1.2)
with the nth-order accuracy. (In this notation, the upper and lower solutions α and β are denoted
by α0 and β0 .) The asymptotics of αn and βn is modified in terms of order n + 1 in the same way
as α and β are modified in the first order, and λ > 0 is a sufficiently small parameter depending
on ε. One can readily show that βn − αn = O(εn ), which, in turn, implies that
Obviously, inequalities similar to (3.2) and (3.3) are valid for α̃ and β̃, since they hold for αn
and βn . An inequality similar to (3.1) can be verified directly. Let us verify it for β̃. By substituting
β̃ into the operator (ε2 Δ − ∂/∂t − L), we obtain
ε2 Δβ̃ − ∂ β̃/∂t − L β̃, β̃, x, ε
By definition, the first term on the right-hand side in this relation is negative and has the small-
ness order O(εn+1 ), the second term is
positive and has the order O(εn ), and the difference between
the third and fourth terms is O ε2(n+1) . By setting λ = ελ0 , where λ0 > 0 is a sufficiently small pa-
rameter independent of ε, we find that the left-hand side of (3.6) is negative for sufficiently small ε.
Likewise, we obtain
ε2 Δα̃ − ∂ α̃/∂t − L (α̃, α̃, x, ε) > 0.
The preceding considerations imply the local uniqueness of the solution of problem (1.1), (1.2) in
a neighborhood of the asymptotics constructed above and the asymptotic stability of that solution
viewed as a stationary solution of problem (1.8), (1.2).
Theorem 3.3. If Conditions I–VI are satisfied, then for sufficiently small ε, there exists a locally
unique asymptotically stable stationary solution us of problem (1.8), (1.2) satisfying the inequality
4. EXAMPLE
By way of example, we consider the problem
2
ε Δu = f (u) + μ u(s)ds (4.1)
Ω
in the unit area disk with the homogeneous boundary conditions (1.2), where
u + 1 if u < −1/2
f (u) = −u if −1/2 < u < 1/2
u − 1 if u > 1/2,
and 0 < μ < 1/2 is a constant. The function f (u) has discontinuities of the derivative at the
point u = ±1/2. To satisfy Condition I, the function f (u) can be smoothed up to an arbitrary
smoothness degree in small (but independent of ε) neighborhoods of these points. However, this
is unnecessary for the construction of the zero-order asymptotics. For each closed curve C0 ≡ ∂Ωi ,
the degenerate equation has a discontinuous solution of the form
−1 + μ (1 − 2Si )/(1 + μ) for x ∈ Ω\Ωi
ϕ (x, C0 ) =
1 + μ (1 − 2Si )/(1 + μ) for x ∈ Ωi
and the continuous solution ϕ2 ≡ 0. One can readily see that all assumptions in Condition II are
satisfied. The function J (x0 ) can be computed in closed form and represented as
J (x, C0 ) = 2μ (2Si − 1)/(1 + μ).
Obviously, Condition III is valid for any domain Ωi such that Si = 1/2. For the degenerate solution
of zero order, we take
−1 for x ∈ Ω\Ωi
ϕ (x, C0 ) = (4.2)
1 for x ∈ Ωi .
DIFFERENTIAL EQUATIONS Vol. 42 No. 5 2006
METHOD OF DIFFERENTIAL INEQUALITIES FOR STEP-LIKE CONTRAST . . . 749
and has only the trivial solution for μ = −1. Condition IV is satisfied. Condition V is also obviously
valid in the case under consideration. Condition VI has the form 1 − μ > 0 in this case and also
holds.
Thus all assumptions of Theorem 3.2 are valid. Consequently, for each closed domain Ωi with
smooth boundary and area 1/2, there exists a step-like solution of problem (4.1), (1.2) with limit
form (4.2).
ACKNOWLEDGMENTS
The work was financially supported by the Russian Foundation for Basic Research (project
no. 04-01-00710).
REFERENCES
1. Nefedov, N.N. and Nikitin, A.G., Zh. Vychisl. Mat. Mat. Fiz., 2001, vol. 41, No. 7, pp. 1057–1066.
2. Nefedov, N.N. and Nikitin, A.G., Differ. Uravn., 2000, vol. 36, no. 10, pp. 1398–1404.
3. Nefedov, N.N., Differ. Uravn., 1995, vol. 31, no. 7, pp. 1132–1139.
4. Vasil’eva, A.B. and Butuzov, V.F., Asimptoticheskie metody v teorii singulyarnykh vozmushchenii
(Asymptotic Methods in Singular Perturbation Theory), Moscow: Nauka, 1990.
