ISSN 0012-2661, Differential Equations, 2006, Vol. 42, No. 7, pp. 954–966.

c Pleiades Publishing, Inc., 2006.
Original Russian Text

c V.B. Andreev, 2006, published in Differentsial’nye Uravneniya, 2006, Vol. 42, No. 7, pp. 895–906.

NUMERICAL METHODS

On the Accuracy of Grid Approximations

to Nonsmooth Solutions of a Singularly Perturbed
Reaction–Diffusion Equation in the Square
V. B. Andreev
Moscow State University, Moscow, Russia
Received March 13, 2006

DOI: 10.1134/S0012266106070044

In the present paper, we consider the Dirichlet problem for a singularly perturbed reaction–
diffusion equation in the unit square. We assume that the coefficient of the equation, its right-
hand side, and the boundary values on the sides of the square are sufficiently smooth functions.
No compatibility conditions are assumed to be satisfied at the corners. Under these assumptions,
the desired solution in the closed square has a very restricted smoothness; namely, it belongs to the
space C 1,λ for λ ∈ (0, 1), so that its second derivatives are not necessarily bounded.
In the considered domain, we introduce a piecewise uniform condensing Shishkin mesh. For the
numerical solution, we use the classical five-point approximation. We show that the approximate
solution

−2 is4 uniformly
 with respect to a small parameter convergent in the Lh∞ -norm at the rate
O N ln N , where N is the number of grid nodes in each coordinate direction.
It is known that the error of a grid solution on a uniform mesh for the Poisson equation under
similar assumptions on the solution smoothness is estimated in Lh∞ as O (h2 ln(1/h)). For our
equation and scheme, the best estimate O (N −2 ) was proved only under the assumption of much

higher smoothness of the desired solution; namely, the solution should belong to the space C 4,λ Ω̄ ,
which is possible only in the case of validity of second-order compatibility conditions

at
 the corners.
In the case of the smoothness considered here, the best known estimate is O N −1/4 .
The results are obtained with the use of a new a priori estimate of the grid solution, which is
derived with the use of the grid Green function in the present paper.

1. INTRODUCTION
Consider the two-dimensional Dirichlet problem for the differential reaction–diffusion equation

L u := −ε2 ∆u + q(x, y)u = f (x, y), (x, y) ∈ Ω, (1.1)


4
u = g(x, y), (x, y) ∈ ∂Ω ≡ Γ = Γk , (1.2)
k=1

where the diffusion coefficient ε2 ∈ (0, 1] is constant and can take arbitrarily small values, Ω = (0, 1)2
is the unit square, and Γk = Γ̄k are its sides numbered in the counterclockwise sense starting from
Γ1 = {(x, y) ∈ Γ | x = 0}. The vertices of the square are denoted by ak , k = 1, 2, 3, 4, and are
numbered in a similar way; moreover, a1 = (0, 0). We assume that the reaction coefficient q(x, y)
satisfies the condition
q(x, y) ≥ 2α2 = const > 0. (1.3)
It is well known (e.g., see [1, p. 65]) that, for small ε, problem (1.1)–(1.3) has an exponential
boundary layer of width O(ε) in a neighborhood of the boundary Γ, which causes certain difficulties
in the numerical solution of this problem. In particular, the numerical solution on a uniform mesh
is a poor approximation to the solution of problem (1.1)–(1.3), and a mesh condensing in the
boundary layer is desired.

954
 ON THE ACCURACY OF GRID APPROXIMATIONS TO NONSMOOTH SOLUTIONS 955

By [2], the presence of corners on the boundary of the domain (the vertices of the square in the
case under consideration) deteriorates the smoothness of the solution and often results in reduced
accuracy of the approximate solution [3].
We assume that the coefficient q(x, y) of Eq. (1.1) and the right-hand side f (x, y) belong to the
Hölder class C 4,λ , 0 < λ < 1, on the closure Ω̄ of the domain Ω :


q(x, y), f (x, y) ∈ C 4,λ Ω̄ . (1.4)

We assume the corresponding smoothness of the restriction of the boundary function g(x, y) to the
sides of the square Γk (recall that, by definition, Γk is a closed interval, i.e., Γk = Γ̄k ) :

gk (s) := g(x, y)|Γk ∈ C 4,λ (Γk ) , k = 1, 2, 3, 4. (1.5)

If the boundary ∂Ω were sufficiently smooth and the function g(x, y) were class C 4,λ on the entire
boundary, then under condition (1.4) (and even under much less restrictive conditions), the solution

 (1.1)–(1.3) would satisfy [4, p. 113 of the Russian translation] the inclusion u(x, y) ∈
of problem
C 4,λ Ω̄ , which is sufficient for justifying the second-order convergence of classical finite-difference
schemes approximating Eq. (1.1) for ε = O(1) on the uniform mesh. In the case of a domain with
corners, there is no straightforward analog of the condition g(x, y) ∈ C 4,λ (∂Ω). But if

g(x, y) ∈ C(∂Ω), (1.6)

then it only follows from the inclusions (1.4) and (1.5) that



u(x, y) ∈ C 1,λ Ω̄ ∩ C 6,λ (Ω), (1.7)


where λ ∈ (0, 1) is an arbitrary number; moreover, in general, u(x, y) ∈ C 2 Ω̄ . To obtain a
higher smoothness of a solution of problem (1.1)–(1.5) than (1.7), one should subject the coeffi-
cients and right-hand sides of the equation and boundary conditions to the so-called
 compatibility
3,λ

conditions [2, 5] at the corners of the boundary. For example, to obtain u ∈ C Ω̄ , in addition to
the continuity of the boundary function at the corners, one should require the validity of Eq. (1.1)
at the corners. For the corner a1 , this condition has the form

κ1,1 := −ε2 (g1 (0) + g2 (0)) + q(0, 0)g(0, 0) − f (0, 0) = 0, (1.8)

where the prime stands for the differentiation with respect to the arc length measured on ∂Ω from
the point (0, 0) in the counterclockwise sense.
Remark 1.1. In fact, relation (1.8) contains two conditions. Since it is not assumed that the
functions q(x, y), f (x, y), and g(x, y) depend on ε and relation (1.8) contains the parameter ε,
it follows that the conditions

g1 (0) + g2 (0) = 0, q(0, 0)g(0, 0) = f (0, 0)

should be satisfied.


