The projection method for the incompressible Navier-Stokes equations:

The pressure near a no-slip wall

A.W. Vreman
AkzoNobel, Research Development & Innovation, Process Technology, P.O. Box 10, 7400 AA Deventer, The Netherlands

Preprint of a paper published in Journal of Computational Physics 263, 353-374 (2014)

Abstract
An explicit staggered projection method for the incompressible Navier-Stokes equations with no-slip
walls is analyzed and used in simulations to address several issues related to the pressure boundary
condition required when the continuity equation is replaced by the standard pressure Poisson equation
(PPE), ∇2 p = ∇ · (−u · ∇u + f ). First, it is shown that a PPE system supplemented with a Neumann
pressure boundary condition derived from the momentum equation can be made consistent with the
Navier-Stokes equations if it is extended with the requirement that ∇ · ∇2 u = 0 is zero near the wall and
the solution is sufficiently smooth. This implies that it is possible to formulate a boundary condition
for the standard PPE without the necessity to resort to Green’s functions, which is interesting for
theoretical reasons. Second, the equivalence is shown between the staggered projection method and the
staggered discretization of above PPE system. The derivation of the equivalence sheds light upon the
so-called PPE paradox and leads to an approximation of the wall value of ∂p/∂n, which is not required
but implied by the staggered projection method. Third, the (near-wall) regularity of a solution of the
Navier-Stokes equations is numerically analyzed by means of Direct Numerical Simulation of turbulent
channel flow performed with the staggered projection method. From the numerical inspection of all
terms of the momentum equation in the near-wall region, it is concluded that the three components of
the momentum equation are satisfied on the wall for t > 0 (for short times, but also in the turbulent
regime). In the limit t → 0, the pressure gradient is observed to converge to the initial pressure gradient in
the L2 -norm, which confirms a disputed theoretical result in literature. Even in the maximum norm, the
pressure gradient appears to converge to the initial pressure gradient. The only discontinuities observed
in the simulations are the discontinuities of the tangential viscous terms and the time derivatives of the
tangential velocities on the wall at t = 0. Thus the numerical results indicate that the regularity of the
solution for turbulent channel flow is stronger than claimed by existing theory.
Keywords: projection method, staggered methods, Navier-Stokes equations, pressure Poisson
equation, pressure boundary condition, DNS of turbulent channel flow

1. Introduction

The behaviour of the pressure near a no-slip wall and in particular the appropriate wall boundary
condition for the pressure Poisson equation deduced from the Navier-Stokes equations have been widely
discussed in literature [1, 2, 3, 4, 5, 6, 7, 8, 9]. The literature on these topics easily causes confusion;
contradicting statements, claims and even contradicting theorems can be found. In the present paper
these issues are reconsidered and interpreted with use of Direct Numerical Simulation (DNS) of turbulent
channel flow.
Before we mention the aims and approach of the present paper in more detail, we specify the Navier-
Stokes equations and mention several pressure Poisson equations that can be found in literature. We
also recapitulate some relevant theoretical results on the regularity of solutions for the Navier-Stokes
equations in a wall-bounded domain.

Email address: [email protected], [email protected] (A.W. Vreman)

 The Navier-Stokes equations are formulated for incompressible plane channel flow in the spatial
domain Ω = [0, Lx ] × (0, Ly ) × [0, Lz ] for time t ≥ 0. The two parallel walls are given by Γ =
[0, Lx ] × {0, Ly } × [0, Lz ], such that Ω = Ω ∪ Γ. Pressure and velocity are denoted by p and u = (u, v, w),
respectively, and depend on location x = (x, y, z) and t. The pressure is defined as a function with a
volume integral of zero. In streamwise direction (x) and spanwise direction (z), periodic boundary con-
ditions are imposed on p and u. In the normal direction, no-slip boundary conditions are imposed to the
velocity. The velocity prescribed on Γ is denoted by uΓ . The Navier-Stokes equations for incompressible
flow in the space-time domain Ω × (0, ∞) read:

∇·u = 0, (1)
2
∂u/∂t = −u · ∇u − ∇p + ν∇ u + f , (2)
Z
u = uΓ on Γ; u, p periodic in x, z; pdV = 0; u(x, 0) = u0 (x). (3)
Ω

Here ν is the kinematic viscosity, f (x, t) a smooth (C ∞ ) forcing term, and u0 (x) a smooth initial velocity
field with ∇ · u0 = 0 in Ω and u0 (x) = uΓ (x, 0) for x ∈ Γ. The velocity uΓ (x, t) is defined as a smooth
vector field on Γ that satisfies the periodic boundary conditions in x and z. The normal velocity on
the wall, vΓ , is a function with the property that its integral over Γ is zero. For the time being the
local value of uΓ is not assumed to be zero, to generalize the formulation to sliding walls, nonpermeable
walls or velocity specified inflow boundary conditions. Equations (1-3) are frequently called the primitive
equations, we will refer to these equations as system NS. It is stressed that in system NS no boundary
condition for the pressure on Γ is specified.
Application of the divergence to Eq. (2) and ∂/∂t to Eq. (1) provides the standard pressure Poisson
equation (PPE)
∇2 p = ∇ · (−u · ∇u + f ). (4)
With the differentation the order of the equation has increased and therefore a boundary condition for
the pressure is required to solve Eq. (4). One procedure to obtain the required boundary condition is the
matrix influence method proposed by Kleiser and Schumann [2] in the context of spectral methods. With
the use of Green’s functions, this method provides a Dirichlet boundary condition for p on Γ by imposing
∇ · u = 0 on Γ at the new time level. The system of equations (1-4) equipped with the matrix influence
Dirichlet boundary condition for p will be called PPE0. The implementation of PPE0 is complicated.
A simpler procedure to specify a wall boundary condition for the pressure Poisson equation is the
derivation of a Neumann pressure boundary condition from the momentum equation in the normal
direction. This procedure is used in system PPE1 [5], defined by (2-4) and

∂p ∂vΓ
= ν∇2 v + fy − uΓ · ∇v − on Γ. (5)
∂y ∂t

The last two terms vanish if uΓ = 0, and also if uΓ does not depend on time and ∇ · u = 0 on Γ (since
then ∇v = 0 on Γ). Gresho and Sani [5] discussed PPE1 in detail and formulated an hypothesis that
NS and PPE1 would be equivalent. However, this hypothesis was falsified by the same authors in later
work [7] and [9]; an example in section 3.10.3 of Ref. [7] shows that PPE1 is ill-posed.
Gresho and Sani [5] also introduced system PPE2, defined by (2-3), Poisson equation

∇2 p = ∇ · (−u · ∇u + ν∇2 u + f ), (6)

and boundary condition (5). The so-called PPE paradox ”if you include it, you do not need it, but if
you do not include it, you do need it” [7] refers to the inclusion of the term ν∇ · ∇2 u, which is zero,
into the right-hand side of this Poison equation. The standard Poisson equation (4) is also called the
simplified Poisson equation [5], while (6) is also called the consistent Poisson equation.
In the derivation of a pressure boundary condition from the momentum equation it is not obvious
why the momentum equation in the normal direction should be chosen and whether the choice of one
of the other momentum equations would imply a different solution. This ambiguity was pointed out in
several papers, see for example [1, 4], and led to a theorem with a corollary that PPE2 would be ill-posed
[8], but according to Ref. [9] this statement is incorrect.

2
 The discussion in literature about the PPE1 and PPE2 prompts research question I: Is there another
way to extend PPE1 such that it becomes consistent, without the term ν∇ · ∇2 u in the Poisson equation,
but with the Neumann boundary condition? In answer to this question a new PPE system is proposed,
PPE3, which consists of equations (2-5) and the additional near-wall requirement

∇ · ∇2 u = 0 in the neighbourhood of Γ, (7)

i.e. the divergence of the Laplacian of the velocity should be zero for x ∈ Ω infinitesimally close to Γ.
Thus system PPE3 is essentially PPE1 extended with condition (7).
The difficult and confusing part of PPE systems is related to the pressure boundary condition of the
PPE at the wall. We are therefore interested in the behaviour of the pressure gradient near a no-slip wall.
For a numerical investigation of the behaviour of the pressure gradient near the wall, it is convenient to
use a pressure-based method which does not require a boundary condition for the pressure on Γ. The
Marker and Cell (MAC) method introduced by Harlow and Welch [10] is such a method, since it is a
direct discretization of system NS (1-3), on a staggered grid. Another well-known method to solve system
NS is the projection method, introduced by Chorin [11] and Temam [12]. In case of explicit treatment
of the convective and implicit treatment of the viscous terms, which introduces a splitting error, the
projection method is also called the fractional step method [1, 13]. The projection of an intermediate
velocity on the space of divergence free functions leads to an implicit equation for the pressure. On a
staggered grid and entirely formulated in discrete space the definition of a pressure boundary condition
can be avoided [13], such that, like the MAC method, the projection method can also be regarded as a
direct discretization of system NS [3, 6, 13, 14]. The formulation of the projection method on staggered
grids in discrete space is appealing, but it is still unclear how we should interpret the circumvention
of the pressure boundary condition, since the discrete implicit equation is very similar to a discretized
Poisson equation for the pressure. This prompts research question II: Which continuous pressure Poisson
equation problem leads to the standard staggered projection method, that apparently does not require a
pressure boundary condition?
In the answers to this and the first research question, PPE3 plays a central role. System PPE3 is
not introduced for computational efficiency but for theoretical purposes, such as the clarification of the
connection between PPE systems and the projection method, the clarification of the PPE paradox, and
the derivation of the wall value of ∂p/∂n implied by the projection method. An appealing theoretical
feature of PPE3 is that it is just the standard Poisson equation supplemented with the Neumann pres-
sure boundary condition and a simple additional condition that can be implemented without a matrix
influence method and Green’s functions.
If the pressure boundary condition for a pressure Poisson equation is based on a component of the
momentum equation, like in Eq. (5), it is important that the terms used in that component of the
momentum equation are continuous in the near-wall limit. A quantity is called discontinuous in the
near-wall limit if the space-time limit approaching some point on the wall does not exist. Possible
discontinuities in the tangential directions are an important reason to derive the boundary condition
from the normal component momentum equation and not from the tangential ones; in Refs. [5, 7] the
use of the normal component has been justified with reference to an analytical result in Heywood and
Rannacher [16], but unfortunately precisely this result is contradicted in later literature by another
analytical argument [8], as will be explained later on. For these reasons we will numerically explore the
regularity of a solution of the Navier-Stokes equation with a focus on the near-wall behaviour of the
pressure gradient and other terms in the momentum equation.
The main reason of the problems and contradictions in literature on PPE systems seems to be that,
despite the smooth initial condition, the solution of NS for wall-bounded flows is not necessarily smooth.
Problems particularly arise in the limit t → 0, since the initial condition does not necessarily satisfy the
compatibility conditions that follow from the requirement of infinite differentiability in space and time
of the solution on Γ at t = 0 [15, 16]. To specify a divergence free initial condition u0 that satisfies the
compatibility conditions seems desirable, but this appears to be very difficult. Consider for example the
conditions that would arise if we would require continuous ∂u/∂t on Γ, which can be expressed as three
boundary conditions for the initial pressure p0 :
∂uΓ
∇p0 = ν∇2 u0 + f 0 − u0 · ∇u0 − (x, 0) on Γ, (8)
∂t

3
in combination with ∇2 p0 = ∇ · (−u0 · ∇u0 + f 0 ), where f 0 = f (x, 0). To find a divergence-free u0 ,
such that a p0 exists that satisfies above three boundary conditions and the Poisson equation, is far from
trivial.
At this point we rephrase two theoretical regularity results for system NS:
• H1980 (Heywood [15]; Theorem 9, see also [17]). There exists a classical solution (u, p) of problem
NS on some time interval (0, T ), such that (a) u is smooth (C ∞ ) on Ω × (0, T ), (b) u is continuous
on Ω×[0, T ) (so also at t = 0 and at the wall Γ), and (c) the first- and higher-order time derivatives
of u are continuous on Ω × (0, T ) (so also on Γ, but not necessarily at t = 0).
• HR1982 (Heywood & Rannacher [16]; Proposition 2.1). Define the pressure p0 as the solution of
the (generalized) Poisson equation in system PPE1 at t = 0. The solution p0 exists, is unique up
to a constant, and is continuously differentiable on Ω. A classical solution u, p of NS satisfies

k∇p − ∇p0 k2 → 0 if t → 0, (9)

where k.k2 denotes the L2 -norm on Ω.

