(0, T; L
2
())
p Q := L
1
(0, T; L
2
0
())
where
L
p
(0, T; X) :=
_
f : (0, T) X 
_
T
0
f
p
X
dt <
_
The existence of solutions was proved by:
R
2
(J.L. Lions)
The data f , u
0
and the smoothness of allow to prove that
u L
2
(0, T; H
2
()) V or u L
(0, T; L
4
()) V
(classical solutions). This is possible if T is small.
It is an open problem to prove:

=k
1
2

u(, t )
2
displays three different regions:
1. For k < k
0
(small), E(k, t ) corresponds to the
macroscopic ow pattern (vortices)
2. For k
0
k k
K
(Kolmogorov) scale, it holds:
E(k, t ) k
5/3
3. For k > k
K
, even if Re is very large ( very small) viscous
effects are able to dissipate these ow scales, which are
considered unresolvable.
Galerkin nite element approximation of the incompressible NavierStokes equations
Statement of the problem
Galerkin nite element approximation of the incompressible NavierStokes equations
Discretization in time using monolithic schemes
Discretization in time using monolithic schemes
Let us write the problem as
t
u + N(u) + L(u) +p = f
u = 0
where N(u) is the convective term and L(u) the rest of terms.
Let 0 = t
0
< t
1
< ... < t
N
= t
n
be a partition of the time interval
and [0, 1]. We shall take the time step size t := t
n+1
t
n
constant for all n. Let
f
n
:= f
n+1
f
n
, f
n
f (t
n
)
f
n+
:= f
n+1
+ (1 )f
n
t
f
n
:=
f
n
t
Galerkin nite element approximation of the incompressible NavierStokes equations
Discretization in time using monolithic schemes
One step approximation to the time derivative
The time discretization is this case leads to the problem: from
known u
n
, nd u
n+1
and p
n+1
such that
t
u
n
+ N
n
(u) + L(u
n+
) +p
n+1
= f
n+
u
n+1
= 0
where N
n
(u) is an approximation of the convective term.
Explicit convection
N
n
(u) = (u
n
)u
n
The resulting scheme is conditionally stable (t must be
small enough).
_
v
u
t
+v (u )u
_
2
_
v [ ((u))]
+
_
v p =
_
v f
Since
_
v [ ((u))] =
_
v : (u) +
_
v (n (u))
=
_
(v) : (u) +
_
v (n (u)),
_
v p =
_
p v +
_
pn v
Galerkin nite element approximation of the incompressible NavierStokes equations
Weak form
we can write
2
_
v [ ((u))] +
_
v p
= 2
_
(v) : (u)
_
p v
_
v (n )
If v is such that v = 0 on
D
and we make use of the Neumann
boundary condition:
d
dt
_
u v +
_
v (u )u + 2
_
(v) : (u)
p v =
_
f v +
_
N
v h
This equation makes sense if u and v are square integrable in
space and have square integrable rst derivatives. The function
space where they have to belong is H
1
().
Galerkin nite element approximation of the incompressible NavierStokes equations
Weak form
Let us introduce the notation
(u, v) =
_
u v (L
2
inner product)
p v = b(p, v)
_
f v +
_
N
v h = l (v)
The weak statement can be written as
d
dt
(u, v) + c(u, u, v) + a(u, v) b(p, v) = l (v)
Galerkin nite element approximation of the incompressible NavierStokes equations
Weak form
with
u(, t ), v H
1
()
u(, t ) = g, v = 0 on
D
u(x, 0) = u
0
(x)
On the other hand, if q is a scalar function
_
q u = b(q, u) = 0
This equation makes sense if q is square integrable, i.e.,
q L
2
().
Galerkin nite element approximation of the incompressible NavierStokes equations
Weak form
Other forms of the viscous and convective terms
If u = 0 and is constant, it is easy to see that
2 [(u)] = u
In this case, a(u, v) can be replaced by
a(u, v) =
_
u : v
but the Neumann boundary condition must be changed.
Also, the convective term can be written as:
(u )u = (u )u +u( u) = (u u)
= (u )u +
1
2
u( u)
The rst expression is the so called conservative form. The
second veries (skewsymmetry): c(u, u
1
, u
2
) = c(u, u
2
, u
1
)
Galerkin nite element approximation of the incompressible NavierStokes equations
Linearization of the convective term
Linearization of the convective term
Let us consider the stationary problem: nd u V and p Q
such that
c(u, u, v) + a(u, v) b(p, v) = l (v) v V
b(q, u) = 0 q Q
What follows can be applied both to the continuous problem
and to the nite element approximation.