To increase the smoothness of the function u(x, y) to C 4,λ Ω̄ , one needs second-order compat-
ibility conditions: for example, at the point a1 , these are
 
κ1,2 := ε4 g1IV (0) − g2IV (0)
 2 
∂ ∂2 (1.9)
+ ε2 2
(−f + qg 2 ) (0, 0) − 2
(−f + qg1 ) (0, 0) = 0.
∂x ∂y

Numerous publications deal with numerical methods for problem (1.1), (1.2) and similar prob-
lems. Here we are interested only in methods convergent uniformly with respect to ε in the grid
norm Lh∞ . We do not consider fitting schemes and one-dimensional problems.

DIFFERENTIAL EQUATIONS Vol. 42 No. 7 2006

956 ANDREEV

Problem

 (1.1)–(1.3) was considered in [6] on a Bakhvalov mesh [7] under the assumption u ∈
C 4,λ Ω̄ , and the estimate  h 

uij − u (xi , yj ) = O N −2
uniform with respect to ε was proved there.
It was also shown there that if the smoothness is
diminished to (1.7), then the estimate is O N −2/11 .
A piecewise uniform mesh (a Shishkin mesh) was used in [1] for the solution of problem
(1.1)–(1.3), the estimate O (N −1 ln N ) was obtained for a sufficient smoothness of the solution,

1/4
and it was shown that if there is no compatibility condition, then the estimate is O (N −1 ln N ) .


Problem (1.1)–(1.3) with u ∈ C 4,λ
Ω̄ was considered
in [8] on a piecewise uniform mesh with
−1
2 
several points of increment change, and the estimate O N ln
ln
. . . ln N was obtained.

 k


The estimate O N −2 ln2 N was obtained in [9] on a Shishkin mesh for u ∈ C 4,λ Ω̄ .
Note also the paper [10], where problem (1.1)–(1.3) was considered in an L-shaped domain (nat-
urally, without the compatibility
conditions,
 since, in this case, the main compatibility condition
is nonlocal). The estimate O N −2/11+δ , δ > 0, was obtained for the approximate solution of this
problem on a piecewise uniform mesh.
Finally, note Bakhvalov’s classical paper [7]. It was the first paper where the use of condensing
meshes for the numerical solution of singularly perturbed equations was justified. This paper is
often cited in connection with a smoothly condensing mesh introduced there and nowadays referred
to as a Bakhvalov mesh. However, in addition to the construction of the mesh, the approximate
solution of the equation ε2 ∂ 2 u/∂x2 + ∂ 2 u/∂y 2 = f (x, y) was investigated there in the square with
the Dirichlet boundary conditions and without compatibility conditions. It was shown there that
the approximate solution (on the Bakhvalov mesh) has the accuracy O (N −2 ln N ). This result was
obtained for the case in which the assumptions on the smoothness of the right-hand side and the
boundary conditions are much weaker than (1.4), (1.5), namely, C 1,1 .
The aim of the present paper is to prove the uniform (with respect to ε) convergence of the
classical five-point finite-difference scheme on a Shishkin piecewise uniform mesh in Lh∞ with an
almost second order for problem (1.1)–(1.3) under assumptions (1.4)–(1.6), i.e., without compat-
ibility conditions at the corners. From the viewpoint of the convergence rate and the absence of
compatibility conditions, this result is close to [7], which deals with more restrictive conditions (1.4)
and (1.5) on the input data and considers a different equation. In the derivation of the convergence
rate, we substantially use some results in [9], which prove to remain valid without compatibility
conditions. The behavior of the solution in a neighborhood of the corners was partly investigated
in [5], and we use these results when estimating the truncation error of our finite-difference scheme.
To adequately take into account the behavior of the above-mentioned error in a neighborhood of
the corners, where it has power-law singularities, we prove a new a priori estimate for the grid
solution in Lh∞ via the Lh1 -norm of the right-hand side. This a priori estimate is justified with the
use of a grid Green function, for which the desired estimate is obtained.
In Section 2, we introduce a Shishkin piecewise uniform mesh and a grid approximation to
the differential problem. Then we represent and refine earlier obtained results on the structure
of the solution in a neighborhood of the corners and estimates for the closeness between various
components of the exact and grid solutions on some subsets of grid nodes. In Section 3, we introduce
the Green function of the auxiliary grid problem, obtain some estimates for it, and prove a new
a priori estimate

−2for 4thegrid solution. We prove the main theorem on the uniform (with respect to ε)
accuracy O N ln N of the grid solution. In Section 4, we represent and discuss numerical
results.
Throughout the following, by c we denote constants independent of the number of grid nodes
and the small parameter.

2. STATEMENT OF THE PROBLEM. EARLIER RESULTS

In the domain Ω̄, we introduce the piecewise uniform Shishkin mesh
Ω̄h = ω̄ (xi ) × ω̄ (yj ) , (2.1)

DIFFERENTIAL EQUATIONS Vol. 42 No. 7 2006

 ON THE ACCURACY OF GRID APPROXIMATIONS TO NONSMOOTH SOLUTIONS 957

where ω̄ is a one-dimensional piecewise uniform mesh condensing near the endpoints of the interval
[0, 1] and given by the relation
ω̄ (si ) = {si = ξ(i/N ), i = 0, 1, . . . , N } , (2.2)
and
  
4σt if 0 ≤ t ≤ 1/4 1 2ε
ξ(t) = σ + 2(1 − 2σ)(t − 1/4) if 1/4 ≤ t ≤ 3/4 σ = min , ln N . (2.3)
1 − 4σ(1 − t) if 3/4 ≤ t ≤ 1, 4 α

By hi = si − si−1 , i = 1, . . . , N , we denote the increments of the nonuniform mesh ω̄ (si ), and

by i = (hi + hi+1 )/2, i = 1, . . . , N − 1, we denote its mean increments. From (2.3) with N being
a multiple of 4, we obtain

h = 4σ/N for i = 1, . . . , N/4, 3N/4 + 1, . . . , N
hi =
H = 2(1 − 2σ)/N for i = N/4 + 1, . . . , 3N/4.