The first result (H1980) is important, but it does not state that the existing solution is unique nor that
it exists for infinite time. Apparently, HR1980 does not rule out ill-posedness of system NS. The second
result (HR1982), which has been used to justify PPE systems with a Neumann boundary condition [5],
is controversial, in the sense that Lemma 3 in Ref. [8] claims the opposite, namely that the pressure
gradient is discontinuous at t = 0 (limt→0 k∇p − ∇p0 k2 > 0).
Reflecting on what is theoretically known about the initial boundary value problem of the Navier-
Stokes equations in wall-bounded domains and the apparent confusion this problem has caused, one may
wonder whether Direct Numerical Simulation (DNS) of wall-bounded incompressible flow started from an
’incompatible’ initial condition has a sufficiently sound mathematical basis. This and the relation of this
topic with PPE systems prompts research question III: What is the near-wall behaviour of a solution of
the Navier Stokes equations obtained in typical DNS, in particular the behaviour of the pressure gradient
near the wall at t = 0 and in the chaotic turbulent regime? To answer this question, we will revisit
DNS of incompressible turbulent plane channel flow at Reτ = 180 [18, 19, 20, 21, 22, 23, 24, 25, 26].
We will precisely define the initial boundary value problem solved by the DNS, and show accurate and
well-validated results that focus on the research questions above.
The same staggered projection method is used for the investigations of both research questions II
and III. This projection method is embedded into a low-storage explicit second-order Runge-Kutta
scheme. The pressure is frequently only first-order accurate in projection or fractional step methods
[13, 14, 27, 28], but it will be shown that the second-order temporal accuracy of the present method is
second-order, provided the pressure at time level n is not identified with the pressure computed in the last
stage of time step n, but with the pressure computed in the first stage of the next time step (specifying the
time level on the Poisson equation with right-hand side r, pn+1 is defined by ∇2 pn+1 = rn+1 = rn+1,0 ;
that Poisson equation is solved in the first stage of the next time step). This result was obtained
independently of a recent and interesting paper of Sanderse & Koren [14], where the same notion can
be found. The result is important in the present context, because it shows that it is possible to use the
projection method to compute the pressure at time zero.
The main new points of the present paper are related to the above three research questions:
1. There exists a pressure Poisson equation system (PPE3), which is based on the standard Poisson
equation (4) with Neumann pressure boundary condition (5). The additional near wall condition
makes PPE3 equivalent to NS for smooth solutions, and a simple implementation of PPE3, which
does not involve Green’s functions, exists.
2. An analytical derivation and a numerical verification that a straightforward discretization of PPE3
on a staggered mesh is equivalent to the staggered discretization of NS (projection method) and to
the staggered discretization of PPE2. The derivation of these equivalences shows how the staggered
projection method, which does not need a pressure boundary condition, is related to PPE systems,
which do need a (derived) pressure boundary condition. The derivation also sheds light on the
PPE paradox and leads to an approximation of the wall value of ∂p/∂n, which is not required but
implied by the projection method.

4
 3. A numerical investigation of the regularity properties described by H1980 by means of accurate and
well-validated DNS of plane channel flow that develops from a clearly prescribed three-dimensional
initial condition into a fully developed turbulent flow. The numerical solution appears to have the
properties of the (possibly non-unique) solution described by H1980, in particular spatial continuity
of the time derivative on the wall (H1980c): the vector equation ∇p = ν∇2 u + f − u · ∇u − ∂u/∂t
on Γ, initially violated, appears to be valid immediately after t = 0, and also in the turbulent
regime. In fact the numerical results indicate stronger regularity than described in H1980; the
only discontinuities observed are those of the tangential viscous terms and the time derivatives of
the tangential velocities on the wall at t = 0. The numerical results also indicate that HR1982
is correct and Lemma 3 in Ref. [8] is not correct, as in the limit t → 0 the pressure gradient is
shown to converge to the initial pressure gradient in the maximum norm, which is even stronger
than HR1982.
The remainder of this paper is organized as follows. In section 2 we will present the discretization
of system NS by an explicit projection method on a staggered grid. In section 3, we will analyze the
equivalences between NS and the PPE systems, first for the continuous case and then more extensively
for the discrete cases (research questions I and II). In sections 4-5, a specific channel flow case will be
considered. In section 4, the simulation cases will be defined, and the spatial and temporal accuracy
of the projection method of section 2 will be validated. In section 5, the discrete equivalence between
PPE3 and NS will be numerically verified and the pressure boundary condition provided by PPE3 will be
evaluated in a DNS (research question II). Furthermore, the near-wall behaviour of the pressure gradient
and other terms of the Navier-Stokes equation will be numerically investigated both in the limit t → 0
and in the turbulent regime, and theorems H1980 and HR1982 will be revisited using the DNS results
(research question III). Conclusions will be presented in section 6.

2. An explicit projection method on a staggered grid

2.1. Temporal discretization
For the time integration in projection methods it is common to use the explicit second-order Adams-
Bashforth method for the convective terms and the implicit second-order Crank-Nicolson method for
the viscous terms, see e.g. [13, 29, 30]. Due to the time splitting error, the pressure is then not
the true pressure but a pseudo-pressure. Several authors [13, 27, 28] demonstrated that the formal
temporal accuracy of the Adams-Bashforth / Crank-Nicolson method is only first-order in wall-bounded
flows. Perot [13] proposed an improvement that made the velocity second-order, but the pseudo-pressure
remained first-order accurate in time. A second-order prediction of the pressure can be obtained after a
correction with a term proportional to the Laplacian of the pseudo-pressure [31].
Two drawbacks of the Adams-Bashforth method are that it is not self-starting (un−1 is required) and
unconditionally unstable in the inviscid limit. These limitations are overcome in a frequently employed
variant, in which the convective terms are treated with an explicit Runge-Kutta method [32, 33, 34],
and according to the latter two references, the pressure at the new time level is then identified with
the pressure computed in the last stage of the time step. Several computations of incompressible flow
have been performed with an explicit Runge-Kutta method applied to both convective and viscous terms
[26, 14]. Runge-Kutta methods are self-starting, have no time-splitting errors if the pressure is computed
in each stage, have in general good stability properties, and can be made higher-order accurate in time,
also for the pressure [14].
The present projection method is embedded in a low-storage Runge-Kutta scheme with Runge-
0
Kutta coefficients αm (m = 0, ..., M − 1), and associated coefficients αm = αm ∆t. Each time step n + 1
n+1 n
corresponds to ∆t = t − t and consists out of M stages. The time levels of the stages are denoted by
0
tn,m , where tn,0 = tn and tn,m+1 = tn,0 + αm for 0 ≤ m < M . Thus tn,M = tn+1 , provided αM −1 = 1. A
quantity superscripted with n, m is evaluated at time tn,m . At the beginning of each time step the start
velocity field un at time tn is stored as un,0 . Then M stages follow. Stage number m + 1 (0 ≤ m < M )
contains three substeps:
1. compute the intermediate velocity
u∗ = un,0 + αm
0
[−un,m · ∇un,m + ν∇2 un,m + f n,m ]; (10)

5
Figure 1: Stability regime for the low-storage three-stage RK method with coefficients α0 = 1/3, α1 = 1/2, and α2 = 1.

2. solve the Poisson equation

∇2 pn,m = (∇ · u∗ )/αm
0
; (11)
3. update the velocity
un,m+1 = u∗ − αm
0
∇pn,m . (12)
The velocity un+1 at the new time level is equal to un,M . The unconventional element in this scheme is
that the pressure computed in stage m is labeled with index m instead of m + 1, although the latter is
common in literature, see for example [33, 34]. In the present paper, the pressure at time tn is defined
by pn = pn,0 . With this definition of pn , not only the velocity, but also the pressure signal satisfies
the formal order of accuracy of the Runge-Kutta method, provided the normal velocity on Γ is time-
independent [14]. The definition implies that at the end of a time step the pressure at level n + 1 is not
known yet; it becomes available in the first stage of the next time step.
For the simulations in this paper, a three-stage Runge-Kutta method (M = 3) was chosen, with
α0 = 1/3, α1 = 1/2 and α2 = 1. This low-storage scheme is third-order accurate for linear and second-
order accurate for nonlinear systems. The M -stage Runge-Kutta method for general coefficients αm
can be found in Ref. [35]. The Taylor expansion and the requirement of M th -order accuracy for linear
systems imply the coefficients αm = 1/(M − m), where M = 2 recovers modified Euler, M = 3 the three-
stage method, and M = 4 the four-stage method in Jameson & Baker [36]. The four-stage method, also
second-order accurate, has successfully been applied in DNS of turbulent flows [37, 38]. The stability
region of the three-stage method is defined by |1 + α2 Z + α1 α2 Z 2 + α0 α1 α2 Z 3 | ≤ 1 and shown in Fig.
1. Since the imaginary axis is inside the stability region for Im(z) ≤ 31/2 , 31/2 is the maximum Courant
number for stability in the inviscid limit. There are several third-order accurate three-stage Runge-Kutta
schemes [14], but the simplicity of this particular three-stage scheme is appealing.

