Beruflich Dokumente
Kultur Dokumente
. A suitable circle pattern from the respective class can be dened on each
,
so that the circle centers (Schramm-pattern) or intersection points (CR-mapping)
approximate the values of u at corresponding sites with an error O(
2
). Moreover,
the convergence is in C
t
u(t, x) = M
x
u(t, x) +f(u(t, x)), (1.1)
u(0, x) = u
0
(x).
The continuous solution u : C
d
is sought on R
2
, where M is a constant
d d-matrix, and f is an analytic function. For x-analytic data u
0
, there exists a
local solution to (1.1) by the Cauchy-Kovalevskaya theorem.
A variety of elliptic equations e.g., the nonlinear Poisson equation (
2
x
+
2
t
)u =
f(u) can be brought into the standard form (1.1). In the case of most interest
here, u is a conformal map. Identifying the (t, x)-plane with the complex numbers,
u(t, x) : C is holomorphic in z = x+it, and hence solves the Cauchy-Riemann-
equations, which is (1.1) with M = i and f 0.
Circle patterns are constructed which approximate u. These are characterized by
discrete functions v
on two-dimensional grids
converge to u in C
. Eventu-
ally, this leads to convergence of the corresponding circle patterns.
The article is organized as follows: in section 2, the theorem about discrete approxi-
mation of the Cauchy problem (1.1) is formulated. The theorem is proven in section
3. Its numerical applicability is discussed in section 4. Convergence theorems for
CR-mappings and Schramm-patterns are proven in sections 5 and 6, respectively.
Acknowledgments. The author would like to thank Alexander Bobenko and Yuri
Suris for many discussions and helpful advice. Further, the author is grateful to
Walter Craig for an entertaining introduction to the topic of abstract Cauchy-
Kovalevskaya theorems.
2. Discrete Approximation of the Cauchy Problem
As usual, a function f : D C
p
C
d
is called analytic, if it is complex dier-
entiable in its p arguments. Moreover, a function u : I
C
d
dened on the real
interval I
(I
) = x C[ dist(x, I
. Solutions u = (u
(1)
, . . . , u
(d)
) : C
d
are sought on diamond-
shaped domains
= (r) = (t, x) R
2
: [x[ +[t[ r.
The Cauchy-Kovalevskaya-Theorem implies
Theorem 2.1. Problem (1.1) has a classical solution u on = (r) for a suitable
r < . This solution is x-analytic and t-smooth, i.e., of class C
.
Replace u by a function v
= (v
(1)
, . . . , v
(d)
) dened on the discrete set
2. For v
(r) C
d
, dierence quotients
q
q
q
q
q
q
q
q
q
q
q
q
q
q
q
q
q
q
q
q
q
q
q
q
q
q
q
q
q
q
q
q
q
q
q
q
q
q
q
q
q
q
q
q
q
q
q
q
q a
a
a
a
a
a
a
a
a
a
a
a
a
a
a
a
a
a
a
a
a
a
a
a
a
a
a
a
a
a
a
a
a
a
a
a s s s s s s s
s s s s s s
-
x
6
t
-
v
(r) and
are marked
and , respectively. Initial data for v
)(t, x) =
1
_
v
(t, x +
2
) v
(t, x
2
)
_
(
t
v
)(t, x) =
1
_
v
(t +
2
, x) v
(t
2
, x)
_
,
are given on the dual lattice
(r) = (t, x) (r
2
) :
2
+x +t Z.
Higher dierence quotients
m
x
n
t
v
are dened on
m+n
(r) =
_
(r (m +n)
2
) if m +n is even,
(r (m+n)
2
) if m +n is odd.
Replace the Cauchy problem (1.1) by the discrete problem
(2.1)
t
v
(t, x) = M
x
v
(t, x) +f
(v
)(t, x) ((t, x)
)
v
(0, x) = v
0
(x) ((0, x)
)
v
2
, x) = v
+
(x) ((
2
, x)
)
For t 0, the nonlinearity f
(v
(v
)(t, x) = F
(v
(t, x
2
), v
(t, x
2
), v
(t
2
, x)).
Theorem 2.2. Assume that u
0
is analytic on I
is analytic on D D D C
3d
for each > 0.
Furthermore, assume that
(2.2) [F
(u
+
, u
, u
) f(
u
+
+u
2
)[ K( +[u
+
u
[)
2
holds for all u
+
, u
, u
0
(x) = u
0
(x) ((0, x)
(r))
v
+
(x) = u
0
(x) +
2
_
M
x
u
0
(x) +f(u
0
(x))
_
((
2
, x)
(r))
are C
m+n
(r)
[
m
x
n
t
u(t, x)
m
x
n
t
v
(t, x)[ C
mn
2
. (2.4)
The function u constitutes a classical solution to the Cauchy problem (1.1).
In particular, Theorem 2.2 implies the classical existence Theorem 2.1.
Example 2.3 (A Nonlinear Elliptic Problem). Rewrite the elliptic initial value
problem
2
t
(t, x) +
2
x
(t, x) = g((x, t)) (2.5)
(0, x) =
0
(x),
t
(0, x) =
+
(x) (2.6)
for the scalar function in the form of problem (1.1):
(2.7)
t
_
_
u
(1)
u
(2)
u
(3)
_
_
=
_
_
0 0 0
0 0 1
0 1 0
_
_
x
_
_
u
(1)
u
(2)
u
(3)
_
_
+
_
_
u
(3)
0
g(u
(1)
)
_
_
,
denoting u
(1)
= , u
(2)
=
x
and u
(3)
=
t
. For functions
0
,
+
analytic on
I
0
(x), u
0(3)
=
+
(x) are
analytic, too. Assuming further that the nonlinearity g is analytic near
0
(0), a
solution to (2.7) exists on some = (r). The constraint
x
u
(1)
= u
(2)
propagates,
so := u
(1)
indeed solves (2.5).