5. Pao, C.V., Nonlinear Parabolic and Elliptic Equations, New York: Springer, 1992.
SHORT COMMUNICATIONS
DOI: 10.1134/S0012266106050144
Let A be a pseudodifferential operator with symbol A(ξ), ξ = (ξ1 , ξ2 ), satisfying the ellipticity
condition
c1 ≤ A(ξ)(1 + |ξ|)−α ≤ c2 , α ∈ R, (1)
where c1 and c2 are positive constants.
Definition. A wave factorization of the elliptic symbol A(ξ) with respect to the cone
C+a = x = (x1 , x2 ) ∈ R2 : x2 > a |x1 | , a > 0
is a representation of the form A(ξ) = A= (ξ)A= (ξ), where the factors A= (ξ) and A= (ξ) have the
following properties:
(1◦ ) in general, A= (ξ) and A= (ξ) are defined only on the set {x ∈ R2 : a2 x22 = x21 };
∗
(2◦ ) A= (ξ) admits an analytic continuation into the radial tubular domain T C a+ with the cone
∗
C a+ = {x ∈ R2 : ax2 > |x1 |}, and this continuation satisfies the estimate
±1 ∗
A (ξ + iτ ) ≤ c(1 + |ξ| + |τ |)±κ ∀τ ∈ C a+ .
=
∗ ∗
The factor A= (ξ) must have a similar property with C a+ replaced by −C a+ and κ replaced by α − κ.
The number κ is referred to as the index of the wave factorization.
Consider the pseudodifferential equation
(Au)(x) = f (x), x ∈ C+a , (2)
with the boundary conditions
(B1j u)| x1 >0 = g1j , (B2j u)| x1 <0 = g2j (3)
x2 −ax1 =0 x2 +ax1 =0
750
ELLIPTIC PROBLEMS AND FINITE-DIFFERENCE EQUATIONS 751
The right-hand side f (x) of Eq. (2) is taken in the space H1s−α C+a of generalized functions
that belong to S C+a and admit a continuation f to R2 such that f ∈ H s−α (R2 ). The norm in
the space H1s−α C+a is given by the formula
where the infimum is taken over all continuations . The right-hand sides grj are taken in the spaces
H s−βrj −1/2 (R+ ), r = 1, 2, j = 0, 1, . . . , n − 1, in accordance with the theorem on the restriction to
a hyperplane [1, p. 40].
By [2, p. 74], the boundary value problem (2), (3) can be reduced with the use of the Fourier
transform to a system of 2n linear integral equations on the real line. It turns out that, by sim-
ple transformations with the subsequent use of the Mellin transform, one can reduce that sys-
(m) (m)
tem to a system of 4n linear finite-difference equations for the unknowns ck (s) and dk (s),
k = 0, 1, . . . , n − 1, m = 0, 1, of the form
n−1
ACKNOWLEDGMENTS
The work was financially supported by the Center of Basic Natural Sciences at St. Petersburg
State University (project no. E02-1.0-161). The first author was also supported by the program
“Universities of Russia” (project no. UR.04.01.023).
REFERENCES
1. Eskin, G.I., Kraevye zadachi dlya ellipticheskikh psevdodifferentsial’nykh uravnenii (Boundary Value
Problems for Elliptic Pseudodifferential Equations), Moscow, 1973.
2. Vasil’ev, V.B., Wave Factorization of Elliptic Symbols: Theory and Applications, Dordrecht, 2000.
SHORT COMMUNICATIONS
DOI: 10.1134/S0012266106050168
1. Bitsadze [1] was the first to study this nonclassical problem. The study was continued by
the present authors (e.g., see [2–5] as well as [6] and the bibliography therein).
The present paper deals with this problem for a parabolic pseudodifferential operator. Unlike [4],
we use a different class of pseudodifferential operators, which can act in the Sobolev spaces Hs,p
with p = 2.
2. By Hs (Rn ) = Hs we denote the Sobolev space with the norm
⎛ ⎞1/2
s
u ≡ us,Rn = ⎝ 1 + |ξ | |F u| dξ ⎠ ,
2 2
Rn
where q ∈ Q = {z ∈ C, α0 ≤ arg z ≤ β0 }. The spaces Hs,q (Ω) and Hs,q (∂Ω) are defined in a similar
way.
