HAL Id: hal-02880030
https://hal.archives-ouvertes.fr/hal-02880030v2
Preprint submitted on 15 Jan 2021
HAL is a multi-disciplinary open access archive for the deposit and dissemination of sci- entific research documents, whether they are pub- lished or not. The documents may come from teaching and research institutions in France or abroad, or from public or private research centers.
L’archive ouverte pluridisciplinaire HAL, est destinée au dépôt et à la diffusion de documents scientifiques de niveau recherche, publiés ou non, émanant des établissements d’enseignement et de recherche français ou étrangers, des laboratoires publics ou privés.
High order homogenization of the Stokes system in a periodic porous medium
Florian Feppon
To cite this version:
Florian Feppon. High order homogenization of the Stokes system in a periodic porous medium. 2021.
�hal-02880030v2�
A PERIODIC POROUS MEDIUM
2
FLORIAN FEPPON† 3
Abstract. We derive high order homogenized models for the incompressible Stokes system in 4
a cubic domain filled with periodic obstacles. These models have the potential to unify the three 5
classical limit problems (namely the “unchanged” Stokes system, the Brinkman model, and the 6
Darcy’s law) corresponding to various asymptotic regimes of the ratioη≡aε/εbetween the radius 7
aεof the holes and the sizeεof the periodic cell. What is more, a novel, rather surprising feature 8
of our higher order effective equations is the occurrence of odd order differential operators when the 9
obstacles are not symmetric. Our derivation relies on the method of two-scale power series expansions 10
and on the existence of a “criminal” ansatz, which allows to reconstruct the oscillating velocity and 11
pressure (uε, pε) as a linear combination of the derivatives of their formal average (u∗ε, p∗ε) weighted 12
by suitable corrector tensors. The formal average (u∗ε, p∗ε) is itself the solution to a formal, infinite 13
order homogenized equation, whose truncation at any finite order is in general ill-posed. Inspired 14
by the variational truncation method of [53,27], we derive, for anyK∈N, a well-posed model of 15
order 2K+ 2 which yields approximations of the original solutions with an error of orderO(εK+3) 16
in theL2norm. Furthermore, the error improves up to the orderO(ε2K+4) if a slight modification 17
of this model remains well-posed. Finally, we find asymptotics of all homogenized tensors in the low 18
volume fraction limitη →0 and in dimensiond≥3. This allows us to obtain that our effective 19
equations converge coefficient-wise to either of the Brinkman or Darcy regimes which arise whenη 20
is respectively equivalent, or greater than the critical scalingηcrit∼ε2/(d−2). 21
Key words. Homogenization, higher order models, porous media, Stokes system, strange term.
22
AMS subject classifications. 35B27, 76M50, 35330 23
1. Introduction. This article is concerned with the high order homogenization
24
of the Stokes system in a periodic porous medium. Let D := (0, L)
dbe a d-dimensional
25
box filled with periodic obstacles ω
ε:= ε( Z
d+ ηT ) ∩ D obtained by rescaling and
26
repeating periodically a unit hole T (the setting is illustrated on Figure 1). The
27
parameter ε denotes the size of the periodic cell, it is equal to ε := L/N where N ∈ N
28
is a large integer and L is the length of the box. The parameter η is the scaling ratio
29
between the radius a
ε:= ηε of the obstacles and the length ε of the cells. The total
30
fluid domain is denoted by D
ε:= D\ω
εand it is assumed to be connected. P = (0, 1)
d31
is the unit cell and Y = P \ηT denotes its fluid component.
32
We consider (u
ε, p
ε) ∈ H
1(D
ε, R
d) × L
2(D
ε)/ R the solution to the Stokes system
33
(1.1)
−∆u
ε+ ∇p
ε= f in D
εdiv(u
ε) = 0 u
ε= 0 on ∂ω
εu
εis D–periodic,
34
where f ∈ C
per∞(D, R
d) (and all its derivatives) is a smooth, D–periodic right hand-
35
side. The goal of this paper is to derive high order effective models for (1.1); i.e. a
36
family of well-posed partial differential equations posed in the homogeneous domain
37
∗Submitted to the editors DATE.
Funding:This work was supported by the Association Nationale de la Recherche et de la Tech- nologie (ANRT) [grant number CIFRE 2017/0024] and by the project ANR-18-CE40-0013 SHAPO financed by the French Agence Nationale de la Recherche (ANR).
† Centre de Math´ematiques Appliqu´ees, Ecole´ Polytechnique, Palaiseau, France (flo- rian.feppon@polytechnique.edu).
1
Y ηT
P = (0, 1)
dD
ε= D\ω
εD ω
εε
Fig. 1.The perforated domainDε=D\ωε and the unit cellY =P\(ηT).
D (without the holes) and whose solutions approximate the macroscopic behavior of
38
(u
ε, p
ε) at any desired order of accuracy in ε as ε → 0.
39
The literature [52, 47, 29, 5, 7, 4, 8] describes the occurrence of different asymp-
40
totic regimes depending on how the size a
ε= ηε of the holes compares to the critical
41
size σ
ε:= ε
d/(d−2)in dimension d ≥ 3 (if d = 2, then these regimes depend on
42
how log(a
ε) compares to −ε
−2, see [7]). In loose mathematical terms, these can be
43
summarized as follows (see e.g. [5, 7] for the precise statements):
44
• if a
ε= o(σ
ε), then the holes have no effect and (u
ε, p
ε) converges as ε → 0
45
to the solution (u, p) of the Stokes equation in the homogeneous domain D:
46
(1.2)
−∆u + ∇p = f in D div(u) = 0
u is D–periodic.