Let Ω1 = (0, σ)2 . By Ω̄h1 := Ω̄h ∩ Ω̄1 we denote the subset of nodes of the mesh Ω̄h in a
neighborhood of the vertex a1 , where Ω̄1 is the closure of Ω1 . Let Ωh1 := Ω̄h1 ∩ Ω1 be the set of
interior nodes of Ω̄h1 , and let ∂Ωh1 = Ω̄h1 \Ωh1 be its boundary. Obviously, Ωh1 is a uniform mesh with
increment h. In a similar way, we denote meshes in neighborhoods of the other vertices ak .
For grid functions defined on the mesh Ω̄h or on a subset Ω̃h , we introduce the norm
 h  h
vij  h h := max vij .
L (Ω̃ ) ∞ (xi ,yj )∈Ω̃h

We approximate problem (1.1), (1.2) on the mesh Ω̄h by the classical difference scheme


L h uh := −ε2 uhx̄x̂ + uhȳŷ ij + qij uhij = fij , (xi , yj ) ∈ Ωh = Ω̄h ∩ Ω,
(2.4)
uhij = g (xi , yj ) , (xi , yj ) ∈ Γh = Ω̄h ∩ Γ,
where, as usual, vx̄,ij := (vij − vi−1 j )/hi , vx̂,ij := (vi+1 j − vij )/i , and so on. In addition, let
vx,ij := (vi+1 j − vij )/hi+1 , and so on.
Our aim is to estimate the rate of convergence of the solution of problem (2.1)–(2.4) to the solu-
tion of problem (1.1)–(1.6). To this end, we need some information on the behavior of the solution
of problem (1.1)–(1.6) in neighborhoods of the vertices ak , k = 1, 2, 3, 4, of the domain Ω and on
the asymptotic properties of the solution as ε → 0.
A formal answer to this problem is given, for example, by Theorem 4.1 in [11]; however, the rep-
resentation of a solution provided by it is not sufficiently specific for using in the analysis of the
truncation error. It would be best to use the representation from Theorem 3.3 in [5] if the proof
of this theorem and the related representation were error-free.1 This theorem is valid only for
k = 1, which is insufficient for our purpose. We formulate the desired part of the above-mentioned
theorem in the corrected form.
Theorem 2.1. If conditions (1.4)–(1.6) are satisfied, then the solution of problem (1.1)–(1.3)
admits the representation
 4  2
2κk,l 4
κk,1
u(x, y) = ϕk,l (x, y) + q (ak ) 2k ϕk,1 (x, y) + ũ(x, y), (2.5)
k=1 l=1
π(2l)! k=1
12


where ũ(x, y) ∈ C 4,λ Ω̄ , κk,l are the constants given by (1.8) and (1.9),


ϕ1,l (x, y) = Im ζ 2l ln ζ , 1 = |ζ| for εζ = x + iy, (2.6)
and the remaining functions ϕk,l (x, y) and k are obtained from ϕ1,l (x, y) and 1 by the replacement
of either x, y, or both x and y by (1 − x) and/or (1 − y).
1
A part of Theorem 3.3 in [5] was cited and used in [12] with no reference to the errors. Fortunately, the error does
not affect the definitive result in [12].

DIFFERENTIAL EQUATIONS Vol. 42 No. 7 2006

958 ANDREEV

To prove the theorem, it suffices to show that ũ(x, y) satisfies the compatibility conditions of
order ≤ 2 and use Theorem 3.2 in [5]. In this connection, note that ε2 ∆ (21 ϕ1,1 ) = 12ϕ1,1 + P̊2 (x, y),
where P̊2 (x, y) is a homogeneous polynomial of degree 2.
Remark 2.1. The representation of the solution in Theorem 3.3 in [5] with k = 2 differs
from (2.5) by
the  absence of terms  the coefficients q (ak ) [the last sum in (2.5)], which belong

with
3,λ
4,λ
only to C Ω̄ rather than C Ω̄ .
Remark 2.2. The validity of the first-order compatibility conditions for problem (1.1), (1.2)
[see (1.8)] implies the vanishing of the coefficients κk,1 from (2.5). If the second-order compatibility
conditions are valid, then κk,2 also vanishes. It was assumed in [9] that all κk,l = 0 in (2.5).
Remark 2.3. In particular, it follows from Theorem 2.1 that the inequality
 4  
∂ u/∂xi ∂y 4−i  ≤ c(εr)−2 , r = x2 + y 2 ,

is valid in a neighborhood of the vertex a1 = (0, 0).

The following assertion answers the second question.
Theorem 2.2 [9]. If conditions (1.4)–(1.6) are satisfied, then the solution of problem (1.1)–(1.3)
admits the representation


4 
4
u(x, y) = v(x, y) + wk (x, y) + zk (x, y), (2.7)
k=1 k=1

where
L v = f, L wk = L zk = 0, k = 1, 2, 3, 4, (2.8)



moreover, v(x, y) ∈ C 4,λ Ω̄ is a regular part of the solution, wk ∈ C 4,λ Ω̄ , k = 1, 2, 3, 4, are expo-
nential boundary functions
specifying the solution in neighborhoods of the edges Γk of the square Ω,


 
and zk ∈ C 1,λ Ω̄ ∩ C 4,λ Ω̄\Ωk are corner layer functions. The functions v(x, y), wk (x, y), and
zk (x, y), k = 1, 2, 3, 4, and their derivatives can be estimated as . . .
Here we represent only the needed part of Theorem 2.2 in [9] as applied to our case. Obviously,
in [9], it was stated with all related details.