By Hs,γ (Rn+1 ) = Hs,γ we denote the Slobodetskii space with the norm
⎛ ⎞1/2
s
us,γ ≡ us,γ,Rn+1 = ⎝ 1 + |ξ | + |ξ0 | |F u|2 dξ ⎠ ,
2 2/γ
Rn+1
where ξ = (ξ0 , ξ ) ∈ Rn+1 , F = Fx→ξ is the Fourier transform, s is an integer multiple of γ, and
s ≥ 0 is an integer.
Just as above, we introduce the spaces Hs,γ (Ω∞ ) and Hs,γ (∂Ω∞ ), where
τ τ
The spaces Hs,γ (Ω∞ ) and Hs,γ (∂Ω∞ ) are defined in a natural way.
757
758 NGUYEN MINH CHUONG, TRAN TRI KIET
Let Es,γ,τ (Rn+1 ) be the space of functions U (q, x ) defined for almost all x ∈ Rn and q, Re q ≥ τ ,
such that the following conditions are satisfied:
(i) the inclusion U (q, x ) ∈ Hs,q (Rn ) is valid for all q with Re q > τ and for almost all q with
Re q = τ ;
(ii) the function U (q, x ) ≡ U (σ + iη, x ) is holomorphic for Re > τ for almost all x;
s
2 2 2
(iii) the norm
τ
is a topological isomorphism of the spaces Hs,γ (Rn+1 ) and Es,γ,τ (Rn+1 ).
3. Let r ∈ Z+ . Consider the operator
⎡ ⎤
1 ⎣
Au (q, x ) = n/2
qβ Kαβ (x , x − y ) Dyα u (y ) dy ⎦ ,
(2π)
|α|+γβ≤2r
where q ∈ Q, Kαβ (x , z ) = Hαβ (z ) + kαβ (x , z ), and the functions Hαβ and kαβ satisfy the
following conditions:
(a) Hαβ (z ) and kαβ (x , z ) are homogeneous functions of z of degree −n;
(b) kαβ (x , z ) ∈ C ∞ (Rn ), and Dxλ kαβ (x , z ) → 0 as |x | → ∞ for all λ;
(c) |z|=1 Hαβ (z ) dσ (z ) = 0 and |z|=1 kαβ (x , z ) dσ (z ) = 0;
2
(d) |z|=1 |Hαβ (z )| dσ (z ) < +∞, and R |z|=1 Dxλ kαβ (x , z ) dσ (z ) dx < +∞ for all λ.
2
Obviously, if conditions (a)–(d) are satisfied, then the operator A is defined on C0∞ (Rn ) and has
the form
Au (q, x ) = ei(x ,ξ ) σA (q, x , ξ ) F u (q, ξ ) dξ ,
Rn
where
σA (q, x , ξ ) = q β gαβ (x , ξ ) ξ α ,
|α|+β≤2r
gαβ (x , ξ ) = gαβ 0
(x , ξ ) + gαβ
1
(ξ ) ,
1
0
gαβ (x , ξ ) = n/2
e−i(x ,ξ ) kαβ (x , z ) dz ,
(2π)
Rn
1
1
gαβ (ξ ) = n/2
e−i(x ,ξ ) Hαβ (z ) dz .
(2π)
Rn
where Dν is a field that can be tangent to the boundary ∂Ω at points of Γ0 but is not tangent
to Γ0 , A is the above-defined operator, Bjk are pseudodifferential operators of order mj − k on the
boundary hyperplane Rn−1 with symbols σjk (x , ξ ), x = (x1 , x2 , . . . , xn−1 ), ξ = (ξ1 , ξ2 , . . . , ξn−1 ),
Dn is the derivative along the inward normal n to ∂Ω, and r is an integer multiple of γ ≥ 1. Here
we use the classification of manifolds Γ0 given in [3]. If Γ0 belongs to the first class, then we add
the conditions
Dnk u (q, x ) = u0,k (q, x ) , k = 0, 1, . . . , r − 1, on Γ0 . (3)
We assume that (A, Bj ) is an elliptic problem and A and Bj are admissible operators; i.e., say,
2r
σA (q, x , 0, ξ , ξn ) = σAj,k (x , ξ ) ξn−1
k
qj ,
k+j=0
where N is the manifold given by the equation z 1 = 0 in the local coordinate system [3]. By Gs,q (∂Ω)
we denote the space of functions defined on ∂Ω with finite norm
The spaces Πs,γ (Ω∞ ) and Πs,γ (∂Ω∞ ) are defined in a similar way with the use of the norms of the
τ τ
spaces Hs,γ (Ω∞ ) and Hs,γ (∂Ω∞ ), respectively.