CONVERGENCE IN DISCRETE CAUCHY PROBLEMS 5
In order to approximate (2.7) by a discrete problem (2.1), choose
F
(v
+
, v
, v
) =
1
2
(v
+
(3)
+v
(3)
, 0, g(v
+
(1)
) +g(v
(1)
))
as nonlinearity, independent of . The assumptions of Theorem 2.2 concerning
analyticity and boundedness of F
(1)
)/2 proves the estimate (2.2):
1
2
(g(v
+
(1)
) +g(v
(1)
)) = g(
V ) +
1
2
(v
+
(1)
v
(1)
)g
V ) +O([v
+
v
[
2
),
Hence, the component v
(1)
of the discrete solutions to problem (2.1) converges to
the smooth solution on (r) with a suitable positive r < in the sense of (2.4).
3. Proof of Theorem 2.2
The proof consists of three parts: in the rst, a smooth solution u to (1.1) is
constructed as the limit of a suitable sequence of discrete solutions v
to (2.1).
Secondly, C
0
-approximation of u is shown, i.e., (2.4) for m = n = 0. Finally, an
induction argument yields (2.4) for arbitrary m, n.
The proof combines elements from discrete approximation theory for hyperbolic
PDE [BMS] with ideas from Walters proof [Wal] of the Cauchy-Kovalevskaya the-
orem, which is based on the classical papers [Nir], [Nag]. Recall the strategy of
the latter. A solution u to (1.1) is constructed such that u(t, x) is x-analytic on
the time-dependent domain B
(t)
(I
(t)
) for any t [T, +T]. The size parameters
(t), (t) decreases as [t[ increases. The motivation is that
x
is a bounded operator
between the spaces of analytic functions on B
(I
) and on B
(I
), respectively,
for any <
and
. The norm of
x
is determined by the classical Cauchy
estimate
sup
xB(I
)
[
x
u(x)[ (
)
1
sup
xB
(I
)
[u(x)[. (3.1)
Introducing suitable time-dependent norms | |
t
for u(t), an a priori estimate of
Gronwall type is derived from (3.1),
|u(t)|
t
|u
0
|
0
+
_
t
0
C(t, s)|u(s)|
s
ds. (3.2)
Eventually, (3.2) enables one to obtain the solution u as the xed point of a con-
tracting map.
Time-dependent norms for discrete functions are introduced in the following. To
prove approximation of u by v
(t) := u(t) v
(t) at time t = n
2
is estimated in terms of w
(t
2
) and w
= [, +] and B
(I
) = z C [ dist(z, I
) .
The crucial quantity that determines most of the following constructions is r > 0,
the appropriate diameter of the domain (r), which has to be found. From r, one
denes
(t) = r [t[,
n
= r [n[
2
, (t) = 3r 2[t[,
n
= 3r [n[.
6 D. MATTHES
Discrete intervals containing m + 1 points are given by
I
m
=
_
m
2
, . . . , , 0, , . . . ,
m
2
if m is even
m
2
, . . . ,
2
,
2
, . . . ,
m
2
if m is odd
For a function v : I
m
C
d
, dene its dierence quotient,
2
-shift, restriction and
linear interpolation:
x
v : I
m1
C
d
, (
x
v)(x) =
1
_
v(x +
2
) v(x
2
)
_
x
v : I
m1
C
d
, (
x
v)(x) = v(x +
2
), (
1
x
v)(x) = v(x
2
)
v : I
m2
C
d
, (v)(x) = v(x)
Ev : I
2
m
C
d
, (Ev)(x) =
x x
v(x
+
) +
x
+
x
v(x
).
Here x
, x
+
I
m
are such that x
x < x
+
and x
+
= x
+ . The restriction
of a function u : (r) C
d
to the discrete domain
. For
the discrete function v :
(r) C
d
, let v
n
be its restriction to time t = n
2
, so
v
n
(x) = v(n
2
, x), which is dened on the interval I
n
where n
n
.
In these notations, the problem (2.1) takes the convenient form (n > 0):
v
n+1
= v
n1
+
x
v
n
+f
n
(v
) (3.3)
f
n
(v
) = F
(
x
v
n
,
1
x
v
n
, v
n1
) (3.4)
where pointwise evaluation of the arguments is understood.
Without loss of generality, assume that [Mu[ [u[ for all u C
d
, and that
u
0
(0) = 0. Indeed, this can be achieved by a dilation of the t-axis and an ane
transformation of the values of u, respectively. For an appropriate choice of U > 0,
f is dened on D
d
(3U) and F
on D
3d
(3U), respectively.
3.2. Discrete norms and their properties. For > 0, dene a functional on
scalar discrete functions v : I
m
C by
|v|
=
n
k=0
k
k!
max
xI
mk
[
k
v(x)[.
For multi-component functions v = (v
(1)
, . . . , v
(p)
) : I
m
C
p
, let
|v|
= max
i=1,...,p
|v
(i)
|
.
These norms share several properties with the maximum-norm of analytic functions
over a complex domain.
Lemma 3.1. The functionals | |
n
C
d
and 0
|u|
. In addition, each | |
has
the following properties:
(1) Absolute bound: [u(x)[ |u|
for all x I
n
.
(2) Submultiplicativity: |u v|
|u|
|v|
+|
x
u|
|u|
+
.
(4) Restriction estimate: If u : I
C
d
extends analytically to B
(I
) and u
is its restriction to I
n
I
,
(3.5) |u
(1 /
)
1
sup
B
(I
)
[u[
CONVERGENCE IN DISCRETE CAUCHY PROBLEMS 7
(5) Analyticity estimate: Let be given functions v
: I
n
C
d
with n
0 and |v
C
d
, so that E
k
x
v
(k)
converges to
k
x
u uniformly for each k 0
and a suitable subsequence (k) 0 of . Moreover, u possesses a complex
extension to B
(I
) which is bounded by C.
This lemma is proven in the appendix.
A norm with properties (1) and (2) will be called submultiplicative norm in the
following. For an arbitrary submultiplicative norm, the following lemma holds:
Lemma 3.2. Let the analytic function f : D
p
(U) C satisfy
[f(u)[ C([u
(1)
[, . . . , [u
(p)
[)
for all u D
p
(U), with some function C 0 that is non-decreasing in each of its
arguments. Then for each > 1 and every discrete function v : I
n
C
p
with
|v|
n
, and
(3.6) |f(v)|
C(|v
(1)
|
, . . . , |v
(p)
|
).
The constant depends on but not on the submultiplicative norm | |
.