2.2. Spatial discretization

The spatial discretization is defined on a staggered grid, where the pressure and velocity divergence are
defined at cell centers and the velocity components at cell faces [10]. Staggered finite difference methods
have several advantages above colocated finite difference and spectral methods. Unlike colocated finite
difference methods, staggered methods lead to a strong coupling of neighbouring pressure points without
additional treatment. For the simulation of flows in complex geometries, finite difference methods are
much more flexible than spectral methods. The staggered second-order scheme [10] has been generalized
to higher order [39, 23], to body-fitted grids [40] and is also frequently used in immersed boundary

6
methods for flows with complex geometrical features [41]. In the present study the staggered second-
order scheme is used [10]. The interpolation of the convective velocities is such that the convective
scheme is also energy-conserving on a nonuniform grid [23].
We consider a staggered grid that is uniform in the x- and z-direction (mesh spacings hx and hz ), but
nonuniform in the wall-normal direction, the y-direction. Let N be the number of cells in the y-direction.
The staggered grid locations where v is defined are given by

ys,j = 21 Ly (1 + tanh(−γ + 2γj/N )/ tanh(γ)), j = 0, ..., N, (13)

where the end locations (j = 0, N ) coincide with a wall. The cell-center locations are defined by

yc,j = 21 Ly (1 + tanh(−γ + 2γ(j − 12 )/N )/ tanh(γ)), j = 0, ..., N + 1, (14)

where j = 1, ..., N refers to the cells inside the domain. The last equation also defines the cell-center
locations of a layer of ghost-cells outside the domain (yc,0 and yc,N +1 ). The grid spacing in the normal
direction is defined by

hc,j = ys,j − ys,j−1 , j = 1, ..., N, (15)

hs,j = yc,j+1 − yc,j , j = 0, ..., N + 1. (16)

The stretched grid is smooth in the entire domain also near the wall, and we will clarify that this property
is crucial for the computation of the near-wall limit.
The discrete u is defined on the entire grid, which means that the tangential velocities u and w are
also defined at the ghost-cells at j = 0 and j = N + 1, and that the normal velocity v is also defined
on the walls (j = 0, N ). For an efficient implementation of the viscous terms in an explicit method, the
concept of ghost-cells is very convenient, since it leads to an identical structure of the discretization at
each internal point. However, it is stressed that the ghost-cells are not used for the pressure pn,m , which
is therefore only defined at internal points (1 ≤ j ≤ N ).
The convective terms are discretized in the momentum-conserving divergence form (derived from the
finite volume approach). For example, the term ∂v 2 /∂y in the v-equation, required at at y-location ys,j ,
is discretized by
[ 14 (vj + vj+1 )2 − 14 (vj−1 + vj )2 ]/hs,j . (17)
It is remarked that the other convective derivatives with respect to y, ∂(vu)/∂y and ∂(vw)/∂y, are
independent of the wall values uΓ and wΓ in case the normal wall velocity vΓ equals zero. The divergence
form of the convective terms requires velocities to be interpolated from one staggered location to the
cell-face of the same or another velocity component. All interpolations to obtain velocities on cell faces
are simple averages over two adjacent points. The weights of these averages are 12 and 12 for the velocities
in the seven derivatives ∂(uu)/∂x, ∂(vu)/∂y, ∂(wu)/∂z, ∂(uw)/∂x, ∂(vw)/∂y, ∂(ww)/∂z, and ∂(vv)/∂y.
In the discretization of ∂(uv)/∂x, v is also interpolated with the weights 21 and 12 , but the convective
velocity u is interpolated by (hc,j uj + hc,j+1 uj+1 )/2hs,j , which can be deduced from the finite volume
formulation in Ref. [23]. The latter interpolation rule is also used for the convective velocity w in
∂(wv)/∂z. The selected interpolations lead to an energy-conserving convective scheme on the present
nonuniform grid (this was verified, see section 4.2).
Turning to the viscous terms, the x- and z-derivatives in the velocity Laplacians are discretized
with the compact second-order stencil (three points). For example, ∂ 2 u/∂x2 is discretized as Bu,x =
(ui−1 − 2ui + ui+1 )/h2x . With respect to the derivatives in the y-direction, ∂ 2 u/∂y 2 is discretized as
uj+1 − uj uj − uj−1
(Bu,y u)j = [ − ]/hc,j , (18)
hs,j hs,j−1

and similarly Bw,y w approximates ∂ 2 w/∂y 2 . Further, ∂ 2 v/∂y 2 is discretized as

vj+1 − vj vj − vj−1
(Bv,y v)j = [ − ]/hs,j . (19)
hc,j+1 hc,j
The subscripts u, v or w in the operator B denote that the result of the discrete operator is defined on
staggered u, v or w locations, respectively.

7
 N Extrapolation (20) Extrapolation (21)
64 -2.00127 -1.5295
128 -2.00032 -1.5147
256 -2.00008 -1.5073

Table 1: Results for discretization (18) applied to y(2 − y) in cell j = 1 for two different extrapolations at the wall. The
analytical value equals −2. Results are shown for three resolutions, in each case γ = 1.4.

Since the wall at ys,N is treated in the same way as the wall at ys,0 , it suffices to limit the pre-
sentation of the velocity boundary condition to the wall at ys,0 . The wall boundary condition for v
is straightforward: v0 = vΓ . Since, unlike the discrete v, the discrete u and w are not defined on the
walls, the discretization of the boundary conditions for u and w on the wall at ys,0 is more complicated
and involves the ghost-cells at yc,0 . The velocities u and w in these ghost-cells are determined by a
fourth-order extrapolation,

u0 = uΓ + β1 (u1 − uΓ ) + β2 (u2 − uΓ ) + β3 (u3 − uΓ ). (20)

The grid-dependent coefficients β1 , β2 and β3 are specified in Appendix 1. With Eq. (20) the approxima-
tion of ∂ 2 u/∂y 2 is locally second-order in the entire domain, provided the stretching function is smooth
and its derivative is nonzero (also across the wall).
The standard second-order (linear) interpolation,

u0 = 2uΓ − u1 and yc,0 = −yc,1 , (21)

is not used, because it leads to zeroth-order local accuracy of ∂ 2 u/∂y 2 near the wall [3, 5]. This is
unsuitable for our purpose to study the near-wall limit. As an example, Table 1 shows results for
∂ 2 u/∂y 2 with u = y(2 − y) in the point yc,1 approximated by (18) using extrapolations (20) and (21)
on three different grids. Both cases seem to converge with grid-refinement, but only extrapolation (20)
converges to the correct value near the wall. It is remarked that the global truncation error (the error
in u and p) is second-order accurate, also if the standard extrapolation (21) is used [42].
With the definition of the convective and viscous schemes the intermediate velocity u∗ can be obtained
at all internal grid points. We proceed with the discretization of the Poisson equation. The discrete
pressure gradient operator is denoted by G = (Gx , Gy , Gz ), while the discrete divergence operator is
denoted by D = (Dx , Dy , Dz ). The discretized form of the Poisson equation reads

D · Gpn,m = (D · u∗ )/αm
0
. (22)

The operators Dx , Dz , Gx , Gz do not interfere with the walls. In the normal direction, the operators
are defined by
(Gy pn,m )j = (pn,m n,m
j+1 − pj )/hs,j , j = 1, ..., N − 1, (23)
and
(Dy v ∗ )j = (vj∗ − vj−1
∗
)/hc,j , j = 1, ..., N. (24)
The discretization of ∂ 2 p/∂y 2 in the Poisson equation equals Dy Gy pn,m . It can be convenient to use a
modified definition of Dy near the walls [6, 13, 14], but we prefer not to do so, to keep the structure of
the discrete Poisson equation as close as possible to the structure of a continuous Poisson equation. As
a consequence, Poisson equation (22) involves (Gy pn,m )Γ and vΓ∗ , the values of v ∗ and Gy pn,m at j = 0
and j = N , which have not yet been defined.
Before we specify the latter two quantities, we discretize the third substep of the projection method,

un,m+1 = u∗ − αm
0
Gpn,m , (25)

which is used to obtain the new velocity un,m+1 at all internal grid points. It is not used to obtain the
normal velocity v n,m+1 on the walls, since these values are directly provided by the prescribed velocity
boundary condition vΓ at tn,m+1 . In the execution of the third substep, the pressure gradient Gpn,m
needs to be computed at internal points only.

8
 The momentum balance in the projection method is expressed by the third substep. A continuous
solution of the velocity acceleration requires the momentum balance to be satisfied on the walls as well.
Since (25) is not required to update the velocity at the wall, it is used to define the relation between the
wall values of Gy p and v ∗ , which implies [3]

(Gy pn,m )Γ = (vΓ∗ − vΓn,m+1 )/αm

0
. (26)

This is the Neumann boundary condition used for the Poisson equation in the projection method formu-
lated with unmodified definition of D. With definition (26), the discrete momentum balance (25) for v
also holds on Γ. Since (25) is now satisfied for 0 ≤ j ≤ N , we can apply the discrete divergence operator
to (25), which yields
D · un,m+1 = D · u∗ − αm 0
D · Gp = 0, (27)
in all internal cells, provided Poisson equation (22) is satisfied in all internal cells.
An essential feature of the explicit projection method is that the numerical solution is independent
of the value of vΓ∗ , see section 6.3.1 in [3]. This is also the case for the present multistage method; if
we write out Poisson equation (22) at yc,1 , move −(Gpn,m )Γ /hc,1 to the right-hand side and substitute
(26), the term with vΓ∗ cancels and we obtain

(pn,m − pn,m ) v ∗ − vΓn,m+1

Dx Gx pn,m + 2 1
+ Dz Gz pn,m = (Dx u∗ + 1 + Dz w∗ )/αm
0
, (28)
hs,1 hc,1 hc,1
which shows that the solution pn,m in the interior of the domain does not depend on vΓ∗ = v0∗ . Equation
(28) clearly shows that the staggered discretization does not need a pressure boundary condition.
Equation (28) can also be directly derived from the staggered discretization of the primitive equations
[3, 6, 13, 14]. It is remarked that (28) has not the form of a Poisson equation, since the left-hand side
does not represent the true value of ∇2 pn,m at yc,1 . Strictly speaking, (28) is not a discrete Poisson
equation, we need to take one step to transform (28) into a discrete formal Poisson equation. This has
not a practical purpose, since (28) is quite convenient to define the coefficient matrix and the right-hand
side of the implicit pressure equation solved in practice. To take this formal step is nonetheless useful,
since it will give us information about the underlying continuous PPE problem. So starting from (28),
how is this last formal step taken? Subtract −(Gy pn,m )Γ /hc,1 from both sides of (28), and substitute
for (Gy pn,m )Γ on the right-hand side a discrete Neumann boundary condition. How is this boundary
condition obtained? In case of the projection method, it is simply derived from the requirement that the
normal momentum equation (25) should be satisfied on the walls, which leads to (26). Thus the reasoning
in the discrete case is entirely analogous to the reasoning in the continuous case. The continuous Navier-
Stokes equations do not need a pressure boundary condition, but the derivation of a Poisson equation
from the Navier-Stokes equations increases the order of the system and therefore additional boundary
conditions are needed, boundary conditions that should be derived from the original Navier-Stokes
problem, or more precisely from suitable regularity conditions imposed on the Navier-Stokes equations.
For example, the Neumann pressure boundary condition follows from requirement that ∂v/∂t should be
continuous in space on Ω (which includes Γ).
Since the numerical solution is independent of vΓ∗ , any value can be chosen for vΓ∗ . For the simulations
in the present paper vΓ∗ = vΓn,m+1 was used. This choice is convenient, since the Poisson equation can
then be solved with the simple homogeneous Neumann condition, (Gy pn,m )Γ = 0. For vΓ∗ = vΓn,m+1 , Eq.
(22) is not only equivalent to Eq. (28), but directly reduces to it. Although the solution in the interior
does not depend on vΓ∗ , (Gy pn,m )Γ does depend on vΓ∗ . Therefore, for most choices of vΓ∗ , including
vΓ∗ = vΓn,m+1 , the corresponding (Gy pn,m )Γ defined by (26) does not represent the physical pressure
gradient on the wall. Which definition of vΓ∗ does provide a physical value of (Gy pn,m )Γ will become
clear in the next section.
The solution (un , pn ) at tn is equal to (un,0 , pn,0 ), which is available when the Poisson equation in
the first stage of a time-step has been solved. If the field at tn is needed for post-processing, this is the
moment to write it to file. Alternatively one could compute pn at tn by solving an additional Poisson
equation at the end of time step n and then skip the computation of pn,0 , which was actually preferred
in Ref. [14]. The present variant has the advantage that memory can be saved, since there is no need to
store the pressure during the computation of the convective and viscous terms.

