GENERAL ITERATIVE APPROXIMATION TO DIFFERENTIAL

EQUATIONS.

CHANG GAO

Abstract. This article provides a general iterative approximation to partial

arXiv:2303.14344v1 [math.AP] 25 Mar 2023

differential equations, and thus establish existence of smooth solution. The

heart of the method is to contract (or expand) the boundary conditions uni-
formly in the domain, then using local and global correspondence to transform
discrete step function to successive integration on the domain. Numerical
scheme is discussed and tested for Navier-Stokes equation.

1. introduction
Let Ω ⊂ Rn be a domian with smooth boundary ∂Ω, let P = P1 , P2 , . . . , Pk be
a partition of Rn . Denote by ui : Ω → Pi unknown vector functions defined on
the chart x, denote yi a one dimensional parameter of the chart y. Let φ be a
coordinate transform from the chart y to x so that
∂φ ∂φ
spanj6=i (∂Ω) = T ∂Ωi spannj=1 (Ω) = T Ω ∪ni=1 T ∂Ωi = T ∂Ω,
∂yj ∂yj
where ∂Ωi has dimension dim(Pi ) − 1. Consider a system of differential equations
that can be written in the form:
∂ui (φy)
=Fi (u, y) ui (∂Ωi ) = w0 i = 1, . . . , l
∂yi
(1.1)
∂uj (φy)
=Fj (u, y) uj (∂Ωj ) = Wj (ui , y) j = l + 1, . . . , k
∂yj
where the functionals Fi , Fj and Wj are completely determined, w0 represents
initial and boundary conditions. Equation (1.1) covers many mathematical physical
equations that appear in nature and industry, such as hydrodynamic equations,
wave equations and statistical equations. For example, the imcompressible Navier-
Stokes equation can be written in the form of equation (1.1):
∂u
= −u · ∇u + µ∆u + f − grad p u(t = t0 ) = u0
∂t
∂p
= q p(∂Ω) = p0
∂yj
∂q ∂q
=− − g(y) q + div (−u · ∇u + µ∆u + f ) q(∂Ω) = q0 (u) k 6= j
∂yj ∂yk
The problem of determining the general conditions for existence of solution to
equation (1.1) is a long standing open problem in mathematics [2]. Besides theo-
retical interests, an iterative algorithm that approximate solution to such equations
accurately and efficiently has potentially a wide range of applications.
Current progress on this problem is too vast to survey thoroughly, to name a
few aspects: weak solutions, finite element methods and various strong solutions
in special cases. Examples of non-uniqueness and finite time blow-up can be found
in literature [1]. The goal of this article is to prove that the conditions specified in
1
2 CHANG GAO

equation (1.1) are necessary and sufficient for the existence of smooth solution, by
constructing an iterative approximation that converge to it.
Theorem 1. The following sequences converge to the solution of equation (1.1) as
lim m, p → ∞:
(1.2)
U (i, 0) =w0
U (i, m + 1) =U (i, 0)+
Z yi
+ lim Fi (U (1, m), . . . , U (l, m), U (l + 1, m, p), . . . , U (k, m, p), y) dyi
p→∞ yi proj ∂Ω
for i = 1, . . . , l. And
U (j, m, 0) =Wj (U (1, m), . . . , U (l, m), y)
U (j, m, p + 1) =U (j, m, 0)+
Z yj
+ Fj (U (1, m), . . . , U (l, m), U (l + 1, m, p), . . . , U (k, m, p), y) dyj
yj proj ∂Ω

for j=l+1,. . . ,k.

Different partition index i < k corresponds to dependent sequences, i.e. each
sequence takes limit in each step of another sequence to converge to the solution
u. Notice that setting U (i, m + 1) = U (i, m) in the inductive expression (1.2)
automatically implies U (i, m) is a solution to equation (1.1), the problem reduce
to proving the convergence of the sequences (1.2).
A corresponding step method for numerical computations is also introduced.
The step method provides a more efficient alternative to finite elements, since each
step takes only linear operations.