 Note that problem (1.1)–(1.3) was considered in [9] under
the assumption that u(x, y) ∈ C 4,λ Ω̄ . However, the analysis of the proof given there shows that
all assertions for the functions v(x, y) and wk (x, y) remain valid under assumption

 (1.6). The only
exception is the functions zk (x, y), which do not belong to the space C 4,λ Ω̄ in the case under our
consideration. Nevertheless, the estimates for zk themselves (rather than their derivatives) remain
valid, and this proves to be sufficient to use the results in [9] in our considerations.
Remark 2.4. Considerations given in [9] after (2.16) and aimed to justify the smoothness of
z1 (x, y) are not
correct
 from our viewpoint. Obviously, under the assumptions accepted in [9],
z1 (x, y) ∈ C 4,λ
Ω̄ ; this can be justified with the use of simple considerations differing from the
argument used in [9]. These considerations reveal difference between the nonsmooth case considered
here and the smooth case in [9].
We represent the solution of problem (2.4) in the form


4 
4
uhij = vij
h
+ h
wk,ij + h
zk,ij , (xi , yj ) ∈ Ω̄h , (2.9)
k=1 k=1

where
L h vijh = fij , L h wk,ij
h
= L h zk,ij
h
= 0, k = 1, 2, 3, 4, (2.10)
h h h
and the functions vij , wk,ij , and zk,ij are grid approximations to the functions v(x, y), wk (x, y), and
zk (x, y) occurring in the representation (2.7).

DIFFERENTIAL EQUATIONS Vol. 42 No. 7 2006

 ON THE ACCURACY OF GRID APPROXIMATIONS TO NONSMOOTH SOLUTIONS 959

Theorem 2.3 [9]. If conditions (1.4)–(1.6) are satisfied, then the components in (2.7) and (2.9)
of the solutions u(x, y) and uhij of problems (1.1)–(1.3) and (2.4) are related by the formula

  
4
 
h
v (xi , yj ) − vij + wk (xi , yj ) − wk,ij
h 
Lh h
∞ (Ω ) Lh h
∞ (Ω )
k=1
(2.11)

4
 
+ zk (xi , yj ) − zk,ij
h 
≤ cN −2 ln2 N.
∞ (Ω \Ωk )
Lh h h

k=1

 conditions of order ≤ 2 are valid at the corners ak ,

If, in addition to (1.6), the compatibility
k = 1, 2, 3, 4, and hence u(x, y) ∈ C 4,λ Ω̄ , then


4
 
zk (xi , yj ) − zk,ij
h 
≤ cN −2 ln N. (2.12)
Lh (Ω̄h ) ∞ k
k=1

Remark 2.5. The assertion (2.12) of Theorem 2.3 remains valid without assuming the validity
of the second-order compatibility conditions [which imply that the coefficients κk,2 in the represen-

tation (2.5) vanish]. In this case, obviously, the function u(x, y) does not belong even to C 4 Ω̄ ,
but to prove the estimate (2.12), this is unnecessary. The matter is that the truncation error of the
finite-difference scheme (2.4) is estimated only via the derivatives ∂ 4 u/∂x4 and ∂ 4 u/∂y 4 ; and, as
it follows from Remark 4 in [2] and is easily shown by straightforward computations, these deriva-
tives, as well as ∂ 4 u/∂x2 ∂y 2 , become
bounded
 even under the first-order compatibility conditions.
The failure of the inclusion u ∈ C Ω̄ in this case is related to the derivatives ∂ 4 u/∂x ∂y 3 and
4

∂ 4 u/∂x3 ∂y, which do have logarithmic singularities at the corners ak , but they are unnecessary for
our considerations. By (2.6), the estimate
 4   4 
∂ zk /∂x4  + ∂ zk /∂y 4  ≤ cε−4 , (2.13)
L∞ (Ω̄k ) L∞ (Ω̄k )

used in the proof of inequality (2.12), is valid.

It follows from Theorem 2.3 that to prove the above-mentioned convergence rate of the difference
h
scheme (2.4), it remains to justify appropriate closeness of zk (xi , yj ) and zk,ij on Ωhk . Obviously,
one can restrict considerations to the case in which k = 1.

3. THE A PRIORI ESTIMATE.

THE ESTIMATE FOR THE CONVERGENCE RATE
Let us analyze problem (2.4) on the uniform mesh Ω̄h1 . Let

L h Uij = Fij , (xi , yj ) ∈ Ωh1 , Uij |∂Ωh = 0. (3.1)

1

The Green function of this problem is defined as follows:


= Q
L h Gh (P ; Q) = h0 −2 ifif P
P = Q,
P := (xi , yj ) ∈ Ωh1 ,
(3.2)
Gh (P ; Q) = 0, P ∈ ∂Ωh1 ,

for each Q := (ξi
, ηj
) ∈ Ω̄h1 . Obviously, the Green function is symmetric, i.e., Gh (P ; Q) =
Gh (Q; P ), and satisfies the same equation (3.2) with respect to the variable Q.
Let us estimate the Green function Gh .