This lemma is proven in the appendix. Two cases of particular interest are
|f(v)|
sup
D
p
(U)
[f[ (3.7)
|f(v
1
) f(v
2
)|
sup
D
p
(U)
[f
[ |v
1
v
2
|
(3.8)
which hold for all v, v
1
, v
2
with |v|
U/ and |v
1
|
, |v
2
|
U/(2), respec-
tively, with > 1 arbitrary and = ().
Estimate (3.7) follows immediately with C sup
D
p
(U)
[f[ in (3.6). To obtain (3.8),
consider the function
f(a, b) := f(a +b, a b) which is analytic in a, b D
p
(U/2),
and let C([a[, [b[) := sup
D
p
(U)
[f
exist on
(r), and a limiting function u can be dened on (r). In fact, r > 0 will be
dened so that the solutions v
on
(r) satisfy
|v
n
|
n
U,
n
=
n
+r = 4r n (3.9)
for all n with
2
[n[ r recall that v
n
(x) = v
(n
2
, x). Choose U > 0 suitable
and let F > 0 be an -independent bound for F
on D
3d
(3U). The basic idea is to
derive (3.9) inductively as follows:
M
n
:= |v
n+1
|
n
+|v
n
|
n
U(
1
2
+
n
2N
) (3.10)
To obtain (3.10) at n = 0, observe that the functions
u
0
and u
0
+
2
(M
x
u
0
+f(u
0
))
are analytic on a complex neighborhood of x = 0. It follows that if r and are
small enough, then the initial data (2.3) satises
|v
0
|
4r
, |v
+
|
4r
U
4
,
taking into account that u
0
(0) = 0 and using property (4) of Lemma 3.1.
8 D. MATTHES
Now suppose (3.10) holds at n 0. Then f
n
(v
) is dened on I
n
, and
M
n+1
|v
n
|
n+1
+|
x
v
n
|
n+1
. .
v
n1
n+1
+
+ |f
n
(v
)|
n+1
. .
F
+|v
n
|
n+1
M
n
+ F.
To estimate
x
v
and f
(v
[ max
n
(
1
n
|v
n
|
n
) U/(2r) (3.11)
[
t
v
[ [
x
v
[ +[f
(v
)[ U/r + F (3.12)
[
2
x
v
[ max
n
(2
2
n
|v
n
|
n
) U/(2r
2
). (3.13)
For each , let v
I
be the restriction of v
to points (x, t)
2
. Note that v
I
is dened on a sublattice of (Z)
2
. Dene the
family v
I
on
(r). By the estimates (3.11) and (3.12), this family is equicontinuous. Hence, the
Arzela-Ascoli theorem applies: there is a sequence
0 such that v
I
converges
to a smooth limit u
I
uniformly on (r). Moreover, by (3.13), it is possible to
choose the sequence
I
converges uniformly to
x
u
I
(cf. the
proof of property (5) in Lemma 3.1). The same procedure is applicable to v
II
, the
restriction to the odd time values t = (2k + 1)
2
; choose a subsequence
of
such that v
II
u
II
and
x
v
II
x
u
II
. Also, f
n
( v
I
) with even n converges to
f(u
I
) uniformly as can be deduced from the estimate (2.2) and respectively for
n odd, f
n
( v
II
) f(u
II
).
Since v
, x) (r),
v
I
(t
, x) = v
I
(t, x) +
_
t
t
(
x
v
II
(s, x) +f
(v
II
)(s, x)) ds +O(
).
Passing to the uniform limit as
0:
u
I
(t
, x) = u
I
(t, x) +
_
t
t
(
x
u
II
(s, x) +f(u
II
)(s, x)) ds.
The roles of u
I
and u
II
can be interchanged. Consequently, both functions are
dierentiable in time and satisfy
t
_
u
I
u
II
_
=
_
0 M
M 0
_
x
_
u
I
u
II
_
+
_
f(u
II
)
f(u
I
)
_
. (3.14)
By property (5) of Lemma 3.1, the estimates |v
n
|
n
U imply for xed t [r, r]
that u
I/II
(t, x) extend x-analytically to B
r+(t)
(I
(t)
), where they are bounded by
U. And since u
I/II
solve the equation (3.14), their analytic continuations do as well.
So u
I
and u
II
are smooth with respect to t, as any t-derivative can be expressed in
terms of x-derivatives and compositions with the analytic function f.
The pair (u
I
, u
II
) is the only solution to (3.14) of that smoothness. This is seen
as follows: In the proof of estimate (2.4) below, u could be any x-analytic and
t-smooth solution to the problem (1.1); no reference to the above construction is
made. In particular, estimate (2.4) applies to (u
I
, u
II
) in place of u and the system
(3.14) in place of equation (1.1), respectively. But only one smooth function can
satisfy (2.4) for all > 0.
CONVERGENCE IN DISCRETE CAUCHY PROBLEMS 9
For symmetry reasons, the unique solution (u
I
, u
II
) to (3.14) with initial conditions
u
I
(0, x) = u
II
(0, x) = u
0
(x) satises u
I
= u
II
=: u on (r). Hence, u solves (1.1)
with u(0, x) = u
0
(x) for x I
r
.
3.4. Approximation in C
0
. For shortness, let u
= [u]
= v
n+1
|
n
+|w
n
|
n
(3.15)
which is dened for
2
[n[ r. Again, only n 0 is considered here, and the
treatment of n 0 is left to the reader. It will be shown that
(3.16) L
n
(1 +B
)L
n1
+C
3
and L
0
D
2
,
leading by the standard Gronwall estimate to
(3.17) L
n
(C
+D
)e
rB
2
which implies (2.4) for m 0 and n = 0.
To prove the estimate (3.16) at an instant of time t = n with n 1, express w
n+1
in terms of u
and v
n1
+
x
w
n
|
n
+|w
n
|
n
(A)
+|f
n
(v
) f
n
(u
)|
n
(B)
+|(M
x
u
n
+f
n
(u
)) (
t
u
)
n
|
n
(C)
The three resulting expressions (A)-(C) as well as the initial conditions
L
0
= |w
0
|
0
+|w
1
|
0
(IC)
are estimated separately in the following.
(A) By property (3), and observing that
n
+ =
n1
,
(A) |w
n
|
n
+|
x
w
n
|
n
+|w
n1
|
n
|w
n
|
n+
+|w
n1
|
n
L
n1
.