47
• if a
ε= cσ
εfor a constant c > 0, then (u
ε, p
ε) converges as ε → 0 to the
48
solution (u, p) of the Brinkman equation
49
(1.3)
−∆u + cF u + ∇p = f in D div(u) = 0
u is D–periodic,
50
where the so-called strange term cF u involves a symmetric positive definite
51
d × d matrix F which can be computed by means of an exterior problem in
52
R
d\T (see [4] and section 5).
53
• if σ
ε= o(a
ε) and a
ε= ηε with η → 0 as ε → 0, then the holes are “large”
54
and (a
d−2εε
−du
ε, p
ε) converges to the solution (u, p) of the Darcy problem
55
(1.4)
F u + ∇p = f in D div(u) = 0 in D
u is D–periodic,
56
where F is the same symmetric positive definite d × d matrix as in (1.3).
57
• if a
ε= ηε with the ratio η fixed, then (ε
−2u
ε, p
ε) converges to the solution
58
(u, p) of the Darcy problem
59
(1.5)
M
0u + ∇p = f in D div(u) = 0 in D
u is D–periodic,
60
where M
0is another positive symmetric d × d matrix (which depends on η).
61
Furthermore M
0/| log(η)| → F if d = 2, and M
0/η
d−2→ F (if d ≥ 3) when
62
η → 0, so that there is a continuous transition from (1.5) to (1.4), see [6].
63
One of the long-term motivations driving this work is the need to lay down theo-
64
retical material that would allow to optimize the design of fluid systems by homog-
65
enization methods similar to those available in the context of mechanical structures
66
[21, 20, 10, 50, 14]. To date, the Brinkman [24, 25, 30] and the Darcy models [56, 51]
67
are commonly used by topology optimization algorithms in order to conveniently in-
68
terpolate the physics of the fluid at intermediate “gray” regions featuring locally a
69
mixture of fluid and solid. However, the above conclusions imply that these models
70
are consistent only in specific ranges of obstacle sizes a
ε: the Brinkman model (1.3)
71
is relevant when there are none or tiny obstacles, while the Darcy models (1.4) and
72
(1.5) should be used at locations where the obstacles are large enough. The arising
73
of these different regimes (1.2)–(1.5) is consequently a major obstacle towards the de-
74
velopment of ‘de-homogenization’ methods [14, 37, 50, 39, 40] for the optimal design
75
of fluid systems, which would enable to interpret “gray” designs as locally periodic
76
“black and white” microstructures (featuring for instance many small tubes or thin
77
plates).
78
It turns out that there is a continuous transition between these regimes which
79
can be captured by higher order homogenized equations, which is the object of the
80
present article. These higher order models are obtained by adding corrective terms
81
scaled by increasing powers of ε to the Darcy equation (1.5); they yield more accurate
82
approximations of (u
ε, p
ε) when ε is “not so small”. For a desired order K ∈ N , the
83
homogenized model of order 2K + 2 reads
84
(1.6)
2K+2
X
k=0
ε
k−2D
kK· ∇
kv
∗ε,K+ ∇q
∗ε,K= f , div(v
∗ε,K) = 0,
v
∗ε,Kis D–periodic,
85
where (v
ε,K∗, q
ε,K∗) is a high order homogenized approximation of (u
ε, p
ε). The coeffi-
86
cient D
kKis a k-th order matrix valued tensor which can be computed by a procedure
87
involving the resolution of cell problems; it makes D
kK· ∇
ka differential operator of
88
order k (the notation is defined in section 2 below). Finally, the high order equation
89
(1.6) encompasses at least the Brinkman and the Darcy regimes in the sense that it
90
converges coefficient-wise to either of (1.3) and (1.4) for the corresponding asymptotic
91
regime of the scaling η (see Remarks 5.6 and 5.7) (the analysis of the subcritical case
92
leading to the Stokes regime (1.2) requires more sophisticated arguments which are
93
to be investigated in future works).
94
A rather striking feature of (1.6) is the arising of odd order differential operators
95
despite the symmetry of the Laplace operator −∆ (these vanish, however, in case
96
the obstacle ηT is symmetric with respect to the cell axes; see Corollary 3.16). This
97
fact is closely related to the vectorial nature of the Stokes system (1.1): the tensors
98
D
kKare symmetric and antisymmetric valued matrices for respectively even and odd
99
values of k. This property ensures that D
kK· ∇
kis a symmetric operator for any
100
0 ≤ k ≤ 2K + 2 (see Remark 3.12). To our knowledge, such terms have so far not
101
been proposed in the literature seeking similar higher order corrections for the Stokes
102
system, although these have been observed in other vectorial contexts [27, 28, 53].
103
Most of the available works have focused on situations with low regularity for f , T
104
and D (see [52, 5]), where the homogenization process can be justified only for the
105
approximation at the leading order in ε. Error bounds for higher order approximations
106
of (u
ε, p
ε) (namely for the truncation of the ansatz (1.7) below) have been obtained
107
in [46, 26], without relating these to effective models. A few additional works have
108
sought corrector terms from physical modelling considerations [35, 18, 17], without
109
considering odd order operators.
110
Our derivation is inspired from the works [19, 53, 15]; it is based on (non-standard)
111
two-scale asymptotic expansions and formal operations on related power series which
112
give rise to several families of tensors and homogenized equations at any order. We
113
extend our previous works [34, 33] where we investigated the cases of the perforated
114
Poisson problem and of the perforated elasticity system. Expectedly, the major dif-
115
ficulty in extending the analysis to (1.6) is the treatment of the pressure variable p
ε 116and of the incompressibility constraint div(u
ε) = 0. Note that the D–periodicity
117
assumption on f and u
εis made in order to eliminate additional difficulties related
118
to the arising of boundary layers (see [43, 22, 23, 11]).