2. convergence and error estimate

Notations for F, W, y and others follow previous section. Let L be a Lipschitz
constant that bounds the functionals W (u, y), F (u, y) and boundary conditions
∂φ ∂u
w0 . Recall that ∂y i
are transvesal to the boundary ∂Ω, so that ∂y j
(∂Ω) j 6= i is
completely determined. Let T ≥ |yi − yi proj ∂Ω| be an integration bound for all i.
The convergence for the inductive sequences (1.2) may now be proved.

proof of theorem 1. For simplicity denote f (m) = |U (m + 1) − U (m)| the maximum

of |U (i, m+1)−U (i, m)| for all partition index i = 1, . . . , k. Substitute the inductive
expression (1.2) into f (m)
Z yi
f (m) =| (F (U (m), y) − F (U (m − 1), y)) dyi |
yi proj ∂Ω
Z yi
≤ L|U (m) − U (m − 1)| dyi
yi proj ∂Ω
1
≤ L T f (m − 1)
2
By induction on m, a bound for f (m) is obtained
Z
1 m yi 1
f (m) ≤ L |F (U (0))| yim dyi ≤ Lm+1 T m+1
m! yi proj ∂Ω (m + 1)!
Since
X
∞
lim U (i, m) ≤ U (0) + f (m) ≤ w0 + eL T ,
m→∞
m=0
 GENERAL ITERATIVE APPROXIMATION TO DIFFERENTIAL EQUATIONS. 3

I now conclude that U (i, m) converge uniformly as m → ∞, by Weierstrass M test.

The same reasoning hold for U (j, m, p). 
P ∞
Denote the error term R(m′ ) = m=m′ +1 f (m) for order m′ , which is bounded
by the Lagrange remainder term:
X∞
1 1 ′
(2.1) R(m′ ) ≤ (L T )m ≤ ′
(L T )m +1 eL T .
m! (m + 1)!
m=m +1
′

Notice that the error may increase initially when L T is large, but eventually con-
verge.

3. step method
To obatin numerical solutions efficiently, there is a corresponding step method
to the iterative approximation (1.2), which sometimes takes less computational
resource than successive integration. The basic idea is that when the step size dy
is small enough, the following may be equalized:
Z y0 +dy
u(y0 + dy) = u(y0 ) + F (u, y) dy ≈ u(y0 ) + F (u(y0 ), y0 ) dy.
y0

Consider a simplified differential system of equation (1.1), ∂u

∂y = F (u, y). Suppose
T
the integration interval is T , and the total step number is N = dy . Denote u(y0 +
m dy) by U (m), the iterative expression for each step is obtained:
U (0) = u(y0 )
(3.1)
U (m + 1) = U (m) + F (U (m), m) dy
with U (N ) the numerical solution sought. To prove convergence, it suffice to prove
that the step method (3.1) is actually equivalent to the iterative approximation
(1.2) when lim dy → 0. Observe that
U (m + 1) = U (m) + F (U (m), m) dy
= U (m − 1) + (F (U (m − 1), m − 1) + F (U (m), m)) dy.
By induction on m
X
m
U (m + 1) = U (0) + F (U (m′ ), m′ ) dy.
m′ =0
Then take dy → 0
Z y
lim U (m + 1) = U (0) + F (U (m), y) dy ,
dy→0 y0

which has exactly the same form as the iterative approximation (1.2). I may con-
clude that the step method (3.1) converges to the iterative approximation (1.2) as
dy → 0 by this local and global correspondence.
error bound. Let L be the Lipschitz constant that bounds F (u) as before. In case
of uniform step size dy, denote by u(m) the solution obtained by step method at
m-th step. Denote by u(m) the actual solution. Denote by R(m) = |u(m) − u(m)|
the error term. Then
Z m+1
R(m + 1) = |u(m) + F (u(m)) dy − u(m) − F (u(y)) dy|
m
≤ R(m) + |F (u(m)) − F (u(m + 1))| dy
≤ R(m) + |F (u(m)) − F (u(m))| dy + |F (u(m)) − F (u(m + 1))| dy
≤ R(m) + L R(m) dy + L2 dy.
4 CHANG GAO

Set R(0) = 0, by induction on m

X
N −1
′ X
N −1
′
R(N ) ≤ (1 + L dy)N R(0) + L2 dy (1 + L dy)m = L2 dy (1 + L dy)m .
m′ =0 m′ =0

This is the error bound for the step method at N -th step. The error increases
further away from the boundary.