9
 In this work, the discrete Poisson equation was solved down to machine precision by a direct method,
FFT for the homogeneous directions and the inversion of the tri-diagonal matrix in the normal direction
by two sweeps of Gauss-elimination. As a result, the divergence of the velocity un,m was zero down to
machine precision (after the wall-boundary condition to v n,m in substep 1 was applied). The spatial
integral of the solution of the Poisson equation was set to zero by subtracting the spatial average. The
discrete integral of the pressure was defined as a summation over all grid cells after multiplication of the
cell-centered value with the local cell-volume hx hc,j hz .

3. Equivalence between NS and PPE systems

To investigate the issues addressed in research questions I and II formulated in the Introduction, we
analyze the smooth and discrete equivalence between NS and several PPE systems, with a focus on PPE
systems with the Neumann pressure boundary condition.

3.1. Smooth equivalence

Although, as explained in the Introduction, solutions of system NS and the PPE systems are not
necessarily smooth, it is instructive to analyze ’smooth equivalence’, i.e. the equivalence of the systems
for smooth solutions. The four PPE systems, PPE0-3 defined in the Introduction, have all been derived
from NS by differentation and combination of the equations and boundary conditions. Therefore, a
smooth solution of NS is also a solution of the four PPE systems.
Reversely, suppose that (u, p) is a smooth solution of one of the four PPE systems. The systems
PPE0, PPE1 and PPE3 have the same Poisson equation (4), but different boundary conditions. For a
smooth solution (u, p) of one of these systems, subtracting (4) from the divergence of the momentum
equation leads to
∂
∇ · u = ν∇2 (∇ · u) (29)
∂t
defined in Ω [7]. For PPE0, this divergence diffusion equation implies ∇ · u = 0 for all t, since PPE0
imposes ∇ · u = 0 on Γ and ∇ · u0 = 0 [7]. Therefore, a smooth solution (u, p) of PPE0 is also a smooth
solution of NS. For PPE1, the divergence diffusion equation does not imply that ∇ · u = 0 remains
zero for all times; nonzero divergence might emerge on the wall and then diffuse into the interior of the
domain. An example of a smooth solution that does not remain divergence free can be found in section
3.10.3 in Ref. [7], for a PPE1 problem based on the Stokes equations. System PPE2 uses another Poisson
equation, Eq. (6). Application of the divergence to the momentum equation (2) and subtracting (6)
leads to ∂(∇ · u)/∂t = 0 for all t. Since ∇ · u = 0 at t = 0 and the solution is assumed to be smooth,
∇ · u = 0 for all t is implied; a smooth solution (u, p) of PPE2 is also a solution of NS [7].
For PPE3, constraint Eq. (7) implies ∇2 ∇·u = 0 in the neighbourhood of Γ. It follows from Eq. (29)
that ∂(∇ · u)/∂t = 0 in that neighbourhood, and hence ∇ · u = 0 for all t in the same neighbourhood,
since also ∇ · u0 = 0. Equation (29) with zero value of ∇ · u in a neighbourhood of Γ and ∇ · u0 = 0
implies ∇ · u = 0 in Ω for all t. So a smooth solution (u, p) of PPE3 is also a solution of NS. We conclude
that PPE3 is smoothly equivalent to NS, like PPE0 and PPE2.
If two systems A and B are equivalent and A has a unique solution (is well-posed), B has also a
unique solution; the equivalence implies that B has at least one solution and if B had multiple solutions,
A would also have multiple solutions, such that the solution of A would not be unique.
Important differences between PPE3 and the other PPE systems can be expressed as follows. Whereas
PPE0 requires Green’s functions to obtain the (Dirichlet) boundary condition of the Poisson equation,
PPE3 uses the Neumann boundary condition, which is much simpler to evaluate. Whereas PPE1 is not
smoothly equivalent to NS, PPE3 is. Whereas PPE2 needs the paradoxical inclusion of the viscous term
in the Poisson equation (PPE paradox), PPE3 does not require this paradoxical term.
As indicated in the Introduction, the formulation of system NS is such that the solution is not
necessarily smooth. Formal equivalence between PPE3 (or PPE0 or PPE2) and NS for non-smooth
solutions seems to be a much more complicated problem than the equivalence for smooth solutions.
Instead of that question, we will address a related but simpler question: Is it possible to prove discrete
equivalence, i.e. equivalence between suitable discretizations of PPE systems and the discretization of
system NS (the projection method of section 2)?

10
3.2. Discrete equivalence
The discrete equivalence between NS and PPE systems will be presented in the following order.
After discussing the constraint on the initial divergence, we will consider a discretization of the primitive
equations (the MAC method), the projection method, a discretization of PPE2, and finally a discrete
version of PPE3. The equivalence of these four discrete systems will be discussed and proven, if required.
The derivation of a PPE involves a spatial differentation of the momentum equation, but also a
temporal differentation of the continuity equation. Due to the latter, a PPE system needs an additional
initial condition, which is ∇ · u = 0 at t = 0 (see the literature on discrete algebraic equations, for
example [7]). Although the physical condition ∇ · u0 = 0 is also assumed for system NS, this is formally
not necessary; without this property, the (discontinuous) solution of NS will still satisfy ∇ · u = 0 for
t > 0, since that equation is imposed for t > 0. We recognize this feature of NS in the projection method:
after one stage the discrete divergence is zero, even if the initial velocity field is not divergence free. For
a PPE system, however, a divergence free velocity for t > 0 can only be obtained if the integration is
started from a divergence-free velocity at t = 0 [7]. If we had defined NS without ∇ · u0 = 0, then the
initial condition ∇ · u0 = 0 in the PPE system would follow from the regularity requirement that ∇ · u
should be continuous on [0, ∞) × Ω. Thus for the discrete PPE systems we require that the discrete
divergence of the initial velocity is zero, D · u0,0 = 0. Unlike the discretization of NS by the projection
method, a discretization of a PPE system needs this to produce a solution with zero discrete divergence
down to machine precision.
The MAC method is a direct discretization of system NS on a staggered grid. It does not involve a
definition of a pressure boundary condition on Γ. Stage m + 1 (0 ≤ m < M ) of the present Runge-Kutta
method applied to the MAC method can be described as follows. First the ghost-cell values of un,m ,
wn,m and the wall values of v n,m are set. Also the wall values of v n,m+1 are set. The discrete continuity
and momentum equations at internal points are:

D · un,m+1 = 0, (30)
n,m+1 0
u − Gpn,m = un,0 + αm [−an,m + bn,m + f n,m ]. (31)

The symbols an,m and bn,m refer to the discretizations of the convective and viscous terms, respectively,
and also include the implementation of the no-slip boundary condition as described in section 2.2. The
vector of unknowns consists out of the internal values of (un,m+1 , pn,m ). Most known terms appear on
the right-hand side, but the known terms in the continuity equation (the boundary value of the normal
velocity vΓn,m+1 ) are kept on the left-hand side, since we prefer the use of an unmodified D, as explained
in section 2.2. Since the continuity equation is defined at the same locations as the pressure, the number
of unknowns in (30-31) is the same as the number of equations; in principle (30-31) can be solved directly.
The structure of stage m + 1 of the projection method is (section 2.2):

u∗ 0
= un,0 + αm [−an,m + bn,m + f n,m ], (32)
n,m
(Gy p )Γ = (vΓ∗− vΓn,m+1 )/αm
0
, (33)
∗ 0
D · Gpn,m = (D · u )/αm , (34)
un,m+1 = u∗ − αm0
Gpn,m . (35)

All known quantities are on the right-hand sides, except vΓ∗ , but un,m+1 and pn,m do not depend on this
quantity. The projection method is equivalent to the MAC method, since it can be derived from the
MAC method and vice-versa [3, 6, 13, 14].
Next we proceed with the PPE systems. For these systems we need to assume D · u0,0 = 0, which is
not required for the MAC and the projection method. The discrete PPE2 on a staggered grid is given
by

d∗ = −an,m + bn,m + f n,m , (36)

(Gy p )Γn,m
= d∗Γ− (vΓn,m+1 − vΓn,0 )/αm
0
, (37)
∗ 0
D · Gpn,m = (D · d )/αm , (38)
0
un,m+1 = un,0 + αm (d∗ − Gpn,m ). (39)

11
If D·un,0 = 0, the application of the discrete divergence operator to the last equation yields D·un,m+1 = 0
at points where the discrete divergence is defined. Since D·u0,0 = 0 by definition, an induction argument
implies D · un,m = 0 for all n, m. Like the projection method does not depend on vΓ∗ , the discrete PPE2
does not depend on (d∗y )Γ , the boundary value of the normal component of d∗ .
System PPE3 discretized on a staggered grid is given by

a∗ = −an,m + f n,m , (40)

(bn,m
y )Γ = (bn,m
y )1 + hc,1 (Dx bx + Dz bz )1 , (41)
n,m
(Gy p )Γ = (bn,m
y )Γ+ a∗Γ − (vΓn,m+1 − vΓn,0 )/αm
0
, (42)
∗ 0
D · Gpn,m = (D · a )/αm , (43)
0
un,m+1 = un,0 + αm (a∗ − Gpn,m + bn,m ). (44)

The second equation prescribes the boundary value of the normal component of the viscous term, which,
in PPE3, is a direct consequence of the discretization of the zero divergence of the velocity Laplacian
near the wall at location yc,1 . A similar equation imposed at yc,N provides the boundary value at ys,N .
Like the projection method and PPE2 do not depend on vΓ∗ and (d∗y )Γ , PPE3 does not depend on the
boundary value (a∗y )Γ . Note that if the velocity on the boundary does not depend on time, the physical
value for (a∗y )Γ simply equals (fy )Γ , since the convective term on the boundary is then zero (see the
remark below (5)).
Application of the discrete divergence to momentum equation (44) and subsequent substitution of
the Poisson equation (43) implies the discrete divergence diffusion equation
0
D · un,m+1 = D · un,0 + αm D · bn,m , (45)

for Runge-Kutta stage m + 1. An important statement of the staggered viscous approximation of the
viscous terms is:

If D · un,m = 0 for 1 ≤ j ≤ N, then D · bn,m = 0 for 2 ≤ j ≤ N − 1. (46)