119
The starting point of the method of two-scale expansions is to postulate an ansatz
120
for the velocity and pressure solution (u
ε, p
ε):
121
(1.7) u
ε(x) =
+∞
X
i=0
ε
i+2u
i(x, x/ε), p
ε(x) =
+∞
X
i=0
ε
i(p
∗i(x) + εp
i(x, x/ε)), x ∈ D
ε,
122
where the functions u
i(x, y) and p
i(x, y) are P –periodic with respect to y ∈ P , and D–periodic with respect to x ∈ D. In (1.7), the oscillating function p
i(x, y) is required to be of zero average with respect to y:
Z
Y
p
i(x, y)dy = 0, ∀i ≥ 0.
The aim of the homogenization process is to obtain effective equations for the formal
123
“infinite order” homogenized averages u
∗εand p
∗εdefined by
124
(1.8) u
∗ε(x) :=
+∞
X
i=0
ε
i+2Z
Y
u
i(x, y)dy, p
∗ε(x) :=
+∞
X
i=0
ε
ip
∗i(x), x ∈ D.
125
In Proposition 3.7 below, we obtain that (u
∗ε, p
∗ε) solves the following formal “infinite-
126
order” homogenized equation,
127
(1.9)
+∞
X
k=0
ε
k−2M
k· ∇
ku
∗ε+ ∇p
∗ε= f , div(u
∗) = 0,
u
∗εis D–periodic,
128
which involves a family of constant matrix-valued tensors (M
k)
k∈N. Classically, trun-
129
cating directly (1.9) yields, in general, an ill-posed model [12]. Several methods have
130
been proposed to address this issue in order to obtain nonetheless well-posed higher
131
order equations [16, 13, 1, 2, 15]. In our case, we adapt an idea from [53], whereby
132
the coefficients D
kKare obtained thanks to a minimization principle (described in sec-
133
tion 4) which makes indeed (1.6) well-posed. It is based on the existence of remarkable
134
identities which relate the oscillating solution (u
ε, p
ε) to its formal average (u
∗ε, p
∗ε):
135
(1.10)
u
ε(x) =
+∞
X
i=0
ε
iN
i(x/ε) · ∇
iu
∗ε(x)
p
ε(x) = p
∗ε(x) +
+∞
X
i=0
ε
i−1β
i(x/ε) · ∇
iu
∗ε(x),
∀x ∈ D
ε,
136
where (N
i(y))
i∈Nand (β
i(y))
i∈Nare different families of respectively matrix valued
137
and vector valued P –periodic tensors (of order i). The ansatz (1.10) is substantially
138
different from (1.7); following [15], we call it “criminal” because the expansions of
139
(1.10) depend on u
∗εwhich is itself a formal power series in ε (eqn. (1.8)).
140
The order of accuracy at which the solution (v
∗ε,K, p
∗ε,K) yields an approximation
141
of the original solution (u
ε, p
ε) is determined by how many leading coefficients of (1.6)
142
and (1.9) coincide (Proposition 4.5). In Proposition 4.10, we show that D
kK= M
kfor
143
0 ≤ k ≤ K, which allows to infer error estimates of order O(ε
K+3) in the L
2(D) norm.
144
It may seem disappointing that one needs to solve an equation of order 2K +2 in order
145
to obtain approximations of order O(ε
K+3) “only”. This shortcoming is related to the
146
zero-divergence constraint: in the scalar and elasticity cases considered in [34, 33], it
147
turns out that K + 1 extra coefficients coincide, namely D
kK= M
kfor 0 ≤ k ≤ 2K + 1,
148
which yields error estimates of order O(ε
2K+4). In the present context devoted to the
149
Stokes system (1.1), the equation obtained by substituting D
kKwith M
kin (1.6) for
150
K + 1 ≤ k ≤ 2K + 1,
151
(1.11)
ε
2KD
2K+2K· ∇
2K+2v b
ε,K∗+
2K+1
X
k=0
ε
k−2M
k· ∇
kv
∗ε,K+ ∇ q b
ε,K∗= f div( v b
ε,K∗) = 0
b v
ε,K∗is D–periodic,
152
corresponds to applying the truncation method of [53] to the mixed variational for-
153
mulation rather than to the minimization problem associated with (1.1) (see Re-
154
mark 4.11). While the minimization principle ensures that (1.6) is well-posed, we do
155
not know whether this is the case for (1.11). However if it is, then Proposition 4.5
156
implies that (1.11) improves the approximation accuracy up to the order O(ε
2K+4).
157
The article outlines as follows. Notation conventions related to tensors and tech-
158
nical assumptions are exposed in section 2.
159
In section 3, we introduce cell problems and their solution tensors (X
k, α
k) which
160
allow to identify the functions u
i, p
∗iand p
iin the ansatz (1.7). We show that the
161
formal average (u
∗ε, p
∗ε) solves the infinite order homogenized equation (1.9) involving
162
the tensors M
k. After defining the tensors N
k(y) and β
k(y), we derive the “criminal”
163
ansatz (1.10) expressing (u
ε, p
ε) in terms of p
∗εand of the derivatives of u
∗ε. Through-
164
out this section, a number of algebraic properties are stated for the various tensors
165
coming at play, such as the symmetry and the antisymmetry of the matrix valued
166
tensors M
kfor respectively even and odd values of k, and the simplifications taking
167
place in case the obstacle ηT is symmetric with respect to the cell axes.
168
Section 4 details the truncation process of the infinite order equation (1.9) leading
169
to the well-posed model (1.6). We then provide an error analysis of the homogenized
170
approximations of (u
ε, p
ε) generated by our procedure: our main result is stated in
171
Corollary 4.15 where we show that the solution (v
∗ε,K, q
∗ε,K) of (1.6) yield approxima-
172
tions of (u
ε, p
ε) in the L
2(D
ε) norm of order K + 3 and K + 1 for the velocity and
173
the pressure respectively. We establish explicit formulas relating the coefficients D
kK 174to the coefficients M
kand we briefly discuss the improvement provided by (1.11) in
175
case it is well-posed.
