J Sci Comput (2018) 74:631–639

https://doi.org/10.1007/s10915-017-0458-z

Convergence Analysis in the Maximum Norm of the

Numerical Gradient of the Shortley–Weller Method

Jiwon Seo1 · Seung-yeal Ha2 · Chohong Min1

Received: 10 April 2017 / Revised: 5 May 2017 / Accepted: 15 May 2017 /

Published online: 27 May 2017
© Springer Science+Business Media New York 2017

Abstract The Shortley–Weller method is a standard central finite-difference-method for

solving the Poisson equation in irregular domains with Dirichlet boundary conditions. It is
well known that the Shortley–Weller method produces second-order accurate solutions and
it has been numerically observed that the solution gradients are also second-order accurate;
a property known as super-convergence. The super-convergence was proved in the L 2 norm
in Yoon and Min (J Sci Comput 67(2):602–617, 2016). In this article, we present a proof for
the super-convergence in the L ∞ norm.

Keywords Shortley–Weller · Finite difference method · Super-convergence · Convergence

analysis

1 Introduction

We consider the Poisson equation in a smooth irregular domain , with Dirichlet boundary
condition on its boundary  = ∂:

−u = f in 
, (1)
u=g on  := ∂
where f and g are given smooth functions.
The Shortley–Weller method is a standard central finite-difference-method for solving
the boundary value problem (1). Its numerical solution is well known to be convergent to
the analytic solution with the second order accuracy [3,8]. For many physical problems of
interest however, the gradient of the solution has more importance than the solution itself.
For example, this is the case of incompressible fluids [2,11,15] and free boundary problems

B Chohong Min
[email protected]
1 Ewha Womans University, Seoul, Republic of Korea
2 Seoul National University, Seoul, Republic of Korea

123
632 J Sci Comput (2018) 74:631–639

[1,4], for which the accuracy of the gradient of (1) determines the accuracy of the solution.
It is therefore an important topic to analyze the convergence of the numerical gradient of the
Shortley–Weller method.
The numerical gradient has been observed to be second-order accurate in the L ∞ norm
in numerical tests (e.g. [10]), but to this day the observation had not been confirmed by a
mathematical proof. In the case of rectangular domains, the second-order convergence in the
L 2 norm appears in a textbook [13], whereas in the case of polygonal domains, the 1.5 order
of convergence in L 2 norm was proved by Li et al. [9]. For arbitrary smooth domains, the
super-convergence was proved in the L 2 norm in [17]. Recently Weynans [14] sketched a
proof in the L ∞ norm based on the discrete Green function by Ciarlet [3,8]. In particular,
Weynans’ proof is based on the commutativity between the discrete Laplacian and the gradient
operators away from the boundary, as well as an error estimate near the boundary. However,
although the conclusion in [14] is correct,
  many important details are omitted. For example,
the numerical solution with error O h 3 does not necessarily implies a O h 2 error for the
numerical gradient because of the division by a local step size h i+ 1 , j near the interface that
2
can be much smaller than h. Likewise, the estimation of Green’s function omits much details
in its derivation. Instead of using the Green function, we focus on the commutative region of
the discrete Laplacian and the discrete gradient, and we leverage the existing error analysis
of our previous work [17] to present a rigorous and complete proof for the super-convergence
in the L ∞ norm.
The rest of this paper after is organized as follows. In Sect. 2, we review the notations and
the existing error analysis, and we discuss the commutative region of the discrete gradient
in Sect. 3. Combining the commutativity and the existing error analysis, we provide a proof
for the second-order convergence in the L ∞ norm in Sect. 4.