Let N
c
, N
a
, N
b
, N
l
, K
a
and K
b
be the constants dened by
c(u, v, w) N
c
u
V
v
V
w
V
a(u, v) N
a
u
V
v
V
, a(v, v) K
a
v
2
V
b(q, v) N
b
q
Q
v
V
, inf
qQ
sup
vV
b(q, v)
q
Q
v
V
K
b
l (v) N
l
v
V
Galerkin nite element approximation of the incompressible NavierStokes equations
Linearization of the convective term
Theorem. The stationary NavierStokes equations have always
a solution u, p V Q. This solution is unique provided
:=
N
c
N
l
K
a
< 1
Assume that we know u
(i 1)
and we want to compute u
(i )
,
where the superescript denotes the iteration counter. Let
u
(i )
= u
(i )
= u
(i 1)
. It holds:
c(u
(i )
, u
(i )
, v) = c(u
(i 1)
, u
(i 1)
, v) + O(u
(i )
)
= c(u
(i 1)
, u
(i )
, v) + O(u
(i )
)
= c(u
(i )
, u
(i 1)
, v) + O(u
(i )
)
= c(u
(i 1)
, u
(i )
, v) + c(u
(i )
, u
(i 1)
, v)
c(u
(i 1)
, u
(i 1)
, v) + O(u
(i )
2
)
Galerkin nite element approximation of the incompressible NavierStokes equations
Linearization of the convective term
We can therefore consider the linearized problem: nd u
(i )
V
and p
(i )
Q such that
c(u
(j )
, u
(k)
, v) + a(u
(i )
, v) b(p
(i )
, v) = l (v) v V
b(q, u
(i )
) = 0 q Q
V
N
l
K
a
, p
(i )
Q
N
l
K
b
_
N
c
N
l
K
2
a
+
N
a
K
a
+ 1
_
2) For all u
(0)
V, if < 1 it holds that
u u
(i )
V
u u
(i 1)
V
p p
(i )
Q
2K
a
+ N
a
K
b
u u
(i 1)
V
and therefore convergence is linear.
Galerkin nite element approximation of the incompressible NavierStokes equations
Linearization of the convective term
Convergence of NewtonRaphsons method
Consider the problem
c(u
(i 1)
, u
(i )
, v) + c(u
(i )
, u
(i 1)
, v) + a(u
(i )
, v) b(p
(i )
, v)
= l (v) + c(u
(i 1)
, u
(i 1)
, v) v V
b(q, u
(i )
) = 0 q Q
Theorem. Assume that < 1 and that
u u
(0)
<
1
C
, C :=
2N
c
K
a
(1 )
,
with < 1. Then, for each i there exists a unique solution to the
linearized problem and it holds that
u u
(i )
V
1
C
2
i
, p p
(i )
Q
C
2
i
,
and therefore convergence is quadratic.
Galerkin nite element approximation of the incompressible NavierStokes equations
Linearization of the convective term
Attraction ball of Newton
Raphsons method
 uu
(0)

Re
Potential
curve
Log E
i
Convergence of Newton
Raphsons method
Galerkin nite element approximation of the incompressible NavierStokes equations
Finite element approximation
Finite element approximation
Let us consider the generalized trapezoidal rule for the time
integration and Picards linearization of the convective term.
The time discrete and linearized problem is
(
t
u
n,i +1
, v) + c(u
n+,i
, u
n+,i +1
, v)
+a(u
n+,i +1
, v) b(p
n+1,i +1
, v) = l (v)
b(q, u
n,i +1
) = 0
for all test functions v V and q Q.
Let {
e
} be a nite element partition of the domain , with
index e ranging from 1 to n
el
. We denote with a subscript h
nite element quantities associated to {
e
}.
Galerkin nite element approximation of the incompressible NavierStokes equations
Finite element approximation
Let V
h
V (at any time level) and Q
h
Q be nite element
spaces to approximate the velocity and the pressure,
respectively.