176
The last section 5 investigates asymptotics of the tensors M
kin the low volume
177
fraction limit where the scaling of the obstacle η converges to zero. Our main result is
178
Corollary 5.5 where we obtain the “coefficient-wise” convergence of the infinite order
179
homogenized equation as well as the one of (1.6) towards either of the Brinkman or
180
Darcy regimes (1.4) and (1.5) when η is respectively equivalent or greater than the
181
critical size η
crit∼ ε
2/(d−2), and towards the Stokes regime (1.3) for η = o(ε
2/(d−2))
182
in the case K = 0. Although our error estimates for (1.6), are a priori not uniform
183
in η, this suggests that our higher order model (1.6) has the potential to yield valid
184
approximations in any regime of size of holes (at least for K = 0 or above the critical
185
scale). Note that our analysis is unfortunately unsufficient to establish the convergence
186
of the high order coefficients ε
k−2M
kwith k > 2 towards 0 as η → 0. Future works
187
will investigate higher order asymptotics of the tensors M
kin the subcritical regime
188
η = o(ε
2/(d−2)) which are required to establish or invalidate such a claim.
189
2. Setting and notation conventions related to tensors. In the sequel, we
190
consider the following two classical assumptions for the distributions of the holes ω
ε 191(we recall the schematic of Figure 1), following [5]:
192
(H1) Y = P \(ηT ) ⊂ P , as a subset of the unit torus (opposite matching faces
193
of (0, 1)
dare identified) is a smooth connected set with non-empty interior.
194
(H2) The fluid component D
ε= D\ω
εis a smooth connected set.
195
Remark 2.1. Assumption (H1) does not necessarily imply (H2), see [3] for a coun-
196
terexample. Assumption (H1) is not very restrictive and can easily be generalized to
197
the case where the subset Y has m connected components with m ∈ N (see Appendix
198
7.5.6 in [33]). Assumption (H2) is stronger, but is also more connected to physical
199
applications. It forbids the existence of isolated fluid inclusions. Most of our deri-
200
vations only assume (H1). However, we rely on both assumptions (H1) and (H2) in
201
order to obtain error bounds section 4, because we use some technical results of [5].
202
Below and further on, we consider scalar and vectorial functions such as
203
(2.1) u : D × P → R
(x, y) 7→ u(x, y) , u : D × P → R
d(x, y) 7→ u(x, y)
204
which are both D and P –periodic with respect to respectively the first and the second
205
variable, and which vanish on the hole D ×(ηT ). The arguments x and y of u(x, y) are
206
respectively called the “slow” and the “fast” or “oscillating” variable. With a small
207
abuse of notation, the partial derivative with respect to the variable y
j(respectively
208
x
j) is simply written ∂
jinstead of ∂
yj(respectively ∂
xj) when the context is clear.
209
The star–“∗”– symbol is used to indicate that a quantity is “macroscopic” in the sense that it does not depend on the fast variable x/ε; e.g. (v
ε,K∗, q
ε,K∗) or (u
∗ε, p
∗ε) in (1.6) and (1.9). In the particular case where a two-variable quantity u(x, y) is given such as (2.1), u
∗(x) always denotes the average of y 7→ u(x, y) with respect to the y variable:
u
∗(x) :=
Z
P
u(x, y)dy = Z
Y
u(x, y)dy, x ∈ D,
where the last equality is a consequence of u vanishing on P\Y = ηT . When a function X : P → R depends only on the y variable, we find occasionally more convenient to write its cell average with the usual angle bracket symbols:
hX i :=
Z
P
X (y)dy.
In all what follows, unless otherwise specified, the Einstein summation convention
210
over repeated subscript indices is assumed (but never on superscript indices). Vectors
211
b ∈ R
dare written in bold face notation.
212
The notation conventions used for tensor related operations are summarized in the
213
nomenclature below. Some of them are not standard; they allow to avoid to system-
214
atically write partial derivative indices (e.g. 1 ≤ i
1. . . i
k≤ d) and to distinguish them
215
from spatial indices (e.g. 1 ≤ l, m ≤ d) associated with vector or matrix components.
216
Scalar, vector, and matrix valued tensors and their coordinates
217
b Vector of R
d218
(b
j)
1≤j≤dCoordinates of the vector b
219
b
kScalar valued tensor of order k (b
ki1...ik
∈ R for 1 ≤ i
1, . . . , i
k≤ d)
220
b
kVector valued tensor of order k (b
ki1...ik∈ R
dfor 1 ≤ i
1, . . . , i
k≤ d)
221
B
kMatrix valued tensor of order k (B
ki1...ik
∈ R
d×dfor 1 ≤ i
1, . . . , i
k≤ d)
222
(b
kj)
1≤j≤dCoordinates of the vector valued tensor b
k(b
kjis a scalar tensor of order
223
k).
224
(B
lmk)
1≤l,m≤dCoefficients of the matrix valued tensor B
k(B
lmkis a scalar tensors of
225
order k).
226
b
ki1...ik,jCoefficient of the vector valued tensor b
k(1 ≤ i
1, . . . i
k, j ≤ d)
227
B
ik1...ik,lm
Coefficients of the matrix valued tensor B
k(1 ≤ i
1, . . . i
k, l, m ≤ d)
228
Tensor products
229
b
p⊗ c
k−pTensor product of scalar tensors b
pand c
k−p:
230
(2.2) (b
p⊗ c
k−p)
i1...ik:= b
pi1...ip
c
k−pip+1...ik
.