The following assertion can be proved in the same way as in [8].
then for sufficiently large |q|, problem (1)–(3) has a unique solution u ∈ Πs,q (Ω). If Γ0 belongs to
the third class, f ∈ Πs−2r,q (Ω), and gj ∈ Gs−mj −1/2,q (∂Ω), then for sufficiently large |q|, problem (1),
(2) has a unique solution u ∈ Πs,q (Ω).
mj
where Aαβ are the pseudodifferential operators in Rn defined in item 3, Bjkβ are pseudodifferential
operators of order mj − k, γ ≥ 1, with symbols σjkβ (x , ξ ), and [m] is the integer part of a
number m. If Γ0 belongs to the first class, then we add the conditions
Obviously, the Laplace transform reduces problem (1 ), (2 ), (2 ), (3 ) to (1)–(3). Following the
same scheme of proof as in [8], one can justify the following assertion.
Theorem 3. Let
all above-mentioned conditions be satisfied. Then for sufficiently large τ, the
operators A, Bj , Dnk determined by problem (1 ), (2 ), (3 ) specify a continuous one-to-one mapping
of the space Πs,γ (Ω∞ ) either onto the space
REFERENCES
1. Bitsadze, A.V., Dokl. Akad. Nauk , 1963, vol. 148, no. 4, pp. 749–752.
2. Borrelli, R., J. Math. Mech., 1966, vol. 16, pp. 51–81.
3. Egorov, Yu.V. and Kondrat’ev, V.A., Mat. Sb., 1969, vol. 78, no. 1, pp. 148–176.
4. Egorov, Yu.V. and Nguyen Minh Chuong, Uspekhi Mat. Nauk , 1998, vol. 53, no. 6, pp. 249–250.
5. Nguyen Minh Chuong and Tran Tri Kiet, On a Priori Estimates in Lp (Rn ), p ≥ 2, for Non-Classical
Oblique Derivative Problem, Preprint / Institute of mathematics, Hanoi, 2000, no. 16.
6. Le Quang Trung, Uspekhi Mat. Nauk , 1989, vol. 44, no. 5, pp. 169–170.
7. Nguyen Minh Chuong, Dokl. Akad. Nauk , 1983, vol. 121(163), no. 1(5), pp. 3–7.
8. Nguyen Minh Chuong, Mat. Zametki, 1984, vol. 35, no. 2, pp. 321–329.
SHORT COMMUNICATIONS
DOI: 10.1134/S001226610605017X
Theorem 1. The minimum number of frequencies needed for plant identification is given by
the formula
(ν + 1)m + n
min = , (4)
2m
where ]x[ is the minimum integer greater than or equal to x, ν is the plant observability index [1],
that is, the minimum number such that rank Oν−1 = rank Oν ; here
T T T T
Oβ = C A C . . . AβT C T .
Theorem 1 is stated under the assumption that the observability index ν and the order n of
the plant are known. If this is not the case, they can be found on the basis of the frequency
parameters (2) of the plant (1). To this end, using the values (2) and (3), we define the matrices
Fα = (I ΩI . . . Ωα I) , Gβ = H ΩH . . . Ωβ H ,
761
762 ORLOV
where
0 −1
Ω = [λE2 + diag (ω1 , ω2 , . . . , ω ) ⊗ J] ⊗ Em , J= ,
1 0
T
I = (Em 0m Em 0m . . . Em 0m ) ,
T
H = (−Φ1 − Ψ1 − Φ2 − Ψ2 . . . − Φ − Ψ ) , ≥ min .
Assertion 1. The minimum number ν such that
rank (Fν | Gν−1 ) = rank (Fν | Gν ) (5)
is the plant observability index.
Assertion 1 permits one to find ν by verifying condition (5). As a result, one forms the block
matrix (Fν | Gν ).
Assertion 2. The order of the plant is given by the formula n = rank (Fν | Gν ) − rank Fν .
Assertions 1 and 2 permit one to find the minimum number (4) of frequencies necessary for the
identification of a plant whose observability index ν and order n are unknown.
To proof the assertions, we make the following decompositions.