See Appendix 2 for a proof of this statement. If D · un,m = D · un,0 = 0 in all cells, then this statement
and (41) imply that D · bn,m = 0 in all cells, such that the discrete divergence diffusion equation (45)
implies D · un,m+1 = 0. Since D · u0,0 = 0 by definition, we have D · un,m = 0 for all n, m.
The equivalences between the MAC method, the projection method, the discrete PPE2 and the
discrete PPE3 can be proven by showing that the solution spaces are the same. We sketch such a proof
for the equivalence between the projection method and PPE3. Suppose (un,m , pn,m ) is a solution of the
projection method. Then D · un,m = 0. Define (40-42) and derive (43-44) from (32-35), (40-42) and
from the fact that D · bn,m = 0, due to (41) and (46). Thus (un,m , pn,m ) is also a solution of PPE3.
Reversely, suppose (un,m , pn,m ) is a solution of PPE3. Then D · un,m = 0 and D · bn,m = 0 (proven
above). Define (32-33) and derive (34-35) from (40-44). Thus the solution of PPE3 is also a solution of
the projection method.
Comparing the projection method, PPE2, and PPE3, we see a remarkable structure. In fact each of
these methods uses projection, but the scope of the projection is different in each case. In the projection
0
method, the convective term plus forcing term plus viscous term plus old velocity divided by αm is
projected to obtain a divergence free velocity. In PPE2, only the convective plus forcing plus viscous
term is projected. In PPE3, only the convective plus forcing term is projected. In each case the discrete
Neumann pressure boundary condition is derived from the discrete normal momentum equation on the
boundary; the requirement that for the v-component (35), (39) or (44) should hold on Γ implies (33),
(37) or (42), respectively. With each limitation of the scope of the projection, a new condition needs
to be added. PPE2 and PPE3 skip the initial velocity from the projection scope, and therefore they
need zero initial divergence of the velocity in order to obtain zero divergence for t > 0. PPE3 skips the
viscous term from the projection scope, and therefore it needs zero divergence of the velocity Laplacian
near the boundary to keep the velocity divergence free.
In each of these three cases only (Gy pn,m )Γ is influenced by the wall value of the projected term, vΓ∗ ,
dΓ and a∗Γ , respectively. The discrete internal solution (un,m , pn,m ) does not depend on the wall value of
∗

the projected term; by substitution of (Gy pn,m )Γ into the Poisson equation not only (Gy pn,m )Γ , but also

12
the wall value of the projected term can be eliminated. The equation after the eliminations is of course
correct and very convenient in practice, but this equation has not the form of a direct discretization of
a Poisson equation, since the left-hand side does not represent the physical ∇2 p (see also the discussion
after Eq. (28)).

3.3. Discussion
Next we discuss in some more detail implications of the previous section for the PPE paradox (why
does PPE2 work while PPE1 does not?) and the physical pressure Neumann condition not required but
implied by the projection method. In addition the pressure correction method (SMAC) will be discussed.
A discretization of PPE1 is obtained if (41) is replaced by
J
X
(bn,m
y )Γ = b∗Γ = ζj (by )j , (47)
j=1

which is a straightforward extrapolation of the discretized ν∇2 v from the interior. Here J is an integer
value, and ζj are coefficients such that a fourth-order approximation for ν∇2 v on the wall is obtained.
After the first stage in the first time step, D · u0,1 is zero for 2 ≤ j ≤ N − 1, which follows from
D · u0,0 = 0, statement (46) and equation (45). However, D · u0,1 will in general be nonzero at yc,1 ,
since statement (46) is in general not valid at j = 1. In the next stage the velocity divergence will then
become nonzero at yc,2 and so forth. Thus nonzero divergence emerges in the internal cells adjacent to
the wall and subsequently diffuses into the interior of the domain. This is the reason why the discrete
PPE1 is not equivalent to the discrete NS, unlike PPE3.
Thus the constraint of zero divergence of the velocity Laplacian in PPE3 is an essential one, it can
not simply be left out if the standard (simplified) Poisson equation (4) is used. The difference between
PPE1 and PPE3 (see also section 5.4) shows that the appearance of the wall value of the viscous term by
in the Neumann pressure boundary condition influences the internal solution. In this respect both PPE1
and PPE3 differ from PPE2 and the projection method, in which the numerical value of the viscous
term in the definition of (Gy p)Γ does not matter. Compared to PPE1 and PPE3, due to the inclusion
of the viscous term in the Poisson equation, PPE2 is more similar to the standard projection method
and does not require an additional constraint on the divergence of the velocity or velocity Laplacian on
or near the wall. In PPE2 (and in the projection method), it does not matter that, for arbitrary (by )Γ ,
D · b is in general nonzero in the cells j = 1 and j = N . In these cases, divergence errors induced by
D · b 6= 0 are automatically corrected in the projection step (when the pressure gradient is substracted).
As indicated in the Introduction PPE2 has been criticized in literature, in the sense that the problems
of PPE1 were also attributed to PPE2 [8]. Whereas this statement was already claimed to be unjustified
in Ref. [9] by arguments that involved weak formulations of the Stokes equations, the presently shown
discrete equivalence between PPE2 and NS can be viewed as a further justification of PPE2.
Another consequence of the equivalence between the discrete PPE3 and the projection method is
that it reveals the discrete physical Neumann pressure boundary condition, which is not required by
the projection method, but can be derived from the projection method. For time-independent velocity
on Γ, the physical approximation of (a∗y )Γ is simply (fy )Γ , and in the more general case, a physical
approximation of this quantity is given by −(ay )Γ +(fy )Γ , where (ay )Γ can be any consistent discretization
of u·∇v on Γ. Then the approximation of (Gy p)Γ in PPE3 is the physical approximation of the boundary
value of ∂p/∂y implied by the standard projection method:

(Gy pn,m )Γ = (bn,m

y )Γ + (fyn,m )Γ − (an,m
y )Γ − (vΓn,m+1 − vΓn,0 )/αm
0
, (48)

with (bn,m
y )Γ defined by (41). Since the projection method is equivalent to the discrete PPE3, the choice
 
vΓ∗ = vΓn,0 + αm
0
(bn,m
y )Γ + (fyn,m )Γ − (an,m
y )Γ (49)

in the projection method and subsequent evaluation of (33) produces the same result as (48).
We finish this subsection with a discussion of the concept of pressure correction, proposed in the
SMAC (simplified marker and cell) method [43]. It appears that the introduction of pressure correction

13
 Simulation System p(tn+1 ) ∆t Spatial resolution
A1 NS pn+1,0 0.002 256 × 128 × 128
A2 NS pn+1,0 0.001 256 × 128 × 128
A3 NS pn+1,0 0.0005 256 × 128 × 128
A3R NS pn+1,0 0.0005 512 × 256 × 256
A4 NS pn+1,0 0.00025 256 × 128 × 128
B1 NS pn,2 0.002 256 × 128 × 128
B2 NS pn,2 0.001 256 × 128 × 128
B3 NS pn,2 0.0005 256 × 128 × 128
B4 NS pn,2 0.00025 256 × 128 × 128
C1 PPE1 pn+1,0 0.002 256 × 128 × 128
D1 PPE3 pn+1,0 0.002 256 × 128 × 128

Table 2: Overview of performed simulations.

does not make the analysis essentially different. Pressure correction can be embedded in the present
projection method by reuse of the pressure pn,m−1 computed in the previous stage after the definition of
tn,−1 = tn−1,M −1 and p0,−1 = 0. Then αm 0
Gpn,m−1 is substracted from u∗ in substep 1, pn,m is replaced
by p0 in substeps 2 and 3, and substep 3 is followed by the computation of pn,m = pn,m−1 + p0 . Clearly,
p0 is an approximation of (tn,m − tn,m−1 )∂p/∂t, except at t = 0. In pressure correction methods it is
common to specify ∂p0 /∂n = 0 on a wall, see for example Ref. [44], although this is not proportional to
the true value of ∂ 2 p/∂n∂t on Γ, because the true (physical) value is in general not zero. However, this
is not a problem, provided the correct relation between the wall values of Gy p0 and v ∗ is specified (Eq.
(26) with pn,m replaced by p0 ). Then the internal pressure does not depend on the wall values of Gy p0
and v ∗ , since these can be eliminated from the equations, like in the original projection method the wall
values of Gy p and v ∗ can be eliminated.

4. Numerical case definition and accuracy check

4.1. Case definition

In this section we define the turbulent plane channel flow simulated by DNS. The dimensions of the
domain are the same as in Ref. [21], Lx = 4π, Ly = 2 and Lz = 4π/3. The no-slip condition is u = 0
on Γ (y = 0 and y = Ly ). The forcing term is given by f = (1, 0, 0). As a consequence ūτ = 1, where
ūτ is defined as (νdū/dy)1/2 evaluated on Γ, where ū(y) is the mean streamwise velocity. The viscosity
equals ν = 1/180 and the Reynolds number equals Reτ = ūτ Ly /2ν = 180. The Reynolds number
based on mean centerline velocity and Ly is approximately 6500. The viscous length-scale is defined by
δν = ν/ūτ = 1/180. The normal coordinate in wall units is defined by y+ = y/δν = 180y.
The initial condition u0 is defined as
2
X 2kπx 2kπy 2kπz
u0 = ay(Ly − y) + b cos sin sin , (50)
Lx Ly Lz
k=1
2
bLy X 2kπx 2kπy 2kπz
v0 = − sin (−1 + cos ) sin , (51)
2Lx Lx Ly Lz
k=1
2
bLz X 2kπx 2kπy 2kπz
w0 = − sin sin cos , (52)
2Lx Lx Ly Lz
k=1

with a = 22.5 and b = a/10. This velocity field is divergence free and satisfies the no-slip condition
on the two walls. The discrete divergence of the initial velocity u0 defined above is only approximately
zero (with a second-order spatial truncation error). Therefore the discrete initial velocity is defined by
ŭ0 = u0 − Gφ, where φ is the solution of D · Gφ = D · u0 , such that D · ŭ0 = 0. It is stressed that the
variable φ is not the initial pressure but a purely numerical quantity. The correction −Gφ converges to

14
 (a) (b)
Figure 2: Time-dependent behaviour of spatiallyR averaged quantities for simulation A1; (a): huiτ (solid), hui at the
centerline divided by 10 (dashed), and bulk value huidy/Ly divided by 10 (dotted); (b) Rii for i = 1 (solid), 2 (dashed),
3 (dotted).