2 Analysis of the Numerical Solution [17]

The domain  ⊂ R2 is approximated by the set of grid nodes h := ∩(hZ)2 , where (hZ)2
denotes the set of grid nodes with uniform step size h. The boundary  := ∂ is approximated
by  h , the set of intersection points between  and the grid lines. As illustrated in Fig. 1, each
grid node (xi , y j ) has four neighboring nodes in h ∪ h , which are denoted by (xi±1 , y j ) and
(xi , y j±1 ). Given a discrete function v h : h ∪  h → R, Shortley and Weller [12] discretize
the Laplacian −h v h : h → R by applying the standard central-finite-differences on each
node and its adjacent neighbors as follows:

 
vi+1,
h
j − vi j
h vihj − vi−1,
h
j
h i+ 1 , j + h i− 1 , j
h vihj = − 2 2
h i+ 1 , j h i− 1 , j 2
 2 2

vi,h j+1 − vihj vihj − vi,h j−1 h i, j+ 1 + h i, j− 1
+ − 2 2
.
h i, j+ 1 h i, j− 1 2
2 2

The numerical approximation u h : h ∪  h → R for the Poisson problem (1) is the solution
of the following linear system:


−h u h = f in h ,
(2)
uh = g on  h .

123
J Sci Comput (2018) 74:631–639 633

Fig. 1 The nodes in h are marked with the symbol open circle, and the nodes in  h with the symbol filled
circle. The interior nodes in 0h are enclosed in the dashed curve. Each grid node in h has four neighboring
nodes in h ∪  h

The consistency and the error vectors are important quantities in the error analysis and are
defined as follows:

(consistency) ch := −h u + u in h
(3)
(error) eh := u h − u in h ∪  h

We recall the following results from the error analysis of [17] and refer the interested
reader to [17] for the details.

Lemma 2.1 Let eh be a error defined by (3). Then, eh satisfies the following relation:
−h eh = ch in h .

Theorem 2.2 (Comparison principle) If v h , w h : h ∪  h → R satisfy −h v h ≤ −h w h

in h and v h ≤ w h on  h , then v h ≤ w h in h .

Theorem 2.3 (Refined error estimate) If h is sufficiently small, then there exists a constant
E = E(, u) such that for any (xi , y j ) ∈ h ,
   
 h
ei j  ≤ E(, u)h 2 dist xi , y j ,  h + min h i± 1 . j , h i, j± 1 .
2 2

3 Commutativity of the Discrete Laplace and Gradient Operators

for Interior Edges

For a discrete function v h sampled at grid nodes, its discrete gradient values are calculated
by central differencing and defined on the corresponding cell faces, following the Marker-
and-Cell configuration [7]. For example, the derivative in the x-direction is calculated as
follows:

123
634 J Sci Comput (2018) 74:631–639

˜ h are marked with the symbol open square, and the interior edges are enclosed by the
Fig. 2 The edges in  
dashed curve. Each interior edge xi+ 1 , y j has 8 neighboring nodes (marked with the symbol open circle)
2
in h

   
˜ h :=
 xi+ 1 , y j | xi , y j , xi+1 , y j ∈ h ∪  h ,
2

vi+1,
h − vi,h j
Dxh v h :=
j ˜ h,
, for each edge xi+ 1 , y j ∈ 
i+ 21 , j h i+ 1 , j 2
2

In this section, we discuss the commutativity of the two discrete operatorsh and h
 Dxnoting
that the commutativity between h and D hy is similar. The calculation of h Dxh v h i+ 1 , j
  2
requires four neighboring values of Dxh v h i+ 1 , j , however the neighboring values may not
2
exist near the boundary. As illustrated in Figs. 1 and 2, we define the set of interior nodes
and the set of interior edges (faces) as:

     
oh := xi , y j | xi±1 , y j and xi , y j±1 ∈ h , and
   
˜ oh :=
 xi+ 1 , y j | xi , y j and xi+1 , y j ∈ oh .
2

˜ h.
Now we show the commutativity, Dxh ◦ h = h ◦ Dxh , for the interior edge set  0
 
Theorem 3.1 (Commutativity) For any discrete function v h : h → R, h ◦ Dxh v h and
 h 
˜ oh , and they are equal.
Dx ◦ h v h are well defined in 

   
˜ oh , the grid nodes xi , y j and xi+1 , y j belong to oh ,
Proof For each edge xi+ 1 , y j ∈ 
2
and all of their neighboring nodes belong to h . As illustrated in Figs. 1 and 2, there exist
 ◦) around
8 neighboring nodes (denoted by the symbol  the edgeand they are  aligned with
uniform step size h. The calculations of h Dxh v h i+ 1 , j and Dxh h v h i+ 1 , j are the
2 2
same and expressed as a function of the 8 neighbors as follows:

123
J Sci Comput (2018) 74:631–639 635

h Dxh v h = Dxh h v h
i+ 21 , j i+ 21 , j
 
1 vi+2,
h +v h +v h +5v h
= 3 j i+1, j+1 i+1, j−1 i, j
.
h −vi−1,
h
j −vi,h j+1 −vi,h j−1 −5vi+1,
h
j

4 Accuracy Analysis of the Discrete Gradient

In this section, we prove the second-order accuracy of the solution’s gradient in L ∞ norm.
We first present several lemmas before we present our main results.
    