The Galerkin nite element approximation to the problem is:
(
t
u
n,i +1
h
, v
h
) + c(u
n+,i
h
, u
n+,i +1
h
, v
h
)
+ a(u
n+,i +1
h
, v
h
) b(p
n+1,i +1
h
, v
h
) = l (v
h
)
b(q
h
, u
n,i +1
h
) = 0
Omitting for simplicity the iteration counter, the matrix form of
the problem is given by
M
t
U
n
+ K(U
n+
)U
n+
D
t
P
n+1
= F
n+
DU
n+1
= 0,
where U and P are the arrays of nodal velocities and pressures,
respectively.
Galerkin nite element approximation of the incompressible NavierStokes equations
Finite element approximation
Let us denote the node indexes with superscripts a, b and the
space indexes with subscripts i , j .
Let N
a
u
be the standard shape function associated to the
velocity node a and N
c
p
the standard shape function associated
to the pressure node c. In general, the velocity and pressure
interpolations may (must) be different. The previous matrices
are:
M
ab
ij
= (N
a
u
, N
b
u
)
ij
(
ij
is the Kronecker ),
K(U
n+
)
ab
ij
= (N
a
u
, u
n+
N
b
u
)
ij
+ (N
a
u
, N
b
u
)
ij
D
cb
j
= (N
c
p
,
j
N
b
u
)
The components of these matrices are obtained by grouping
together the left indexes in the previous expressions and the
right indexes.
Galerkin nite element approximation of the incompressible NavierStokes equations
Finite element approximation
The method of lines
Using the Galerkin method, it follows that
Time discretization + Space dicretization
= Space dicretization + Time discretization
The rst approach is more convenient to introduce stabilized
formulations. The second (known as method of lines) starts by
writing the weak form of the continuous problem as: nd
u(, t ) H
1
0
(), with u = u
0
for t = 0, and p(, t ) L
2
() such
that
d
dt
(u, v) + c(u, u, v) + a(u, v) b(p, v) = l (v) v H
1
0
()
b(q, u) = 0 q L
2
()
Galerkin nite element approximation of the incompressible NavierStokes equations
Finite element approximation
As before, the (Galerkin) FE approximation consists of
replacing the spaces H
1
0
() and L
2
() by nite dimensional
subspaces V
h
and Q
h
. The discrete problem is: nd
u
h
(, t ) V
h
, u
h
= u
0
h
for t = 0 and p
h
(, t ) Q
h
such that
d
dt
(u
h
, v
h
) + c(u
h
, u
h
, v
h
) + a(u
h
, v
h
) b(p
h
, v
h
) = l (v
h
)
b(q
h
, u
h
) = 0
for all v
h
V
h
and q
h
Q
h
. In matrix form:
M
d
dt
U + K(U)U D
t
P = F
DU = 0
where U = U(t ), P = P(t ). This is a classical system of DAEs.
Once discretized, we recover the same fully discrete system as
in the previous approach.
Galerkin nite element approximation of the incompressible NavierStokes equations
Finite element approximation
Spacetime nite element approximation
The nite element interpolation in time must be discontinuous.
Otherwise the solution in time would be coupled.
Let
0 = t
0
< t
1
< ... < t
N
= T (partition of [0, T])
I
n
= (t
n
, t
n+1
), Q
n
= I
n
, Q
e
n
=
e
I
n
Integrating the NavierStokes equations multiplied by the test
functions over Q
n
rather than over and considering the test
function v
h
time dependent:
_
I
n
_
u
h
t
, v
h
_
dt +
_
I
n
c(u
h
, u
h
, v
h
) dt +
_
I
n
a(u
h
, v
h
) dt
_
I
n
b(p
h
, v
h
) dt =
_
I
n
l (v
h
) dt
Galerkin nite element approximation of the incompressible NavierStokes equations
Finite element approximation
The temporal term can be integrated by parts as
_
I
n
_
u
h
t
, v
h
_
dt =
_
I
n
d
dt
(u
h
, v
h
) dt
_
I
n
_
u
h
,
v
h
t
_
dt
=
_
u
h
(t
n+1
), v
h
(t
n+1
)
_
_
u
h
(t
+
n
), v
h
(t
+
n
)
_
_
I
n
_
u
h
,
v
h
t
_
dt
Prescribing weakly the condition u
h
(t
+
n
) = u
h
(t
n
):
_
I
n
_
u
h
t
, v
h
_
dt =
_
I
n
_
u
h
,
v
h
t
_
dt
+
_
u
h
(t
n+1
), v
h
(t
n+1
)
_
_
u
h
(t
n
), v
h
(t
+
n
)
_
Galerkin nite element approximation of the incompressible NavierStokes equations
Finite element approximation
The nal fully discrete problem is:
_
I
n
_
u
h
,
v
h
t
_
dt +
_
u
h
(t
n+1
), v
h
(t
n+1
)
_
_
u
h
(t
n
), v
h
(t
+
n
)
_
+
_
I
n
[c(u
h
, u
h
, v
h
) + a(u
h
, v
h
) b(p
h
, v
h
)] dt =
_
I
n
l (v
h
) dt
_
I
n
b(q
h
, u
h
) = 0
Remarks:
2
l (u
h
) u
h
H
1f
H
1
Adding up the momentum equation and the continuity equation
with v
h
= u
h
and q
h
= p
h
we obtain
u
h
2
u
h
H
1f
H
1
Galerkin nite element approximation of the incompressible NavierStokes equations
Finite element approximation
From these estimates we see that:
H
1 b(p
h
, v
h
)
or, equivalently,
inf
p
h
Q
h
sup
v
h
V
h
b(p
h
, v
h
)
p
h
v
h
H
1
where > 0 is independent of h.