(B) With the help of estimate (3.5), one concludes
|f
(v
) f
(u
)|
n
|F
(
x
v
n
,
1
x
v
n
, v
n1
) F
(
x
u
n
,
1
x
u
n
, u
n1
)|
n
sup [DF
[ max(|v
n
u
n
|
n
, |v
n1
u
n1
|
n
)
B
L
n1
.
The constant B
is uniformly bounded on D
p
(U) independently of because of (2.2),
and so are its derivatives. Hence (B) B
L
n1
.
(C) Dene the functions
A
x
(t, x) =
x
u(t, x)
1
(u(t, x +
2
) u(t, x
2
)) (3.18)
A
t
(t, x) =
t
u(t, x)
1
(u(t +
2
, x) u(t
2
, x)) (3.19)
A
f
(t, x) = f(u(x, t)) F
(u(t, x +
2
), u(t, x
2
), u(x
2
, t)). (3.20)
10 D. MATTHES
For t [r, +r], all three functions possess an analytic extension for x B
(t)+r
(I
(t)
).
As u is smooth, one trivially has
[A
x
(x, t)[ C
x
2
and [A
t
(x, t)[ C
t
2
. (3.21)
Furthermore, because of estimate (2.2)
[A
f
(x, t)[ K( +[u(t, x +
2
) u(t, x
2
)[)
2
+sup[Df[
(u(t, x +
2
) +u(t, x
2
))/2 u(t, x)
C
f
2
.
Subtracting 0 = M
x
u + f(u)
t
u from the expression for (C),
(C) = |M
x
[u]
n
+f
n
([u]
)
t
[u]
n
[M
x
u +f(u)
t
u]
n
|
n
|
x
[u]
n
[
x
u]
n
|
n
+|
t
[u]
n
[
t
u]
n
|
n
+|f
n
([u]
) [f(u)]
n
|
n
|[A
x
]
n
|
n
+|[A
t
]
n
|
n
+ |[A
f
]
n
|
n
(C
x
+C
t
+C
f
)
2
= C
2
.
For the conclusive estimate, property (4) of Lemma 3.1 has been used.
(IC) Obviously, w
0
= 0 by (2.3). And v
1
= v
+
is the restriction of u
+
:= u
0
+
t
u
0
,
which is an x-analytic function on B
4r
(I
r
2
) and satises
[u
+
(x) u(
2
, x)[
2
8
sup
0<s</2
[
2
t
u(s, x)[.
Exploiting the estimate (3.5) yields
L
0
|w
1
|
r
D
2
.
3.5. Proof of Smooth Approximation. Since u is smooth on (r), derivatives
and dierence quotients may be interchanged at the cost of O(
2
):
sup
m+n
(r)
[
m
x
n
t
u
m
x
n
t
[u]
[ C
(1)
mn
2
.
This is a consequence of the more general formula (5.19) given in Lemma 5.5 in
section 5. Hence, to prove (2.4) for given m, n, it suces to show that
[
m
x
n
t
w
(x, t)[ C
(2)
mn
2
. (3.22)
Submultiplicative norms for functions v :
(r) R
d
are given by
|v|
(N)
nN
m=1,2,...
m+n
m!n!
sup
(x,t)
m+n
(r)
[
m
x
n
t
v(x, t)[.
Dene positive numbers
N
=
0
/(N +1), where
0
< r is suitably chosen later. It
is inductively shown that
|w
|
(N)
N
C
(3)
N
2
(3.23)
with appropriate constants C
(3)
N
. Obviously, (3.23) implies (3.22) with the choice
C
(2)
mn
= C
(3)
n
m!n!
(m+n)
n
.
Firstly, some properties of the norms are summarized. In contrast to the previously
considered norms | |
n
, the
| |
(N)
|
(N)
max
nN
sup
|t|r
sup
xD(I)
[
n
t
u(t, x)[, (3.24)
where depends on the ratio / only.
For N = 0, inequality (3.23) follows from the previous discussion,
|w
|
(0)
0
max
n
|w
n
|
n
= O(
2
)
because
0
< r
n
. Now assume (3.23) for N 0. By denition of | |
(N)
, and
since v
|
(N+1)
N+1
|w
|
(N)
N+1
+
N+1
N+1
|
t
w
|
(N)
N+1
|w
|
(N)
N+1
+
N+1
N+1
|
x
w
|
(N)
N+1
+
0
N
2
|f
(u
) f
(v
)|
(N)
N
+
0
N
2
|
|
(N)
N
The Cauchy estimate is applied to the the sum of the rst two terms, and inequality
(3.8) to the third term. The norm of
= M
x
[u]
+f
([u]
)
t
[u]
= [A
x
]
+ [A
t
]
+ [A
f
]
with A
x
, A
t
and A
f
dened as in (3.18)(3.20), is estimated using (3.24),
|
|
(N)
N
max
nN
sup
x,t
([
n
t
A
x
[ +[
n
t
A
t
[ +[
n
t
A
f
[) C
(4)
N
2
.
In conclusion,
|w
|
(N+1)
N+1
|w
|
(N)
N
+
0
N
2
sup [Df[ |w
|
(N)
N
+
0
N
2
C
(4)
N
2
_
(1 +
0
N
2
sup [Df[)C
(3)
N
+C
(4)
N
_
2
. (3.25)
As inequality (3.8) has been applied to estimate f
(u
) f
(v
), it needs to be
checked that |u
|
(N)
N
, |v
|
(N)
N
U. To verify the bound for v
, formally estimate
along the same lines as above
|v
|
(N+1)
N+1
|v
|
(N)
N+1
+
N+1
N+1
|
x
v
|
(N)
N+1
+
0
N
2
|f
(v
)|
(N)
N+1
|v
|
(N)
N
+
0
N
2
F |v
|
(0)
0
+
0
F
nN
1/n
2
.
For simplicity, assume that |v
|
(0)
0
<
1
2
U, possibly after diminishing r. As the sum
n=0
1/n
2
is nite, it can be achieved that |v
|
(N)
N
<
3
4
U for all N 0 by choosing
0
small enough. For <
N
so small that Q
N
2
< U/4,
|u
|
(N)
N
|v
|
(N)
N
+|w
|
(N)
N
U.
This justies estimate (3.25) and nishes the proof.
