POLONICI MATHEMATICI LXIX.1 (1998)
On the method of lines for a non-linear heat equation with functional dependence
by H. Leszczy´ nski (Gda´ nsk)
Abstract. We consider a heat equation with a non-linear right-hand side which de- pends on certain Volterra-type functionals. We study the problem of existence and con- vergence for the method of lines by means of semi-discrete inverse formulae.
1. Introduction. Let a > 0, τ
0, τ
1, . . . , τ
n∈R
+and [−τ, τ] = [−τ
1, τ
1]×
. . . × [−τ
n, τ
n]. Define E = [0, a] × R
n, E
0= [−τ
0, 0] × R
n, E
+= (0, a] × R
nand B = [−τ
0, 0] × [−τ, τ]. If u : E
0∪ E → R and (t, x) ∈ E, then we define the Hale-type functional u
(t,x): B → R by u
(t,x)(s, y) = u(t + s, x + y) for (s, y) ∈ B. Because we also take into account the functional dependence on the gradient D
xu = (D
x1u, . . . , D
xnu), we write
(D
xu)
(t,x)= ((D
x1u)
(t,x), . . . , (D
xnu)
(t,x)).
Denote by ∆ the Laplacian, that is, ∆ = D
x1x1+ . . . + D
xnxn. Define Ω := E × C(B, R) × C(B, R
n).
Given f : Ω → R and φ : E
0→ R, we consider the Cauchy problem D
tu(t, x) = ∆u(t, x) + f (t, x, u
(t,x), (D
xu)
(t,x)), (1)
u(t, x) = φ(t, x) for (t, x) ∈ E
0. (2)
Two specific examples of (1) are equations with Volterra integral and delayed (deviated) dependence:
D
tu(t, x) = ∆u(t, x) + e f t, x,
\
B
u(t + s, x + y) dy ds,
0\
τ0
D
xu(t + s, x) ds (integral dependence), D
tu(t, x) = ∆u(t, x) + e f t, x, u
12t, x
, D
xu(t − τ
0, x + τ )
(delays), where e f : E × R × R
n→ R.
1991 Mathematics Subject Classification: Primary 65M20; Secondary 35K05, 35R10.
Key words and phrases : method of lines, stability, consistency.
[61]
Define
f [u](t, x) := f (t, x, u
(t,x), (D
xu)
(t,x)).
The following system of integral-functional equations is equivalent to the differential-functional problem (1), (2):
u(t, x) = L
0[u](t, x), (3)
D
xu(t, x) = L
′[u](t, x), (4)
for (t, x) ∈ E
0∪ E, where L
′= (L
1, . . . , L
n), and L
0[u](t, x) :=
\
Rn
H(t, x − y)φ(0, y) dy
+
t
\
0
\
Rn
H(t − s, x − y)f[u](s, y) dy ds, L
i[u](t, x) :=
\
Rn
H(t, x − y)D
yiφ(0, y) dy
+
t
\
0
\
Rn
D
xiH(t − s, x − y)f[u](s, y) dy ds, for (t, x) ∈ E
+:= (0, a] × R
n(i = 1, . . . , n), where
H(t, x) =
1 (2 √
πt)
nexp
− kxk
24t
for (t, x) ∈ (0, ∞) × R
n,
0 for (t, x) ∈ (−∞, 0] × R
n,
is the Green function H : R
1+n→ R, and
L
0[u](t, x) := φ(t, x), L
i[u](t, x) := D
xiφ(t, x) (i = 1, . . . , n), for (t, x) ∈ E
0.
We intend to formulate a semi-discrete problem which corresponds to (1), (2), and next to find a semi-discrete version of (3), (4). This seems to be a new approach to the error analysis of the method of lines. We study so-called C
0,1solutions to (1), (2), that is, u ∈ C(E
0∪ E, R) satisfying (2) in E
0and (3) in E with the continuous gradient D
xu. However, it will oc- cur that consistency requirements lead to classical, even sufficiently regular, solutions.
In [12] we obtained some existence results by means of the Banach con- traction principle and discussed the question of their continuous differen- tiability in the set E
+, getting C
0,1and classical solutions to the Cauchy problem. Fundamental notions, ideas and existence results in the theory of parabolic equations can be found in [8, 10].