Lemma 4.1 For any xi , y j ∈ h ∪  h \ ◦h , dist xi , y j ,  ≤ h.
   
Proof Since xi , y j ∈/ 0h , either xi , y j ∈  h or one of its neighboring nodes, at a distance
  
at most h away, belong to  h . Since  h ⊂ , in either case we have dist xi , y j ,  ≤ h
Lemma 4.2 The following estimate holds true:
 
 h h
Dx e  ∞ ˜ h ˜ h ≤ 2E · h 2 .
L  \ 0
   
Proof Since xi+ 1 , y j ∈/ ˜ h , either xi , y j or xi+1 , y j belongs to h ∪  h − ◦h . By
0
2      
the above lemma, dist xi , y j ,  ≤ 2h and dist xi+1 , y j ,  ≤ 2h. Using Theorem
2.3, we have
  
 h
u i, j − u xi , y j  ≤ h i+ 1 , j · h 2 · 2E, and
   2
 h
u i+1, j − u xi+1 , y j  ≤ h i+ 1 , j · h 2 · 2E.
2
 h h
Combining the two inequalities, we conclude that Dx e  ∞ ˜ h ˜ h ≤ 2E · h 2 .

L  \ 0

Lemma 4.3 Let K := 105

4 · max,|α|≤5
¯ |∂ α u|, then
 
 h h
Dx c  ∞ h ≤ K · h 2 .
˜
L  0

   
˜ h , xi , y j and xi+1 , y j belongs to ◦h , and all of their adjacent
Proof Since xi+ 1 , y j ∈ 
2 0
nodes are aligned on the uniform grid. Thus Dxh ch is given as
i+ 21 , j

4u i+1, j − u i+2, j − u i, j − u i+1, j+1 − u i+1, j−1 u i+1, j

Dxh ci+
h
1
,j
= +
2 h3 h
4u i, j − u i+1, j − u i−1, j − u i, j+1 − u i, j−1 u i, j
− − . (4)
h3 h
Substituting each term with its Taylor expansion at xi+ 1 , y j gives the following result.
2
See the details in Appendix.
     
 h h    2 3 5 1 5
 D c 1  ≤ max ∂ α u  · h 2·
5
+2· 5 +4·6· +2·2·
5
 x i+ 2 , j  ˜ 120 2 2 2 2
,|α|≤5

= K · h2.

123
636 J Sci Comput (2018) 74:631–639

Lemma 4.4 Consider the function p h : ˜ h → R defined as −h p h = 1 in 

˜ h and p h = 0
0
˜ \
on  h ˜ . There exists a constant D = D () such that 0 ≤ p ≤ D in ,
h h ˜ whenever
 0
h ≤ D. 6

Proof Let p (x) be the analytic solution of −p = 2 in  with p = 0 on . Let us sample
˜ h and calculate −h p h in 
p (x) in  ˜ h . A simple Taylor expansion shows that
0

h2 ∂ 4 p   ∂4 p
−h p xi+ 1 , y j = 2 + ξ 1 , y j + x 1 , ξ 2 ,
2 12 ∂ x 4 ∂ x 4 i+ 2
 4   4 
   
for some ξ1 ∈ (xi , xi+1 ) and ξ2 ∈ y j− 1 , y j+ 1 . Now, let D = max¯ | p| ,  ∂∂ x p4  ,  ∂∂ y p4  ,
 2 2

then whenever h ≤ D6 ,
2
˜h
−h p h = 1 ≤ 2 − h12 2D ≤ −h p in 0
h ˜
p = 0 ≤ p in  \ 0 .
h ˜ h