zero if the grid size converges to zero. The derived boundary condition for φ is also different from that
of the pressure. The boundary condition is Gy φ = 0 on the walls, which is implied by the requirement
v̆ 0 = v 0 on the walls.
Table 2 shows an overview of the performed simulations. The simulations A1-4 and A3R were
performed with the projection method described in section 2, which is a discretization of system NS. For
simulations B1-4 the same method was used, except that the pressure at the new time level was defined
as the pressure computed in the last stage of the time step. Simulations C1 and D1 were based on the
discretization of two nonequivalent PPE systems, PPE1 and PPE3, respectively. All simulations were
carried out in standard double precision (8-byte floating point). The grid stretching parameter γ was set
to 1.4 in all cases. In all cases except A3R, the grid size in wall units was given by hx+ = 8.8, hz+ = 5.9,
and 0.98 ≤ hy+ ≤ 4.5. The center of the first grid cell adjacent to the wall was at yc,1+ = 0.49. These
values were two times smaller in the case with the refined grid (A3R). The time step in simulation A1
was ∆t = 0.002. In case A1 the (initial) maximum Courant number was about 1.0, while the maximum
diffusion number equalled 4ν∆t/h2c,1 ≈ 1.5, well within the stability region shown in Fig. 1.
The average over the spatially homogeneous directions h.i is a function of y and t. It was implemented
as a uniform average over all the grid points in xz-planes either at cell center location yc or at staggered
location ys , dependent on the definition of the quantity averaged. The statistical mean, denoted by an
overbar, is a function of y only, obtained by application of the average h.i and subsequent averaging
over time, based on samples between t = 10 and t = 30 (with a time interval of 0.1) from both sides of
the channel (using the appropriate mirror condition with respect to the centerline). Temporal averaging
over 20 time units was sufficient for the present purposes; tests with averaging over 90 time units in case
A1 did not lead to noticeable differences on the scale of the present figures. The wall friction velocity
was approximated by
ūτ = [(u1 − u0 )/(yc,1 − yc,0 )]1/2 , (53)
where Eq. (20) was applied to the mean streamwise velocity u. The numerical value of ūτ after averaging
over 20 time units was 1.000 and 0.999, for case A1 and A3R, respectively. Since fx = 1 and total
momentum is conserved the numerical value of ūτ should converge to 1 in each case if the averaging time
approaches infinity.
Figure 2 shows the temporal development of the flow, which undergoes a transition from the simple
initial condition defined above to fully developed turbulent flow. Fig. 2a shows the quantity huiτ =
(ν∂hui/∂y)1/2 evaluated at y = 0, which is a friction velocity that depends on time; it is initially 0.5, it
increases and attains a value above 1.0 around t = 2, and afterwards it decreases and fluctuates around

15
 (a) (b)

Figure 3: (a) Mean flow u and (b) velocity fluctuations (u2i − u2i )1/2 for x-direction i=1 (circles), y-direction i=2 (squares),
and z-direction i=3 (triangles). Simulation A1 (dashed), A3R (solid), and Moser, Kim & Mansour [21] (symbols).

1.0, the statistical value of uτ = huiτ . The center line and bulk values of hui are also shown in Fig.
2a (divided by a factor 10). Fig. 2b shows the temporal development of the cross-sectionally averaged
diagonal Reynolds stresses based on h.i, Rii (t) = (hu2i i − hui i2 )dy/Ly , where i denotes the spatial
R

direction. These quantities peak around t = 2 and afterwards they saturate. For t > 10 fluctuations
were similar to those shown between t = 5 and t = 10. For an integration in the normal direction
the midpoint rule was used for quantities defined at yc locations and the trapezoidal rule for quantities
defined at ys locations.

4.2. Numerical accuracy

Standard turbulence statistics of the simulations A1 and A3R are shown in Fig. 3 and compared
with the data reported in Ref. [21]. The effect of the combined spatial and temporal truncation error is
illustrated by the difference between cases A1 and A3R for a given statistical quantity. Since the method
is second-order and the grid ratio equals 2 in each direction, estimates based on the Taylor expansion
show that the truncation error in A3R is about 1/3 of the difference between A1 and A3R, and that
the truncation error in A1 is about 4/3 times the difference. This and later comparisons between A1
and A3R confirm that the accuracy of A3R is good and the accuracy of A1 sufficient for the present
purposes.
Figure 4 shows the evolution of the root of the average pressure variance Rpp (t) = (hp2 i−hpi2 )dy/Ly
R

for different cases. Here the spatial and temporal truncation errors in the various cases cause visible
differences after t ≈ 1.5, and these differences show apparently chaotic behaviour after t > 3, due to the
developed turbulence. The curves of simulations A1 and D1, which have the same truncation errors, but
different round-off errors, are on top of each other until much later time (until t ≈ 7). Since round-off
errors are much smaller than truncation errors, physical instabilities triggered by the round-off errors
take more time to grow to a significant level.
The energy produced by the convective scheme (integral of innerproduct of convective term and
u) was verified to be zero to machine precision at all times. Although the viscous scheme could not
be proven to be formally negative-definite, no stability problems were encountered (except in case C1,
discussed in section 5.4). The robustness of the present method was further tested by coarsening the
grid of case A1. Simulations were performed on all grids 2k+1 × 2k × 2k with 1 ≤ k ≤ 7 and k an integer
value. All these simulations were numerically stable. In another test the physical viscosity was set to
zero, and also that simulation was stable.
We finish this subsection with a verification of the order of the Runge-Kutta method. Usually the
temporal order of numerical methods is shown for relatively simple test cases solved over a short time. We

16
 (a) (b)
1/2
Figure 4: Average pressure fluctuation Rpp . Simulation A1 (dashed), A3 (dash-dotted), A3R (solid), and D1 (dashed).
The second plot shows that the curves of A1 and D1 coincide until t = 7.

will investigate to which extent we can verify the order of temporal accuracy of a case that undergoes a
transition to fully developed turbulent flow. For this purpose we performed simulations A1-4, performed
with time step reduction of a factor 2 in consecutive cases. In simulations A1-4 (and A1S, A3R, C1)
the pressure at the end of a time step (tn+1 ) is identified with the pressure obtained in the first stage
of the next time step tn+1,0 . However, simulations B1-4 were performed with the usual approach, the
identification of the pressure at time tn+1 with the pressure computed in the last stage of the time step
tn,2 . For a quantity q computed in three simulations with time steps ∆t, ∆t/2 and ∆t/4, resulting in
q1 , q2 and q3 respectively, we define an approximation of the order of accuracy by
2
Q = sign[(q1 − q2 )/(q2 − q3 )] × log |(q1 − q2 )/(q2 − q3 )|. (54)

If the truncation error of q is proportional to (∆t)β , then Q equals β. Figure 5 shows Q for 2 quantities
and 4 triplets of simulations, (A1, A2, A3), (A2, A3, A4), (B1, B2, B3) and (B2, B3, B4).
Fig. 5a shows Q obtained for the quantity q(t) = huiτ (t). Since the simulations only differ with
respect to the interpretation of the pressure, results of series A and B coincide in Fig. 5a. For very short
times Q ≈ 3. Around t = 0.1 we see negative Q for a very short time. Negative or very large values often
indicate the change of the ordering q1 < q2 < q3 < q4 to the ordering q1 > q2 > q3 > q4 , or vice-versa.
This was also the case at t = 0.1. For t > 0.1 the order of the method appears to be around 2. For t > 1
the behaviour of Q shows more noise, probably because temporal and spatial truncation errors start to
interfere when the flow becomes turbulent. For this reason Q is not shown for t > 2.
Fig. 5b shows Q obtained for a quantity based on the pressure, q(t) = Rpp (t). It clearly indicates
that the temporal order of accuracy of series A is higher than of series B. The value Q(Rpp ) for series A
is about 3 for t < 0.1, almost 2 for t < 0.35, then increases to a higher value again. The conclusion is
that the pressure in the present projection method is indeed second-order accurate in time. However, the
results for series B in Fig. 5b (Q(Rpp ) ≈ 1) show that the conventional method, in which the pressure
at the new time level is identified with the pressure computed in the last stage of the time step, is only
first-order accurate, see also Ref. [14].
In the simulations presented in this paper the velocity on Γ did not depend on time. However, short-
time simulations of cases A1-4 modified for time-dependent velocity on the walls were also performed.
For a time dependent tangential velocity component on the walls, the temporal accuracy of the pressure
remained second-order. However, for a time-dependent normal velocity component on the walls, the
pressure was found to be only first-order accurate in time. If the normal component of the velocity on a
Dirichlet boundary condition is time dependent and the continuity equation is rewritten as an equation

17
 (a) (b)
Figure 5: Convergence ratios Q(huiτ ) (a) and Q(Rpp ) (b) for simulation triple (A1, A2, A3) (small circles), (A2, A3, A4)
(large circles), (B1, B2, B3) (small triangles), (B2, B3, B4) (large triangles).

with unknowns at the left- and known quantities on the right-hand side, then the right-hand side depends
on time and this usually leads to reduction of the temporal order of the pressure, see Ref. [14] and the
discussion on differential algebraic equations in Ref. [7], section 3.16.

5. The pressure gradient near the wall

5.1. Comparison of NS, PPE1 and PPE3

In this subsection results are presented for the discrete PPE1 and PPE3 systems and compared with
results of system NS. PPE1 and PPE3 have been implemented (a) to show, by comparison with PPE1,
that the PPE3 condition of zero divergence of the velocity Laplacian near the wall is essential, (b) to
verify the discrete equivalence between PPE3 and the projection method proven in section 3, (c) to
show that the effect of the accumulation of precision errors on the divergence in PPE systems that use
an exact solver for the Poisson equation is negligible, and (d) to compute the physical value of ∂p/∂y
on the wall, which naturally appears in PPE3. It is remarked that the outcome of (c) is not evident,
since in contrast to the projection method, zero divergence of the velocity in a PPE system relies on
the induction principle, as explained in section 3. To know the physical wall value of ∂p/∂y can be
important in applications, for example if the solution is used to track the motion of Lagrangian particles
that depend on the pressure gradient force, which requires accurate interpolation of the pressure gradient
at any particle location. However, the purpose of the present implementation of PPE3 is to illustrate the
theory in section 3 and certainly not to provide a practical alternative for the more convenient projection
method. If desired, the latter can be extended with a computation of the physical wall value of ∂p/∂y
provided by PPE3 (see section 3.3).
The equivalence of the discrete versions of PPE3 and the projection method is demonstrated by Fig.
4b, where the curves of D1 and A1 coincide up to t = 7, and by Fig. 6a, which shows that the absolute
velocity divergence remained below 10−8 in case D1 for an extended simulation time (until t = 100).
Since the flow is turbulent, equivalent methods do not necessarily show the same time dependence.
Indeed, Fig. 4b shows visible deviations between A1 (projection method) and D1 (PPE3), but only after
t = 7, when the flow has been turbulent for some time.
The catastrophic growth of the divergence in simulation C1 (Fig. 6a, logarithmic scaling of both
axes) demonstrates that the near-wall condition of zero divergence of the velocity Laplacian, which is
missing in PPE1, is a crucial feature of PPE3. Comparison of cases D1 and A1 in Fig. 6a shows that
the accumulation of divergence round-off errors in case D1 remained negligible.

18
 (a) (b)
Figure 6: (a) Maximum |∇ · u| as function of time for simulation A1 (projection method, solid) and C1 (PPE1, dashed)
and D1 (PPE3, dotted). (b) Normal component of the pressure gradient on the walls for simulation D1: (h(∂p/∂y)2 i −
h∂p/∂yi2 )1/2 (solid) and h∂p/∂yi (dashed).

The no-slip condition u = 0 on Γ implies −u · ∇u = ∂u/∂t = 0 on Γ. Thus convective and time

derivatives in the definition of the Neumann pressure boundary condition vanish. In addition fy = 0 in
this case, such that the wall boundary condition for the pressure Poisson equation in PPE3 reduces to

∂p/∂y = ν∇2 v on Γ. (55)

Fig. 6b shows the average and rms of the physical value of ∂p/∂y on the walls, defined with respect to
the operator h.i (on the walls an average over the periodic directions and the two walls). The rms value
is nonzero, but the average is zero. The latter is consistent with the averaged Navier-Stokes equations
for this case, since the continuity equation implies < v >= 0 in the entire domain, and therefore Eq.
(55) implies h∂p/∂yi = 0 on the wall.