4. A Remark on Numerical Applicability
The iteration (3.3) solving the discrete problem (2.1) is tailored to numerical im-
plementation. Assume, the values of v
n1
and v
n
are known. Then:
v
n+1
(x) = v
n1
(x) +M(v
n
(x +
2
) v
n
(x
2
)) +f
n
(v
)(x). (4.1)
12 D. MATTHES
To investigate the eect of round-o errors, consider the perturbed quantity v
n
(x) = v
n+2
(x) +M( v
n+1
(x +
2
) v
n+1
(x
2
)) +f
n
( v)(x) +
n
(x).
Usually, only an absolute bound on the round-o error is known, [
n
(x)[ 10
q
.
One should think of q as the number of precise digits, which is xed a priori. To
estimate the deviation of v
L
n
(1 +B
L
n+1
+C
3
+E
,
where E
= |
n
|
n
10
q
exp(r/). (4.2)
As before, r is the diameter of the domain . The relation (4.2) predicts instability
of the discrete equation (4.1). A ner mesh size does not necessarily correspond
to a better approximation. One expects that there is an
0
> 0 depending on the
initial data u
0
and radius r > 0 such that the approximation error
decays like
2
for
0
,
grows dramatically like e
1/
for 0 and <
0
.
From (4.2) it is suggested that
0
r/q. The pictures in the following sections
were calculated with r 1, 0.1 and q 10.
5. Cross-Ratio-Equation
The cross-ratio of four mutually distinct complex numbers q
1
, . . . , q
4
C is
CR(q
1
, q
2
, q
3
, q
4
) =
(q
1
q
2
)(q
3
q
4
)
(q
2
q
3
)(q
4
q
1
)
.
Consider (R) as a subset of C, i.e, identify each point (t, x) (R) with the
complex number z = x +it C.
Denition 5.1 ([NQC]). A CR-mapping is a lattice-function
(R) C such
that for all z
(R) C
(5.1) CR
_
(z
+
2
),
(z
+i
2
),
(z
2
),
(z
2
)
_
= 1.
CR-mappings are suitable discrete analogues of holomorphic functions in the sense
that any holomorphic function can be locally approximated.
Theorem 5.2. Assume : (R) C is a holomorphic function with
(0) ,= 0.
Then there are positive constants r < R and C
mn
, so that for each > 0, a CR-
mapping
is dened on
:
sup
z
m+n
(r)
[
m
x
n
t
(z)
m
x
n
t
(z)[ C
mn
2
. (5.2)
Proof: As a rst step, an equation of type (2.1) is derived for an arbitrary CR-
mapping
(R) C. Denote by T
h)(z) = h(z +
2
(i 1)).
The edges
= (T
+
)/,
= (T
)/ (5.3)
CONVERGENCE IN DISCRETE CAUCHY PROBLEMS 13
Figure 2. An Airy function is approximated by a CRmapping.
naturally satisfy the closure condition
+ T
+
+ T
. (5.4)
Let further the function Q
.
In terms of these variables, the cross-ratio-equation (5.1) implies T
+
= (Q
)
1
T
,
and the following two formulas are obtained:
T
=
1 Q
1 +Q
, T
+
=
1 Q
1 +Q
T
+
Q
. (5.6)
Their compatibility condition T
+
(T
) = T
(T
+
(z +i
2
)
Q
(z i
2
)
=
1 +Q
(z +
2
)
1 Q
(z +
2
)
1 Q
(z
2
)
1 +Q
(z
2
)
for all z
(R). (5.7)
Now introduce v
(R) C by
(5.8) i exp(v
(z)) = Q
(z i
2
).
Expressing the relation (5.7) in terms of v
, one obtains
t
v
= i
x
v
+F
(v
) (5.9)
which is in the form of the discrete equation (2.1) with M = i and
F
(v
+
, v
, v
) =
1
2
(g(v
+
) g(v
)) (5.10)
g(V ) = log
_
i exp(V )
1 +i exp(V )
1 i exp(V )
_
. (5.11)
Note that v
is dened on
(R) instead of
up to Euclidean
motions and a homothety.
14 D. MATTHES
Proof of Lemma 5.3. Let z be the bottom of
(r), i.e., z = ir
(r) with
r
2
< r
r. Assign
C. Equation
(5.8) denes Q
from v
on
) = T
(T
+
), because Q
= Q
. Assign
an arbitrary number
C to
on
is indeed
a CR-mapping, combine equations (5.6) and (5.4). Variations of A
correspond to
rigid rotations and dilations of the whole lattice
, variations of
to translations.
is
(5.12) u(z) =
1
2
(z)
(z)
,
which is dened on some (R) with R > 0 small enough (since
(0) ,= 0 by
hypothesis). To justify this ansatz, dene
a
= (T
+
)/, b
= (T
)/ (5.13)
as analogues of
and
(z) = i exp(u(z +i
2
))a
(z) +O(
2
) (5.14)
T
(z) =
1 i exp(u(z +i
2
))
1 +i exp(u(z +i
2
))
i exp(u(z +i
2
))a
(z) +O(
2
), (5.15)
corresponding to (5.5) and (5.6), respectively. The Cauchy-Riemann equation for
u,
(5.16)
t
u = i
x
u,
corresponds to M = i and f 0 in problem (1.1). Since g(0) = g
(0) = g
(0) = 0
by (5.11), one easily veries [F
(v
+
, v
, v
)[ K[v
+
v
[, so the function F
in
(5.10) satises the estimate (2.2). Let v
corresponds to a CR-mapping
. Fix
by choosing
( z) =
( z) and
( z) = (T
+
( z) ( z))/. It remains to be shown that approximation of
u by v
implies approximation of by
|
(N)
=
m+nN
1
m!n!
max
z
m+n
(r)
[
m
x
n
t
w
(z)[. (5.17)
This norm is submultiplicative, so Lemma 3.2 applies. The other essential ingredi-
ent is
Lemma 5.4. For v :
(r) C
d
, and N 0, one has
|v|
(N+1)
C
r
([v(z
1
)[ +[v(z
2
)[ +|
x
v|
(N)
+|
t
v|
(N)
), (5.18)
where z
1
, z
2
(r)
[v(z)[ +|
x
v|
(N)
+|
t
v|
(N)
.
