Asymptotic Methods in Fluid Mechanics by Steinruck PDF

~ SpringerWienNewYork
&,60&2856(6$1'/(&785(6
6HULHV(GLWRUV
7KH5HFWRUV
*LXOLR0DLHU0LODQ
)UDQ]*5DPPHUVWRUIHU:LHQ
-HDQ6DOHQoRQ3DODLVHDX
7KH6HFUHWDU\*HQHUDO
%HUQKDUG6FKUHÁHU3DGXD
([HFXWLYH(GLWRU
3DROR6HUDÀQL8GLQH
7KHVHULHVSUHVHQWVOHFWXUHQRWHVPRQRJUDSKVHGLWHGZRUNVDQG
SURFHHGLQJVLQWKHÀHOGRI0HFKDQLFV(QJLQHHULQJ&RPSXWHU6FLHQFH
DQG$SSOLHG0DWKHPDWLFV
3XUSRVHRIWKHVHULHVLVWRPDNHNQRZQLQWKHLQWHUQDWLRQDOVFLHQWLÀF
DQGWHFKQLFDOFRPPXQLW\UHVXOWVREWDLQHGLQVRPHRIWKHDFWLYLWLHV
RUJDQL]HGE\&,60WKH,QWHUQDWLRQDO&HQWUHIRU0HFKDQLFDO6FLHQFHV
,17(51$7,21$/&(175()250(&+$1,&$/6&,(1&(6
&2856(6$1'/(&785(61R
$6<03727,&0(7+2'6
,1)/8,'0(&+$1,&6
6859(<$1'5(&(17$'9$1&(6
(',7(' %<
+(5%(5767(,15h&.
9,(11$81,9(56,7<2)7(&+12/2*<$8675,$
SpringerWienNewYork
7KLVYROXPHFRQWDLQVLOOXVWUDWLRQV
7KLVZRUNLVVXEMHFWWRFRS\ULJKW
$OOULJKWVDUHUHVHUYHG
ZKHWKHUWKHZKROHRUSDUWRIWKHPDWHULDOLVFRQFHUQHG
VSHFLÀFDOO\WKRVHRIWUDQVODWLRQUHSULQWLQJUHXVHRILOOXVWUDWLRQV
EURDGFDVWLQJUHSURGXFWLRQE\SKRWRFRS\LQJPDFKLQH
RUVLPLODUPHDQVDQGVWRUDJHLQGDWDEDQNV
E\&,608GLQH
3ULQWHGLQ,WDO\
63,1
$OOFRQWULEXWLRQVKDYHEHHQW\SHVHWE\WKHDXWKRUV
,6%16SULQJHU:LHQ1HZ<RUN
PREFACE
Rational asymptotic methods developed in the fifties and sixties of

the last century have played an important role in theoretical physics,
mechanics and in particular in fluid mechanics. Among the most
powerful methods used in fluid mechanics are the method of matched
asymptotic expansions and multiple scales methods. Matched asymp-
totic expansions are based on the idea of Prandtl’s boundary-layer
theory. In case of high Reynolds number flows the flow field can
be approximated by an inviscid flow with the exception of a thin
boundary-layer along the wall where the viscosity has to be taken
into account. Both approximations have to match in an interme-
diate region. In some cases the inviscid flow and the viscous flow in
a sub-layer have to be determined simultaneously. Thus one speaks of
interacting boundary-layers. An introduction to triple deck problems
and recent applications to internal flows, external sub- or supersonic
flows, thermal flows and free surface flows will be presented.
Another fruitful application is the theory of separated laminar in-
compressible flows. Various examples of fluid flows involving sep-
aration will be considered, including self-induced separation of the
boundary-layer in supersonic gas flows, and incompressible flow sep-
aration at the leading edge of an aerofoil. A characteristic feature of
a multiple scales problem is that the solution exhibits almost periodic
structures whose properties vary on a large scale. Recently, multiple
scales methods have been applied to problems in meteorology. Thus
well established ad hoc approximations have been verified by applying
the method of multiple scales to the basic equations of fluid flow in
the atmosphere. It will be demonstrated how a large collection of well-
established models of theoretical meteorology can be recovered system-
atically, how new insight into scale interaction processes is gained,
and how the asymptotic analyses provide hints for the construction
of accurate and efficient numerical methods. The known limitations
of the approach are also discussed.
Many problems in fluid mechanics involve asymptotic expansions
in the form of power series. Such expansions necessarily fail to pro-
vide terms which are exponentially smaller than all terms in the se-
ries. Although small, these missing terms are often of physical im-
portance. How to find such exponentially small terms, using as the
main tool matched asymptotic expansions in the complex plane and
Borel summation will be discussed. The techniques will be developed
in the context of model problems related to the theory of weakly non-
local solitary waves which arise in the study of gravity-capillary waves
and also for internal waves.
This volume comprises the lecture notes of a course with the title
“Asymptotic Methods in Fluid Mechanics - Survey and Recent Ad-
vances” held at the Centre for Mechanical Sciences in Udine, Septem-
ber 21-25, 2009. Also included are contributed papers presented at a
workshop embedded in the course.
The organizer of the course thanks all lectures and participants
of the workshop for their valuable contributions and their coopera-
tion. My personal thanks are to former rector of CISM Prof. Wil-
helm Schneider who suggested this course and for his advice during
the preparation. Thanks also to the staff of CISM for the perfect
organization and the support in producing these lecture notes.
Herbert Steinrück
CONTENTS
Introduction to Matched Asymptotic Expansions

by H. Steinrück . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1
Asymptotic Methods for PDE Problems in Fluid Mechan-
ics and Related Systems with Strong Localized Perturba-
tions in Two Dimensional Domains
by M. J. Ward, M.-C. Kropinski . . . . . . . . . . . . . . . . . . . . . . . . . 23
Exponential Asymptotics and Generalized Solitary Waves
by R. Grimshaw . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 71
Exponential Asymptotics and Stokes Line Smoothing for
Generalized Solitary Waves
by P. Trinh . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 121
Multiple Scales Methods in Meteorology
by R. Klein, S. Vater, E. Paeschke, D. Ruprecht . . . . . . . . . 127
Multiple Scales Analysis of the Turbulent Hydraulic Jump
by H. Steinrück . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 197
Modern Aspects of High-Reynolds-Number Asymptotics of
Turbulent Boundary Layers
by B. Scheichl . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 221
Interactive Boundary Layers
by P.-Y. Lagrée. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 247
Interaction Mechanisms in Mixed Convection Flow past a
Horizontal Plate
by H. Steinrück . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 287
Asymptotic Theory of Separated Flows
by A. Ruban . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 311
Weakly 3D Effects Upstream a Surface Mounted Obstacle
in Transonic Flows
by A. Kluwick, M. Kornfeld . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 409
Self Similar Blow-up Structures in Unsteady Marginally
Separated Flows
by M. Aigner, S. Braun . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 415
Introduction to Matched Asymptotic
Expansions
Herbert Steinrück*
*
Vienna University of Technology, Institute of Fluid Mechanics and Heat
Transfer, Vienna, Austria
Abstract The method of matched asymptotic expansions will be
presented by applying it to three examples showing the wide appli-
cability of the method.
1 Introduction
The governing equations describing a flow field are in general a set of non-
linear partial differential equations. Only in few situation exact solutions
mostly in the form of similarity solutions exist. Thus asymptotic expan-
sions with respect to an appropriate dimensionless parameter (e.g. Reynolds
number, Mach number, thickness ratio, ...) which tends to a limiting value
(zero or infinity) are sought. Let φ(x, ε) with x ∈ D ⊂ R3 be a function of
a variable x depending on a small, positive parameter ε with 0 < ε 1.
We call
[φ](n) = δ1 (ε)φ1 (x) + δ2 (ε)φ2 (x) + · · · + δn (ε)φn (x) (1)
a n-term asymptotic series of φ with respect to ε 1 if the gauge functions
δk (ε) form an asymptotic series, i.e. δk+1 (ε) = o(δk (ε)) for k = 1, · · · , n − 1
and φ(x, ε) − [φ(n) ] = o(δn (ε)).
Note a function f (ε) is called a small ‘o’ of the function g(ε), f (ε) =
o(g(ε)) if limε→0 f (ε)/g(ε) = 0 holds, see Van Dyke (1975). The expansion
(1) is called uniformly valid if there exist constants c1 ,...,cn independent of
x with
φ(x, ε) − [φ](m)
< cm , m = 1, ..., n. (2)
δm+1 (ε)
However, the solutions of many perturbation problems in fluid mechan-
ics do not permit an approximation by an asymptotic series of type (1).
Such problems are called singularly perturbed. Most of these problems are
characterized by two different (length) scales. For example consider the at-
tached high Reynolds number flow. A regular (outer) expansion fails near a
H. Steinrück (ed.), Asymptotic Methods in Fluid Mechanics: Survey and Recent Advances
© CISM, Udine 2010
2 H. Steinrück
solid surface and a local variable has to be introduced to describe the local
behavior near the wall and a local (inner) expansion of the flow field can
be found. To get a uniformly valid approximation of the entire flow region
both expansions have to match. Thus one speaks of matched asymptotic
expansion.
Boundary layers have been first introduced by Prandtl (1904) by explain-
ing the role of viscosity in large Reynolds number flows. As a mathemati-
cal tool the method of matched asymptotic expansions has been developed
systematically in between the 50s and 70s of the last century, see Kaplun
(1967), Lagerström and Van Dyke (1975), Fraenkel (1969)
Matched asymptotic expansions are used if a regular asymptotic expan-
sion fails near located singularities. Then the problem has to be rescaled
appropriately by using local variables before expanding its solution asymp-
totically. Both expansion have to agree in some overlap region, i.e. a region
where both expansions hold. We will demonstrate the method by consider-
ing three typical examples showing the wide applicability of the method.
Inviscid potential flow around a thin profile. The flow potential is