Galerkin nite element approximation of the incompressible NavierStokes equations
Finite element approximation
The previous condition is known as infsup or
LadyzhenskayaBabukaBrezzi condition, since:
H
1 +p p
h
L
2
C
_
inf
v
h
V
h
u v
h
H
1 + inf
q
h
Q
h
p q
h
L
2
_
C
_
h
k
u
+ h
k
p
+1
_
Galerkin nite element approximation of the incompressible NavierStokes equations
Finite element approximation
Admissible velocitypressure interpolations
Some of the velocitypressure pairs that satisfy the LBB
condition are:
Discontinuous pressure interpolations
Q
1
/P
0
. Multilinear velocity, piecewise constant pressure. It
is believed that, except in some particular situations, this
element satises the infsup condition. It gives good results
except for the possible appearance of a spurious pressure
mode which can be ltered out.
2
/P
0
. Serendipid second order velocity, piecewise
constant pressure.
Q
2
/P
1
. Multiquadratic velocity, piecewise linear pressure.
P
2
/P
0
. Quadratic velocity, piecewise constant pressure.
P
+
2
/P
0
. Quadratic velocity enriched with an internal bubble,
piecewise linear pressure.
Galerkin nite element approximation of the incompressible NavierStokes equations
Finite element approximation
Galerkin nite element approximation of the incompressible NavierStokes equations
Finite element approximation
Continuous pressure interpolations
P
+
1
/P
1
. Linear velocity enriched with an internal bubble,
linear pressure (minielement).
P
2
/P
1
. Quadratic velocity, linear pressure (TaylorHood).
Q
2
/Q
1
. Multiquadratic velocity, multilinear pressure
(TaylorHood).
Galerkin nite element approximation of the incompressible NavierStokes equations
Finite element approximation
Stabilization methods
Stabilization methods
Contents:
2.4. A general framework: the subgrid scale concept
2.5. Stabilization techniques applied to the NavierStokes
equations
Stabilization methods
A general framework: the subgrid scale concept
A general framework: the subgrid scale concept
Model problem
System of convectiondiffusionreaction equations:
L(U) = F in
L(U) :=
d
i =1
x
i
(A
i
U)
d
i ,j =1
x
i
_
K
ij
U
x
j
_
+SU
U = 0 on
U and F are vectors of D unknowns and A
i
, K
ij
and S are
D D matrices (i , j = 1, ..., n
sd
).
Let W := (H
1
())
D
, W
0
= (H
1
0
())
D
. Weak form of the
problem: nd U W
0
such that
B(U, V) = L(V) V W
0
where
Stabilization methods
A general framework: the subgrid scale concept
B(U, V) :=
_
V
t
x
i
(A
i
U) +
_
V
t
x
i
K
ij
U
x
j
+
_
V
t
SU
L(V) :=
_
V
t
F
Galerkin nite element approximation: nd U
h
W
h,0
such that
B(U
h
, V
h
) = L(V
h
) V
h
W
h,0
Description of the method
Let W = W
h,0
W. The continuous problem is equivalent to
B(U
h
, V
h
) + B(
U, V
h
) = L(V
h
)
B(U
h
,
V) + B(
U,
V) = L(
V)
V
h
W
h,0
,
V
W, where U = U
h
+
U and U
h
W
h,0
,
U
W. The functions
U are called subscales or subgrid scales.