CONVERGENCE IN DISCRETE CAUCHY PROBLEMS 15
Write z
1
= (t
1
, x
1
), z
2
= (t
2
, x
2
) and let z = (t, x)
p=1
[
x
v(z
i
+p
2
)[ +
Q
q=1
[
t
v(z iq +i
2
)[
[v(z
i
)[ + 2r(|
x
v|
(N)
+|
t
v|
(N)
).
So (5.18) holds with C
r
= 1 + 2r.
Proof of Theorem 5.2, continued: At this point, it is sensible to introduce the
notion u
= O
(
p
) for a family of smooth functions u
: (r) C
d
, meaning that
| u
|
C
N
()
< C
N
p
for all N 0, with suitable constants C
N
> 0.
Lemma 5.5. Let u : (r) C
d
be an arbitrary smooth function. For natural
numbers m, n, one has
(5.19) [
m
x
n
t
u]
=
m
x
n
t
[ u]
+O
(
2
).
Moreover, for
x
,
t
R, dene u
(x, t) = u(x
x
, t
t
). Then
u
+
+ u
= 2 u +O
(
2
), (5.20)
u
+
u
= 2(
x
x
u +
t
t
u) +O
(
3
). (5.21)
The identity (5.19) is proven in the appendix. The derivation of the properties
(5.20), (5.21) is analogous.
It is clear from Theorem 2.2 and Lemma 5.5 that |[u]
|
(N)
= O(
2
) for arbitrary
N. It is now shown that this implies
|
|
(N)
= O(
2
). (5.22)
Introduce the functions g
, G
by
(5.23) (z) = exp(g
(z))a( z).
Combining the formulas in (5.6) yields the simple relation
= T
1
(T
+
) =
1
(
x
Q
)(T
1
, so
x
g
= (T
+
v
+v
).
Adopting the notion O
(
2
), and using the rules (5.20) and (5.21),
x
G
(z) =
T
+
T
+
(z) =
(z +
1+i
2
2
) +O
(
2
)
= 2u(z +
1+i
2
2
) +O
(
2
) = (T
+
u +u)(z) +O
(
2
).
Similar identities are derived for
t
g
and
t
G
, so that
|
x
(g
)|
(N)
+|
t
(g
)|
(N)
4|[u]
|
(N)
+O(
2
) = O(
2
).
As
( z) = a
( z) = T
( z) + O(
2
). Hence also g
( z) = G
( z) = 0 and T
( z) =
T
( z) +O(
2
). Lemma 5.4 gives
|G
|
(N)
C
N
|[u]
|
(N)
.
By means of the composition estimate (3.8),
|
|
(N)
| exp(g
) exp(G
)|
(N)
[( z)[ = O(
2
),
16 D. MATTHES
and this implies (5.22). An analogous argument gives the respective approximation
of b
by
.
In the second step,
and a
, b
, respectively.
By denition of these quantities,
(z) =
(z
2
)
(z +
2
),
x
[]
(z) = a
(z
2
) b
(z +
2
),
(z) =
(z i
2
) +
(z +
2
),
t
[]
(z) = a
(z i
2
) +b
(z +
2
).
By construction of
, one has
( z) = T
+
( z) + O(
2
)
because
( z) = a
( z) + O(
2
). Applying Lemma 5.4 to []
and
gives |[]
|
(N)
= O(
2
) and thus the desired nal result (5.2).
6. Schramms Orthogonal Patterns
In [Sch], Schramm-patterns (or orthogonal circle patterns) are proposed as discrete
analogues of conformal maps. A Schramm-pattern C
(Z)
2
a circle C
(x, t) in R
2
C. The dening condition is that circles
belonging to neighboring vertices intersect orthogonally, and circles assigned to
opposite corners of an (Z)
2
-square are tangent. So C
(x, t) and C
(x
, t
) intersect
orthogonally and are tangent, respectively, if [x x
[ +[t t
[ = and if [x x
[ =
[t t
-approximation of
conformal maps. The obtained local result diers in nature from the C
0
-convergence
theorem presented in [Sch], which deals with global boundary value problems.
For notational simplicity, the domain for an orthogonal circle pattern C
is
(r) =
(r) (Z)
2
,
containing half of the grid points of
k
(r), the sets
mn
(r)
m+n
(r) are introduced so that the central dierence quotient
m
x
n
t
of a function
on
mn
(r). Note that
mn
and
nm
do not coincide in general.
Figure 3. An Airy function is approximated by a Schramm-pattern.
By the results in [Sch], a pattern C
(t, x).
CONVERGENCE IN DISCRETE CAUCHY PROBLEMS 17
Theorem 6.1. Given a conformal map : (R) C with
x
(0) ,= 0, there is
a positive r < R, and there exists a family of orthogonal circle patterns C
dened
on
,
sup
(x,t)
mn
(r)
[
m
x
n
t
(x, t)
m
x
n
t
(x, t)[ C
mn
2
. (6.1)
Proof. The function log is harmonic, i.e., satises Laplaces equation
2
x
(log ) +
2
t
(log ) = 0. (6.2)
In terms of u
(1)
=
x
(log ) and u
(2)
=
t
(log ), harmonicity reads
t
_
u
(1)
u
(2)
_
=
_
0 1
1 0
_
x
_
u
(1)
u
(2)
_
. (6.3)
For the radius function
)(
2
t
)(
2
x
)(
2
t
)
(
)
2
=
(
2
x
) + (
2
t
) + (
2
x
) + (
2
t
)
(
2
x
)
1
+ (
2
t
)
1
+ (
2
x
)
1
+ (
2
t
)
1
.
Recall that
x
and
t
denote the
2
-shift in the x- and t-direction, respectively, so
that
2
x
and
2
t
are shifts by . Equation (6.4) is satised by a positive function
(r) R
+
if and only if it is the radius function of a Schramm-pattern.