expanded with respect to a small aspect ratio of the profile. At first glance
one might think that a regular expansion will be sufficient. However, it
turns out that the tentative regular expansion is not uniformly valid near
the leading and the trailing edge. Thus local expansions turn out to be
necessary to obtain a uniformly valid solution, cf. Van Dyke (1975).
Flow between two rotating discs: Ekman-layer. A common rea-

son for the necessity to introduce a local expansion is that the perturbation
parameter multiplies the highest derivative of the unknown function in a dif-
ferential equation. As a consequence the solution of the limiting differential
equation cannot satisfy all required boundary conditions. By introducing a
local variable the small coefficient of the highest derivative can be rescaled
and the local expansion can satisfy all boundary conditions.
As a representative of that class of problems the flow between two ro-
tating discs in the limit of a small Ekman number will be discussed, see
Ungarish (1993).
Model equation: turbulent pipe flow. Here the asymptotic behav-

ior off an ordinary differential equation is analyzed modeling turbulent pipe
flow in the limit of large Reynolds numbers. The limiting differential equa-
tion is of the same order as the perturbed one. However, the coefficient
of the highest derivative vanishes only at the boundary where a boundary
Introduction to Matched Asymptotic Expansions 3
condition has to be satisfied. Thus again a local expansion turns out to be

necessary. The matching of the two expansions will be discussed and an
short introduction to turbulence asymptotics will be given.
2 Flow around a thin elliptical airfoil ε 1

radius of curva-
U∞ = 1 ture at apex ε2
y
ε y = ±εT (x)
-ε x
-1 1
Figure 1. Elliptical thin airfoil
As a first example we consider the two-dimensional inviscid irrotational

flow past a thin airfoil. For simplicity we consider a symmetric profile in
a uniform free stream parallel to its center line. We place the x-axis of
a coordinate system at the centerline of the airfoil such that leading and
trailing edge are at x = ∓1 in dimensionless coordinates, respectively. The
contour of the profile is given by y = ±εT (x), where ε is the thickness of
the airfoil assumed to be small, see figure 1.
The dimensionless flow field can be described by a flow potential φ =
φ(x, y) where the dimensionless velocity components in x and y direction
are given by u = φx and v = φy . Thus φ is the solution of the potential
equation
φxx + φyy = 0 (3)
subject to the kinematic boundary condition at the surface of the airfoil
φy (x, ±εT (x)) = ±εT (x)φx (x, ±εT (x)), −1 < x < 1 (4)
and the incident flow condition
φ → x, for x2 + y 2 → ∞. (5)
2.1 Asymptotic expansion of the flow potential (regular expan-

sion)
In order to find an asymptotic expansion of the flow potential with re-
spect to a small thickness parameter ε 1 a regular expansion in terms of
4 H. Steinrück
powers of ε is employed. In the limiting case ε = 0 the undisturbed parallel

flow (φ = x) is obtained. Thus we try to determine a regular expansion in
powers of the perturbation parameter ε
φ = x + εφ1 + ε2 φ2 + . . . , (6)
where φi are the solutions of the potential equation. In order to satisfy
the kinematic boundary condition derivatives of the flow potential have to
be evaluated at y = ±εT (x). However, the evaluation at y = εT (x) is
approximated by a Taylor expansion of the corresponding quantity around
(x, 0+), i.e.
φx (x, εT (x)) ∼ 1 + εφ1,x (x, 0+)+
(7)
+ ε2 (φ1,xy (x, 0+)T (x) + φ2,x (x, 0+)) + ...
Thus the expansion of the kinematic boundary conditions yields condi-
tions for the perturbation potentials φk
⎧
⎨ ±T (x), k = 1,
φk,y (x, ±0) = ±(T (x)φ1,x (x, 0±))x , k = 2, − 1 < x < 1 (8)
⎩
±(T (φ2,x + 12 T T ))x , k = 3,
The incident flow condition requires that flow field vanishes for large x2 +
y 2 → ∞.
φi (x, y) → 0, x, y → ∞. (9)
(q)
Note the flow potential of a source at the origin of strength q is φ =
q
2π ln x2 + y 2 . The perturbation potentials φk can be obtained by placing
distributed sources along the centerline of the airfoil on the interval (−1, 1)
and one can verify that the corresponding velocity fields can be represented
by:

1 1 x−ξ
uk (x, y) = φk,x (x, y) = φk,y (ξ, 0) dξ, (10a)
π −1 (x − ξ)2 + y 2

1 1 y
vk (x, y) = φk,y (x, y) = φk,y (ξ, 0) dξ, (10b)
π −1 (x − ξ)2 + y 2
cf. Van Dyke (1975). Using the perturbation potentials φi the surface
velocity us has the expansion

us (x) = φ2x (x, ±εT (x)) + φ2y (x, ±εT (x)) ∼

(11)
2 1 2
+εφ1 (x, 0) + ε φ2x (x, 0) + T (x)T (x) + T (x) + · · · .
2
ïWHUPVRXWHUH[S ε = 0.3

WHUPVLQQHUH[SDQVLRQ
us
ïWHUPRXWHUH[SDQVLRQ
FRPSRVLWHH[SDQVLRQ

WHUPVRXWHUH[SDQVLRQ

ï ï ï ï ï ï
x
Figure 2. Asymptotic expansion of surface velocity
To be more
√ specific we consider an elliptical airfoil with the shape func-
tion T (x) = 1 − x2 . Using the notation of complex variables z = x + iy
the perturbation potentials φ1 , φ2 are

φ1 = φ2 =
z − z 2 − 1 , (12)
where
z denotes the real part of a complex number z. In order to make
the square root unique the complex plane is sliced along the interval (-1,1).
We have φ1,x (x, ±0) = φ2,x (x, ±0) = ±1. Thus the expansion of the surface
velocity
ε2 x2
us (x) ∼ 1 + ε − + ··· (13)
2 1 − x2
turns out to be not uniformly valid. If x is close to the leading or trailing
edge, say |x + 1| ε2 the second order correction term will become larger
than the first, (see figure 2).
It is interesting to note that the flow potential of a source or sink flow
at the leading or trailing edge can be added to the perturbation potential
φ1 . In particular the flow potential

2
C z+1
φ1 =
z − z − 1 − ln (14)
2π z − 1
satisfies all required conditions for an arbitrary constant C.
2.2 Local expansion at leading/trailing edge

Since the expansion presented previously fails near the leading edge we
introduce local coordinates to describe the flow field there. A natural length
6 H. Steinrück
scale near the leading edge is the radius of curvature of the profile, which
in case of the elliptical airfoil is ε2 . Thus we define the local coordinates
1+x y
X= , Y = , (15)
ε2 ε2
and the local flow potential by Φ(X, Y ) by
φ(x, y) = φ(−1, 0) + ε2 Φ(X, Y ). (16)
The local potential satisfies again the Laplace equation. The contour of
the airfoil written in local coordinates is given by

√ ε2
2 2
Y = ± 2X + ε X ∼ 2X 1 + X + ... . (17)
4
Thus the kinematic boundary condition in local coordinates reads