˜ h.
Then by the comparison principle 2.2, we have 0 ≤ p h ≤ p ≤ D in 

Lemma 4.2 states that Dxh u h is second-order accurate to Dxh u in  ˜h \ 

˜ ◦h . Now we
˜ ◦ in the lemma to obtain our main theorem.
investigate the excluded region  h

Theorem 4.5 (Main theorem) Let u (x, y) be the analytic solution of the Poisson problem in
 ∪ , and u i,h j be the discrete solution in h ∪  h obtained by the Shortley–Weller method,
then there exists a constant C = C (, u), independent of h, such that
 
 h h 
 D u 1 − ∂u x 1 , y j 
 ∞ ˜h ≤ C · h .
2
 x i+ 2 , j ∂ x i+ 2
L 

˜ h , we have
Proof Using the commutativity principle in  0

˜ 0h .
−h Dxh eh = Dxh −hx eh = Dxh ch in 

Lemmas 4.2 and 4.3 state that

−2E · h 2 ≤ Dxh eh ≤ 2Eh 2 on  ˜h −

˜ 0h , and
˜ 0h .
−K · h 2 ≤ Dxh ch ≤ K h 2 in 

Using the function p h in Lemma 4.4 and the commutativity principle, we have
 
− K h 2 ph + 2Eh 2 ≤ Dxheh ≤ K h 2p h + 2Eh 2 on ˜h ˜h
  2−h0 
− − K h p + 2Eh
h 2 h 2 ≤ − Dx e ≤ − K h p + 2Eh 2 in 
h h h h ˜ h.
0

Applying the comparison principle, we obtain:

˜ h,
− K h 2 p h + 2Eh 2 ≤ Dxh eh ≤ K h 2 p h + 2Eh 2 in 

while invoking Lemma 4.4 gives

 
 h h
Dx e  ≤ h 2 (K · D + 2E) .
˜h
L∞ 

123
J Sci Comput (2018) 74:631–639 637

The error of the central finite difference is bounded as follows:

  
 u x , y  − u x , y  ∂u  h i+
2  3 
∂ u 
  1
2,j
max  3  .
i+1 j i j
 − x 1 , yj  ≤
 h i+ 1 , j ∂ x i+ 2  24 ¯
 ∂x
2

Finally, by the triangle inequality, we have

   3 
 h h   
 D u 1 − ∂u x 1 , y j  ≤ h 2 K · D + 2E + 1 max  ∂ u  .
 x i+ 2 , j ∂ x i+ 2  24 ¯  ∂ x 3 

5 Analysis in Three Spatial Dimensions

In the previous sections, we have analyzed the convergence in two spatial dimensions. Since
the Shortley–Weller method follows a dimension-by-dimension approach, the analysis in
three spatial dimensions is similar, except for a few exceptions that we detail next.
The domain  ⊂ R3 is approximated by the set of grid nodes h :=  ∩ (hZ)3 and  h ,
i.e. the intersection of  and the grid lines of (hZ)3 . The discrete Laplacian is defined as:
 h 
vi+1, j,k − vihjk vihjk − vi−1,h
j,k
h i+ 1 , j,k + h i− 1 , j,k
 vi jk =
h h
− / 2 2
h i+ 1 , j,k h i− 1 , j,k 2
 h 2 2

vi, j+1,k − vihjk vihjk − vi,h j−1,k h i, j+ 1 ,k + h i, j− 1 ,k
+ − / 2 2
h i, j+ 1 ,k h i, j− 1 ,k 2
 h 2 2

vi, j,k+1 − vihjk vihjk − vi,h j,k−1 h i, j,k+ 1 + h i, j,k− 1
+ − / 2 2
,
h i, j,k+ 1 h i, j,k− 1 2
2 2

and we can derive the corresponding main theorem, as in the previous sections.
Theorem 5.1 (Main theorem in R3 ) Let u (x, y, z) be the analytic solution of the Poisson
problem in  ∪ , and u i,h j,k be the discrete solution in h ∪  h obtained by the Shortley–
Weller method, then there exists a constant C = C (, u), independent of h, such that
 
 h h ∂u 
D u 1 
 x i+ 2 , j,k − ∂ x xi+ 21 , y j , z k  ∞ ˜ h ≤ C · h .
2
L 

The only notable difference is the constant in Lemma 4.3.