Theoretical search for some iterative methods, esp. monotone iterative
techniques (cf. [3, 9]), reveals certain advantages of Chaplygin’s method (see
[2, 4]). This method can guarantee the second-order convergence under some natural assumptions. However, one would like to avoid using it because of the inevitable necessity to solve quasi-linear Cauchy problems with functional dependence at each stage of the iterative method. There are well known and frequently applied finite difference methods (FDM) (cf. [5, 6, 11, 13–16]).
Explicit FDMs for parabolic equations require a very specific condition on the time and space steps, whereas there appear large non-linear systems at each stage of any implicit FDM.
Enormous progress in parallel software has pointed to semi-discrete methods such as the Rothe method and the method of lines. The latter gives a large-scale structured system of non-linear ordinary differential equations which can be solved by means of an effective Runge–Kutta method (see [17]). Since the early sixties, the method of lines has become very attractive for many mathematicians and engineers. We draw the reader’s attention to some references on parabolic differential and differential-functional equa- tions such as [7, 18, 19]. Because these papers develop a sort of maximum principle as a main tool of their convergence proof, the present paper essen- tially differs from them. Namely our method can be extended not only to parabolic equations with functional dependence at spatial derivatives, but also to strongly coupled systems of parabolic differential-functional equa- tions. The assumptions in the present paper correspond to those in the existence results.
Now, we introduce a natural mesh and a family of semi-discrete schemes.
First, we take the steps h = (h
1, . . . , h
n) ∈ R
n+. Define x
β= h ⋆ β :=
(h
1β
1, . . . , h
nβ
n) for β ∈ Z
n. Let
Z
h= {x
β| β ∈ Z
n}.
Define some discrete sets associated with the sets E
0and E:
E
h0= [−τ
0, 0] × Z
h, E
h= [0, a] × Z
h, E e
h= E
h0∪ E
h, E
h+= (0, a] × Z
h. If u : E
0∪ E → R, then we write u
β(t) = u(t, x
β) for (t, x
β) ∈ e E
h.
We also need some further notation. Let e
l= (δ
1,l, . . . , δ
n,l), where δ
j,lis the Kronecker symbol. Define the difference operators
∆
h= (∆
1,h, . . . , ∆
n,h), ∆
2h= ∆
21,h+ . . . + ∆
2n,h, as follows:
∆
l,hu
β(t) = (2h
l)
−1(u
β+el(t) − u
β−el(t)),
∆
2l,hu
β(t) = h
−2l(u
β+el(t) − 2u
β(t) + u
β−el(t)) (l = 1, . . . , n),
for (t, x
β) ∈ e E
h.
Suppose that we are given the interpolation operators T
h: C( e E
h, R) → C(E
0∪ E, R) and T
h′= (T
h1, . . . , T
hn) : C( e E
h, R
n) → C(E
0∪ E, R
n), where
(T
hu)(t, x) = X
xβ∈Zh
u
β(t)p
βh(x),
(T
h′U )(t, x) = ((T
hu
1)(t, x), . . . , (T
hu
n)(t, x))
for (t, x) ∈ E
0∪ E and u ∈ C(E
0∪ E, R), U = (u
1, . . . , u
n) ∈ C(E
0∪ E, R
n), where p
βh(t) ∈ C(R
n, R) for h ∈ R
n+and β ∈ Z
n.
We formulate the method of lines:
d
dt u
β(t) = ∆
2hu
β(t) + f (t, x
β, (T
hu)
(t,xβ), (T
h′∆u)
(t,xβ)) on E
h+, (5)
u
β(t) = φ
β(t) on E
h0, (6)
where φ : e E
h→ R is a discrete perturbed counterpart of the function φ.
Finite difference schemes for parabolic problems were considered in [1, 11, 16]. The convergence theorems were proved there by means of difference inequalities or a sort of maximum principle. In [13] we prove a convergence theorem for finite difference schemes that approximate unbounded solutions to parabolic problems with differential-functional dependence by means of a comparison lemma, which was possible in absence of functionals acting on partial derivatives. Nevertheless, there were some technical problems.
The present paper shows new ways to solve parabolic equations with more complex functional dependence, such as delay and Volterra type integrals, esp. acting also on partial derivatives.
Define the set F
pβfor p = 1, 2, . . . and β ∈ Z
nas follows:
σ ∈ F
pβif σ = (σ
1, . . . , σ
p) and σ
1, . . . , σ
p∈ {±e
k| k = 1, . . . , n}.