Lemma 1 (a modified F G-decomposition). One has
(Fα | Gβ ) = N F2−1 P̄α | Q̄β ,
where
N = diag (N1 , N2 , . . . , N ) ,
Nk = E2 ⊗ Re P̄ −T (sk ) + J ⊗ Im P̄ −T (sk ) , k = 1, . . . , ;
P̄α = P̄α,0 P̄α,1 . . . P̄α,α ,
T
P̄α,i = 0m,im P̄ [0] P̄ [1] . . . P̄ [μ] 0m,(2−μ−i−1)m , i = 0, . . . , α,
Q̄β = −Q̄β,0 − Q̄β,1 . . . − Q̄β,β ,
T
Q̄β,i = 0r,im Q̄[0] Q̄[1] . . . Q̄[μ] 0r,(2−μ−i−1)m , i = 0, . . . , β.
Remark 1. The eliminating 2m × (α + 1)m + (β + 1)r matrix P̄α | Q̄β has been constructed
from the blocks of the description [equivalent to (1)]
P̄ (s)z = u, y = Q̄(s)z, t ≥ t0 ,
of the plant (z ∈ Rm is an auxiliary state), whose matrix polynomials
P̄ (s) = P̄ [0] + P̄ [1] s + · · · + P̄ [μ] sμ , Q̄(s) = Q̄[0] + Q̄[1] s + · · · + Q̄[μ] sμ
have the special structure
[0] [1] [μ −1] μi −1
p̄ii (s) = p̄ii + p̄ii s + · · · + p̄ii i s + sμi , i = 1, . . . , m,
[μ̌ −1]
p̄ij ij sμ̌ij −1 ,
[0] [1]
p̄ij (s) = p̄ij + p̄ij s + ··· + i = j, i = 1, . . . , m, j = 1, . . . , m,
[0] [1] [μ ]
q̄ij (s) = q̄ij + q̄ij s + · · · + q̄ij j sμj , i = 1, . . . , r, j = 1, . . . , m,
where μ̌ii = μi , μ̌ij = min (μi , μj + 1) for i < j, μ̌ij = min (μi , μj ) for i > j, the μi , i = 1, . . . , m,
are the Kronecker (controllability) indices [1] of the plant, that is, the minimum numbers such that
the columns Aμi coli B of the controllability matrix
C = col1 B col2 B . . . colm B | A col1 B A col2 B . . . A colm B | . . . |
An−1 col1 B An−1 col2 B . . . An−1 colm B
can be linearly expressed via the preceding columns of this matrix, and μ = max (μ1 , μ2 , . . . , μm ).
where
⎛ ⎞
ΣT
⎜ 0(α+1)m,μm−n ⎟
⎜ colμm+1:(α+μ+1)m ΠT 0(α+μ+1)m,(2−α−μ−1)m ⎟
Lα = ⎜
⎜ 0(α+1)m,n
⎟,
⎟
⎝ ΣT ⎠
0(2−α−μ−1)m,(α+μ+1)m E(2−α−μ−1)m
⎛ ⎞
Σ row1:μm P̄α
⎜ ⎟
⎜ 0μm−n,(α+1)m ⎟ T
Γα = ⎜ ⎟, Θβ =
Ōβ ,
⎜ ⎟
⎝ Δα ⎠ 02m−n,(β+1)r
0(2−α−μ−1)m,(α+1)m
⎛ ⎞
Tβ
⎜ E(α+1)m ⎟
Rα,β = ⎝ 0(α−β)m,(β+1)r ⎠ .
0(β+1)r,(α+1)m −E(β+1)r
Here
ΣT = (Σ1 Σ2 . . . Σm ) , ΣT = (Σ1 Σ2 . . . Σm ) ,
Σi = coli Eμm coli+m Eμm coli+2m Eμm . . . coli+(μi −1)m Eμm ,
Σi = coli+μi m Eμm coli+(μi +1)m Eμm . . . coli+(μ−1)m Eμm , i = 1, . . . , m;
coli ΠT = col[i−1+(α+μ[i−1] mod m+1 +1)m]mod(α+μ+1)m+1 E(α+μ+1)m , i = 1, . . . , (α + μ + 1)m;
min(α,μ)
ŌβT = C̄ T ĀT C̄ T . . . ĀβT C̄ T ; Δα = i
Iα+1 ⊗ P̄ {μ−i}T ,
i=0
β
Tβ = i
Iβ+1 ⊗ Q̄{μ}(i)T ,
i=0
where
0α,1 Eα
Iα+1 = ; Q̄{μ}(i)T = P̄ −{μ}T Q̄{μ−1}(i−1)T ,
0 01,α
Q̄{j}(i)T = Q̄{j−1}(i−1)T − P̄ {j}T Q̄{μ}(i)T ,
Q̄{0}(i)T = −P̄ {0}T Q̄{μ}(i)T , j = 1, . . . , μ − 1, i = 1, . . . , β;
{μ}(0)T −{μ}T {μ}T {j}(0)T {j}T
Q̄ = P̄ Q̄ and Q̄ = Q̄ − P̄ {j}T Q̄{μ}(0)T , j = 0, . . . , μ − 1.