Lemma 5.2 Let K := 257
30 · max,|α|≤5
¯ |∂ α u|, then
 
 h h
Dx c  ∞ h ≤ K · h 2 .
˜
L  0

6 Conclusion

We have presented a formal mathematical proof that the discrete gradient obtained from
the Shortley–Weller method is second-order accurate in the L ∞ norm. This work can thus
provide useful guidelines on how to compute the gradient of the Poisson solution at every grid

123
638 J Sci Comput (2018) 74:631–639

nodes. In turn, this can be used in a plethora of applications in computational fluid dynamics
and free boundary problems where the gradient drives the accuracy of the simulations.
Elliptic problems exhibit a strong dependence of the solution from any grid location to the
entire region in space. As a result, an adaptive grid with finer resolution in some regions is
not as efficient in elliptic problems as they are in hyperbolic and parabolic problems. Finite
difference methods are simple to implement and easy to use, and are therefore very attractive
for solving elliptic problems. However, formal proof of accuracy analysis are rare, as opposed
to what is found in the finite element community.
We hope this work will incite other researchers to attempt many unsolved problems of
finite difference methods. For example, we list below two open problems that, to the best of
our knowledge, are still unsolved. In his seminal work [6], Gustafsson conjectured that only
modified-ILU preconditioner decreases the condition number by one less order than the other
ILU-type preconditioners. Gustafsson’s conjecture was proved in the case of a symmetric
finite difference method [16], but not in the case of the Shortley–Weller method. Gibou and
Min [5] introduced a finite difference method for solving fluid-solid-interaction, and recently
proved that the solid velocity is at least first-order accurate [18]. Numerical results for the
solid velocity however strongly suggests that second-order convergence is obtained, a proof
of which has not yet been provided.

Appendix: Detailed Calculations in Lemma 4.3

 
In this section, we provide detailed calculations that lead to the estimate  Dxh ch  ≤
L ∞ ˜oh
105
4 max,|α≤5|
˜ |∂ α u| · h 2 in Lemma 4.3.
Using the Taylor series of u (x, y) at xi+ 1 , y j , the terms that sum up Dxh ch in (4) are
2

expanded. For notational conveniences, the local coordinates centered at xi+ 1 , y j are used
  2
in the calculations. For example, u i+1, j+1 is denoted by u h2 , h . The Taylor expansions are
listed below with remainders.
2 3 4 5 ±
u ± 3h 3h 9h 9h 27h 81h
2 , 0 = u ± 2 u x + 8 u x x ± 16 u x x x + 128 u x x x x ± 1280 u x x x x x ξ1 , 0
2 3 4 5 ±
u ± h2 , 0 = u ± h2 u x + h8 u x x ± h48 u x x x + 384
h u h
x x x x ± 3840 u x x x x x ξ2 , 0
  2   3  
u ± h2 , 0 = u ± h2 u x x x + u x yy + h8 u x x x x + u x x yy ± h48 u x x x x x + u x x x yy ξ4± , 0

When the above expansions are inserted into the summation of Dxh ch , canceled out all the
terms but the remainders.

123
J Sci Comput (2018) 74:631–639 639

⎛
 5           ⎞
− 23 ⎛u x x x x x ξ1+ , 0 +u x x x x x ξ1− , 0 + 255 u x x x x x ξ2+ , 0 +u x x x x x ξ2− , 0⎞
⎜   1 +,+ + ⎟
⎜ u x x x x x +u x x x x y +u x x x yy +u x x yyy +u x yyyy +u yyyyy 2 ξ3 , ξ3 ⎟
⎜ ⎜  1 +,− − ⎟ ⎟
⎟
⎜ ⎜  ⎟
2 ⎜
h ⎜  1 5 ⎜ + u x x x x x −u x x x x y +u x x x yy −u x x yyy +u x yyyy −u yyyyy 2 ξ3 , ξ3 ⎟ ⎟ ⎟
Dxh ci+
h
= − 2 ⎜ ⎜  ⎟
120 ⎜ ⎟⎟
1
2,j ⎜ + u x x x x x −u x x x x y +u x x x yy −u x x yyy +u x yyyy −u yyyyy 21 ξ3−,+ , ξ3+ ⎟ ⎟
⎜ ⎜
⎜ ⎝   ⎠⎟ ⎟
⎝ + u x x x x x +u x x x x y +u x x x yy +u x x yyy +u x yyyy +u yyyyy 21 ξ3−,− , ξ3− ⎠
   +     − 
+ 2 u x x x x x + u x x x yy ξ4 , 0 + u x x x x x + u x x x yy ξ4 , 0
5