Set
C
0,hβ=
1 for β = 0,
0 for β 6= 0, C
p,hβ= X
σ∈Fpβ
h
−2σ1. . . h
−2σp(p = 1, 2, . . .), and
[h]
2= X
n j=1h
−2jfor h ∈ R
n+. Define
(7)
(8) H
pβ(t) =
( C
p,hβexp(−2tn[h]
2)t
p/p! for t > 0, p = 0, 1, . . . , 0 for t ≤ 0, p = 0, 1, . . . , If u : e E
h→ R satisfies the equation
(9) d
dt u
β(t) = ∆
2hu
β(t) + g
β(t) on E
h+,
where g : E
h+→ R, then we can rewrite (9) as follows:
d
dt (u
β(t) exp(2tn[h]
2))
= n X
nl=1
h
−2l(u
β+el(t) + u
β−el(t)) + g
β(t) o
exp(2tn[h]
2) and next integrating from 0 to t we obtain
u
β(t) = u
β(0) exp(−2tn[h]
2) +
t
\
0
n X
nl=1
h
−2l(u
β+el(s) + u
β−el(s)) + g
β(s) o
exp(−2(t − s)n[h]
2) ds, or, in explicit form,
u
β(t) = X
∞ p=0X
η∈Zn
H
pβ−η(t)u
η(0) + X
∞ p=0X
η∈Zn t
\
0
H
pβ−η(t − s)g
η(s) ds on E
h, where H
pβ(t) are defined by (7)–(8).
If we take g
β(t) := f (t, x
β, . . .) in (9), then we get the following discrete inverse formula for the scheme (5), (6):
u
β(t) = X
∞ p=0X
η∈Zn
H
pβ−η(t)φ
η(0)
+ X
∞ p=0X
η∈Zn t
\
0
H
pβ−η(t − s)f
η[u](s)ds on E
h, where
f
β[u](t) = f (t, x
β, (T
hu)
(t,xβ), (T
h′∆u)
(t,xβ)).
Define the residual expression (10) Θ
β[u; h](t) = d
dt u
β(t) − ∆
2hu
β(t) − f
β[u](t)
for (t, x
β) ∈ E
h+and u ∈ C( e E
h, R) differentiable with respect to t. Observe that f
β[u](t) is a semi-discrete version of the Nemytski˘ı’s operator.
2. Existence and convergence results. We use the symbol C
Bto indicate classes of bounded continuous functions. Write
X [φ] := {(u, U) ∈ C
B(E
0∪ E, R
1) × C
B(E
0∪ E, R
n) |
u(t, x) = φ(t, x), U (t, x) = D
xφ(t, x) for (t, x) ∈ E
0}.
Denote by k · k
0the supremum norm.
Assumption 1. Suppose that:
1) φ, φ ∈ C
B(E
0, R), D
xφ, D
xφ ∈ C
B(E
0, R), f (·, ·, 0, 0) ∈ C
B(E, R).
2) There are L
1, L
2∈ R
+such that
|f(t, x, w, W ) − f(t, x, w, W )| ≤ L
1kw − wk
0+ L
2kW − W k
0for (t, x, w, W ), (t, x, w, W ) ∈ Ω (recall that Ω = E × C(B, R) × C(B, R
n)).
Assumption 2. For every step h and for all x
β∈ Z
hwe have p
βh∈ C(R
n, R), and there are λ ≥ 1 and M ∈ R
+such that
p
βh(x) = 0 for kx − x
βk
0> hλ, X
xβ∈Zh
p
βh(x) = 1 and kp
βh(x)k
0≤ M on R
n.
Assumption 3. Suppose that v ∈ C
B(E
0∪E, R) is a classical solution to the problem (1), (2) such that (v, D
xv) ∈ X [φ] and D
xixiv ∈ C
B(E
0∪ E, R) for i = 1, . . . , n. Assume that the function D
xv is uniformly continuous with respect to x in E
0∪ E, and D
xixiv is uniformly continuous with respect to x
ifor i = 1, . . . , n. Denote their moduli of continuity by σ
xand σ
xx, respectively.
First, we formulate a lemma on global estimates for a system of integral equations.