1 − ε2 X
ΦY X, 2X − ε2 X 2 − √ ΦX X, 2X − ε2 X 2 = 0. (18)
2X − ε2 X 2
We expand the local solution with respect to ε asymptotically
Φ(X, Y ) = Φ0 + εΦ1 + ε2 Φ2 + ... (19)
and obtain for the first two terms the kinematic boundary condition
√ 1 √
Φi,Y X, 2X − √ Φi,X X, 2X = 0, i = 0, 1. (20)
2X
This can be interpreted as the kinematic boundary condition for the inviscid
flow around a parabola. Due to symmetry the stagnation point is in the
apex of the parabola. The flow potential can be determined by conformal
mapping, cf. Betz (1964). It is given by
√
Φi = Uloc,i
Z − 1 + 1 − 2Z , i = 0, 1, (21)
where the velocities of the free stream Uloc,i with i = 1, 2 are unknown. They
have to be determined by matching with the outer (global) expansion. The
expansion of the local surface velocity is given by

2X
Us ∼ (Uloc,0 + εUloc,1 ) . (22)
1 + 2X
y
Y =
ε2
√
Y = ± 2X + ε2 X 2
Uloc
1+x
X=
ε2
Figure 3. Airfoil in local coordinates
2.3 Matching procedure

We have now determined asymptotic expansions on two different length
scales: the outer (global) length scale and a local expansion around the lead-
ing edge of the profile, where the radius of curvature is the reference length
scale. Both expansions are not uniformly valid in the entire flow domain.
The basic hypothesis is that there exists (asymptotically) an overlap where
both expansions are valid. Thus we take the outer (global) expansion and
rewrite it in the inner (local) variable. For the matching procedure we use
the velocity field instead of the flow potential.
We introduce an intermediate variable ξ(ε) such that
z(ε) = −1 + ε2 ξ(ε) → −1, Z(ε) = ξ(ε) → ∞ as ε → 0, (23)
and insert it into the global and local expansion, respectively.

The outer expansion of the velocity field in the overlap region is:

z C 1 1
φ ∼ 1 − ε√ + − ··· ∼ (24)
z 2 − 1 2π 1 + z z−1

−1 + ε2 ξC 1 C
1+ε 1− + 2 ∼1+ε+ √ + . (25)
2 4
−2ε ξ + ε ξ 4 ε ξ −2ξ εξ
The local expansion of the velocity field in the overlap region is given by

1
Φ ∼ (Uloc,0 + εUloc,1 ) 1 + √
1 − 2Z
(26)
1
∼ Uloc,0 1 + √ + εUloc,1 .
−2ξ
8 H. Steinrück
Thus both expansions agree in the overlap region if
Uloc,0 = Uloc,1 = 0, and C = 0 (27)
holds.
Van Dyke (1975) has formalized the matching procedure in the match-
ing principle. Fraenkel (1969) discussed criteria on the inner and outer
expansion for the validity of the matching principle. For example when the
gauge functions in the outer and inner expansion are powers of the expan-
sion parameter, which is defined as the ratio of the scales of the inner and
outer variable, the matching principle holds. Problems may arise when the
gauge functions are a combination of powers and logarithmic terms of the
perturbation parameter.
Matching principle: n-terms of the outer expansion rewritten in the

inner variable and expanded into m terms must agree with m terms of the
inner expansion rewritten in the outer variable and expanded into n terms.

(m)
(n)
(n) (m)
φ = φ . (28)
out in in out
We demonstrate the Matching Principle at the surface velocity of a thin

airfoil. We start with the 3 term outer expansion and rewrite it in the local
(inner) variables
(3) ε2 x2 1 (−1 + ε2 X)2

[us ]out = 1 + ε − =1+ε− . (29)
2 1−x 2 X 2 − ε2 X
Expanding the above expression into two terms yields

(3)
(2) 1
us =1− + ε. (30)
out in 4X
On the other hand two terms of the inner expansion rewritten in the outer
variables gives

(2) (2)
(2) 2X ε2
us = (1 + ε) =1+ε− . (31)
in out 1 + 2X 4(1 + x)
out
Thus both expressions agree.

2.4 Composite approximation

In order to get a uniformly valid approximation one has to combine the
inner and outer expansion. This can be done by adding both expansions.
Doing so the overlap region is represented twice. Thus the common part of
both expansion has to be subtracted.

(3)
(3)
(2) (3)
us ∼ [us ]in + [us ]out − φ = (32)
out in

2X 2 x2 1
= (1 + ε) − 2
+ .
1 + 2X 2 1−x 4X
In figure 2 the outer, the inner expansion and the composite approxima-
tion of the surface velocity are shown for ε = 0.3 .
3 Flow between rotating discs - Ekman layers

In many applications local expansions have to be introduced since the so-
lution of the limiting problem cannot satisfy all boundary conditions. Of-
ten this is due the fact that the small perturbation parameters multiplies
the highest derivative of the unknown function in the governing differential
equation. As a representative example we study here the incompressible flow
between two infinite parallel discs, which rotate coaxially but at different
speeds in the limit of a small Ekman number.
z̃
Ω̃ − ω̃
ũ
r̃
2L̃
Ekman-layers Ω̃ + ω̃
Figure 4. Flow between two coaxially rotating discs
The distance between the two discs is 2L̃. Here and in the following
we denote dimensional quantities with a tilde. The upper disc rotates with
speed Ω̃ − ω̃ and the lower with Ω̃ + ω̃. We choose a cylindrical coordinate
system with the axis of rotation as the z-axis and its origin in the mid-
10 H. Steinrück
dle between the two plates. The fluid between the discs is assumed to be
incompressible and its kinematic viscosity ν̃ to be constant.
Thus two independent dimensionless group can be formed: The Ekman
number Ek = ν̃/Ω̃L̃2 and the Rossby number Ro = ω̃/Ω̃. If the Rossby
number is zero both discs rotate at the same angular speed. In that case the
fluid between the two discs would do the same. Thus the Rossby number is
a measure of the deviation of the solid body rotation of the fluid. Here we
will assume that the Rossby number is small.
The Ekman number can be interpreted as the reciprocal value of a
Reynolds number based on the reference velocity Ω̃L̃. Here we are interested
in the limit of small Ekman numbers.
The governing Navier-Stokes equation written in cylindrical coordinates
can be found in Schlichting (2000). We use the dimensionless vertical coor-
dinate Z = L̃z̃ and introduce the similarity ansatz
ũ = r̃ ω̃U (Z), ṽ = r̃ Ω̃ + r̃ ω̃V (Z), w̃ = L̃ω̃W (Z), (33a)

1 2 2 1
p̃ = ρ̃Ω̃ r̃ + ρ̃Ω̃ω̃r̃2 A + ρ̃Ω̃ω̃ L̃2 B(Z). (33b)
2 2
Thus the Navier-Stokes equation reduce to a set of nonlinear ordinary dif-
ferential equations for the secondary flow induced by the difference of the
angular velocities.
Ro (U 2 − V 2 + W UZ ) = 2V − A + Ek UZZ , (34a)
Ro (2U V + W VZ ) = − 2U + Ek VZZ , (34b)

Ro W WZ = −BZ + Ek WZZ , (34c)
2U + WZ = 0. (34d)
At the two discs the no slip boundary conditions have to be satisfied.
U (±1) = 0, V (±1) = ∓1, W (±1) = 0. (35)
These are a set of ordinary differential equations for the velocity profiles
U , V , W , the pressure profile B and the constant A. At first glance one
might think that the six no-slip boundary conditions are not enough. But
we have to consider equation (34a), (34b) as second order equations for U
and V , respectively. The continuity equation (34d) can be considered as
first order equation for the vertical velocity profile W and equation (34c)
can be considered as an algebraic equation for the vertical pressure gradient
BZ . Thus in total six boundary conditions are needed to determine U , V ,
W , BZ and A .
Assuming a small difference in the speeds of rotation of the two discs

(small Rossby number Ro) we can neglect the nonlinear terms and obtain
a linear set of ordinary differential equations with constant coefficients.
Ek UZZ = A − 2V, Ek VZZ = 2U, WZ − 2U, BZ = Ek WZZ . (36)
We remark that for positive Ro-numbers the solution can be expanded into
a regular power series with respect to powers of Ro.
3.1 Small Ekman numbers

Of course the set of ordinary differential equations (36) can be solved an-
alytically. Here we want to demonstrate how to find an asymptotic solution
in the limit Ek → 0.
3.2 Core region

We expand the constant A = A0 +Ek α A1 +... and the solution (U, V, W )
in the core region into powers of the Ekman number Ek
⎛ ⎞ ⎛ ⎞ ⎛ ⎞
U (Z; Ek) Ū0 (Z) Ū1 (Z)
⎝ V (Z; Ek) ⎠ = ⎝ V̄0 (Z) ⎠ + Ek α ⎝ V̄1 (Z) ⎠ + .... (37)
W (Z; Ek) W̄0 (Z) W̄1 (Z)
and insert it into (36). Comparing like powers we obtain
Ai
Ūi = 0, V̄i =, W̄i = Wi , i = 0, 1. (38)
2
However, the constants A0 , A1 , W0 and W1 remain undetermined yet. Un-
fortunately (38) cannot satisfy all boundary conditions for V and W .
3.3 Boundary layers