Now, we prove the lemma.

 
 h h   3 5  1 5
 D c 1  ≤ max ˜ α h2
,|α|≤5 |∂ u| · 2· +2· +4·6· +2·2·
5 5
 x 
i+ 2 , j 120 2 25 2 2
= 105
4 max,|α≤5|
˜ |∂ α u| · h 2 .

References
1. Caffarelli, L.A., Gilardi, G.: Monotonicity of the free boundary in the two-dimensional dam problem.
Annali della Scuola Normale Superiore di Pisa-Classe di Scienze 7(3), 523–537 (1980)
2. Chorin, A.J.: A numerical method for solving incompressible viscous flow problems. J. Comput. Phys.
135(2), 118–125 (1997)
3. Ciarlet, P.G., Miara, B., Thomas, J.M.: Introduction to Numerical Linear Algebra and Optimisation.
Cambridge Texts in Applied Mathematics. Cambridge University Press, Cambridge (1989)
4. Friedman, A.: Variational principles and free-boundary problems. Courier Corporation, North Chelmsford
(2010)
5. Gibou, F., Min, C.: Efficient symmetric positive definite second-order accurate monolithic solver for
fluid/solid interactions. J. Comput. Phys. 231, 3245–3263 (2012)
6. Gustafsson, I.: A class of first order factorization methods. BIT 18, 142–156 (1978)
7. Harlow, F.H., Welch, J.E.: Numerical calculation of time-dependent viscous incompressible flow of fluid
with a free surface. Phys. Fluids 8(3), 2182–2189 (1965)
8. Iserles, A.: A First Course in the Numerical Analysis of Differential Equations. Cambridge Texts in
Applied Mathematics. Cambridge University Press, Cambridge (1996)
9. Li, Z.-C., Hu, H.-Y., Wang, S., Fang, Q.: Superconvergence of solution derivatives of the shortley-weller
difference approximation to poisson’s equation with singularities on polygonal domains. Appl. Numer.
Math. 58(5), 689–704 (2008)
10. Ng, Y.T., Chen, H., Min, C., Gibou, F.: Guidelines for poisson solvers on irregular domains with dirichlet
boundary conditions using the ghost fluid method. J. Sci. Comput. 41(2), 300–320 (2009)
11. Peskin, C.S.: Flow patterns around heart valves. In: Proceedings of the Third International Conference
on Numerical Methods in Fluid Mechanics, pp. 214–221. Springer (1973)
12. Shortley, G.H., Weller, R.: The numerical solution of Laplace’s equation. J. Appl. Phys. 9(5), 334–348
(1938)
13. Strikwerda, J.C.: Finite Difference Schemes and Partial Differential Equations. SIAM, New Delhi (2004)
14. Weynans, L.: A proof in the finite-difference spirit of the superconvergence of the gradient for the Shortley–
Weller method. Ph.D. thesis, INRIA Bordeaux (2015)
15. Xiu, D., Karniadakis, G.E.: A semi-Lagrangian high-order method for Navier–Stokes equations. J. Com-
put. Phys. 172(2), 658–684 (2001)
16. Yoon, G., Min, C.: Analyses on the finite difference method by Gibou, et al. for poisson equation. J.
Comput. Phys. 280, 184–194 (2015)
17. Yoon, G., Min, C.: Convergence analysis of the standard central finite difference method for Poisson
equation. J. Sci. Comput. 67(2), 602–617 (2016)
18. Yoon, G., Min, C., Kim, S.: A stable and convergent method for hodge decomposition of fluid-solid
interaction. (2016). arXiv:1610.03195

123