Lemma 1. Suppose that ε
0, ε
1, P , Q, L, L
′, Q
j, S
j, S
j′∈ R
+for j = 1, . . . , n. If
L
′S < 1 for S := max
j=1,...,n
S
jQ + 2 S
j′Q
3j, and W
0, W
j∈ C([0, a], R
+), where
W
0(t) = ε
0+
t
\
0
{P + LW
0(s) + L
′W
1(s)} ds, (11)
W
j(t) = ε
1+
t
\
0
{S
je
−Q(t−s)+ S
j′(t − s)
2e
−Qj(t−s)} (12)
× {P + LW
0(s) + L
′W
1(s)} ds (j = 1, . . . , n), then
W
0(t) ≤ ε
0+ e ε L
exp
tL 1 − L
′S
− 1
,
W
j(t) ≤ ε
1+ e ε
S
jexp
1−LtL′S− exp(−Qt)
(1 − L
′S)Q + L + 2S
j′exp
1−LtL′SQ
j+
1−LL′S 3(j = 1, . . . , n),
where ε = P + Lε e
0+ L
′ε
1.
P r o o f. Define
f W (t) = P + LW
0(t) + L
′max
j=1,...,n
W
j(t) (t ∈ [0, a]).
Then f W satisfies the integral inequality f W (t) ≤ e ε + max
j=1,...,n
n
\t0
W (s){L + L f
′(S
je
−Q(t−s)+ S
j′(t − s)
2e
−Qj(t−s))} ds o
≤ e ε +
t
\
0
f W (s)L ds
+ max
j=1,...,n
n
t\0
W (t)L f
′(S
je
−Q(t−s)+ S
j′(t − s)
2e
−Qj(t−s)) ds o
≤ e ε +
t
\
0
f W (s)L ds + f W (t)L
′max
j=1,...,n
∞\0
S
jQ e
−ξdξ +
∞
\
0
S
j′Q
3jξ
2e
−ξdξ
. Hence (by the Gronwall lemma) we get
(13) W (t) ≤ f ε e
1 − L
′S exp
tL 1 − L
′S
. From (11)–(13), we get the assertions of our lemma.
Lemma 2 (Existence). If Assumptions 1–2 are satisfied, then there exists a unique bounded and continuous solution to (5), (6).
P r o o f. The right-hand side of the system satisfies the Lipschitz condi- tion in the Banach space of all bounded continuous functions. Apply the Banach contraction principle.
We say that a particular method of lines (e.g. (5)) is stable if small per- turbations of its right-hand side and initial data result in a correspondingly small variation of its solutions. The method (5) will be called consistent with the differential equation if, given a regular solution v to (1), we get
|f
β[v](t) − f[v](t, x
β)| ≤ C
h′for all (t, x
β) ∈ E
h+, where R
+∋ C
h′→ 0 as khk
0→ 0.
Given K ∈ R
+and h ∈ R
n+, define
I
K(h) = {h ∈ R
n+| h ≤ h, h
j/h
l≤ K (j, l = 1, . . . , n)}.
Lemma 3 (Stability). Suppose that u, v ∈ C
B( e E
h, R) and there are C
h, C
h, P
h∈ R
+such that
|v
β(t) − u
β(t)| ≤ C
h→ 0 on E
0h,
k∆
hv
β(t) − ∆
hu
β(t)k
0≤ C
h→ 0 on E
0h,
Θ
β[u; h](t) = 0, |Θ
β[v; h](t)| ≤ P
h→ 0 on E
h+. If Assumptions 1–2 are satisfied and K ≥ 1, h ∈ R
n+, then
sup
β∈Zn
k(v
β− u
β, ∆
h(v − u)
β)k
0→ 0 as khk
0→ 0, h ∈ I
K(h).
P r o o f. Set ω
β(t) := v
β(t) − u
β(t). Then
|Θ
β[v; h](t)| ≤ P
hon E
h+,
|ω
β(t)| ≤ C
h, k∆
hω
β(t)k
0≤ C
hon E
h0, and
ω
β(t) = X
∞ p=0X
η∈Zn
H
pβ−η(t)ω
η(0) (14)
+ X
∞ p=0X
η∈Zn t
\
0
H
pβ−η(t − s)
× {(f
η[v](s) − f
η[u](s)) + Θ
η[v; h](s)} ds,
∆
j,hω
β(t) = X
∞ p=0X
η∈Zn
∆
j,hH
pβ−η(t)ω
η(0) (15)
+ X
∞ p=0X
η∈Zn t
\
0
∆
j,hH
pβ−η(t − s)
× {(f
η[v](s) − f
η[u](s)) + Θ
η[v; h](s)} ds for (t, x
β) ∈ E
hand j = 1, . . . , n. Observe that
(16)
X
∞ p=0X
η∈Zn
H
pβ−η(t) = 1 on E
h+,
(17)
X
∞ p=0X
η∈Zn
|∆
j,hH
pβ−η(t)|
≤ h
−1jX
∞ p=0exp(−2t[h]
2) t
pp!