Introduce functions v
(1)
, v
(2)
by
exp(v
(1)
) = (
x
)/(
1
x
), exp(v
(2)
) = (
t
)/(
1
t
). (6.5)
Since
is given on
(r), v
(1)
and v
(2)
are a priori dened on dierent subsets of
(1)
and
x
v
(2)
coincide,
and the same is true for
x
v
(1)
and
t
v
(2)
. Equation (6.4) and the compatibility
condition
x
v
(2)
=
t
v
(1)
imply formally
t
_
v
(1)
v
(2)
_
=
_
0 1
1 0
_
x
_
v
(1)
v
(2)
_
+
_
0
G
_
, (6.6)
with G
=
1
2
log(H
/H
)
H
(v
+
, v
, v
) = e
v
+
(1)
+e
v
(1)
+e
v
(2)
e
(v
+
(1)
v
(1)
v
(2)
)
.
Suppose v
(r) R
2
is a solution to (6.6), then the equations (6.5) are com-
patible. From the components v
(1)
and v
(2)
, a solution
(r) R
+
to (6.5)
can be constructed and is uniquely determined up to a global scalar factor. (Note
that v
(1)
and v
(2)
are dened at more points than needed to calculate
. Actually,
any solution v
sat-
ises the exponential Laplace equation (6.4), hence is the radius function of some
Schramm-pattern C
on
(r).
H is an analytic function with respect to v
+
, v
, v
=0
log(H
/H
) =
_
0 for k = 0, 1, 2
(v
+
(1)
v
(1)
) h(v) for k = 3
,
18 D. MATTHES
where h is some analytic expression of the vs. Hence [G
[ K[v
+
v
[, implying
the estimate (2.2) for F. Now let v
to u
on a suitable (r).
Make the respective solution
from v
) one obtains C
-convergence of
to .
The radius function
(r) R
+
is accompanied by a function
(r) C,
such that C
(x, t) C.
Theorem 6.2. Under the hypotheses of Theorem 6.1 and for every > 0 small
enough, there exists a Schramm-pattern C
on
:
sup
(x,t)
mn
(r)
[
m
x
n
t
(x, t)
m
x
n
t
(x, t)[ C
mn
2
. (6.7)
Proof. As pointed out before, the radius function alone determines the pattern C
and hence
from and
= exp(i),
where
(x, t) :=
x
(x, t) is holomorphic with respect to the complex variable
z = x +it. The Cauchy-Riemann equations for
read
x
=
t
log = u
(2)
(6.9)
t
=
x
log = u
(1)
, (6.10)
with u dened as before, u
(1)
=
x
(log ), u
(2)
=
t
(log ).
Analogous quantities and relations are now given for an arbitrary Schramm-pattern
C
and d
on
1,0
(r) by
= d
exp(i
), (6.11)
with d
(x+
2
, t)
and
(x +
2
, t), and
0+
,
+0
00
) = (,
0+
,
++
) (
00
,
+0
, )
= arctan
_
++
0+
_
arctan
_
00
+0
_
.
Introducing v
= g
(
t
v
(1)
,
1
t
v
(1)
) (6.12)
g
(v
+
, v
) =
1
_
arctanexp(v
+
) arctanexp(v
)
_
.
From the sketch on the right, it follows that
(
+0
00
,
00
0
) = (
4
,
00
,
3
) (
1
,
00
,
2
)
= arctan
_
00
_
arctan
_
0+
00
_
.
CONVERGENCE IN DISCRETE CAUCHY PROBLEMS 19
oo
+o
o+
++
oo
o
+o
o+
3
4
Figure 4. Relations between centers of adjacent circles.
Expressed in terms of
and v
, this yields:
= g
(
t
v
(2)
,
1
t
v
(2)
). (6.13)
With formulas (6.12), (6.13) at hand, the argument of the proof for Theorem 6.1 is
continued. Let
(r) R
+
be the radius function approximating the conformal
factor of , and v
be
the function of center positions of the unique Schramm-pattern determined by the
radius function
(0) = (0)
and
(0,
2
) = (0,
2
).
The same norms (5.17) are used as in the proof of Theorem 5.2. To prove C
-
convergence of
(v
+
, v
) (v
+
+v
)/2[ K( +[v
+
v
[)
2
.
Restating (6.10) in the form (recall the notion O
(
2
) from section 5)
t
[w]
(x, t) =
1
2
(u(x, t +
2
) +u(x, t
2
)) +O
(
2
),
one obtains, using Lemma 3.2,
|
t
([w]
)|
(N)
|g
(
t
v
,
1
t
v
)
1
2
(
t
v
,
1
t
v
)| +|O
(
2
)|
(N)
K( +|
t
v
|
(N)
)
2
+|[u]
|
(N)
+O(
2
).
Similarly, |
x
([]
)|
(N)
= O(
2
). Now Lemma 5.4 gives |[]
|
(N)
= O(
2
).
Completely analogous estimates hold for the approximation of and
dened by
t
= exp(i ) (6.14)
=
d
exp(i
). (6.15)
Equations (6.12),(6.13) change in the obvious way, and the choice
2
, 0) = (
2
, 0)
implies that
(0,
2
) = /2 + (0,
2
) + O(
2
). Another application of Lemma 5.4
yields |
[]
|
(N)
= O(
2
) and |
[ ]
|
(N)
= O(
2
).
Since the distance between the centers of two orthogonally intersecting circles with
radii
1
and
2
is
d =
_
2
1
+
2
2
=
1
_
1 + (
2
/
1
)
2
,
20 D. MATTHES
the quantities d
and
d
are given by
d
=
_
1+exp(2v
(1)
)
2
(
1
x
),
d
=
_
1+exp(2v
(2)
)
2
(
1
t
). (6.16)
Now observe that
1
=
_
2
=
_
1
x
2
+x
2
2
+O
(
2
) =
_
1+(x/
1
x
)
2
2
(
1
x
) +O
(
2
)
=
_
1+exp(2u
(1)
)
2
(
1
x
) +O
(
2
).
Hence, with Lemma 3.2 it follows
|
x
(
[]
)|
(N)
C
N
(|
[]
|
(N)
+|
[]
|
(N)
) +O(
2
),
|
t
(
[]
)|
(N)
C
N
(|
[]
|
(N)
+|
[ ]
|
(N)
) +O(
2
).
An application of Lemma 5.4 nishes the argument.
7. Appendix
Proof of Lemma 3.1. As a nite sum of norms, | |
n
C,
|uv|
=
n
k=0
k
k!
sup
xI
nk
[
k
x
(uv)(x)[
kn
k
k!
k
=0
_
k
_
sup
xI
n
[
x
u(x)[ sup
yI
nk+
[
k
x
v(y)[
=0
n
m=0
+m
! m!
sup
xI
n
[
x
u(x)[ sup
yI
nm
[
m
x
v(y)[
|u|
|v|
.