The matrices P̄ {i} and Q̄{i} , i = 0, . . . , μ, are obtained by the multiplication of P̄ (s) and Q̄(s) by
S̄μ∗ (s) = diag sμ−μ1 , sμ−μ2 , . . . , sμ−μm :
P̄ (s)S̄μ∗ (s) = P̄ {0} + P̄ {1} s + · · · + P̄ {μ} sμ , Q̄(s)S̄μ∗ (s) = Q̄{0} + Q̄{1} s + · · · + Q̄{μ} sμ .
of the plant with the blocks Āij and b̄ij , i = 1, . . . , m, j = 1, . . . , m, of the matrices Ā and B̄ having
the following special structure: Āii is a matrix with unit superdiagonal,
[0] [1] [μ −1]
rowμi Āii = −āii − āii . . . − āii i ,
and the remaining entries are zero; Āij (i = j) is a matrix with the last row
[0] [1] [μ −1]
rowμi Āij = −āij − āij . . . − āij ij 0 ... 0 , μij = min (μi , μj ) ,
T
and the remaining entries are zero; b̄ii = [0 . . . 0 1]T , b̄ij = 0 . . . 0 − b̄ij for i < j, and
[0] [1] [μ −1]
b̄ij = [0 . . . 0]T for i > j. The matrix C̄ consists of the blocks c̄ij = c̄ij c̄ij . . . c̄ij j ,
i = 1, . . . , r, j = 1, . . . , m, and D has no special structure.
Proofof Assertion 1. The nondegeneracy of the matrices N (which is block diagonal,
det N = k=1 det Nk = 0, since det P̄ −T (sk ) = 0, k = 1, . . . , ), F2−1 [which is a block matrix of
Vandermonde structure, det F2−1 = 0, by (3)], Lα (which is a permutation matrix, det Lα = ±1),
and Rα,β (which is a triangular block Toepliz matrix, det Rα,β = ±1) provides the validity of the
chain of relations
rank (Fα | Gβ ) = rank N F2−1 P̄α | Q̄β = rank P̄α | Q̄β
= rank [Lα (Γα | Θβ ) Rα,β ] = rank (Γα | Θβ ) (7)
= (α + 1)m + rank Oβ , α ≥ β.
The last equality in (7) is valid since the ((α + 1)m × (α + 1)m) block of Δα is nondegenerate:
rank (Γα | Θβ ) = rank Δα + rank Ōβ , where rank Δα = (α + 1)m and rank Ōβ = rank Oβ .
The relation rank Oν−1 = rank Oν implies that rank (Γν | Θν−1 ) = rank (Γν | Θν ), and accord-
ingly,
rank (Fν | Gν−1 ) = rank (Fν | Gν ) .
On the other hand, the inequality rank Oν−2 < rank Oν−1 implies that
rank (Γν−1 | Θν−2 ) < rank (Γν−1 | Θν−1 ) ,
and accordingly, rank (Fν−1 | Gν−2 ) < rank (Fν−1 | Gν−1 ), which completes the proof of Assertion 1.
Proof of Assertion 2. Let us write out a chain of relations similar to (7),
rank Fα = rank N F2−1 P̄α = rank P̄α = rank (Lα Γα ) = rank Γα = rank Δα = (α + 1)m,
and subtract it from (7): rank Oβ = rank (Fα | Gβ ) − rank Fα , α ≥ β. If α = β = ν, then
rank Oν = n = rank (Fν | Gν ) − rank Fν .
Proof of Theorem 1. The entries of the matrices A, B, C, and D of the canonical form (1)
are implicitly determined by the system of frequency identification equations [2]
m
νk
[j]
r ν̌
ki −1
[j]
Ω j
ii qki + Ωj hi pki = −Ωνk hk , k = 1, . . . , r, (8)
i=1 j=0 i=1 j=0
can be linearly expressed via the preceding (upper) rows of this matrix. The matrix of system (8)
consists of (ν + 1)m + n linearly independent columns of the matrix (Fν | Gν ). Their number is
minimal in system (8), since ν = max (ν1 , ν2 , . . . , νr ). The number 2m of rows of the matrix of
system (8) should be not less than the number (ν + 1)m + n of columns, 2m ≥ (ν + 1)m + n,
which readily implies the estimate (4).