× n
2
n−1([h]
2)
p+ p(p − 1)(2[h]
2− h
−2j)
p−2h
−4j− X
n l=1, l6=jh
−4lo
= h
−1jn
2
n−1exp(−t[h]
2) + t
2exp(−th
−2j) h
−4j− X
n l=1, l6=jh
−4lo
on E
h+(j = 1, . . . , n).
Define M
λ= M (2λ + 1)
nand kγk
0(t) = sup
s≤t, xη∈Zh
kγ
η(s)k
0for γ ∈ C
B( e E
h, R
k).
It follows from Assumption 2 that k(T
hω)
(t,xβ)k
0= sup
(s,y)∈B
|(T
hω)(s + t, y + x
β)|
≤ X
xβ∈Zh
|ω
β(s)| |p
ηh(s + t, y + x
β)|
≤ M(2λ + 1)
nsup
s, xη∈Zh
|ω
η(s)| = M
λkωk
0(t).
Hence
k(T
h∆
j,hω)
(t,xβ)k
0≤ M
λk∆
hωk
0(t) (j = 1, . . . , n),
for (t, x
β) ∈ e E
h. Taking supremum on both sides of (14), (15) and applying (16), (18), we obtain the integral estimates
kωk
0(t) ≤ C
h+
t
\
0
{P
h+ M
λL
1kωk
0(s) + M
λL
2k∆
hωk
0(s)} ds,
k∆
j,hωk
0(t) ≤ C
h+
t
\
0
h
−1j{P
h+ M
λL
1kωk
0(s) + M
λL
2k∆
hωk
0(s)}
× n
2
n−1exp(−(t − s)[h]
2) + t
2exp(−(t − s)h
−2j) h
−4j−
X
n l=1, l6=jh
−4lo ds
(j = 1, . . . , n), for 0 ≤ t ≤ a. Take W
0, W
j(j = 1, . . . , n) given by (11), (12) with
L = M
λL
1, L
′= M
λL
2, Q = [h]
2, P = P
h, ε
0= C
h, ε
1= C
h, Q
j= h
−2j, S
j= h
−1j2
n−1,
S
j′= h
−1jh
−4j− X
n l=1, l6=jh
−4l(j = 1, . . . , n).
There exists h
′∈ I
K(h) such that θ
K(h) := M
λL
2max
j=1,...,n
h
−1j2
n−1[h]
2+ 2h
5jh
−4j−
X
n l=1, l6=jh
−4l< 1
for each h ∈ I
K(h), h ≤ h
′. In view of Lemma 1, there is a global estimate for the comparison problem, thus we have the estimates
(18) kωk
0(t) ≤ W
0(t) ≤ C
h+ C e
hM
λL
1exp
tM
λL
11 − M
λL
2θ
K(h)
− 1
, (19) k∆
j,hωk
0(t)
≤ W
j(t) ≤ C
h+ e C
hh
−1j2
n−1exp
1−MtMλL1λL2θK(h)
− exp(−t[h]
2) (1 − M
λL
2θ
K(h))[h]
2+ M
λL
1+ 2h
−1jexp
1−MtMλL1λL2θK(h)
h
−2j+
1−MMλL1λL2θK(h)
3h
−4j−
X
n l=1, l6=jh
−4l(j = 1, . . . , n),
for h ∈ I
K(h
′), where
(20) C e
h= P
h+ L
1C
h+ L
2C
h.
It is clear that kW
jk
0→ 0 as khk
0→ 0, h ∈ I
K(h
′) (j = 0, . . . , n), which completes the proof.
Lemma 4 (Consistency). If Assumptions 1–3 are satisfied, then the scheme (5) is consistent with the differential-functional problem.