Thus we expect that the solution will vary rapidly near the boundary
in order to satisfy the boundary conditions. In order to capture this rapid
variation we introduce local variables near the boundaries at Z = ±1.
1−Z 1+Z
η= , ζ= (39)
Ekβ Ek β
The independent variable will be stretched with the factor Ekβ . The expo-
nent β will be determined appropriately later.
Setting U (Z) = Û (η) and similarly V and W and inserting into (36) we
obtain the differential equations describing the local behavior of the flow
near the lower disc.
Ek 1−2β Û = A0 − 2V̂ , Ek 1−2β V̂ = 2Û , Ek−β Ŵ = ∓2Û . (40)
12 H. Steinrück
We remark that for the local behavior near the upper disc we obtain a sim-
ilar differential equation. Only the sign in front of Ŵ has to be changed.
Inspecting the local differential equation and setting Ek = 0 we see imme-
diately that a nontrivial differential equation is obtained if β = 1/2. Thus
we expand
⎛ ⎞ ⎛ ⎞ ⎛ ⎞
U Û0 (ζ) √ Û1 (ζ)
⎝ V ⎠ ∼ ⎝ V̂0 (ζ) ⎠ + Ek ⎝ V̂1 (ζ) ⎠ + · · · . (41)
W Ŵ0 (ζ) Ŵ1 (ζ)
From the boundary condition at the lower disc (Z = −1) we obtain the
boundary condition for the local expansion
Û0 (0) = Û1 (0), V̂0 (0) = 1, V̂1 (0) = 0, Ŵ0 (0) = Ŵ1 (0). (42)
Inserting the expansion (41) into (40) and after some elementary manipu-
lations we obtain a fourth-order differential equation for V̇0
(iv)
V̂0 + 4V̂0 = 2A0 , V̂0 (0) = 1, V̂0 (0) = 0, (43)
with the solution

A0 A0
V̂0 (ξ) = + 1− e−ξ cos ξ + c1 sinh ξ cos ξ + c2 cosh ξ sin ξ. (44)
2 2
For the radial velocity component we obtain from Û0 = − 12 V̂0

A0
Û0 (ξ) = 1 − e−ξ sin ξ − 2c1 cosh ξ sin ξ + 2c2 sinh ξ cos ξ (45)
2
and for the vertical component Ŵ0 = 0 and

A0
Ŵ1 (ξ) = 1 − [1 − e−ξ (sin ξ + cos ξ)]− (46)
2
c1 − c2 c 1 + c2
sinh ξ sin ξ − (cosh ξ cos ξ − 1)
2 2
follows.
3.4 Matching
Applying the matching principle yields that all quantities in the overlap
region between the core layer and the boundary layer have to be constant.
Thus we conclude that c1 = c2 = 0. Furthermore we obtain
A0
Ū0 = 0, W̄0 = −1 + . (47)
2

Ro = 0
Ek = 0.1 Ro = 1

Ek = 0.001
U

E
k
ï =
0.0
ï 1
ï
ï
ï ï ï ï ï
Z
a) radial velocity component

Ek = 0.1 Ro = 0

Ro = 1

Ek = 0.01
V
ï
ï
ï Ek = 0.001
ï
ï
ï ï ï ï ï
Z
b) azimuthal velocity component

Ro = 0
ï Ro = 1
W Ek = 0.1
ï
ï
ï
ï
ï
Ek = 0.001 Ek = 0.01
ï ï ï ï ï
Z
c) vertical velocity component
Figure 5. Velocity profiles between two rotating discs (solution of differen-
tial eq. (33)) for Ek = 0.1, 0.01, 0.001 and Ro = 0, 1.
14 H. Steinrück
The analysis of the boundary layer at Z = 1 follows the same lines as
indicated above. Only the sign in front of the first derivatives and in the
boundary condition for the azimuthal velocity V has to be changed.
After matching with the upper layer with the core region we obtain
A0
Ū0 = 0, W̄0 = −1 − . (48)
2
Thus we conclude that A0 = 0 and we have determined all leading order
terms of the asymptotic expansion of the velocity profile.
Combining the expansion of the core region with that of the Ekman-
layers we obtain the uniformly valid approximation.
U ∼ e−ξ sin ξ − e−η sin η, V ∼ e−ξ cos ξ − e−η cos η, (49a)

√
W ∼ Ek −1 + e−ξ (sin ξ + cos ξ) + e−η (sin η + cos η) . (49b)
In figure 5 numerical solutions of the similarity equations (33) are shown for
different Ekman- and Rossby numbers. As expected, the velocity profiles
exhibit Ekman-layers near the rotating walls.
4 Model equation for fully developed turbulent

channel flow
As the third example of matched asymptotic expansions we will study the
fully developed turbulent channel flow in the limit of large Reynolds num-
bers. We expect that the reader is familiar with the basic terminology in
turbulent flows, namely the Reynolds decomposition of the flow and pressure
field into mean flow, mean pressure and fluctuating velocities and pressure,
respectively.
Assuming a fully developed flow, i.e. all averaged flow quantities do not
depend on the coordinate x̃ in flow direction the mean velocity ũ and the
shear stress τ̃ are function of the lateral coordinate ỹ only.
The momentum balance yields an equilibrium between the pressure gra-
dient dp̃/dx̃ and the ỹ-derivative of the shear stress τ̃
dp̃ dτ̃
0=− + . (50)
dx̃ dỹ
The averaged shear stress τ̃ is the sum of the Reynolds shear stress τ̃t =
−ρ̃(ũ ṽ ) and the viscous stress
dũ
τ̃ = τ̃t + μ̃ . (51)
dỹ
At the channel wall ỹ = 0 the no slip boundary condition ũ = 0 holds and at

the centerline ỹ = d˜ due to symmetry the shear stress τ̃ vanishes. Here we
consider the center line velocity ũc as a given quantity and want to determine
the velocity profile ũ = ũ(ỹ) and the pressure gradient, respectively. A more
physical boundary condition is that the mean velocity (or volume flux) is
prescribed, but for the sake of simplicity of the analysis we prescribe here
the center line velocity.
In order to close the problem, a relation between the mean flow ũ or
its derivatives and the turbulent shear stress is missing. There is a vast
literature and several approaches how such a closure can be accomplished.
We assume here a simple turbulence model for wall bounded shear flows,
namely the mixing length model

dũ dũ
τ̃t = ρ̃l̃ 2 . (52)
dỹ dỹ
In case of a channel flow an expression for the mixing length l̃ = ˜l(ỹ) = dl(y)
˜
as a function of the dimensionless distance y = ỹ/l̃ from the wall can be
found in Schlichting (2000)
κ κ
l(y) = c0 − 2c0 − (1 − y)2 − − c0 (1 − y)4 . (53)
2 2
Note that the mixing length l̃ vanishes at the wall ỹ = 0 and that
l(y) ∼ κy + l2 y 2 /2 + O(y 3 ) for y 1 holds.
We introduce dimensionless variables by referring the velocity to the
center line velocity ũc , the shear stress to the double stagnation pressure
ρ̃ũ2c , the unknown pressure gradient to ρ̃ũ2c /d.
˜
We define
d˜ dp̃ τ̃w ũ2 1 μ̃

γ2 = − 2
= 2 = τ2 , ε= = , (54)
ρ̃ũc dx̃ ρ̃ũc ũc Re ρ̃ũc d˜
where we have made use of the force balance −dp̃/dx̃ = τ̃w /d˜ for a fully
developed
flow and the definition of the wall shear stress velocity ũτ =
τ̃w /ρ̃ with τ̃w denoting the wall shear stress. The dimensionless equations
reduce to the stress balance
dτ
0 = γ2 + , (55)
dy
and stress relation 2
du du
τ =ε + l(y)2 , (56)
dy dy
16 H. Steinrück
subject to the boundary conditions

u(0) = 0, u(1) = 1, τ (1) = 0. (57)
Integration of the momentum balance yields
τ (y) = γ 2 (1 − y). (58)
and it remains to solve the first order ordinary differential equation (56) for
the velocity profile u and the dimensionless wall shear stress (or negative
pressure gradient) γ 2 .
4.1 Defect Layer

We expand the solution with respect to small values of ε (large Reynolds
numbers). However, we also have to determine γ whose order of magnitude
as a function of ε is not obvious. We anticipate that γ = o(1) as ε → 0.
From the stress relation we deduce that du/dy = O(γ). Thus we expand u
(D) (D) (D)
u(y, ε) ∼ u0 (y) + γ(ε)u1 (y) + εu2 (y). (59)
(D)
Inserting into the stress relation we obtain that is a constant. However,
u0
(D)
we have two contradicting boundary conditions to determine u0 . For the
(D)
next order term u1 we obtain from the stress relation