4. example: Navier-Stokes equations

To demonstrate intuitively the accuracy of the step method (3.1), consider the
imcompressible Navier-Stokes equations:
∂u
= −u · ∇u − grad p + ∆u
(4.1) ∂t
∆p = div(−u · ∇u + ∆u) ,
where the Poisson equation for p is obtained by taking the derivative of incompress-
ible condition div(ut ) = 0 and substituted into the first equation. Finding numerical
solution to Navier-Stokes equation is then equivalent to iterate step function (3.1)
for Poisson equation of p at each step of u. The step function for u is
u(i + 1) = u(i) + (−u(i) · ∇u(i) − grad p(i) + ∆u(i)) dt.
To compare intuitively the accuracy of the step method, consider for example
the Taylor-Green vortex, an analytic solution to Navier-Stokes equation
u = (sin(x) cos(y) e−2t , − cos(x) sin(y) e−2t )
1
p = cos(2x) cos(2y) e−4t .
4
Assuming the solution to Poisson equation of p is known for each u, the step
function for u may be iterated from t = 0, see figure 1.

3 3

2 2

1 1

0 0

-1 -1

-2 -2

-3 -3

-3 -2 -1 0 1 2 3 -3 -2 -1 0 1 2 3

(a) Analytic solution at time 1 ex- (b) Step function at time 1 with
hibits saddle connections. step size 0.25 .

Figure 1. Taylor-Green vortex.

The central difference between solving ODEs’ and PDEs’ with step method is
that PDEs’ take derivatives at each step, while ODEs’ are purely numerical. One
way to optimize this algorithm is to find the closed-form solution to the step func-
tion. This is achieved by treating the differential operators as linear coefficients,
and then solve the corresponding linear difference equation. Though the specific
computation time for a given goal depends on the form of the equation, the shape
 GENERAL ITERATIVE APPROXIMATION TO DIFFERENTIAL EQUATIONS. 5

of boundary and boundary conditions, substituting derivatives into closed form

significantly reduces computation time.
Consider the step function for Poisson equation of p with circular boundary
conditions, denote f (j) = div(−u(i) · ∇u(i) + ∆u(i)) the right hand side of poisson
equation at j-th step away from boundary
p(j + 1) = p(j) + q(j) dr
q(j) 1 ∂ 2 p(j)
q(j + 1) = q(j) + (f (j) + − ) dr
r(j) r(j)2 ∂θ2
where q = ∂p∂r . The right hand side of q is just laplacian in polar coordinates, where
∂
the differential operators satisfy ∂r = √ −x2
∂
2 ∂x
+ √ −y
2
∂
2 ∂y
∂
and ∂θ ∂
= −y ∂x ∂
+x ∂y .
x +y x +y
Another optimization approach is to consider stable solution of the corresponding
linear system, which will not be addressed in this article. The step method intro-
duced above adapts to systems of arbitrary dimensions and order, should enough
boundary conditions be supplmented.

5. conclusions
The iterative approximation described in this article estabishes the existence of
smooth solution to a wide class of, if not all, partial differential equations. Addition-
aly, the corresponding step method provides a potentially more efficient numerical
scheme for solving equations than existing methods, if the algorithm is properly
optimized. No doubt the method opens up new paths to the understanding of
turbulence bifurcations phenomena.
One of the key element of the method is choosing the appropriate differentiation
transformation, as described at the beginning of this article, so the entire bound-
ary continuously contract to a single point. The local and global correspondences
connects discrete steps of this contraction to successive integration on the entire
domain.

References
[1] Tristan Buckmaster, Theodore Drivas, Steve Shkoller, and Vlad Vicol, Formation and devel-
opment of singularities for the compressible Euler equations., Proceedings of the International
Congress of Mathematicians 2022, DOI 10.4171/ICM2022/210.
[2] Charles L. Fefferman, Existence and Smoothness of the Navier-Stokes
Equation, The Millennium Prize Problems, Clay Mathematics Institute,
https://www.claymath.org/sites/default/files/navierstokes.pdf (2006).