231
a
p⊗ b
k−pTensor product of a scalar tensors a
pand a vector valued tensor b
k−p:
232
(2.3) (a
p⊗ b
k−p)
i1...ik:= a
pi1...ip
b
k−pip+1...ik
.
233
B
p⊗ C
k−pTensor product of matrix valued tensors B
pand C
k−p:
234
(2.4) (B
p⊗ C
k−p)
i1...ik,lm:= B
ip1...ip,lj
C
ik−pp+1...ik,jm
.
235
Hence a matrix product is implicitly assumed in the notation B
p⊗C
k−p.
236
B
p: C
k−pTensor product and Frobenius product of matrix tensors B
pand C
k−p:
237
(2.5) (B
p: C
k−p)
i1...ik:= B
pi1...ip,lm
C
ik−pp+1...ik,lm
.
238
b
p· c
k−pTensor product and inner product of vector valued tensors b
pand c
k−p:
239
(2.6) (b
p· c
k−p)
i1...ik:= b
pi1...ip,m
c
k−pip+1...ik,m
.
240
B
p· c
k−pTensor product of a matrix tensor B
pand a vector tensors c
k−p:
241
(2.7) (B
p· c
k−p)
i1...ik,l:= B
ip1...ip,lm
c
k−pip+1...ik,m
.
242
Hence a matrix-vector product is implicitly assumed in B
p· c
k−p.
243
Contraction with partial derivatives
244
b
k· ∇
kDifferential operator of order k associated with a scalar tensor b
k:
245
(2.8) b
k· ∇
k:= b
ki1...ik
∂
ik1...ik
.
246
b
k· ∇
kDifferential operator of order k associated with a vector tensor b
k: for
247
any smooth vector field v ∈ C
per∞(D, R
d),
248
(2.9) b
k· ∇
kv = b
ki1...ik,l∂
ik1...ikv
l.
249
B
k· ∇
kDifferential operator of order k associated with a matrix valued tensor
250
B
k: for any smooth vector field v ∈ C
per∞(D, R
d),
251
(2.10) (B
k· ∇
kv)
l= B
ik1...ik,lm
∂
ki1...ik
v
m.
252
Special tensors
253
(e
j)
1≤j≤dVectors of the canonical basis of R
d.
254
e
jScalar valued tensor of order 1 given by e
j,i1:= δ
i1j(with 1 ≤ j ≤ d).
255
δ
ijKronecker symbol: δ
ij= 1 if i = j and δ
ij= 0 if i 6= j.
256
I Identity tensor of order 2:
I
i1i2= δ
i1i2.
The identity tensor is another notation for the Kronecker tensor and it
257
holds I = e
j⊗ e
jwith summation on the index 1 ≤ j ≤ d.
258
J
2kTensor of order 2k defined by:
J
2k:=
ktimes
z }| { I ⊗ I ⊗ · · · ⊗ I .
259
With a small abuse of notation, we consider zeroth order tensors b
0to be constants
260
(i.e. b
0∈ R if b
0is scalar) and we still denote by b
0⊗ c
k:= b
0c
kthe tensor product
261
with a k-th order tensor c
k. The same convention also applies to vector valued and
262
matrix valued tensors.
263
In all what follows, a k-th order tensor b
k(scalar, vector or matrix valued) truly makes sense when contracted with k partial derivatives, as in (2.8)–(2.10). Therefore all the tensors considered throughout this work are identified to their symmetrization:
b
ki1...ik≡ 1 k!
X
σ∈Sk
b
iσ(1)...iσ(k),
where S
kis the permutation group of order k. Consequently, the order in which the
264
(derivative) indices i
1, . . . , i
kare written in b
ki1...ik
does not matter.
265
Finally, in the whole work, we write C, C
Kor C
K(f ) to denote universal constants
266
that do not depend on ε but whose values may change from lines to lines (and which
267
may depend on η or on the obstacle T ).
268
Remark 2.2. In a limited number of places, the superscript or subscript indices
269
p, q ∈ N are used. Naturally, these are not to be confused with the pressure variables
270
p
εor q
εintroduced in (1.1).
271
3. Infinite order homogenized equation and criminal ansatz. We start
272
by identifying the two-scale structure of (u
ε, p
ε) which arise in the form of the ansatz
273
(1.7). Because it helps emphasizing the arising of Cauchy products, we assume, in
274
this section only, that the right-hand side f can be formally decomposed into a power
275
series in ε:
276
(3.1) ∀x ∈ D, f (x) =
+∞
X
i=0
ε
if
i(x).
277
3.1. Identification of the “classical” ansatz: tensors (X
k, α
k). Inserting
278
(1.7) into the Stokes system (1.1) yields the following cascade of equations:
279
(3.2)
−∆
yyu
i+2+ ∇
yp
i+2= f
i+2− ∇
xp
∗i+2− ∇
xp
i+1+ ∆
xyu
i+1+ ∆
xxu
i, div
y(u
i+2) = −div
x(u
i+1),
u
−2= u
−1= 0, p
−1= 0, u
i(x, ·) = 0 on ∂(ηT )
u
i(x, ·) is P –periodic for any x ∈ D, u
i(·, y) is D–periodic for any y ∈ P ,
280
for any i ≥ −2, where the operators −∆
yy, −∆
xy, −∆
yyare defined by
−∆
xx= −div
x(∇
x·), −∆
xy= −div
x(∇
y·) − div
y(∇
x·), −∆
yy:= −div
y(∇
y·).