Remark 3. By duality, our results can readily be generalized to plant identification in the
canonical form (6). For example, the minimum number of frequencies necessary for plant identifi-
cation is equal to ¯min = ]((μ + 1)r + n)/(2r)[.
REFERENCES
1. Kailath, T., Linear Systems, Englewood Cliffs: Prentice-Hall, 1980.
2. Orlov, Yu.F., Differ. Uravn., 2006, vol. 42, no. 3, pp. 425–428.
SHORT COMMUNICATIONS
DOI: 10.1134/S0012266106050181
In the present paper, we develop the research in [1–4] and consider a boundary value problem
(which is the adjoint of the problem considered in [2]) in the vertical half-strip
{0 < x < 1, 0 < y < ∞}
with nonlocal boundary conditions for two degenerate elliptic equations. We prove the unique solv-
ability of the nonlocal boundary value problem and obtain necessary conditions for the unique
solvability. The paper consists of two sections, which correspond to two different equations.
1. Consider the degenerate elliptic equation
y m uxx + uyy − b2 y m u = 0, 0 < x < 1, y > 0, m > −2, (1)
with the boundary conditions
ux (1, y) = ux (0, y), u(1, y) = 0, y > 0, (2)
and the condition
u(x, 0) = f (x), f ∈ C 2+α [0, 1], f (0) = f (1), f (1) = 0. (3)
If b > 0, then the following assertion is valid.
Theorem 1. Problem (1)–(3) has a unique solution in the class of functions u ∈ C 0 ([0, 1] ×
[0, ∞]) ∩ C 2 ((0, 1) × (0, ∞)) tending to zero at infinity. The solution can be represented by the series
∞
∞
u(x, y) = 2(1 − x)u0 (y) + un (y) × 4(1 − x) cos(2πnx) + vn (y) × 4 sin(2πnx), (4)
n=1 n=1
where
1/(2q) 1
−1 1 b √
u0 (y) = 2Γ f (x)dx y K1/(2q) by q q −1 , (5)
2q 2q
0
1/(2q) 1
1 Φn √
un (y) = 2Γ−1 f (x) cos(2πnx)dx y K1/(2q) Φn y q q −1 , (6)
2q 2q
0
1/(2q) 1
1 Φn √
vn (y) = 2Γ−1 f (x) cos(2πnx)dx y y q q −1 K1/(2q) Φn y q q −1
2q 2q
0
1/(2q) 1
1 Φn √
+ 2Γ−1 f (x) sin(2πnx)dx y K1/(2q) Φn y q q −1 , (7)
2q 2q
0
q = (m + 2)/2, Φn = (2πn)2 + b2 . (8)
766
THE SOLUTION OF A NONLOCAL BOUNDARY VALUE PROBLEM 767
1
un (y) = u(x, y) cos(2πnx)dx,
0
1
vn (y) = u(x, y)x sin(2πnx)dx, n = 1, 2, . . . ,
0
1
u0 (y) = u(x, y)dx.
0
The function u(x, y) is continuously differentiable arbitrarily many times for y > 0 and x ∈ [0, 1] by
virtue of conditions (2). Therefore, by twice differentiating the functions un (y), vn (y), and u0 (y)
in the integrand and by taking into account (1)–(3), we obtain the boundary value problems
1
un (y) −y m 2
(2πn) + b 2
un (y) = 0, un (0) = f (x) cos(2πnx)dx for un , (11)
0
1
vn (y) −y m 2
b + (2πn) 2
vn (y) = −4πny un (y),
m
vn (0) = f (x)x sin(2πnx)dx for vn , (12)
0
Problems (11), (12) and (13), (14) have the unique solutions (6), (7), and (5), respectively [5].
It readily follows from (5)–(8) that the solution is unique, since if f (x) = 0 on [0, 1], then un (y) = 0
for n = 0, 1, . . . and vn (y) = 0 for n = 1, 2, . . . on (0, ∞). Consequently, since system (9) is complete,
we have u(x, y) = 0 for all x ∈ [0, 1] and y ∈ (0, ∞). The proof of the uniqueness of the solution is
complete.