Stabilization methods
A general framework: the subgrid scale concept
Notation:
_
:=
n
el
e=1
_
e
,
_
:=
n
el
e=1
_
e
Integrating by parts within each element:
B(U
h
, V
h
)+
_
U
t
n
i
K
ij
V
h
x
j
d+
_
U
t
L
(V
h
) = L(V
h
)
_
V
t
n
i
K
ij
x
j
(U
h
+
U) d +
_
V
t
L(
U) =
_
V
t
[F L(U
h
)]
Adjoint of the convectiondiffusionreaction operator:
L
(V
h
) = A
t
i
V
h
x
i
x
i
_
K
t
ij
V
h
x
j
_
+S
t
V
h
Stabilization methods
A general framework: the subgrid scale concept
Since diffusive uxes must be continuous across interelement
boundaries, the second equation is equivalent to:
L(
U) = F L(U
h
)+V
h,ort
in
e
U =
U
ske
on
e
for e = 1, ..., n
el
. We call
U
ske
the skeleton of U.
Main problems:
How to choose V
h,ort
?
How to approximate
U
ske
?
V
h,ort
= 0 (this denes the space for
U).
U
ske
= 0
Stabilization methods
A general framework: the subgrid scale concept
It only remains to dene how to solve for
U. Different
approaches, yielding different stabilization methods, are
possible. In any case, the problem to be solved is:
B(U
h
, V
h
)+
_
U
t
L
(V
h
) = L(V
h
) V
h
L(
U) = F L(U
h
) in
e
U = 0 on
e
Algebraic approximation to the Greens function
The problem for the subscales is
L(
U) = F L(U
h
) in
e
U = 0 on
e
The solution can be written using the formula:
U(x) =
_
e
g(x, y) [F L(U
h
)] (y)dy
Stabilization methods
A general framework: the subgrid scale concept
where g(x, y) is the Greens function of the problem (which
now is a matrix) and the integration arguments have been
explicitly indicated.
Let us consider the approximation
g(x, y) (x)(x y)
where (x y) is the Dirac delta function. Using this
approximation:
U(x)
_
e
(x y)(x) [F L(U
h
)] (y)dy
= (x) [F L(U
h
)] (x)
A possible way to determine is to impose that moments of
g(x, y) and (x)(x y) coincide. In particular, if (x) is taken
constant,
=
1
meas(
e
)
_
e
_
e
g(x, y)dxdy
Stabilization methods
A general framework: the subgrid scale concept
In general, g(x, y) is unknown and must be approximated by
other means.
If is taken constant, we call it algebraic approximation to the
subscales. In this case, the stabilized method is:
B(U
h
, V
h
)+
_
[L
(V
h
)]
t
[L(U
h
) F]
= L(V
h
) V
h
Equivalently, the stabilized problem can be written:
B
asgs
(U
h
, V
h
) = L
asgs
(V
h
) V
h
where
B
asgs
(U
h
, V
h
) = B(U
h
, V
h
) +
_
[L
(V
h
)]
t
[L(U
h
)]
L
asgs
(V
h
) = L(V
h
) +
_
[L
(V
h
)]
t
[F]
Stabilization methods
Stabilization techniques applied to the NavierStokes equations
Stabilization techniques applied to the NavierStokes
equations
Problem statement
Incompressible NavierStokes equations with Coriolis forces
and permeability term:
(u )u +
1
2
( u)u u +p+ u + u = f in
p + u = 0 in
u = 0 on
The objective is to formulate a FE method
_
0 0 0
0 0 0
0 0 0
0 0 0 0
_
_
, A
i
=
_
_
a
i
0 0
i 1
0 a
i
0
i 2
0 0 a
i
i 3
i 1
i 2
i 3
0
_
_
S =
_
_
3
2
0
3
1
0
2
1
0
0 0 0
_
_
Functional space: W = (H
1
0
())
n
sd
Q.
Finite element spaces: V
h
(H
1
0
())
n
sd
, Q
h
Q.
Stabilization methods
Stabilization techniques applied to the NavierStokes equations
Adjoint of the operator L for the linearized NavierStokes
equations:
L
(a; V
h
) L
(V
h
) =
_
L
1
(V
h
)
L
2
(V
h
)
_
L
1
(V
h
) = v
h
(a )v
h
q
h
v
h
+ v
h
L
2
(V
h
) = q
h
v
h
Suppose that = 0. If we take
p = p/ and multiply the
continuity equation by , the unit coefcients in the A
i
matrices
become , and the velocity does not change.