Lemma 3.1. If condition (1.3) is satisfied, then the Green function Gh (P ; Q) is nonnegative
and satisfies the estimate
Gh (P ; Q) ≤ cε−2 ln N. (3.3)

DIFFERENTIAL EQUATIONS Vol. 42 No. 7 2006

960 ANDREEV

Proof. Since, under condition (1.3), the maximum principle is valid for L h , we have Gh ≥ 0.
h
Let L0 := L q≡0 , i.e., L0 v := −ε2 (vx̄x + vȳy ). The Green function Gh0 of this operator is also
h h


nonnegative, and since Lh0 Gh0 − Gh = qGh ≥ 0, it follows that Gh ≤ Gh0 . We represent the
function Gh0 in the form of expansion in eigenfunctions of the operator L0h :

 µkl (P )µkl (Q)

N/4−1

Gh0 (P ; Q) = , P, Q ∈ Ω̄h1 ,
k,l=1
λkl

where (e.g., see [13, p. 288])

 
2 kπxi lπyj 4ε2 2kπh lπh
µkl (P ) = sin sin , λkl = 2 sin + sin2 .
σ σ σ h 2σ 2σ

Since [13, p. 284] λkl ≥ (2ε/σ)2 (k2 + l2 ), we have

1 
N/4−1
1 c
Gh0 (P ; Q) ≤ 2 2 2
≤ 2 ln N.
ε k,l=1 k + l ε

This, together with preceding estimates, implies assertion (3.3). The proof of the lemma is complete.
Remark 3.1. In fact, the Green function admits the sharper estimate

Gh (P ; Q) ≤ cε−2 ln(1 + N ε)

(cf. [14]), which can be proved in the same way but with Gh0 replaced by G2α2 , that is, the Green
function of the operator L0h + 2α2 [see (1.3)].

Theorem 3.1. If condition (1.3) is satisfied, then the solution of problem (3.1) admits the
a priori estimate  h
U  h h ≤ cε−2 ln N F Lh (Ωh ) , (3.4)
L (Ω ) ∞ 1 1 1

where 
F Lh1 (Ωh1 ) := |Fij | h2 .
(xi ,yj )∈Ωh
1

Proof. By using (3.1) and (3.2) and the second finite-difference Green formula [13, p. 256],
we obtain 
U (P ) = Gh (P ; Q)F (Q)h2 .
Q∈Ωh
1

This implies the estimate

|U (P )| ≤ maxh Gh (P ; Q)F Lh1 (Ωh1 ) .
Q∈Ω1

Now, by taking into account Lemma 3.1, we obtain inequality (3.4). The proof of the theorem is
complete.
h
We use the a priori estimate (3.4) to estimate the difference between z1 (x, y) and z1,ij on Ωh1 .
Let
h
eij := z1,ij − z1 (xi , yj ) (3.5)
be the considered difference. It follows from (2.8) and (2.10) that


L h eij = ψij , (xi , yj ) ∈ Ωh1 , eij |∂Ωh = z1h − z1 ∂Ωh , (3.6)
1 1

where
ψ := ψij [z1 ] := ε2 (z1x̄x + z1ȳy − ∆z1 )ij (3.7)

DIFFERENTIAL EQUATIONS Vol. 42 No. 7 2006

 ON THE ACCURACY OF GRID APPROXIMATIONS TO NONSMOOTH SOLUTIONS 961

is the truncation error. Since, by Theorem 2.2, the function z1 (x, y) satisfies the first- and second-
order compatibility conditions at the corners a2 , a3 , and a4 , it follows from Theorem 2.1 that

z1 (x, y) = π −1 κ1,1 ϕ1,1 (x, y) + (12π)−1 κ1,2 ϕ1,2 (x, y) + 12−1 κ1,1 21 ϕ1,1 (x, y) + z̃1 (x, y), (3.8)

where

z̃1 (x, y) ∈ C 4,λ Ω̄ . (3.9)
4
We represent the truncation error (3.7) in the form ψij = l=1 ψl,ij , where, by the representa-
tion (3.8),
κ1,1 κ1,2
ψ1,ij = ψij [ϕ1,1 ] , ψ2,ij = ψij [ϕ1,2 ] ,
π 12π
κ1,1   (3.10)
ψ3,ij = ψij 21 ϕ1,1 , ψ4,ij = ψij [z̃1 ] .
12
In a similar way, we represent the solution of problem (3.6) in the form


4
eij = el,ij ; (3.11)
l=1

moreover,
L h el,ij = ψl,ij , (xi , yj ) ∈ Ωh1 , l = 1, 2, 3, 4,
(3.12)
el,ij |∂Ωh = 0, l = 1, 2, 3, e4,ij |∂Ωh = eij |∂Ωh .
1 1 1

By construction, eij |∂Ωh ∩∂Ωh = 0, and, by virtue of assertion (2.11), we have

1

eLh∞ (∂Ωh1 \∂Ωh ) ≤ cN −2 ln2 N.

This, together with Assertion 3.2 in [9] and relations (3.7), (3.9), and (3.10), implies that the
solution e4,ij given by (3.12) satisfies the estimate

e4 Lh∞ (Ωh1 ) ≤ cN −2 ln2 N. (3.13)

By Remark 2.5, a similar estimate is valid for e2 as well:

e2 Lh∞ (Ωh1 ) ≤ cN −2 ln2 N. (3.14)

Since  4 2   

∂ 1 ϕ1,1 /∂x4  + ∂ 4 21 ϕ1,1 /∂y 4 L∞ (Ω̄1 ) = O ε−4
L∞ (Ω̄1 )

[cf. (2.13)], we also have

e3 Lh∞ (Ωh1 ) ≤ cN −2 ln2 N. (3.15)

Let us estimate e1,ij . The problem for e1,ij in (3.12) satisfies the assumptions of Theorem 3.1,
which implies that
e1 Lh∞ (Ωh ) ≤ cε−2 ln N ψ1 Lh (Ωh ) . (3.16)
1 1 1

It follows from (3.7), (3.10), and (2.6) that

ψ1,ij = (ψ1,1 )ij + (ψ1,2 )ij , (3.17)

where
κ1,1 h2 ξij yj κ1,1 h2 xi ηij
(ψ1,1 )ij =
2  , (ψ1,2 )ij =
2  ,
3π ξij + yj2 2 3π xi + ηij 2 2

and
ξij ∈ (xi−1 , xi+1 ) , ηij ∈ (yj−1 , yj+1 ) . (3.18)