5.2. Short-time near-wall behaviour

In this subsection the discontinuity of the velocity acceleration at t = 0 is considered in more detail
for the channel flow case. Each component of the velocity acceleration vector consists out of 3 non-trivial
contributions, convective (−u · ∇u), pressure gradient (−∇p), and viscous term (ν∇2 u). The fourth
contribution is the trivial force term f , which is 1 for the streamwise component and 0 for the other
components.
Profiles of the magnitude of the three components of the convective term at t = 0, t = 0.004 and
t = 0.02 are shown in Fig. 7a. The time-dependent magnitude of a quantity c is defined by hc2 i1/2 , which
is a function of y and t. Due to the no-slip condition and the multiplication with u in the definition of
the convective terms, the magnitude of the convective term tends to zero if the wall is approached. The
magnitude of the pressure gradient vector, shown in Fig. 7b, does not tend to zero for y → 0, but it
appears to be smooth in space and also in time, since the profiles for the three distinct times are close
to each other.
Profiles of the three components of the viscous term are shown in Fig. 8. In contrast to Fig. 7, we do
see a strong effect within a few time steps, but only for the tangential components (note that the scale
on the vertical axis in Fig. 8b was divided by 10). We define the wall values of these quantities as the
limits from the interior, also for t = 0, since the initial velocity is a smooth function of x by definition.
Thus we can conclude from Fig. 8 that, on the wall, the Laplacians of the streamwise and spanwise
velocity components are discontinuous functions of time at t = 0. In contrast to the Laplacians of the
tangential velocities, the Laplacian of the normal component does not display discontinuous behaviour.

19
 (a) (b)

Figure 7: Short-time behaviour of (a) convective terms h(u · ∇ui )2 i1/2 , and (b) pressure terms h(∂p/∂xi )2 i1/2 , for stream-
wise component i = 1 (circles), normal component i = 2 (squares), spanwise component i = 3 (triangles). Profiles are
shown for t = 0 (solid), t = 0.004 (dashed) and t = 0.02 (dotted). Simulation A1 (symbols) and simulation A3R (lines).

The total acceleration of the three velocity components is shown in Fig. 9. The behaviour is consistent
with Figs. 7-8. Since by definition ∂u/∂t is zero on the wall, the tangential components appear to be
discontinuous functions of y at y = 0 and t = 0. It is not possible to extend ∂u/∂t (nor ∂w/∂t) to a
continuous function in space and time; the space-time limit approaching a point on the wall at t = 0 does
in general not exist, because according to Fig. 9 the limit at t = 0 (y → 0) differs from the limit at y = 0
(t → 0), which is zero by definition. However, for t > 0 the discontinuous profiles have evolved into a
smooth profiles by viscous diffusion. The discontinuous behaviour at t = 0 is caused by the fact that no
pressure can be obtained that satisfies the momentum equation on the wall in all three directions at t = 0,
not for the present initial condition, nor for any other initial condition that can easily be expressed into
simple analytical functions. However, the normal velocity acceleration appears to be continuous on the
wall at t = 0. That continuity is observed in the normal direction, but not in the tangential ones, shows
that the solution still satisfies the normal momentum equation if not all three momentum equations can
be satisfied on the wall. This confirms that in case a component of the momentum equation is used to
derive a pressure boundary condition for Poisson equation (4), this should be the normal component and
not one of the tangential components, see Refs. [5, 7, 16].

5.3. The near-wall limit in the turbulent regime

The statistical near-wall behaviour of the pressure gradient and the sum of pressure and viscous
terms in the turbulent regime is shown in Fig. 10. Results are shown for simulation A1 and the refined
simulation A3R. The magnitude of each component of the pressure term (pressure gradient) appears to
converges to a nonzero value on the wall. The magnitude of the sum of the pressure, viscous and force
term converges to zero on the wall, for each component. This indicates that in each direction the nonzero
pressure gradient is balanced by a nonzero viscous plus force term on the wall. In addition it was verified
that the magnitude of the convective term approaches zero in the near-wall limit. Therefore, Fig. 10b
indicates that in the turbulent regime ∂u/∂t is continuous in the near-wall limit. In other words, in the
turbulent regime each component of the momentum equation is satisfied on the wall.

5.4. H1980 and HR1982

The results shown in the previous two subsections indicate that the computed velocity satisfies the
regularity of the solution described in H1980 (see Introduction). First, the velocity appears to be smooth
for 0 < t < T in the interior of the domain (Ω). Second, the velocity is continuous at t = 0 and on Γ.

20
 (a) (b)

(c)

Figure 8: Short-time behaviour of viscous terms h(ν∇2 ui )2 i1/2 for (a) i = 1, (b) i = 2, and (c) i = 3. Profiles are shown
for t = 0 (solid), t = 0.004 (dashed) and t = 0.02 (dotted). Note that the vertical scale in (b) is 10 times smaller than in
(a) and (c). Simulation A1 (symbols) and simulation A3R (lines).

21
Figure 9: Short-time behaviour of acceleration terms h(∂ui /∂t)2 i1/2 for i = 1 (circles), 2 (squares), 3 (triangles). Profiles
are shown for t = 0 (solid), t = 0.004 (dashed) and t = 0.02 (dotted). Simulation A1 (symbols) and simulation A3R (lines).

(a) (b)

Figure 10: Near-wall behaviour in turbulent regime, magnitudes ( qi2 )1/2 for (a) qi = −∂p/∂xi , and (b) qi = −∂p/∂xi +
ν∇2 ui + fi . Simulation A1 (symbols) and A3R (solid). Streamwise direction i=1 (circles), normal direction i=2 (squares)
and spanwise direction i=3 (triangles).

22
Figure 11: Short time behaviour of k∇p − ∇p0 k2 /k∇p0 k2 (solid) and k∇p − ∇p0 k∞ /k∇p0 k∞ (dotted). Symbols are from
simulation A1, lines are from simulation A3R.

Non-smoothness is also observed; the tangential components of ∂u/∂t are discontinuous in space and
time on Γ at t = 0.
In fact, the numerical solution seems to obey even stronger regularity than described in H1980: (a)
The normal component of the time derivative and all components of the pressure gradient look smooth,
both in space and time on [0, T ) × Ω, which includes both t = 0 and Γ; (b) All quantities, including
the tangential components of ∂u/∂t, seem to be smooth in time at t = 0 in the interior of the domain;
time discontinuity occurs only on the walls; (c) The velocity and pressure display spatial smoothness for
0 ≤ t < T , also on Γ.
As an illustration of HR1982 (see Introduction), we consider the convergence of the pressure in the
limit of t → 0 in more detail. Figure 11 shows the behaviour of the L2 -norm and L∞ -norm of ∇p − ∇p0
for small times. Both norms tend to zero if t → 0. This result supports HR1982, but does not support
Lemma 3 in Ref. [8]. In fact, HR1982 claimed convergence in the L2 -norm, while the numerical result
suggests stronger convergence, in L2 and in L∞ . Like results in previous subsections, this result was
validated by grid refinement; Fig. 11 shows very similar curves for simulation A1 and refined simulation
A3R. The L2 -normR of a vector R g = (gRx , gy , gz ) that is discretely available at staggered locations is an
approximation of ( gx2 dx + gy2 dx + gz2 dx)1/2 , where the first and third integral are approximated
with midpoint and the second integral with trapezoidal integration over the domain.
In HR1982, p0 is defined as the solution of the pressure Poisson equation at t = 0 with a boundary
condition for the normal component of the pressure gradient. It is also possible to solve the pressure
Poisson equation at t = 0 with a tangential component of the pressure gradient prescribed on Γ. Say
that q 0 is the solution for a tangential boundary condition. A numerical example in Ref. [8] shows that
in general q 0 6= p0 . Then the logical consequence of HR1982 can only be that in general ∇p does not
converge to ∇q 0 if t → 0.

6. Conclusions

We considered an explicit projection method for the incompressible Navier-Stokes equations with no-
slip walls, discretized on a staggered grid. The method was analyzed and used in simulations to address
several issues related to the formulation of the wall boundary condition for the standard pressure Poisson
equation (PPE), ∇2 p = ∇ · (−u · ∇u + f ). Three research questions were formulated in the Introduction
and the answers to these questions are presented below.
The first question was whether the standard pressure Poisson equation with Neumann boundary
derived from the normal component of the momentum equation, ∂p/∂n = n·(ν∇2 u+f −u·∇u−∂u/∂t),
could be made consistent with the Navier-Stokes equations. It was shown that if the requirement that

23
∇ · ∇2 u = 0 is zero near the wall is added to the system, the PPE problem is equivalent to the Navier-
Stokes equations for sufficiently smooth solutions, without the need to introduce Green’s functions. This
PPE system was called PPE3.
The second research question was which continuous PPE system is in discrete form equivalent to
the staggered projection method. An answer would clarify the relation between PPE systems on the
one hand and the staggered projection method on the other hand. Such a clarification was desirable
since the term pressure Poisson equation is often used to describe the pressure equation solved in the
staggered projection method. Two discrete PPE systems were found to be equivalent to the staggered
projection method: above mentioned PPE3 and a system which uses a seemingly vanishing viscous term
on the right-hand side (PPE2). Thus PPE3 had several theoretical purposes. Apart from providing an
answer to the first research question, PPE3 turned out to be useful to clarify the relation between the
projection method and PPE systems, to clarify the so-called PPE paradox by a discrete analysis, and to
derive the physical wall value of ∂p/∂n implied (not required) by the projection method.
The third research question was about the near-wall behaviour of a solution of the Navier-Stokes
equations, for all t, but in particular for t → 0. This topic was considered because the Neumann
condition for the PPE can only be valid in combination with appropriate regularity of the solution of the
Navier-Stokes equations near and on the wall. Direct Numerical Simulation of turbulent channel flow at
Reτ =180 was performed, starting from a precisely prescribed smooth initial condition. The staggered
projection method, which is known to be equivalent to a direct discretization of the Navier-Stokes
equation (no pressure boundary condition), was used in the simulation. Numerical inspection of the
separate terms of the momentum equation in the near-wall region indicated that the three components
of the momentum equation were valid on the wall for t > 0 (verified for short times and in the fully
developed turbulent regime). In the limit t → 0, the pressure gradient appeared to converge to the
initial pressure gradient in the L2 -norm, which confirmed a disputed theoretical result in literature. The
numerical results also indicated stronger regularity than claimed by existing theorems. Even in the
maximum norm, the pressure gradient appeared to converge to the initial pressure gradient. In fact the
only discontinuities observed in the simulations were discontinuities of the tangential viscous terms and
the time derivatives of the tangential velocities on the wall at t = 0.

Acknowledgement
An example of an initial perturbation for turbulent channel flow, similar to the one used in this
paper, was kindly provided by N.D. Sandham. The author is grateful to J.G.M. Kuerten for interesting
discussions about numerical methods and to the Reviewers for useful comments.