Remark: The argument of the scaling of the pressure is not
valid when > 0.
Expression for proposed:
= diag(
1
,
2
)
Stabilization methods
Stabilization techniques applied to the NavierStokes equations
From the convergence analysis:
= diag(
1
,
1
,
1
,
2
)
Terms to be added to the bilinear form of the Galerkin method:
(V
h
)
t
L(U
h
) d
=
_
1
_
v
h
+ (a )v
h
+ v
h
v
h
+q
h
_
_
u
h
+ (a )u
h
+ u
h
+ u
h
+p
h
_
d
+
_
2
(q
h
+ v
h
)(p
h
+ u
h
) d
Stabilization methods
Stabilization techniques applied to the NavierStokes equations
From the analysis of a 1D convection  diffusion  reaction
problem and from the numerical analysis:
1
=
_
c
1
h
2
+
c
2
a
h
+ c
3
  +
_
1
2
helps to improve the control on the divergence of the velocity.
From the convergence analysis:
2
= c
4
+ c
5
a
h + c
6
 h
2
Choice of the constants: c
1
= 4, c
2
= 2, c
3
= 1, c
4
= 4, c
5
= 2
and c
6
= 1 for linear elements. h can be taken as half the
element size for quadratics. In this case c = c
= 1.
Fractional step schemes for the time integration
Fractional step schemes for the time integration
Contents:
3.1. Classical projection methods
3.2. Second order methods for the continuous problem
3.3. Algebraic approach
Fractional step schemes for the time integration
Classical projection methods
Classical projection methods
Description of the method
Theorem (Hodge decomposition)
Given a vector eld w, there exists a unique pair v, such that
w = v + & v = 0
with n v = 0 on the boundary.
We can write
w = P
div
(w) + P
grad
(w)
Fractional step schemes for the time integration
Classical projection methods
The NavierStokes equations can be written as
u
t
+p = (u )u + u +f
u = 0
Since (u/t ) = 0, we have that
u
t
= P
div
[u (u )u +f ]
The idea of the fractional step method (Chorin & Temam) is:
1. Solve
u
t
=
u (
u )
u +f
2. Project
u
t
= P
div
_
u
t
_
Fractional step schemes for the time integration
Classical projection methods
that is, nd and
u
t
such that
u
t
+ =
u
t
u = 0
At the numerical level, there will be an error due to the
uncoupling of the two steps. There are two possibilities:
To treat
u as an independent variable.
To apply the previous decomposition at each time step.
The second option is usually adopted. If the backward Euler
time discretization is employed, it leads to: given u
n
, nd
u
n+1
,
u
n+1
and such that
Fractional step schemes for the time integration
Classical projection methods
1
t
(
u
n+1
u
n
) =
u
n+1
(
u
n+1
)
u
n+1
+f
1
t
(u
n+1
u
n
) + =
1
t
(
u
n+1
u
n
)
1
t
(u
n+1
u
n+1
) + = 0
u
n+1
= 0
Clearly, the variable can be identied with the pressure at the
time step under consideration. Thus, the scheme can be
written as
1
t
(
u
n+1
u
n
) =
u
n+1
(
u
n+1
)
u
n+1
+f
1
t
(u
n+1
u
n+1
) +p
n+1
= 0
u
n+1
= 0
Fractional step schemes for the time integration
Classical projection methods
The stability and convergence estimates can expressed using
the following notation. Let {f
n
} be a sequence of functions. We
dene
{f
n
}
(L
2
) f
n
C < n
{f
n
}
p
(L
2
)
N
n=0
t f
n
p
< , 1 p <
{f
n
}
2
(H
1
)
N
n=0
t f
n
2
1
C <
where is the standard L
2
norm and
1
the H
1
norm.