Discrete Cauchy estimate: For u : I
n
C
d
,
|u|
+|
x
u|
=
n
k=0
k
k!
max
xI
nk
[
k
x
u(x)[ +
n1
k=0
k
k!
max
xI
nk1
[
k+1
x
u(x)[
k=0
k
k!
(1 +
k
) max
xI
nk
[
k
x
u(x)[
n+1
k=0
( +)
k
k!
max
xI
nk
[
k
x
u(x)[ = |u|
+
because for k = 0, 1, . . ., one has 1 +k/ (1 +/)
k
.
Restriction estimate: With the classical Cauchy estimate
sup
xI
[
k
x
u(x)[ k!(1/
)
k
sup
xB
(I)
[u(x)[
1
The reason formulas (6.16) are employed instead of the symmetric representation d =
q
2
1
+
2
2
is that the former is an analytic expression in , u on some D(U) with arbitrarily
large U as 0, whereas the latter has a singularity at
1
=
2
= 0.
CONVERGENCE IN DISCRETE CAUCHY PROBLEMS 21
it follows that
|[u]
=
n
k=0
k
k!
sup
xI
nk
[
k
x
[u]
(x)[
n
k=0
k
k!
sup
xI
[
k
x
u(x)[
_
n
k=0
(/
)
k
_
sup
xB
(I)
[u(x)[.
Analyticity estimate: For simplicity, assume that all n
are odd. |v
C implies
[
s
v
(x)[ Cs!
s
for all x I
n
and all s n
is equicontinuous.
At s = 0, the Arzela-Ascoli theorem yields a subsequence (0) 0 of so that
Ev
(0)
converges uniformly to a continuous function u. From here, proceed in-
ductively: Assume that E
s
x
v
(s)
converges uniformly to
s
x
u. Apply the Arzela-
Ascoli theorem at s +1 to obtain an innite subsequence (s +1) of (s) for which
E
s+1
x
v
(s+1)
converges uniformly to some u
(s+1)
. To show that u
(s+1)
is indeed
the s + 1-st derivative of u, consider the identity
(
s
x
u
(s+1)
)(x) = (
s
x
u
(s+1)
)(0) +
0j<J
(
s+1
x
u
(s+1)
)(
2
+j)
with arbitrary x = J I
s
. This implies for the interpolated functions
(E
s
x
u
(s+1)
)(x) = (E
s
x
u
(s+1)
)(0) +
_
x
0
(E
s+1
x
u
(s+1)
)(z) dz +O(
)
at arbitrary x I. Pass to the limit
0 on both sides. u
(s+1)
is seen to be the
x-derivative of
s
x
u, so u C
s
(I). From the theorem on dominated convergence, it
follows for x I
s=0
s
s!
[
s
x
u(x)[ C.
So u possesses a convergent Taylor expansion around all x I, with convergence
radius , and the analytic extension is bounded by C.
Proof of Lemma 3.2. As [u
(i)
()[ |u|
n
. The estimate
(3.6) is derived as follows: Since g : D
p
(U) C
d
is analytic, it has a power series
representation
g(u) =
N
p
0
g(0)
!
u
, u
p
j=1
u
j
j
, (7.1)
where = (
1
, . . . ,
p
) denotes a multi-index. For the coecients in (7.1), the
Cauchy integral representation yields
[
g(0)[
!
=
1
(2)
p
_
|1|=m1
d
1
_
|p|=mp
d
p
g(
1
, . . . ,
p
)
1+1
1
p+1
p
F(m
1
, . . . , m
p
)
m
1
1
m
p
p
,
22 D. MATTHES
the m
i
being arbitrary numbers with 0 < m
i
< U. Choosing m
i
= |u
(i)
|
, it
follows from submultiplicativity and the basic norm properties of | |
that
|g(u)|
g(0)
!
|u
F(m
1
, . . . , m
p
)
_
_
u
(1)
m1
_
1
_
u
(p)
mp
_
p
_
||
F(|u
(1)
|
, . . . , |u
(p)
|
).
This proves the claim and also shows that := (1 1/)
p
is independent of the
norm | |
.
Proof of Lemma 5.5. One starts with the representation
m
x
n
t
[u]
(x, t) =
1
m+n
_
[
2
,+
2
]
m
d
m
_
[
2
,+
2
]
n
d
n
m
x
n
t
u(x +
, t + )
of arbitrary partial dierence quotients, where = (
1
, . . . ,
m
) = (
1
, . . . ,
n
) and
the notations
=
m
i=1
i
and =
n
j=1
j
have been used. Then,
(
m
x
n
t
[u]
m
x
n
t
u)(x, t) =
_
d d
m+n
_
1
0
ds
()
m
x
n
t
u(x +s
, t +s )
=
_
d d
m+n
_
1
0
ds
_
()
m
x
n
t
u(x, t)
. .
A(x,t)
+s
_
1
0
ds
2
()
m
x
n
t
u(x +s
, t +s
s )
. .
B(x,t)
_
.
Above,
()
= (
)
x
+ ( )
t
. The integral over A vanishes because
_
d
m
_
d
n
()
f = 0
for an arbitrary , -independent function f. As B(x, t) is an (x, t)-smooth func-
tion, one concludes (5.19).
References
[BMS] A.I. Bobenko, D. Matthes, Yu.B. Suris Discrete and smooth orthogonal systems: C
-
approximation. Int. Math. Res. Not. 45 (2003), 24152459.
[BP1] A.I. Bobenko, U. Pinkall, Discrete isothermic surfaces, J. Reine Angew. Math. 475 (1996),
187208.
[BP2] A.I. Bobenko, U. Pinkall Discretization of surfaces and integrable systems. in: Discrete in-
tegrable geometry and physics (Vienna, 1996), Oxford Lecture Ser. Math. Appl., 16, Oxford
Univ. Press, New York, 1999.
[DHR] P. Doyle, Z.X. He, B. Rodin Second derivatives of circle packings and conformal mappings.
Discrete Comput. Geom. 11 (1994), no. 1, 3549.
[HS] Z.X. He, O. Schramm The C