In order to solve (3.2), we introduce a family of respectively vector valued tensors
281
(X
kj(y))
1≤j≤dand scalar valued tensors (α
kj(y))
1≤j≤ddefined by induction as the
282
unique solutions in H
per1(Y, R
d) × L
2(Y )/ R to the following cell problems:
283
( −∆
yyX
0j+ ∇
yα
0j= e
jin Y, div
y(X
0j) = 0 in Y (3.3)
284
( −∆
yyX
1j+ ∇
yα
1j= (2∂
lX
0j− α
0je
l) ⊗ e
lin Y div
y(X
1j) = −(X
0j− hX
0ji) · e
l⊗ e
lin Y, (3.4)
285
( −∆
yyX
k+2j+ ∇
yα
k+2j= (2∂
lX
k+1j− α
k+1je
l) ⊗ e
l+ X
kj⊗ I in Y
div
y(X
k+2j) = −(X
k+1j− hX
k+1ji) · e
l⊗ e
lin Y ∀k ≥ 0.
(3.5)
286 287
Equations (3.3)–(3.5) are supplemented with the following boundary conditions:
288
(3.6)
Z
Y
α
kjdy = 0
X
kj= 0 on ∂(ηT) (X
kj, α
kj) is P –periodic
∀k ≥ 0.
289
290
Remark 3.1. In view of the notation conventions of section 2, the non bold sym- bols ⊗e
land ⊗I indicate the arising of extra partial derivatives indices. For instance, the first line of (3.5) must be understood as
−∆
yyX
k+2j,i1,...ik+2
+∇α
k+2j,i1...ik+2
= 2∂
ik+2X
k+1j,i1...ik+1
−α
k+1j,i1...ik+1
e
ik+2+X
kj,i1...ikδ
ik+1ik+2.
We introduce the k-th order matrix valued tensors X
kwhose columns are the vector valued tensors (X
kj):
(X
ijk(y))
1≤i,j≤d:=
X
k1(y) . . . X
kd(y)
, ∀y ∈ Y, ∀k ≥ 0.
We also denote by α
kthe k-th order vector valued tensor whose coordinates are the scalar tensors α
kj:
α
k(y) := (α
kj(y))
1≤j≤d, ∀y ∈ Y, ∀k ≥ 0.
Following the conventions of section 2, we use a star notation to denote the average
291
of respectively the tensor X
kand of the vector fields u
i:
292
(3.7) X
k∗:=
Z
Y
X
k(y)dy, ∀k ≥ 0, u
∗i(x) :=
Z
Y
u
i(x, y)dy, ∀x ∈ D, ∀i ≥ 0.
293
The tensors X
kand α
kenable to solve the cascade of equations (3.2):
294
Proposition 3.2. Assume (H1). The solutions u
i(x, y), p
i(x, y) of the cascade
295
of equations (3.2) are given by
296
(3.8)
u
i(x, y) =
i
X
k=0
X
k(y) · ∇
k(f
i−k(x) − ∇p
∗i−k(x))
p
i(x, y) =
i
X
k=0
α
k(y) · ∇
k(f
i−k(x) − ∇p
∗i−k(x)),
297
where the functions p
∗iare uniquely determined recursively as the solutions to the
298
following elliptic system: for any i ≥ 0,
299
(3.9)
−div
x(X
0∗∇
xp
∗i) = −div
x(X
0∗f
i)
−
i
X
k=1
div(X
k∗· ∇
k(f
i−k− ∇
xp
∗i−k)) in D
ε, Z
D
p
∗idx = 0
p
∗iis D–periodic.
300
Recognizing Cauchy products, the identities (3.8) and (3.9) rewrite formally in terms
301
of equality of formal power series:
302
u
ε(x) =
+∞
X
i=0
ε
i+2X
i(x/ε) · ∇
i(f (x) − ∇p
∗ε(x)), (3.10)
303
p
ε(x) = p
∗ε(x) +
+∞
X
i=0
ε
i+1α
i(x/ε) · ∇
i(f (x) − ∇p
∗ε(x)), (3.11)
304
div(u
∗ε(x)) = 0 where u
∗ε(x) =
+∞
X
i=0
ε
i+2X
i∗· ∇
i(f (x) − ∇p
∗ε(x)).
(3.12)
305 306
Proof. The result is proved by induction. The case i = −1 is straightforward
307
thanks to the convention u
−1= p
−1= 0. In this proof we use the short-hand
308
notation h
i(x) = f
i(x) − ∇p
∗i(x). Assuming (3.8) and (3.9) hold till rank i + 1 with
309
i ≥ −2, we compute, substituting (3.8) into (3.2):
310
(3.13)
(−∆
yyu
i+2+ ∇
yp
i+2)(x, y)
= h
i+2,j(x)e
j+ (2∂
lX
0j(y) − α
0j(y)e
l) ⊗ e
l) · ∇h
i+1,j(x) +
i
X
k=0
((2∂
lX
k+1j(y) − α
k+1j(y)e
l) ⊗ e
l+ X
kj(y) ⊗ I) · ∇
k+2h
i−k,j(x)
div
y(u
i+2)(x, y) = −
i+1
X
k=0
(X
kj(y) · e
l⊗ e
l) · ∇
k+1h
i+1−k,j(x).
311
The system (3.13) admits a unique solution (u
i+2, p
i+2) with R
Y
p
i+2(x, y)dy = 0 if and only if the following compatibility condition (the so-called “Fredholm alterna- tive”) holds (for any i ≥ −1):
Z
Y
div
y(u
i+2)(x, y)dy = −
i+1
X
k=0
[hX
kji · e
l⊗ e
l] · ∇
k+1h
i+1−k,j(x) = 0.
The above equation determines p
∗i+1given the values of p
∗kfor 0 ≤ k ≤ i:
(hX
0ji · e
l)∂
l(f
i+1,j− ∂
jp
∗i+1) = −
i+1
X
k=1
[hX
kji · e
l⊗ e
l] · ∇
k+1(f
i+1−k,j− ∂
jp
∗i+1−k), which is (3.9) at order i + 1. This identity allows to rewrite div
y(u
i+2) as
312
(3.14) div
y(u
i+2)(x, y) = −
i+1
X
k=0
[(X
kj(y) − hX
kji) · e
l⊗ e
l] · ∇
k+1h
i+1−k,j(x).