Appendix 1: Extrapolation coefficients

The coefficients of the fourth-order extrapolation Eq. (21) are specified here. For brevity the center
grid locations yc,j are denoted by yj . For a wall located at y = 0, the coefficients are

β1 = (s11 y0 + s12 y02 + s13 y03 )/s, (56)

β2 = (s21 y0 + s22 y02 + s23 y03 )/s, (57)
β3 = (s31 y0 + s32 y02 + s33 y03 )/s, (58)

where sij is given by the matrix

 2 2
y2 y3 (y3 − y2 ) y2 y3 (y22 − y32 ) y2 y3 (y3 − y2 )

 y32 y12 (y1 − y3 ) y3 y1 (y32 − y12 ) y3 y1 (y1 − y3 )  , (59)
y12 y22 (y2 − y1 ) y1 y2 (y12 − y22 ) y1 y2 (y2 − y1 )

and
s = y1 y2 y3 (y2 y3 (y3 − y2 ) + y3 y1 (y1 − y3 ) + y1 y2 (y2 − y1 )). (60)
For completeness coefficients for lower order extrapolations are also provided:
y0 (y2 − y0 ) y0 (y1 − y0 )
β1 = , β2 = , β3 = 0, (61)
y1 (y2 − y1 ) y2 (y1 − y2 )

24
for a third-order extrapolation, and β1 = −y0 /(y1 − y0 ), β2 = β3 = 0, for a second-order extrapolation.
For the example given in Table 1, this second-order extrapolation with y0 given by Eq. (14) produces
the same results as Eq. (21).

Appendix 2: Proof of statement (46)

In this appendix we prove (46), which states that D · bn,m = 0 for 2 ≤ j ≤ N − 1 if D · un,m = 0 for
1 ≤ j ≤ N . For brevity, we omit the superscripts m, n and write the components of b as

bx = νBu u, by = νBv v, bz = νBw w, (62)

where Bu , Bv and Bw refer to the discrete Laplace operator at staggered u-locations, v-locations and
w-locations, respectively, for example Bu = Bu,x + Bu,y + Bu,z . In addition we define a discrete Laplace
operator on center locations, Bc = Bc,x + Bc,y + Bc,z . On Cartesian grids, the following is true for cell
centers that are not adjacent to the wall (2 ≤ j ≤ N − 1):

D · b = Dx Bu,x u + Dx Bu,y u + Dx Bu,z u +

Dy Bv,x v + Dy Bv,y v + Dy Bv,z v +
Dz Bw,x w + Dz Bw,y w + Dz Bw,z w
= Bc,x Dx u + Bc,y Dx u + Bc,z Dx u +
Bc,x Dy v + Bc,y Dy v + Bc,z Dy v +
Bc,x Dz w + Bc,y Dz w + Bc,z Dz w
= Bc (D · u). (63)

The nontrivial part here are the nine ’commutations’, for example

Dy Bv,y v = Bc,y Dy v, (64)

which is not really a commutation, since the operator Bv,y is not precisely the same as Bc,y . To show
the validity of (64), we write
1 h 1  vj+1 − vj vj − vj−1  1  vj − vj−1 vj−1 − vj−2 i
Dy Bv,y v = − − −
hc,j hs,j hc,j+1 hc,j hs,j−1 hc,j hc,j−1
1 h 1   1  i
= (Dy v)j+1 − (Dy v)j − (Dy v)j − (Dy v)j−1
hc,j hs,j hs,j−1
= Bc,y Dy v. (65)

Similarly, one can derive Dx Bu,x = Bc,x Dx , Dx Bu,y = Bc,y Dy , etc. If D · u = 0 for 1 ≤ j ≤ N , Eq. (63)
implies that D · b = 0 for 2 ≤ j ≤ N − 1.

References
[1] P. Moin and J. Kim, On the numerical solution of time-dependent viscous incompressible fluid flows involving
solid boundaries, J. Comput. Phys. 35 (1980) 381-392.

[2] L. Kleiser and U. Schumann, Treatment of Incompressibility and Boundary Conditions in 3-D Numerical
Spectral Simulations of Plane Channel Flows, in: E.H. Hirschel (ed.), Proc. 3rd GAMM Conf. on Numerical
Methods in Fluid Mechanics, Vieweg-Verlag, Braunschweig, 1980, pp. 165-173.

[3] R. Peyret and T.D. Taylor, Computational methods for fluid flow, Springer-Verlag, New York, 1983.

[4] S.A. Orszag, M. Israeli and M.O. Deville, Boundary conditions for incompressible flows, J. Sci. Comput. 1
(1986) 75-111.

[5] P.M. Gresho and R.L. Sani, On pressure boundary conditions for the incompressible Navier-Stokes equations,
Int. J. Numer. Methods Fluids 7 (1987) 1111-1145.

25
 [6] A.E.P. Veldman, ”Missing” boundary conditions? Discretize first, substitute next, combine later, SIAM J.
Sci. Stat. Comput. 11 (1990) 82-91.

[7] P.M. Gresho and R.L. Sani, Incompressible flow and the finite element method, John Wiley and Sons,
Chichester, 1998.

[8] D. Rempfer, On boundary conditions for incompressible Navier-Stokes problems, Appl. Mech. Rev. 59 (2006)
107-125.

[9] R.L. Sani, J. Shen, O. Pironneau and P.M. Gresho, Pressure boundary condition for the time-dependent
incompressible Navier-Stokes equations, Int. J. Numer. Meth. Fluids 50 (2006) 673-682.

[10] F.E. Harlow and J.E. Welch, Numerical calculation of time-dependent viscous incompressible flow of fluid
with free surface, Phys. Fluids 8 (1965) 2182.

[11] A.J. Chorin, Numerical solution of the Navier-Stokes equations, Math. Comput. 22 (1968) 745-762.

[12] R. Temam, Sur 10 approximation de la Solution des équations de Navier-Stokes par la méthode des pas
fractionnaires I, Arch. Ration. Mech. Anal. 32 (1969) pp. 135-153.

[13] J.B. Perot, An Analysis of the Fractional Step Method, J. Comput. Phys. 108 (1993) 51-58.

[14] B. Sanderse and B. Koren, Accuracy analysis of explicit Runge-Kutta methods applied to the incompressible
Navier-Stokes equations, J. Comput. Phys. 231 (2012) 3041-3063.

[15] J.G. Heywood, The Navier-Stokes equations: on the existence, regularity and decay of solutions, Indiana
Univ. Math. J. 29 (1980) 639-680.

[16] J.G. Heywood and R. Rannacher, Finite element approximation of the nonstationary Navier-Stokes problem.
I. Regularity of solutions and second-order error estimates for spatial discretization, SIAM Soc. Ind. Appl.
Math. J. Numer. Anal. 19 (1982) 275-311.

[17] R. Temam, Behaviour at time t=0 of the solutions of semi-linear evolution equations, J. Diff. Equ. 43 (1982)
73-92.

[18] P. Moin and J. Kim, Numerical investigation of turbulent channel flow, J. Fluid Mech. 118 (1982) 341-377.

[19] J. Kim, P. Moin and R. Moser, Turbulence statistics in fully developed channel flow at low Reynolds number,
J. Fluid Mech. 177 (1987) 133-166.

[20] N.D. Sandham, Resolution requirements for Direct Numerical Simulation of near-wall turbulent flow using
finite differences, Queen Mary and Westfield College Report QMW-EP-1097 (1994).

[21] R.D. Moser, J. Kim, N.N. Mansour, Direct numerical simulations of turbulent channel flow up to Reτ = 590,
Phys. Fluids 11 (1999) 943-945.

[22] J.C. del Álamo and J. Jiménez, Spectra of the very large anisotropic scales in turbulent channels, Phys.
Fluids 15 (2003) L41-43.

[23] R.W.C.P. Verstappen, A.E.P. Veldman, Symmetry-preserving discretization of turbulent flow, J. Comp.
Physics 187 (2003) 343-368.

[24] H. Abe, H. Kawamura and Y. Matsuo, Surface heat-flux fluctuations in a turbulent channel flow up to
Reτ = 1020 with P r = 0.025 and P r = 0.71, Int. J. Heat and Fluid Flow 25 (2004) 404-419.

[25] Z.W. Hu, C.L. Morfey and N.D. Sandham, Wall pressure and shear stress spectra from direct numerical
simulations of channel flow up to Reτ = 1440, AIAA Journal 44 (2006) 1541-1549.

[26] J. Meyers and P. Sagaut, Is plane-channel flow a friendly case for the testing of large-eddy simulation
subgrid-scale models? Phys. fluids 19 (2007) 048105.

[27] J.C. Strikwerda and Y.S. Lee, The accuracy of the fractional step Method, SIAM Soc. Ind. Appl. Math. J.
Numer. Anal. 37 (1999) 37-47.

[28] S. Armfield and R. Street, The fractional-step method for the Navier-Stokes equations on staggered grids:
the accuracy of three variations, J. Comput. Phys. 153 (1999) 660-665.

26
[29] J. Kim and P. Moin, Application of a fractional-step method to incompressible Navier-Stokes equations, J.
Comput. Phys. 59 (1985) 308- 323.

[30] T.M. Burton and J.K. Eaton, Analysis of a fractional-step method on overset grids, J. Comput. Phys. 177
(2002) 336-364.

[31] D.L. Brown, R. Cortez, and M.L. Minion, Accurate projection methods for the incompressible Navier-Stokes
equations, J. Comput Phys. 168 (2001) 464-499.

[32] N.D. Sandham and L. Kleiser, The late stages of transition in turbulent channel flow, J. Fluid Mech. 245
(1992) 319-348.

[33] K. Akselvoll and P. Moin, An efficient method for temporal integration of the Navier-Stokes equations in
confined axisymmetric geometries, J. Comput. Phys. 125 (1996) 454-463.

[34] J.G.M. Kuerten, An accurate numerical method for DNS of turbulent pipe flow, in: V. Armenio, B. Geurts,
J. Fröhlich (Eds.), Direct and Large-Eddy Simulation VII, ERCOFTAC Series Volume 13, 2010, pp. 131-136.

[35] A. Jameson, Success and challenges in computational aerodynamics, AIAA Paper, No. 87-1184 (1987), 1-35.

[36] A. Jameson, and T.J. Baker, Solution of the Euler equations for complex configurations, AIAA Paper, No.
83-1929 (1983), 293-302.

[37] B. Vreman, B. Geurts and H. Kuerten, Large-eddy simulation of the turbulent mixing layer, J. Fluid. Mech.
339 (1997) 357-390.

[38] A.W. Vreman, N.D. Sandham, K.H. Luo, Compressible mixing layer growth rate and turbulence character-
istics, J. Fluid. Mech. 320 (1996) 235-258.

[39] Y. Morinishi, T.S. Lund, O.V. Vasilyev and P. Moin, Fully conservative higher order finite difference schemes
for incompressible flow. J. Comput. Phys. 143 (1998) 90-124.

[40] P. Wesseling, Principles of computational fluid dynamics, Springer, Berlin, 2000.

[41] G. Tryggvason and R. Scardovelli, Direct numerical simulations of gas-liquid multiphase flows, Cambridge
University Press, Cambridge, 2011.

[42] T.Y. Hou and B.R.R. Wetton, Second-order convergence of a projection scheme for the incompressible
Navier-Stokes equations with boundaries, Siam J. Numer. Anal. 30 (1993), 609-629.

[43] A.A. Amsden and F.H. Harlow, The SMAC method, Los Alamos Scientific Lab. Rep. No. LA-4370 (1970).

[44] Q.H. Deng and G.F. Tang, Special treatment of pressure correction based on continuity conservation in a
pressure-based algorithm, Num. Heat Transfer, Part B. 42 (2002) 73-92.

27