Let us now study the existence. The solution u(x, y) can be represented by the biorthogonal
series (4); moreover, this series converges in L2 (0, 1) for each y ∈ (0, ∞). Therefore, it is obvious
that, for y > 0, the series (4), together with all derivatives, converges uniformly, since the function
K1/(2q) (Φn y q q −1 ) is decreasing as y → ∞; therefore, the series (4) satisfies Eq. (1) for y > 0 and
the boundary condition (2). To prove that condition (3) holds, it suffices to justify the uniform
convergence of the series (4) as y ≥ 0, which will readily imply the continuity of u(x, y) in the
closed domain x ∈ [0, 1], y ∈ [0, ∞).
Let us now estimate the Fourier coefficients in (5)–(7) from above:
1 1 1
f (x) f
(x) C
f (x) cos(2πnx)dx = sin(2πnx)dx = cos(2πnx)dx ≤ 2+α . (15)
2πn 2
(2πn) n
0 0 0
The last inequality follows from the inclusion f (x) ∈ C[0, 1] and the decay rate of the Fourier
coefficients in the trigonometric system. In a similar way, one can estimate another Fourier coeffi-
cient:
1 1
f (x)x + f (x)
f (x)x sin(2πnx)dx = cos(2πnx)dx
2πn
0
1
0
(16)
sin(2πnx) C
= 2
(f (x)x + 2f (x)) dx ≤ 2+α .
(2πn) n
0
and
q q 1/(2q)
q −1
K1/(2q−1) Φn y q y (y n)
|1/(2q−1)| −1 −q |1/(2q−1)| q q 1/(2q)
= K1/(2q−1) Φn y q q −1 Φn y q q −1 qΦn y y (y n)
⎧ ⎫ (18)
⎪
⎨ ny 2q , 1/(2q) ≥ 1 ⎪ ⎬
≤ C[| ln ny| + 1] y q
(y q
n)
1/q ≤ C[| ln ny| + 1]n
⎪
⎩ , 1/(2q) < 1 ⎪ ⎭
ny q
as n → ∞. The estimates (14)–(17) guarantee the uniform convergence of the series (4) in the
closed domain x ∈ [0, 1], y ∈ [0, T ] for each T > 0. The proof of the theorem is complete.
Remark 1. If b = 0, then for the unique solvability of problem (1)–(3), it is necessary that
1
f (x)dx = 0. (19)
0
Remark 2. If the solution is sought in the class of bounded functions, then condition (19) is
unnecessary for the unique solvability of problem (1)–(3) even for b = 0.
lim y 2p u(x, y) = f (x), f ∈ C 2+α [0, 1], f (0) = f (1), f (1) = 0. (21)
y→+0
Theorem 2. Problem (20), (2), (21) has a unique solution in the class of functions
tending to zero at infinity. The solution can be represented by the series (4), where
p+1/2 1
−1 1 b
u0 (y) = y 1/2−p
× 2Γ f (x)dx K|1/2+p| (by),
2 2
0
p+1/2 1
1 Φn
un (y) = 2Γ−1 p + f (x) cos(2πnx)dx y 1/2−p K|p+1/2| (Φn y) ,
2 2
0
p+1/2 1
3/2−p −1 1 Φn
vn (y) = 2πny 2Γ p+ f (x) cos(2πnx)dx K|p−1/2| (Φn y)
2 2
0
p+1/2 1
1 Φn
+ y 1/2−p × 2Γ−1 p + f (x)x sin(2πnx)dx K|p+1/2| (Φn y) ,
2 2
0
Proof. The general scheme of the proof of Theorem 2 is similar to the proof of Theorem 1.
Remark 3. If p = 0, then Remarks 1 and 2 apply.
REFERENCES
1. Moiseev, E.I., Differ. Uravn., 1999, vol. 35, no. 8, pp. 1094–1100.
2. Moiseev, E.I., Differ. Uravn., 2001, vol. 37, no. 11, pp. 1565–1567.
3. Lerner, M.E. and Repin, O.A., Differ. Uravn., 1999, vol. 35, no. 8, pp. 1087–1093.
4. Lerner, M.E. and Repin, O.A., Differ. Uravn., 2001, vol. 37, no. 11, pp. 1562–1564.
5. Erdélyi, A., Magnus, W., Oberhettinger, F., and Tricomi, F.G., Higher Transcendental Functions
(Bateman Manuscript Project), New York: McGraw-Hill, 1953, vol. II. Translated under the title Vysshie
transtsendentnye funktsii, Moscow: Nauka, 1974, vol. 2.