313
By linearity, (3.13) and (3.14) and the definitions of (X
kj, α
kj) through the cell problems
314
(3.3)–(3.5) imply the result at rank i + 2.
315
Remark 3.3. The truncation of the series (3.12) at first order yields the well-
316
known Darcy’s law [52]. The next terms of the series have been obtained in [46, 26],
317
at least up to the order i = 1.
318
Remark 3.4. The ansatz (3.10) is already non-standard (when compared to (1.7))
319
because it features p
∗εwhich is a formal power series in ε (recall (1.8)); in other words
320
it is not an ansatz but rather a compact way of writing the Cauchy products (3.8).
321
The next proposition establishes the symmetry and antisymmetry of the matrices
322
X
k∗(eqn. (3.7)) for respectively odd and even values of k. We note that similar
323
identities have been found for the Poisson [34] or the wave equation [1].
324
Proposition 3.5. For any k ≥ 0 and 0 ≤ p ≤ k, 1 ≤ i, j ≤ d, the following
325
identity holds for the matrix valued tensor X
k∗:
326
(3.15) X
ijk∗= (−1)
pZ
Y
((−∆
yyX
pi+∇α
pi)·X
k−pj+∇α
jk−p·X
pi−X
k−p−1j·X
p−1i⊗I)dy
327
with X
−1i= 0 by convention. In particular, for any k ≥ 0, X
2k∗and X
2k+1∗take
328
values respectively in the set of d × d symmetric and antisymmetric matrices:
329
(3.16) X
ij2k∗= (−1)
kZ
Y
(∇X
ki: ∇X
kj+ ∇α
ki· X
kj+ ∇α
kj· X
ki− X
k−1i· X
k−1j⊗ I)dy
330
331
(3.17)
X
ij2k+1∗= (−1)
kZ
Y
(X
ki· ∇X
kj− X
kj· ∇X
ki+ α
kiX
kj− α
kjX
ki) · e
l⊗ e
ldy + (−1)
kZ
Y
(X
k−1j· X
kj− X
k−1i· X
kj)dy.
332
Proof. The result holds for p = 0 because X
ijk∗=
Z
Y
X
kj· e
idy = Z
Y
X
kj· (−∆
yyX
0i+ ∇α
0i)dy.
Assuming now that (3.15) holds till rank p with k > p ≥ 0, we prove the result at
333
rank p + 1. We write, after an integration by parts and by using (3.3)–(3.5):
334
X
ijk∗= (−1)
pZ
Y
[−X
pi· ∆X
k−pj− α
pidiv(X
k−pj) − α
k−pjdiv(X
pi)
335
− X
k−p−1· X
p−1i⊗ I]dy
= (−1)
pZ
Y
(X
pi· (2∂
lX
k−p−1j− α
jk−p−1e
l) ⊗ e
l+ X
k−p−2j⊗ I − ∇α
k−pj) · X
pi336
+ α
piX
k−p−1j· e
l⊗ e
l+ α
k−pjX
p−1i· e
l⊗ e
l− X
k−p−1j· X
p−1i⊗ I dy
337
= (−1)
pZ
Y
− X
k−p−1j· ((2∂
lX
pi− α
pie
l) ⊗ e
l+ X
p−1i⊗ I) + α
k−p−1jdiv(X
p+1i)
338
− ∇α
k−pj· X
pi− α
k−pjdiv(X
pi) + X
k−p−2j· X
pi⊗ I dy
339
= (−1)
pZ
Y
[−X
k−p−1j· (−∆
yyX
p+1i+ ∇α
ip+1)
340
− ∇α
k−p−1j· X
p+1i+ X
k−p−2j· X
pi⊗ I]dy,
341342
whence (3.15) at rank p + 1. Finally, the expression (3.16) for X
ij2k∗is obtained by
343
setting k ← 2k and p ← k in (3.15). The expression for X
ij2k+1∗is obtained by setting
344
k ← 2k + 1 and p ← k and performing an integration by parts.
345
3.2. Derivation of the infinite order homogenized equation and of the
346
criminal ansatz. We now proceed on the derivation of the infinite order homogenized
347
equation (1.9). Let us recall the classical positive definiteness of the Darcy tensor X
0∗.
348
Corollary 3.6. Assume (H1). The matrix X
0∗= (X
ij0∗)
1≤i,j≤d(defined in
349
(3.7)) is positive symmetric definite.
350
Proof. See [52] or Corollary 7.8 in [33].
351
Hence, the following definition of the tensors (M
k)
k∈Nmakes sense.
352
Proposition 3.7. Let M
kbe the tensor of order k defined by induction as fol-
353
lows:
354
(3.18)
M
0= (X
0∗)
−1M
k= −(X
0∗)
−1k−1
X
p=0
X
k−p∗⊗ M
p, ∀k ≥ 1.
355
Then the source terms f
i(eqn. (3.1)) can be expressed in terms of the averaged
356
summands u
∗i(x) and p
∗i(x) ( (1.8) and (3.7)) through the following identity:
357
(3.19) ∀i ≥ 0, f
i(x) − ∇p
∗i(x) =
i
X
k=0
M
k· ∇
ku
∗i−k(x).
358
Recognizing a Cauchy product, (3.19) and (3.12) rewrite formally as the “infinite
359
order” homogenized system (1.9) for the formal average (u
∗ε, p
∗ε) defined in (1.8).
360
Proof. The proof is identical to the one of Proposition 5 in [33], it amounts to
361
average the first line of (3.8) with respect to y and to solve the resulting triangular
362
system determining f
i−k− ∇p
∗i−kin terms of u
∗i.