(D) 2
2 du1
1 − y = l(y) . (60)
dy
Integration yields
(D) 1 (D)
u1 (y) = ln y + F (D) (y) + u1 (1), (61)
κ
where √
y
1 − y 1
F (D) (y) = − dy (62)
1 l(y ) κy
is a smooth bounded function of y on the interval (0, 1).
(D)
Now it is obvious to see that u1 is smooth at the centerline y = 1 and
(D) (D)
thus u0 = 1 and u1 (1) = 0. The velocity profile deviates only by a
small velocity defect of order γ from its maximum value at the center line.
Therefore this layer is called defect layer. Near the wall y = 0 the velocity
(D)
component u1 (y) is singular. Its asymptotic behavior is given by
(D) 1 κ + l2
u1 (y) ∼ ln y +CD − y, with CD = F (D) (0) as y → 0. (63)
κ 2κ2
For the term of order ε we obtain the equation

(D) (D) (D)
du1 du1 du2
0= + 2l2 (y) . (64)
dy dy dy
Integration yields
y
(D) dy 1 l2
u2 =− ∼ 2 + 3 ln y + CD,2 + · · · (65)
1 2l2 (y) 2κ y 2κ
with l2 = l (0).
4.2 Viscous wall layer

At the wall y = 0 the no slip boundary condition cannot be satisfied by
the defect expansion (59). Thus we introduce a local variable η
(v) (v) y
u = γu1 (η) + γσu2 (η) + · · · , η= (66)
σ(ε)
with a stretching σ(ε) which will be determined appropriately. Inserting

yields

(v) (v) 2
du1 2 2 du1 (v)
1= +κ η , u1 (0) = 0, (67)
dη dη

(v) 2 (v) (v) (v)
3 du1 du2 du du2 (v)
−η − 2l2 κη = + 2κ2 η 2 1 , u2 (0) = 0, (68)
dη dη dη dη
with σ(ε)γ(ε) = ε. Integration of (67) yields
(v) 1 1
u1 (η) = ln η + ln 2κ + 4κ2 + 1/η 2 . (69)
κ κ
In order to match the viscous layer to the defect layer we consider the
(v) (v)
asymptotic behavior of u1 (η) and u2 (η) for η → ∞.
(v)
du1 1 1 1
∼ − + 3 3 + ··· , (70a)
dη κη 2κ2 η2 8κ η
(v) 1 1 1
u1 (η) = ln η + CV + 2 − + ··· ,
κ 2κ η 16κ2 η 2
(70b)
1
with CV = (ln 4κ − 1)
κ
18 H. Steinrück
du(v)
du2
(v) −η − κl η3 ( 1 )2 1 l2 l2 3l2
dη
=− (v)
∼− − 2+ 3 − 4 2 (70c)
dη du 2κ 2κ 2κ η 8κ η
1 + 2κ2 η 2 1
dη

(v) 1 l2 l2 3l2
u2 (η) ∼ − + 2 η + CV,2 + 3 ln η + , (70d)
2κ 2κ 2κ 16κ4 η
where CV,2 is an appropriate constant.
4.3 Matching
Finally it remains to match the velocity profile of the viscous layer
with that of the defect layer. Applying the matching principle we have
to care when counting the number of terms in the asymptotic expansions.
In Fraenkel (1969) it had been shown that the Van Dykes matching princi-
ple is still valid when the asymptotic expansion contain besides powers of
the perturbation parameter products of powers of ε and and powers of ln ε.
Than all logarithmic term multiplied by the same power of ε have to be
considered as one term.
In the present example it will turn out that γ(ε) ∼ O(1/ ln(1/ε)). Thus
(D) (D)
the two term expansion of the defect layer is 1 + γU1 + εu2 . Expanding
it in the viscous layer variable into two terms and using γσ = ε we obtain
(2) (2)
(2) (D) (D)
[u]D = 1 + γu1 + εu2 =
V V

1 1 1 1
=1+γ ln σ + ln η + CD + 2 + (71)
κ κ 2κ η

κ + l2 l2 l2
+ε CD,2 − η + ln η + ln σ .
2κ2 2κ3 2κ3
(v) (v)
Taking two terms of the inner (viscous)-layer expansion γu1 + γσu2 ,
rewriting it in the outer variables and expanding it into two term yields
(2) (2)
(2) (v) (v)
[u]V = γu1 + γσu2 =
D D

1 1 κ + l2
=γ ln y + CV − ln σ − y + (72)
κ κ 2κ2

l2 σ 1
+γσ CV,2 + 3 ln y − 3 ln σ + 2 .
2κ 2κ 2κ y
Both expressions agree if the matching condition

1 1 l2
= − ln σ + CV − CD − σ ln σ + CD,2 − CV,2 (73)
γ κ 2κ3
is satisfied. Taking only the first order terms of both expansions the well
known friction law
1 1 γ
= ln + CV − CD (74)
γ κ ε
is obtained. It can be interpreted as a relation between the dimensionless
wall shear stress γ 2 and the Reynolds number ε−1 .

u

YLVFRXVVXEOD\HUSURILOH
ï GHIHFWOD\HUSURILOH
ï YHORFLW\SURILOH
ï

y
Figure 6. solution of model problem and viscous and defect layer approx-
imation for ε = 10−4
In figure 6 velocity profile as the solution of the force balance with the
mixing length model (53) for ε = 10−4, the approximation in the viscous
sub-layer and the defect layer is shown in a logarithmic plot. It can be clearly
seen that in the overlap region (here from 0.02 to 0.2) viscous and defect
expansion agree. In the overlap region both expansions can be represented
by a logarithmic velocity profile.
4.4 Turbulence asymptotics

Here we have considered a very simple turbulent shear flow and have
made used of a simple turbulence model to reveal the asymptotic structure
of flows near the wall. However, the weak point of this approach is the
assumption of a turbulence model.
The traditional approach for the limit of large Reynolds numbers is that
one considers the shear rate in the overlap region of the viscous and the
20 H. Steinrück
defect layer. From dimensional analysis one obtains
ỹ dũ ỹ ũτ ỹ
= Φ(y, y+ ), with y+ = , y= . (75)
ũτ dỹ ν̃ d˜
˜ = Reτ . To get the behavior of Φ in the overlap

Note that y+ /y = ũτ d/ν̃
region we have to consider the double limit y+ → ∞, y → 0. Following
von Karman (1930) we assume that this limit exists and its value is the
reciprocal value of the von Karman constant,
1
lim Φ(y, y+ ) = Φ(∞, 0) = . (76)
y+ →∞,y→0 κ
Integration of (76) yields the logarithmic velocity profile in the overlap re-
gion. We emphasize that the existence of the limit (76) from a theoretical
point of view is a nontrivial assumption (similarity of the first kind, see
Barenblatt (1996). However, the logarithmic law, if interpreted correctly,
is in excellent agreement with measured velocity profiles. Thus it can be
considered as an empirical fact. On the other hand there are authors , e. g.
Barenblatt (1996), who question the logarithmic law. Barenblatt (1996)
considers that the limit (76) does not exist, but that the function Φ is a
sophisticated power function of the Reynolds number (similarity of the sec-
ond kind). Instead of the logarithmic velocity profile these authors obtain
a power-law with an Reynolds-number dependent exponent. Although ac-
cording to Barenblatt (1996) the power law seems to reproduce some data
even better than the log-law it is a dead end from the asymptotic point of
view since it does not comply with the requirements of a rational asymptotic
expansion.
In modern papers concerning turbulence asymptotics the order of argu-
ments is reversed, see Walker (1998), Kluwick and Scheichl (2009). Usually
the dimensionless wall shear stress velocity γ = uτ /Uref is considered as
a small perturbation parameter and the existence of a viscous sub-layer
together with the log-law in the overlap region is postulated.
5 Conclusions
We have given an introduction to the method of matched asymptotic expan-
sion by analyzing three different problems of fluid mechanics. Characteristic
to all examples is the appearance of different length scales and that a uni-
formly valid asymptotic approximation can be constructed employing the
matching principle.
Bibliography
G. I. Barenblatt. Scaling, self-similarity and intermediate asymptotics.
Cambridge Univ. Press, 1996.
A. Betz. Konforme Abbildungen. Springer, 2nd edition, 1964.
L. E. Fraenkel. On the method of matched asymptotic expansions, part i:
A matching principle. Proc. Camb. Phil. Soc., 65:209–231, 1969.
S. Kaplun. In P.A. Lagerstrom, L. N. Howard, and C. S. Liu, editors, Fluid
Mechanics and Singular Perturbation. Academic Press, 1967.
A. Kluwick and B. Scheichl. High-Reynolds-Number Asymptotics of Tur-
bulent Boundary Layers: From Fully Attached to Marginally Separated
Flows. In A. Hegarty, N. Kopteva, E. O’Riordan, and M. Stynes, editors,
BAIL 2008 - Boundary and Interior Layers, volume 69 of Lecture Notes
in Computational Science and Engineering, pages 3–22, 2009.
L. Prandtl. Über Flüssigkeitsbewegung bei sehr kleiner Reibung. In Ver-
handl. d. III. Intern. Mathem. Kongresses, Heidelberg, pages 484–491,
1904.
K. Schlichting, H. Gersten. Boundary-layer theory. Springer, 8th edition,
2000.
M. Ungarish. Hydromechanics of Suspensions. Springer, 1993.
M. Van Dyke. Perturbation Methods in Fluid Mechanics. The Parabolic
Press, Stanford, annoted edition edition, 1975.
Th. von Karman. Mechanische Ähnlichkeit und Turbulenz. Ges. Wiss.
Göttingen, Math.-Phys. Kl., pages 58–76, 1930.
J. D. A. Walker. Turbulent boundary layers ii: Further developments. In
A. Kluwick, editor, Recent Advances in Boundary Layer Therory, volume
390 of CISM Courses and Lectures, pages 145–230. Springer, 1998.
Asymptotic Methods For PDE Problems In
Fluid Mechanics and Related Systems With
Strong Localized Perturbations In
Two-Dimensional Domains
Michael J. Ward*‡ and Mary-Catherine Kropinski†‡