The forcing term must satisfy
N
n=0
t f
n
2
1
C < ,
Fractional step schemes for the time integration
Classical projection methods
Theorem (Stability)
The solution of the projection scheme satises:
{u
n
}
(L
2
)
{
u
n
}
(L
2
)
2
(H
1
)
{p
n
}
2
(L
2
)
Theorem (Convergence)
Suppose that the solution u, p of the continuous problem
exists, is unique and u C
0
(0, T, H
1
), p C
0
(0, T, L
2
). Dene
e
n
u
:= u
n
u(t
n
),
e
n
u
:=
u
n
u(t
n
) and e
n
p
= p
n
p(t
n
). Then:
e
n
u
0 in
(L
2
) with order O(t )
e
n
u
0 in
2
(H
1
) with order O(t )
e
n
p
0 in
2
(L
2
) with order O(t
1/2
)
Fractional step schemes for the time integration
Classical projection methods
Pressure Poisson equation
In the projection step:
1
t
(u
n+1
u
n+1
) +p
n+1
= 0
u
n+1
= 0
Taking the divergence of the rst equation and using the
second yields:
p
n+1
=
1
t
u
n+1
There are 2 advantages of doing this:
The calculation of u
n+1
and p
n+1
is uncoupled.
At the discrete level (in space), the continuity equation is
modied. This can be explained considering the matrix version
of the problem:
Fractional step schemes for the time integration
Classical projection methods
1
t
M
_
U
n
U
n
_
+ GP = 0
G
t
U
n
= 0
t LP G
t
U
n
= 0
Eliminating
U
n
from the rst equation and inserting it into the
last one yields:
G
t
U
n
+ t
_
L G
t
M
1
G
_
P = 0
This modication provides additional stability to the pressure. In
particular, in some cases it is even possible to use equal
interpolation for u and p (which does not satisfy the infsup
condition).
Fractional step schemes for the time integration
Classical projection methods
Boundary conditions
Suppose that u =
u on is prescribed for the original
continuous problem. The classical strategy for fractional step
methods is to prescribe
u =
u on
n u = n
u on
p
n
= 0 on
However, experience shows that it is possible to prescribe
numerically u =
u on . The explanation to this fact and the
study of methods that allow this is a current research problem.
In particular, it can be explained by looking at the problem from
a purely algebraic point of view.
The (wrong) pressure boundary condition produces a spurious
pressure boundary layer which does not prevent convergence
in
2
(L
2
).
Fractional step schemes for the time integration
Second order methods for the continuous problem
Second order methods for the continuous problem
Second order monolithic discretizations
Second order accuracy has to be obtained by using:
A second order discretization in time.
A second order splitting error.
Concerning the rst point, let us consider monolithic schemes
of the form:
u
n+1
u
n
t
+ N
n
(u) + L(u
n+1/2
) +p
n+1
= f
n+1/2
u
n+1
= 0
where N
n
(u) is an approximation of the convective term, L(u)
is the viscous term and u
n+1/2
= (u
n
+u
n+1
)/2. The most
common schemes used in practice are:
Fractional step schemes for the time integration
Second order methods for the continuous problem
Implicit convection (CrankNicolson)
N
n
(u) = (u
n+1/2
)u
n+1/2
The resulting scheme is unconditionally stable.
Explicit convection (AdamsBashforth/CrankNicolson)
N
n
(u) =
1
2
_
3(u
n
)u
n
(u
n1
)u
n1
_
The resulting scheme is conditionally stable (t must be
small enough).
Second order splitting
Also a splitting strategy with a second order error is needed.
Consider for example the ABCN scheme and dene the
following split:
Fractional step schemes for the time integration
Second order methods for the continuous problem
u
n+1
u
n
t
+ L(u
n+1/2
) + p
n
= f
n+1/2
+ N
n
(u)
u
n+1
u
n+1
t
+p
n+1
p
n
= 0
u
n+1
= 0
where the only values of interest of are 0 and 1. The splitting
error comes in when L(u
n+1/2
) is replaced by L(
u
n+1/2
), where
u
n+1/2
= (
u
n+1
+u
n
)/2. Formally, from the second equation it
is seen that
u
n+1
u
n+1
= O(t
1+
)
provided the pressure converges and p
n+1
p
n
= O(t ).
A further computational advantage is obtained by taking the
divergence of the second equation and using the third. This
yields the problem:
Fractional step schemes for the time integration
Second order methods for the continuous problem
u
n+1
u
n
t
+ L(
u
n+1/2
) + p
n
= f
n+1/2
+ N
n
(u)
t (p
n+1
p
n
) =
u
n+1
u
n+1
u
n+1
t
+p
n+1
p
n
= 0
Observe that only scalar (uncoupled) equations need to be
solved.