363
The definition (3.18) essentially states that P
+∞k=0
ε
k−2M
k· ∇
kis the inverse of the
364
formal power series P
+∞k=0
ε
k+2X
k∗· ∇
k. In this spirit, it is even possible to write a
365
fully explicit formula (see [34], Proposition 6 and Remark 2 for the proof):
366
Proposition 3.8. For any k ≥ 1, the tensor M
kis explicitly given by
367
(3.20) M
k=
k
X
p=1
(−1)
pX
i1+···+ip=k 1≤i1,...,ip≤k
(X
0∗)
−1⊗ X
i1∗⊗ · · · ⊗ (X
0∗)
−1⊗ X
ip∗⊗ (X
0∗)
−1.
368
We now introduce matrix valued tensors N
kand vector valued tensors β
kwhich allow
369
to obtain the “criminal ansatz” (1.10) expressing the velocity and pressure (u
ε, p
ε) in
370
terms of their formal average (u
∗ε, p
∗ε).
371
Proposition 3.9. Let N
kand β
kbe respectively the k−th order matrix valued and vector valued tensors defined for any k ∈ N by
N
k(y) :=
k
X
p=0
X
k−p(y) ⊗ M
p, β
k(y) :=
k
X
p=0
(−1)
pM
p· α
k−p(y), ∀y ∈ Y.
Then the summands u
i(x, y) and p
i(x, y) of (3.10) and (3.11) are given for any i ≥ 0
372
in terms of the averages u
∗i(eqn. (3.7)) and p
∗ias follows:
373
(3.21) u
i(x, y) =
i
X
k=0
N
k(y) · ∇
ku
∗i−k(x), p
i(x, y) =
i
X
k=0
β
k(y) · ∇
ku
∗i−k(x).
374
Recognizing Cauchy products, the identities (3.21) can be rewritten formally as the
375
“criminal ansatz” (1.10).
376
Proof. The result is obtained by substituting (3.19) into (3.8) which yields u
i(x, y) =
i
X
p=0 i−p
X
q=0
X
p(y) ⊗ M
q· ∇
p+qu
∗i−p−q(x)
=
i
X
k=0 k
X
p=0
(X
p(y) ⊗ M
p−k) · ∇
ku
∗i−k(x) (change of indices k = p + q) from where the identity (3.21) for u
i(x, y) follows by inverting the summation. Simi- larly, we obtain
p
i(x, y) =
i
X
k=0 k
X
p=0
((M
p−k)
T· α
p(y)) · ∇
ku
∗i−k(x),
hence (3.21) by using (M
p−k)
T= (−1)
p−kM
p−k(see Corollary 3.11 below).
377
In what follows, we denote by (N
jk)
1≤j≤dand by (β
jk)
1≤j≤drespectively the column vectors and the coefficients of N
k(y) and β
k(y):
∀1 ≤ i, j ≤ d, N
jk:= N
ke
jand β
kj:= β
k· e
j.
In addition, the convention N
j−1= 0 is assumed. We shall in the sequel use several
378
times the following properties of (N
jk, β
jk) which are dual to those of (X
kj, α
kj).
379
Proposition 3.10. The k-th order tensors N
k, (N
jk)
1≤j≤d, β
kand (β
jk)
1≤j≤d380
satisfy:
381
(i) R
Y
N
0(y)dy = I and R
Y
N
k(y)dy = 0 for any k ≥ 1;
382
(ii) R
Y
β
k(y)dy = 0 for any k ≥ 0;
383
(iii) For any k ≥ −2 and 1 ≤ j ≤ d,
384
(3.22)
( −∆
yyN
jk+2+ ∇β
jk+2= (2∂
lN
jk+1− β
jk+1e
l) ⊗ e
l+ N
jk⊗ I + M
k+2e
j, div(N
jk+2) = −(N
jk+1− hN
jk+1i) · e
l⊗ e
l;
385
Proof. (i) and (ii) are straightforward consequences of (3.18).
386
(iii) is obtained by writing, for k ≥ 0 (implicit summation on the repeated index
387
j assumed):
388
−∆
yyN
jk+2+ ∇β
jk+2= −∆
yy k+2X
p=0
X
k+2−pi(y) ⊗ M
ijp! + ∇
k+2
X
p=0
α
k+2−pi(y) ⊗ M
ijp!
=
k
X
p=0
h
(2∂
lX
k+1−pi− α
k+1−pie
l) ⊗ e
l+ X
k−pi⊗ I i M
ijp+ (2∂
lX
0i− α
0ie
l)M
ijk+1+ M
ijk+2e
i= (2∂
lN
jk+1− β
k+1ie
l) ⊗ e
l+ N
jk⊗ I + M
k+2e
j. div(N
jk+2) =
k+2
X
p=0
div(X
k+2−pi)M
ijp= −
k+1
X
p=0
M
ijp(X
k+1−pi− < X
k+1−pi>) · e
l⊗ e
l. The proof is identical for k = −1 and k = −2.
389
The identity (3.22) allows to infer important properties characterizing the tensors M
k390
which are similar to those of Proposition 3.5.
391
Corollary 3.11. For any 1 ≤ p ≤ k − 1, it holds M
ijk= (−1)
p+1Z
Y
((−∆
yyN
ip+ ∇β
ip) · N
jk−p+ ∇β
jk−p· N
ip− N
ip−1· N
jk−p−1⊗ I)dy.
Consequently, for any k ≥ 0,
392
• M
2kis a symmetric matrix valued tensor, and the following identities hold:
M
ij0= Z
Y
∇N
i0: ∇N
j0dy,
∀k ≥ 1, M
ij2k= (−1)
k+1Z
Y