*
Department of Mathematics University of British Columbia, Vancouver, B.C.,
Canada
†
Department of Mathematics, Simon Fraser University, Burnaby, B.C., Canada
1 Introduction
The method of matched asymptotic expansions is a powerful systematic an-
alytical method for asymptotically calculating solutions to singularly per-
turbed PDE problems. It has been successfully used in a wide variety
of applications (cf. Kevorkian and Cole (1993), Lagerstrom (1988), Dyke
(1975)). However, there are certain special classes of problems where this
method has some apparent limitations.
In particular, for singular perturbation PDE problems leading to infinite
logarithmic series in powers of ν = −1/ log ε, where ε is a small positive pa-
rameter, it is well-known that this method may be of only limited practical
use in approximating the exact solution accurately. This difficulty stems
from the fact that ν → 0 very slowly as ε decreases. Therefore, unless
many coefficients in the infinite logarithmic series can be obtained analyti-
cally, the resulting low order truncation of this series will typically not be
very accurate unless ε is very small. Singular perturbation problems in-
volving infinite logarithmic expansions arise in many areas of application
in two-dimensional spatial domains including, low Reynolds number fluid
flow past bodies of cylindrical cross-section, low Peclet number convection-
diffusion problems with localized obstacles, and the calculation of the mean
first passage time for Brownian motion in the presence of small traps, etc.
In this article we survey consider various singularly perturbed PDE prob-
lems in two-dimensional spatial domains where hybrid asymptotic-numerical
methods have been formulated and implemented to effectively ‘sum’ infinite
logarithmic expansions. Some of the problems considered herein directly re-
late to fluid mechanics, whereas other problems arise in different scientific
© CISM, Udine 2010
24 M.J. Ward and M.-C. Kropinski
contexts. One primary goal of this chapter is to highlight how a common

analytical framework can be used to treat a diverse class of problems having
strong localized perturbations in two-dimensional domains.
2 Infinite Logarithmic Expansions: Simple Pipe Flow

We first consider the simple model problem of Titcombe and Ward (1999)
to illustrate some main ideas for treating PDE problems with infinite log-
arithmic expansions. We consider steady, incompressible, laminar flow in
a straight pipe containing a thin core. Both the pipe and the core have
a constant cross-section of arbitrary shape, and thus the problem is two-
dimensional. With these assumptions, the pipe flow is unidirectional and
the velocity component w in the axial direction satisfies (cf. Ward-Smith
(1980))
w = −β , x ∈ Ω\Ωε , (1a)
w = 0, x ∈ ∂Ω , (1b)
w = 0, x ∈ ∂Ωε . (1c)
Here Ω ∈ R2 is the dimensionless pipe cross-section and Ωε is the cross-

section of the thin core. We assume that Ωε has radius O(ε) and that Ωε →
x0 uniformly as ε → 0, where x0 ∈ Ω. We denote the scaled subdomain
that results from an O(ε−1 ) magnification of the length scale of Ωε by
Ω1 ≡ ε−1 Ωε . In (1a), β is defined in terms of the dynamic viscosity μ of
the fluid and the constant pressure gradient dp/dz along the channel by
β ≡ −μ−1 dp/dz. In terms of w, the mean flow velocity w̄ is defined by

1
w̄ ≡ w dx . (2)
AΩ Ω\Ωε
Here AΩ is the cross-sectional area of the pipe-core geometry. For laminar

flow in pipes of non-circular cross-section, with or without cores, the friction
coefficient f is expressed in terms of w̄ by f ≡ −L(dp/dz)/(2ρw̄ 2 ) (cf. Ward-
Smith (1980)). As a remark, the Reynolds number is defined by Re ≡
w̄Lρ/μ, where ρ is the density of the fluid. Laminar flow occurs for Reynolds
numbers in the approximate range 0 < Re < 2000. In the definition of Re,
L is a characteristic diameter defined by L = 4AΩ /PΩ , where PΩ is the
wetted perimeter of the pipe and the core.
The asymptotic solution to (1) is constructed in two different regions:
an outer region defined at an O(1) distance from the perturbing core, and
an inner region defined in an O(ε) neighborhood of the thin core Ωε . The
Asymptotic Methods for PDE Problems 25
analysis below will show how to calculate the sum of all the logarithmic
terms for w in in the limit ε → 0 of small core radius.
In the outer region we expand the solution to (1) as
w(x; ε) = W0 (x; ν) + σ(ε)W1 (x; ν) + · · · . (3)
Here ν = O(1/ log ε) is a gauge function to be chosen, and we assume that

σ ν k for any k > 0 as ε → 0. Thus, W0 contains all of the logarithmic
terms in the expansion. Substituting (3) into (1a) and (1b), and letting
Ωε → x0 as ε → 0, we get that W0 satisfies
W0 = −β , x ∈ Ω\{x0 } , (4a)
W0 = 0 , x ∈ ∂Ω , (4b)
W0 is singular as x → x0 . (4c)
The matching of the outer and inner expansions will determine a singularity
behavior for W0 as x → x0 .
In the inner region near Ωε we introduce the inner variables
y = ε−1 (x − x0 ) , v(y; ε) = W (x0 + εy; ε) . (5)
If we naively assume that v = O(1) in the inner region, we obtain the
leading-order problem for v that y v = 0 outside Ω1 , with v = 0 on ∂Ω1
and v → W0 (x0 ) as |y| → ∞, where y denotes the Laplacian in the y
variable. This far-field condition as |y| → ∞ is obtained by matching v to
the outer solution. However, in two-dimensions there is no solution to this
problem since the Green’s function for the Laplacian grows logarithmically
at infinity. To overcome this difficulty, we require that v = O(ν) in the
inner region and we allow v to be logarithmically unbounded as |y| → ∞.
Therefore, we expand v as
v(y; ε) = V0 (y; ν) + μ0 (ε)V1 (y) + · · · , (6a)

where we write V0 in the form
V0 (y; ν) = νγvc (y) . (6b)