Theorem (Second order convergence)
Suppose that the solution u, p of the continuous problem
exists, is unique and u C
0
(0, T, H
1
), p C
0
(0, T, L
2
). Dene
e
n
u
:= u
n
u(t
n
),
e
n
u
:=
u
n
u(t
n
) and e
n
p
= p
n
p(t
n
). Then:
e
n
u
0 in
(L
2
) with order O(t
2
)
e
n
u
0 in
2
(H
1
) with order O(t
2
)
e
n
p
0 in
2
(L
2
) with order O(t
1
)
Fractional step schemes for the time integration
Second order methods for the continuous problem
Viscosity splitting methods
Some splitting techniques also involve the viscous term.
Suppose that the boundary condition is u = 0. For example,
one of the possibilities is the threestep scheme (Natarajan):
First step:
u
u
n
1
t
u
+p
= f + (1 )u
n
(u
n
)u
n
u
= 0
u
= 0 on
Second step:
u
2
t
(1 )u
+ (u
)u
= f + u
= 0 on
Fractional step schemes for the time integration
Second order methods for the continuous problem
Third step:
u
n+1
u
1
t
u
n+1
+p
n+1
= f + (1 )u
(u
)u
u
n+1
= 0
u
n+1
= 0 on
This scheme has the following properties:
The rst and third steps are Stokes problems.
The second step is a Burgers equation.
The scheme is unconditionally stable.
1
,
2
and can be chosen so as to obtain a second order
scheme.
The endofstep velocity satises the correct boundary
conditions. This is the main motivation of viscositysplitting
techniques.
Fractional step schemes for the time integration
Algebraic approach
Algebraic approach
Monolithic time discretization
M
t
U
n
+ K(U
n+
)U
n+
+ GP
n+1
= F
n+
DU
n+1
= 0
Fractional step schemes The original equations can be written
as:
M
1
t
(
U
n+1
U
n
) + K(U
n+
)U
n+
+ GP
n
= F
n+
M
1
t
(U
n+1
U
n+1
) + G(P
n+1
P
n
) = 0
DU
n+1
= 0
Fractional step schemes for the time integration
Algebraic approach
Basic approximation:
K(U
n+
)U
n+
K(
U
n+
)
U
n+
where
U
n+
:=
U
n+1
+ (1 )U
n
.
This leads to:
M
1
t
(
U
n+1
U
n
) + K(
U
n+
)
U
n+
+ GP
n
= F
n+
t DM
1
G(P
n+1
P
n
) = D
U
n+1
M
1
t
(U
n+1
U
n+1
) + G(P
n+1
P
n
) = 0
It is easy to show that this approximation can be thought of as
an incomplete LU decomposition of the system matrix.
Fractional step schemes for the time integration
Algebraic approach
To avoid dealing with DM
1
G, it can be approximated by
DM
1
G L, L
ab
= (N
a
, N
b
)
Matrix L is the standard approximation to the Laplacian
operator.
Final problem to be solved:
M
1
t
(
U
n+1
U
n
) + K(
U
n+
)
U
n+
+ GP
n
= F
n+
t L(P
n+1
P
n
) = D
U
n+1
M
1
t
(U
n+1
U
n+1
) + G(P
n+1
P
n
) = 0
Formally, it is easy to see that the splitting error is of order
O(t ) when = 0 and of order O(t
2
) when = 1 (observe
that O(U
n+1
U
n+1
) = t O(P
n+1
P
n
) in any norm ).
Fractional step schemes for the time integration
Algebraic approach
Equivalent stabilized monolithic formulations
The intermediate velocity is:
U
n+1
= U
n+1
+ t M
1
G(P
n+1
P
n
)
Eliminating it, it is found that:
M
1
t
(U
n+1
U
n
) + K(U
n+
)U
n+
+ E(U
n+
)
+GP
n+1
= F
n+1
DU
n+1
+ t (DM
1
GL)(P
n+1
P
n
) = 0
where E(U
n+
) is the splitting error, of order O(t
1+
):
E(U
n+
) := K(S
n+
)U
n+
+ K(U
n+
)S
n+
+ K(S
n+
)S
n+
S
n+
:= t M
1
G(P
n+1
P
n
)
Fractional step schemes for the time integration
Algebraic approach
The stabilization effect now comes from the term
(1 )t BP
n+1
+ O(t
2
), B := DM
1
GL
Theorem
B is positive semidenite.