P r o o f. Take v as in Assumption 3. Suppose that σ
xand σ
xxare the moduli of continuity for D
xv and D
xixiv respectively. Let (t, x
β) ∈ E
h+and (s, y) ∈ B. Then we can use the Taylor expansion at x = x
β+ y to derive (T
hv)
(t,xβ)(s, y) − v(t + s, x)
= X
xη∈Zh
p
ηh(x)(v
η(t + s) − v(t + s, x))
= X
xη∈Zh
p
ηh(x)
1
\
0
D
xv(t + s, ζx
η+ (1 − ζ)x) ◦ (x
η− x) dζ, where z ◦ z denotes the scalar product in R
n, and
(T
h∆
j,hv)
(t,xβ)(s, y) − D
xjv(t + s, x)
= X
xη∈Zh
p
ηh(x){[(∆
j,hv)
η(t + s) − D
xjv(t + s, x
η)]
+ [D
xjv(t + s, x
η) − D
xjv(t + s, x)]}
= X
xη∈Zh
p
ηh(x) n
[D
xjv(t + s, x
η) − D
xjv(t + s, x)]
+ 1 2h
j1
\
0
(1 − ζ)[D
xjxjv(t + s, x
η+ ζh
je
j)h
2j− D
xjxjv(t + s, x
η− ζh
je
j)h
2j] dζ o for j = 1, . . . , n. Thus, we get
k(T
hv)
(t,xβ)− v
(t,xβ)k
0≤ M
λkD
xvk
0nλkhk
0, k(T
h′∆
hv)
(t,xβ)− (D
xv)
(t,xβ)k
0≤ M
λn σ
x(λkhk
0) + 1 2
1
\
0
(1 − ζ)σ
xx(2khk
0)khk
0dζ o . Finally, we obtain
|f
β[v](t) − f[v](t, x
β)|
≤ M
λL
1kD
xvk
0nλkhk
0+ L
2σ
x(λkhk
0) + σ
xx(2khk
0) khk
04
and
|∆
2j,hv
β(t) − D
xjxjv(t, x
β)| ≤ σ
xx(h
j) (j = 1, . . . , n).
These estimates complete the proof.
Theorem 1 (Convergence result). Suppose that Assumptions 1–3 are satisfied , and there are C
h, C
h∈ R
+such that
|φ
β(t) − φ
β(t)| ≤ C
h→ 0 on E
h0, k∆
hφ
β(t) − ∆
hφ
β(t)k
0≤ C
h→ 0 on E
h0.
Let K ≥ 1, h ∈ I
K(h) and θ
K(h) < 1. If u ∈ C
B( e E
h, R) is a solution to (5), (6), then
(21) sup
β∈Zn
k(v
β− u
β, ∆
h(v − u)
β)k
0→ 0 as khk
0→ 0, h ∈ I
K(h).
P r o o f. The existence of bounded solutions to (5), (6) is a consequence of Lemma 2. It is obvious that C
h→ 0 and C
h→ 0 as khk
0→ 0. In view of Lemma 4, we can define
P
h= σ
xx(khk
0) + M
λL
1kD
xvk
0nλkhk
0(22)
+ L
2σ
x(λkhk
0) + σ
xx(2khk
0) khk
04
.
Then P
hand e C
h, defined by (20), tend to 0 as khk
0→ 0. Assertion (21)
follows immediately from estimates (18), (19) in the proof of Lemma 3 and
from the evident fact that kD
xv − ∆
hvk
0→ 0 as khk
0→ 0.
The right-hand sides of estimates (18), (19), show that the convergence rate depends on C
h, C
h, P
h. We formulate a higher-order convergence state- ment.
Corollary 1. Suppose that the assumptions of Theorem 1 are satisfied, and
1) C
h/khk
20and C
h/khk
20are uniformly bounded.
2) For every x ∈ R
nand for every h ∈ R
n+, we have X
xβ∈Zh
p
βh(x)(x
β− x) = 0
3) There are bounded and continuous derivatives D
xxxv in E
0∪ E and D
xjxjxjxjv in E (j = 1, . . . , n).
Then the second-order convergence of the method of lines holds true.
P r o o f. We verify the estimates of the two terms which appear in the proof of Lemma 4, that is,
X
xη∈Zh
p
ηh(x){D
xjv(t + s, x
η) − D
xjv(t + s, x)}
= D
xjxv(t + s, x) ◦ (x
η− x)
+
1\
0