DIFFERENTIAL EQUATIONS Vol. 42 No. 7 2006

962 ANDREEV

Let us now estimate, say, ψ1,1 . By using the inclusions (3.18) and the inequality 2|ab| ≤ αa2 +α−1 b2 ,
we obtain   |κ | h2
  1,1 xi+1 yj
(ψ1,1 )ij  ≤
2 2
3π xi−1 + yj2
⎧ 3
2 ⎨
⎪ 2h/yj for i = 1
|κ1,1 | h 2xi yj
≤ 2 for i = 2, . . . , N/4 − 1
3π ⎪ ⎩ (x /2)2 + y 2
i j

2 |κ1,1 | h2 h/y 3
for i = 1
≤
2 j 2 −1
3π xi + y j for i = 2, . . . , N/4 − 1.
This implies that
∞ 
2 |κ1,1 | h2  −3 

N/4−1
−1
ψ1,1 Lh (Ωh ) ≤ j + i2 + j 2 ≤ ch2 ln N.
1 1 3π j=1 i,j=1

A similar estimate is valid for ψ1,2 as well, which, together with relations (3.17) and (3.16), implies
the estimate
e1 Lh∞ (Ωh ) ≤ c(h/ε)2 ln2 N.
1

But, by (2.2) and (2.3), h/ε ≤ cN −1 ln N , and consequently,

e1 Lh∞ (Ωh1 ) ≤ cN −2 ln4 N.

By combining this estimate with the estimates (3.13)–(3.15), for eij in (3.11), we have

eLh∞ (Ωh1 ) ≤ cN −2 ln4 N.

In a similar way, one can obtain the same estimates for

 h 
z − zk (xi , yj ) h h , k = 2, 3, 4.
k,ij L (Ω ) ∞ k

By combining these estimates with the estimate (2.11), we obtain the following assertion.

Theorem 3.2. If conditions (1.4)–(1.6) are satisfied, then the solution uhij of problem (2.1)–(2.4)
converges uniformly with respect

−2to ε in the sense of the Lh∞ -norm to the solution u(x, y) of prob-
lem (1.1)–(1.3) at the rate O N ln4 N ; i.e.,
 h 
uij − u (xi , yj ) h h ≤ cN −2 ln4 N. (3.19)
L (Ω )
∞

4. NUMERICAL RESULTS
Let us give results of the numerical solution of a sample problem with the use of the finite-
difference scheme (2.4) and analyze them from the viewpoint of the above-proved Theorem 3.2
for various values of N and ε. Following [9], for the error indicator, we take the Lh∞ (Ωh )-norm of
h/2
the difference eij = uhij − ũij of the solution of problem (2.4) on the mesh Ωh from (2.1)–(2.3)
and the solution of the same problem but on the mesh Ω̃h/2 , where the number of nodes in each
h

Ω soas to ensure
direction is doubled
compared with
that Ωh⊂ Ω̃h/2 and the additional nodes of
h/2
the mesh Ω̃ are xi−1/2 , yj , xi , yj−1/2 , and xi−1/2 , yj−1/2 , where xi−1/2 = (xi−1 + xi )/2 and
yj−1/2 = (yj−1 + yj )/2 for i, j = 1, . . . , N .

DIFFERENTIAL EQUATIONS Vol. 42 No. 7 2006

 ON THE ACCURACY OF GRID APPROXIMATIONS TO NONSMOOTH SOLUTIONS 963

Table 1. Problem (4.1)

N
ε
32 64 128 256
−5 −6 −6
3.9611 × 10 9.9374 × 10 2.4876 × 10 6.2203 × 10−7
4.0562 × 10−2 4.0704 × 10−2 4.0757 × 10−2 4.0765 × 10−2
1 1.1704 × 10−2 9.7871 × 10−3 8.4000 × 10−3 7.3515 × 10−3
−3 −3 −3
3.3770 × 10 2.3533 × 10 1.7312 × 10 1.3257 × 10−3
9.7438 × 10−4 5.6585 × 10−4 3.5680 × 10−4 2.3908 × 10−4
2.6516 × 10−4 6.8052 × 10−5 1.7184 × 10−5 4.3101 × 10−6
2.7152 × 10−1 2.7874 × 10−1 2.8154 × 10−1 2.8247 × 10−1
2−2 7.8345 × 10−2 6.7023 × 10−2 5.8026 × 10−2 5.0939 × 10−2
−2 −2 −2
2.2606 × 10 1.6116 × 10 1.1959 × 10 9.1862 × 10−3
6.5226 × 10−3 3.8750 × 10−3 2.4648 × 10−3 1.6566 × 10−3
4.7904 × 10−3 1.3080 × 10−3 3.3597 × 10−4 8.4599 × 10−5
4.9054 5.3576 5.5045 5.5443
−4
2 1.4154 1.2882 1.1345 9.9984 × 10−1
−1 −1 −1
4.0840 × 10 3.0975 × 10 2.3382 × 10 1.8031 × 10−1
1.1784 × 10−1 7.4480 × 10−2 4.8189 × 10−2 3.2516 × 10−2
2.4722 × 10−2 1.0249 × 10−2 3.9219 × 10−3 1.3335 × 10−3
2.5315 × 10 4.1980 × 10 6.4256 × 10 8.7392 × 10
−6
2 7.3045 1.0094 × 10 1.3243 × 10 1.5760 × 10
2.1076 2.4271 2.7294 2.8421
6.0813 × 10−1 5.8359 × 10−1 5.6253 × 10−1 5.1254 × 10−1
2.5548 × 10−2 1.0441 × 10−2 4.0039 × 10−3 1.3562 × 10−3
2.6161 × 10 4.2766 × 10 6.5600 × 10 8.8880 × 10
−12
2 7.5485 1.0283 × 10 1.3520 × 10 1.6028 × 10
2.1780 2.4726 2.7865 2.8905
6.2845 × 10−1 5.9453 × 10−1 5.7429 × 10−1 5.2126 × 10−1