Here γ = γ(ν) is a constant to be determined with γ = O(1) as ν → 0, and
we assume that μ0 ν k for any k > 0 as ε → 0. Substituting (5) and (6)
into (1a) and (1c), and allowing vc (y) to grow logarithmically at infinity,
we obtain that vc (y) satisfies
y vc = 0 , y∈
/ Ω1 ; vc = 0 , y ∈ ∂Ω1 , (7a)
vc ∼ log |y| , as |y| → ∞ . (7b)
The unique solution to (7) has the following far-field asymptotic behavior:
p·y
vc (y) ∼ log |y| − log d + + ··· , as |y| → ∞ . (7c)
|y|2
The constant d > 0, called the logarithmic capacitance of Ω 1 , depends on
the shape of Ω1 but not on its orientation. The vector p is called the
dipole vector. Numerical values for d for different shapes of Ω1 are given in
Ransford (1995), and some of these are reproduced in Table 1. A boundary
integral method to compute d for arbitrarily-shaped domains Ω1 is described
and implemented in Dijkstra and Hochstenbach (2008).
Table 1. The logarithmic capacitance, or shape-dependent parameter, d,

for some cross-sectional shapes of Ω1 = ε−1 Ωε .
Shape of Ω1 ≡ ε−1 Ωε Logarithmic Capacitance d

circle, radius a d=a
ellipse, semi-axes a, b d = a+b 2
√ 3
3 Γ( 13 ) h
equilateral triangle, side h d= 8π 2 ≈ 0.422h
3/4 1 2
3 Γ( ) h
isosceles right triangle, short side h d = 27/2 π43/2 ≈ 0.476h
2
Γ( 1 ) h
square, side h d = 4π43/2 ≈ 0.5902h
The leading-order matching condition between the inner and outer so-
lutions will determine the constant γ in (6b). Upon writing (7c) in outer
variables and substituting into (6b), we get the far-field behavior
v(y; ε) ∼ γν [log |x − x0 | − log(εd)] + · · · , as |y| → ∞ . (8)
Choosing
ν(ε) = −1/ log(εd) , (9)
and matching (8) to the outer expansion (3) for W , we obtain the singularity
condition for W0 ,
W0 = γ + γν log |x − x0 | + o(1) , as x → x0 . (10)
The singularity behavior in (10) specifies both the regular and singular
part of a Coulomb singularity. As such, it provides one constraint for the
determination of γ. More specifically, the solution to (4) together with
(10) must determine γ, since for a singularity condition of the form W0 ∼
S log |x − x0 | + R for an elliptic equation, the constant R is not arbitrary

but is determined as a function of S, x0 , and Ω.
The solution for W0 is decomposed as
W0 (x; ν) = W0H (x) − 2πγνGd (x; x0 ) . (11)
Here W0H (x) is the smooth function satisfying the unperturbed problem
W0H = −β , x ∈ Ω; W0H = 0 , x ∈ ∂Ω . (12)
In (11), Gd (x; x0 ) is the Dirichlet Green’s function satisfying
Gd = −δ(x − x0 ) , x ∈ Ω ; Gd = 0 , x ∈ ∂Ω , (13a)
1
Gd (x; x0 ) = − log |x − x0 | + Rd (x0 ; x0 ) + o(1) , as x → x0 . (13b)
2π
Here Rd00 ≡ Rd (x0 ; x0 ) is the regular part of the Dirichlet Green’s function
Gd (x; x0 ) at x = x0 . This regular part is also known as either the self-
interaction term or the Robin constant (cf. Bandle and Flucher (1996)).
Upon substituting (13b) into (11) and letting x → x0 , we compare the
resulting expression with (10) to obtain that γ is given by
W0H (x0 )
γ= . (14)
1 + 2πνRd00
Therefore, for this problem, γ is determined as the sum of a geometric
series in ν. The range of validity of (14) is limited to values of ε for which
2πν|Rd00 | < 1. This yields,
1
0 < ε < εc , εc ≡ exp [2πRd00 ] . (15)
d
We summarize our result as follows:
Principal Result 1: For ε 1, the outer expansion for (1) is
2πνW0H (x0 )
w ∼ W0 (x; ν) = W0H (x) − Gd (x; x0 ) , for |x − x0 | = O(1) ,
1 + 2πνRd00
(16a)
and the inner expansion with y = ε−1 (x − x0 ) is
νW0H (x0 )
w ∼ V0 (y; ν) = vc (y) , for |x − x0 | = O(ε) . (16b)
1 + 2πνRd00
Here ν = −1/ log(εd), d is defined in (7c), vc (y) satisfies (7), and W0H sat-
isfies the unperturbed problem (12). Also Gd (x; x0 ) and Rd00 ≡ Rd (x0 ; x0 )
are the Dirichlet Green’s function and its regular part satisfying (13).
This formulation is referred to as a hybrid asymptotic-numerical method

since it uses the asymptotic analysis as a means of reducing the original
problem (1) with a hole to the simpler asymptotically related problem (4)
with singularity behavior (10). This related problem does not have a bound-
ary layer structure and so is easy to solve numerically. The numerics re-
quired for the hybrid problem involve the computation of the unperturbed
solution W0H and the Dirichlet Green’s function Gd (x; x0 ). In terms of Gd
we then identify its regular part Rd (x0 ; x0 ) at the singular point. From the
solution to the canonical inner problem (7) we then compute the logarithmic
capacitance, d. The result (16a) then shows that the asymptotic solution
only depends on the product of εd and not on ε itself. This feature allows for
an asymptotic equivalence between holes of different cross-sectional shape,
based on an effective ‘radius’ of the cylinder. This equivalence is known as
Kaplun’s equivalence principle (cf. Kaplun (1957), Kropinski et al. (1995)).
An advantage of the hybrid method over the traditional method of
matched asymptotic expansions is that the hybrid formulation is able to
sum the infinite logarithmic series and thereby provide an accurate approx-
imate solution. From another viewpoint, the hybrid problem is much easier
to solve numerically than the full singularly perturbed problem (1). For the
hybrid method a change of the shape of Ω1 requires us to only re-calculate
the constant d. This simplification does not occur in a full numerical ap-
proach.
We now outline how Principal Result 1 can be obtained by a direct
summation of a conventional infinite-order logarithmic expansion for the
outer solution given in the form
∞

W ∼ W0H (x) + ν j W0j (x) + μ0 (ε)W1 + · · · , (17)
j=1
with μ0 (ε) ν k for any k > 0. By formulating a similar series for the inner
solution, we will derive a recursive set of problems for the W0j for j ≥ 0
from the asymptotic matching of the inner and outer solutions. We will
then sum this series to re-derive the result in Principal Result 1.
In the outer region we expand the solution to (1) as in (17). In (17),
ν = O(1/ log ε) is a gauge function to be chosen, while the smooth function
W0H satisfies the unperturbed problem (12) in the unperturbed domain.
By substituting (17) into (1a) and (1b), and letting Ωε → x0 as ε → 0, we
get that W0j for j ≥ 1 satisfies
W0j = 0 , x ∈ Ω\{x0 } , (18a)

W0j = 0 , x ∈ ∂Ω , (18b)
W0j is singular as x → x0 . (18c)
The matching of the outer and inner expansions will determine a singularity
behavior for W0j as x → x0 for each j ≥ 1.
In the inner region near Ωε we introduce the inner variables
y = ε−1 (x − x0 ) , v(y; ε) = W (x0 + εy; ε) . (19)
We then pose the explicit infinite-order logarithmic inner expansion

∞

v(y; ε) = γj ν j+1 vc (y) . (20)
j=0
Here γj are ε-independent coefficients to be determined. Substituting (20)

and (1a) and (1c), and allowing vc (y) to grow logarithmically at infinity, we
obtain that vc (y) satisfies (7) with far-field behavior (7c).
Upon using the far-field behavior (7c) in (20), and writing the resulting
expression in terms of the outer variable x − x0 = εy, we obtain that
∞

v ∼ γ0 + ν j [γj−1 log |x − x0 | + γj ] . (21)
j=1
The matching condition between the infinite-order outer expansion (17) as

x → x0 and the far-field behavior (21) of the inner expansion is that
∞
∞

W0H (x0 ) + ν j W0j (x) ∼ γ0 + ν j [γj−1 log |x − x0 | + γj ] . (22)
j=1 j=1
The leading-order match yields that
γ0 = W0H (x0 ) . (23)
The higher-order matching condition, from (22), shows that the solution
W0j to (18) must have the singularity behavior
W0j ∼ γj−1 log |x − x0 | + γj , as x → x0 . (24)
The unknown coefficients γj for j ≥ 1, starting with γ0 = W0H (x0 ), are

determined recursively from the infinite sequence of problems (18) and (24)
for j ≥ 1. The explicit solution to (18) with W0j ∼ γj−1 log |x − x0 | as

x → x0 is given explicitly in terms of Gd (x; x0 ) of (13) as
W0j (x) = −2πγj−1 Gd (x; x0 ) . (25)
Next, we expand (25) as x → x0 and compare it with the required

singularity structure (24). This yields

1
−2πγj−1 − log |x − x0 | + Rd00 ∼ γj−1 log |x − x0 | + γj , (26)
2π
where Rd00 ≡ Rd (x0 ; x0 ). By comparing the non-singular parts of (26), we

obtain a recursion relation for the γj , valid for j ≥ 1, given by
γj = (−2πRd00 ) γj−1 , γ0 = W0H (x0 ) , (27)
which has the explicit solution
γj = [−2πRd00 ]j W0H (x0 ) , j ≥ 0. (28)
Finally, to obtain the outer solution we substitute (25) and (28) into
(17) to obtain
∞
∞

w − W0H (x) ∼ ν j (−2πγj−1 ) Gd (x; x0 ) = −2πνGd (x; x0 ) ν j γj
j=1 j=0
∞

∼ −2πνW0H (x0 )Gd (x; x0 ) [−2πνRd00 ]j
j=0
2πνW0H (x0 )
∼− Gd (x0 ; x0 ) . (29a)
1 + 2πνRd00
Equation (29a) agrees with equation (16a) of Principal Result 1. Similarly,
upon substituting (28) into the infinite-order inner expansion (20), we obtain
∞
j νW0H (x0 )
v(y; ε) = νW0H (x0 )vc (y) [−2πRd00 ν] = vc (y) , (30)
1 + 2πνRd00
j=0
which recovers equation (16b) of Principal Result 1.