We use the sample problem in [9]:


−ε2 ∆u + 1 + x2 y 2 u = 0, (x, y) ∈ (0, 1)2 ,
u(x, 0) = (1 − x)2 , 0 ≤ x ≤ 1,
(4.1)
u(0, y) = 1 − y, 0 ≤ y ≤ 1,
u(x, 1) = u(1, y) = 0, 0 ≤ x, y ≤ 1.
One can readily see that the first-order compatibility conditions for problem (4.1) [see (1.8)] are
satisfied at the vertices a3 and a4 and fail at the vertices a1 and a2 . The coefficient κ2,1 in (2.5) for
problem (4.1) is O (ε2 ), and the insufficient smoothness of the solution in a neighborhood of the
vertex a2 has a small influence on the accuracy of the approximate solution for small values of ε,
while κ1,1 = O(1), and in a neighborhood of the vertex a1 , the low smoothness of the solution can
make a substantial contribution to the error of the numerical method.
In Table 1, we represent the following quantities for given values of N and ε : eLh∞ (Ωh ) in
the first row of each section of the table, eN 2 in the second row, eN 2 / ln N in the third row,
eN 2 / ln2 N in the fourth row, and eN 2 / ln3 N in the fifth row. Regardless of expectations, the
first three sections of Table 1 (for ε ≥ 2−4 ) show that the approximate solution (2.4) of problem (4.1)

DIFFERENTIAL EQUATIONS Vol. 42 No. 7 2006

964 ANDREEV

Table 2. The Laplace equation, a uniform mesh

N
8 16 32 64
−3 −4 −5
e 1.1270 × 10 3.0506 × 10 7.8795 × 10 2.0017 × 10−5
eN 2 7.2128 × 10−2 7.8095 × 10−2 8.0686 × 10−2 8.1990 × 10−2
−2 −2 −2
eN / ln N
2
3.4686 × 10 2.8167 × 10 2.3281 × 10 1.9714 × 10−2
N
128 256 512
−6 −6
e 5.0452 × 10 1.2664 × 10 3.1725 × 10−7
eN 2 8.2661 × 10−2 8.2995 × 10−2 8.3165 × 10−2
eN 2 / ln N 1.7036 × 10−2 1.4967 × 10−2 1.3331 × 10−2

converges at the rate O (N −2 ) (the error norms multiplied by N 2 and represented in the second
rows are stabilized better than other quantities as N grows), which, surely, does not contradict
Theorem 3.2 but raises a question concerning its precise statement. Indeed, many details in the
behavior of the considered error can readily be explained. First, by (2.3) and (4.1), if ε ≥ 2−4
(the first three sections in Table 1), and ε is not very large yet, then the mesh Ωh remains uniform,
and two unnecessary logarithms in (3.19) are removed with the use of considerations performed in
the proof of Theorem 3.2. Second, the a priori estimate (3.4) cannot be improved from the viewpoint
of the norms used in it but is somewhat coarse for the precise estimate of the convergence rate
owing to the logarithmic factor. To consider the last logarithm, for comparison, we consider the
Laplace equation with the solution ϕ1,1 (x, y). The solution errors uhij − ϕ11 (xi , yj ) of this problem
by the scheme (2.4) with ε = 1 and q(x, y) ≡ f (x, y) ≡ 0 on a uniform mesh for various values of N
are given in Table 2. It follows from [7] that the convergence rate of this solution is not worse than
O (N −2 ln N ). However, the analysis of Table 2 shows the convergence rate O (N −2 ). It is quite
reasonable to assume2 that for a sufficient smoothness of input data, the numerical solution of the
Poisson equation on a uniform mesh in a rectangle does converge at the rate O (N −2 ) provided
that the boundary function is only continuous at the corners. In this connection, we note that the
low smoothness of a solution in [7] was caused by the low smoothness (much less than that in the
present paper) of the input data rather than the presence of corners. Therefore, there is no reason
to doubt the impossibility of improving the estimate in [7] in order.
Let us continue the analysis of Table 1. If ε ≤ 2−6 , then the mesh becomes nonuniform,
and the solution error is rapidly stabilized as ε diminishes, which implies the uniform convergence of
the solution of problem (2.1)–(2.4) with respect to this parameter. The investigation of
the solution 
error as N grows for these values of ε justified that the convergence is better than O N −2 ln4 N :
eN 2 / ln2 N remains increasing as N grows, but eN 2 / ln3 N is already decreasing. On the basis
of the data of Table 1,
it is difficult to clarify whether

the convergence
 rate of the solution of
problem (2.1)–(2.4) is O N −2 ln2 N rather than O N −2 ln3 N ; although some specific features of
the behavior of eN 2 / ln2 N suggest this. To eliminate any doubt, we consider one more sample
problem
−ε2 ∆u + (1 + x + y)u = (1 − x)(1 − y), (x, y) ∈ (0, 1)2 ,
u(x, 0) = (1 + x), 0 ≤ x ≤ 1,
(4.2)
u(0, y) = 1 + y, 0 ≤ y ≤ 1,
u(x, 1) = u(1, y) = 0, 0 ≤ x, y ≤ 1.
For this problem, the first-order compatibility conditions are valid at all corners; by virtue of
Remark 2.5, the smoothness of the solution of problem (4.2) is sufficient for the convergence the-
2
It is unnecessary to perform a numerical investigation of the convergence

rate of the grid solution of the Laplace
equation in a rectangle, since its convergence at the rate O N −2 was proved in the paper [Wasow, W., J. Res.
Nat. Bur. Standarts, 1952, vol. 48, pp. 345–348] under the assumption that the boundary function is continuous
at the corners and has bounded third derivatives on Γk , k = 1, 2, 3, 4.

DIFFERENTIAL EQUATIONS Vol. 42 No. 7 2006

 ON THE ACCURACY OF GRID APPROXIMATIONS TO NONSMOOTH SOLUTIONS 965

Table 3. Problem (4.2)

N
ε
32 64 128 256
−6 −7 −8
1.3389 × 10 3.3555 × 10 8.3906 × 10 2.0978 × 10−8
1.3710 × 10−3 1.3744 × 10−3 1.3747 × 10−3 1.3748 × 10−3
1 3.9560 × 10−4 3.3048 × 10−4 2.8333 × 10−4 2.4793 × 10−4
−4 −5 −5
1.1415 × 10 7.9463 × 10 5.8394 × 10 4.4711 × 10−5
3.2935 × 10−5 1.9107 × 10−5 1.2035 × 10−5 8.0630 × 10−6
4.4883 × 10−5 1.1248 × 10−5 2.8139 × 10−6 7.0359 × 10−7
4.5960 × 10−2 4.6072 × 10−2 4.6103 × 10−2 4.6110 × 10−2
2−2 1.3261 × 10−2 1.1078 × 10−2 9.5018 × 10−3 8.3154 × 10−3
−3 −3 −3
3.8264 × 10 2.6637 × 10 1.9583 × 10 1.4996 × 10−3
1.1041 × 10−3 6.4048 × 10−4 4.0361 × 10−4 2.7043 × 10−4
6.4022 × 10−4 1.6822 × 10−4 4.2390 × 10−5 1.0647 × 10−5
6.5559 × 10−1 6.8903 × 10−1 6.9452 × 10−1 6.9776 × 10−1
2−4 1.8916 × 10−1 1.6568 × 10−1 1.4314 × 10−1 1.2583 × 10−1
−2 −2 −2
5.4581 × 10 3.9837 × 10 2.9501 × 10 2.2692 × 10−2
1.5749 × 10−2 9.5787 × 10−3 6.0801 × 10−3 4.0922 × 10−3
3.2043 × 10−3 1.3989 × 10−3 5.0645 × 10−4 1.6803 × 10−4
3.2812 5.7299 8.2977 1.1012 × 10
−6 −1
2 9.4676 × 10 1.3777 1.7101 1.9859
2.7318 × 10−1 3.3128 × 10−1 3.5246 × 10−1 3.5813 × 10−1
−2 −2 −2
7.8822 × 10 7.9656 × 10 7.2642 × 10 6.4583 × 10−2
3.2288 × 10−3 1.4132 × 10−3 5.1199 × 10−4 1.6987 × 10−4
3.3063 5.7885 8.3884 1.1133 × 10
−12 −1
2 9.5399 × 10 1.3918 1.7289 2.0076
2.7526 × 10−1 3.3466 × 10−1 3.5632 × 10−1 3.6205 × 10−1
−2 −2 −2
7.9425 × 10 8.0470 × 10 7.3436 × 10 6.5291 × 10−2


orem in [9] [i.e., the estimate O N −2 ln2 N ] to be valid for the error of its numerical solution.
Table 3, whose structure is similar to that of Table 1, represents the results of numerical solution
of problem (4.2) and their processing. An analysis shows that the behavior
−2 of eN2 / ln2 N and
eN / ln N as N grows is the same as in Table 1. Since the estimate O N ln2 N is proved in
2 3

this case, we find that the results in Table 1 should imply the same estimate.

ACKNOWLEDGMENTS
The author is grateful to I.G. Belukhina for aim in performing computations and preparing the
paper for the publication.

REFERENCES
1. Shishkin, G.I., Setochnye approksimatsii singulyarno vozmushchennykh ellipticheskikh i parabolicheskikh
uravnenii (Grid Approximations of Singularly Perturbed Elliptic and Parabolic Equations), Yekaterin-
burg: Ural Branch of RAS, 1992.
2. Volkov, E.A., Tr. Mat. Inst. Steklova, 1965, vol. 77, pp. 89–112.
3. Andreev, V.B., Differ. Uravn., 1980, vol. 16, no. 7, pp. 1172–1184.

DIFFERENTIAL EQUATIONS Vol. 42 No. 7 2006

966 ANDREEV

4. Gilbarg, D. and Trudinger, N.S., Elliptic Partial Differential Equations of Second Order , Berlin:
Springer-Verlag, 1983. Translated under the title Ellipticheskie differentsial’nye uravneniya s chast-
nymi proizvodnymi vtorogo poryadka, Moscow: Nauka, 1989.
5. Han, H. and Kellogg, R.B., SIAM J. Math. Anal., 1990, vol. 21, no. 2, pp. 394–408.
6. Shishkin, G.I., Zh. Vychisl. Mat. Mat. Fiz., 1987, vol. 27, no. 9, pp. 1360–1374.
7. Bakhvalov, N.S., Zh. Vychisl. Mat. Mat. Fiz., 1969, vol. 9, no. 4, pp. 841–859.
8. Shishkin, G.I., Proc. Steklov Inst. Math., 2003, Suppl. 2, pp. S138–S147.
9. Clavero, C., Gracia, J.L., and O’Riordan, E., Math. Comp., 2005, vol. 74, pp. 1743–1758.
10. Shishkin, G.I., Mat. Model., 1999, vol. 11, no. 11, pp. 75–90.
11. Miersemann, E., Math. Nachr., 1988, vol. 135, pp. 239–274.
12. Andreev, V.B., Math. Model. Anal., 2005, pp. 207–212.
13. Samarskii, A.A. and Andreev, V.B., Raznostnye metody dlya ellipticheskikh uravnenii (Finite-Difference
Methods for Elliptic Equations), Moscow: Nauka, 1976.
14. Belukhina, I.G., Vestn. Moskov. Univ. Ser. XV. Vychisl. Mat. Kibernet., 2002, no. 1, pp. 13–18.

DIFFERENTIAL EQUATIONS Vol. 42 No. 7 2006