Next, we compare the results of the hybrid method with results obtained
either analytically or numerically from the full perturbed problem (1).
We consider flow in a circular pipe Ω of cross-sectional radius r0 that
contains a concentric core Ωε of various cross-sectional shapes centered at
the origin. We use Table 1 for the logarithmic capacitance d for a square
core, an elliptical core, and an equilateral triangular core. Using the no-
tation in the table, we set the major and minor semi-axes of the ellipse as
a = 2 and b = 1, and both the side of the square and the equilateral triangle
as h = 1. To compute the hybrid method solution, we readily calculate that
the Green’s function is Gd = −(2π)−1 log(r/r0 ) and that the unperturbed
solution is W0H = β(r02 − r2 )/4. The outer hybrid method solution, as
obtained from (16a) of Principal Result 1, is simply

β 2 2 2 log(r0 /r)
w(x; ε) = r − r − r0 , r = |x| . (31)
4 0 log(r0 /[εd])
From (31), we can compute the asymptotic mean flow rate using (2).
0.42 0.45
Hybrid Hybrid
Full Numerical Exact
0.4 0.44
0.38 Equilateral Triangle 0.43
Square
W
W
0.36 0.42
0.34 0.41
Ellipse
0.32 0.4
0.3 0.39
0.01 0.02 0.03 0.04 0.05 0.06 0.07 0.08 0.09 0.1 0.01 0.02 0.03 0.04 0.05 0.06 0.07 0.08 0.09 0.1
ε ε
(a) Concentric Geometry: w̄ vs. ε (b) Eccentric Geometry: w̄ vs. ε
Figure 1. The mean flow velocity w̄ versus the cross-sectional ‘radius’ of

an inner core pipe located inside a circular pipe of cross-sectional radius
r0 = 2. Left figure: (Concentric annulus geometry). Plots of w̄ vs. ε for
three different cross-sectional shapes of the inner core pipe. The discrete
points are the full numerical results. Right figure (Eccentric Geometry).
Plots of w̄ versus the circular pipe core cross-sectional radius ε when the
inner pipe is centered at x0 = (−1, 0). The hybrid and exact results are the
dotted and solid curves, respectively.
To validate the asymptotic results for w̄, we compare them with corre-
sponding direct numerical results computed from the full problem (1) using
the PDE Toolbox of MATLAB (1996). For a circular pipe of radius r0 = 2
containing a concentric core and with β = 1, Fig. 1(a) contains curves of
mean flow velocity, w̄, versus ε, a measure of the size of the core, for three
different cross-sectional shapes of the core. In the hybrid method solution,

the change in shape and size of the core requires only that we vary the
product εd, which allows us to compute the entire ε curve very easily. In
contrast, for each change of shape and size of the core in the direct nu-
merical solution, we had to re-create the solution geometry and re-mesh
the solution grid when using the PDE Toolbox of MATLAB (1996). For a
core of elliptic cross-section, the figure shows that the hybrid method re-
sults agree very well with those of the direct numerical solution. The slight
discrepancy in comparing the results for the other two core cross-sectional
shapes, the square and equilateral triangle, could be due to the inability of
the numerical method to resolve the non-smooth boundary of the core.
Next, we consider flow in a circular pipe Ω of radius r0 > 1 that contains
a circular core Ωε of radius ε centered at x0 = (−1, 0). For this case,
the exact mean flow velocity w̄E for this eccentric annulus geometry can
be written as a complicated infinite series as in Ward-Smith (1980). In
contrast, we need only calculate three specific quantities for our hybrid
formulation in (16). Firstly, the unperturbed solution is again given by
W0H (r) = β(r02 − r2 )/4. Next, since the inner core cross-section is a circle
of radius ε, then the logarithmic capacitance is d = 1, so that ν = −1/ log ε.
Finally, using the method of images, we solve (13) analytically to obtain
the Green’s function

1 |x − x0 |r0
Gd (x; x0 ) = − log . (32)
2π |x − x0 ||x0 |
Here the image point x0 of x0 in the circle of radius r0 lies along the ray
containing x0 and satisfies |x0 ||x0 | = r02 . Comparing (32) with (13b), we
can then calculate the self-interaction term as

1 r0
Rd00 ≡ Rd (x0 ; x0 ) = − log . (33)
2π |x0 − x0 ||x0 |
Substituting (32), (33), ν = −1/ log ε, and W0H (r), into (16a) we obtain
the outer solution for the hybrid method. This solution is then used in (2)
with AΩ ∼ πr02 to compute the mean flow velocity for the hybrid method.
The integral in (2) is obtained from a numerical quadrature. For an eccentric
annulus with pipe radius r0 = 2, and with β = 1, in Fig. 1(b) we plot the
mean flow velocity w̄ versus the circular core radius ε as obtained from
the exact solution and from the hybrid solution. This plot shows that the
hybrid method results compare rather well with the exact results.
We remark that for an inner pipe core of an arbitrary shape centered at
x0 = (−1, 0), the hybrid method solution as obtained above for the eccentric
annulus still applies, provided that we replace ν = −1/ log ε in (16a) with
ν = −1/ log(εd), where d is to be computed from (7). In particular, if there

is an ellipse with semi-axes ε and 2ε centered at x0 = (−1, 0) instead of the
circle of radius ε, then from Table 1 we get d = 3/2. Hence, the plot in
Fig. 1(b) for the hybrid solution still applies provided that we replace the
horizontal axis in this figure by 3ε/2.
3 Some Related Steady-State Problems in Bounded

Singularly Perturbed Domains
In this section we extend the analysis of §2 to treat some related steady-
state problems. The problem in 3.1, which concerns the distribution of
oxygen partial pressure in muscle tissue, involves multiple inclusions in a
two-dimensional domain. In §3.2 we show how to extend the method of §2
to a nonlinear problem.
3.1 Oxygen Transport From Capillaries to Skeletal Muscle

The analytical study of tissue oxygenation from capillaries dates from
the original work of Krogh (1919). In the oxygen distribution process of the
micro-circulation, oxygen binds to its carrier, haemoglobin, in red blood
cells, which transports it through the arterioles, branching to the capillary
networks, to the collecting venules. In the capillaries, the oxygen is released
from its carrier and diffuses into the surrounding tissue. The reviews of
Popel (1989), Fletcher (1978), and the references in Titcombe and Ward
(2000), provide substantial information on theoretical research in this area.
x2
x3
x1
Capillary
Cross-section
2-D cut
Figure 2. Mathematical idealization of capillary blood supply in skeletal

muscle tissue
In this section, we show how to determine the steady-state oxygen partial

pressure distribution in a two-dimensional domain representing a transverse
section of skeletal muscle tissue that receives oxygen from an array of cap-
illaries of small but arbitrary cross-sectional shape (see Fig. 2). Following
the approach of many authors (e.g. Popel (1989)), we model the transport
of oxygen from capillaries to tissue by a passive diffusive process. Assuming
Fick’s law, J = −D∇c, relating the oxygen flux J to the gradient of oxygen
concentration c, and Henry’s law, c = αp the dimensionless steady-state
oxygen partial pressure p satisfies
N
p = M , x ∈ Ω\Ωp Ωp ≡ ∪ Ωεj , (34a)
j=1
∂n p = 0, x ∈ ∂Ω . (34b)
ε∂n p + κj (p − pcj ) = 0 , x ∈ ∂Ωεj , j = 1, . . . , N . (34c)
The condition (34c) models the capillary wall as a finitely permeable mem-
brane, where κi > 0 is the permeability coefficient of the ith capillary and
pci is the oxygen partial pressure within the ith capillary (assumed con-
stant). In (34c) and (34b), ∂n is the outward normal derivative to the
tissue domain. In deriving (34) we have assumed that the oxygen diffusiv-
ity is spatially constant, and so the oxygen consumption rate M has been
normalized by this constant value. To incorporate skeletal muscle tissue het-
erogeneities, such as localized oxygen-consuming mitochondria, we assume
that M is spatially-dependent and has the form
m
|x − xi |2
M(x) = M0 + Mi exp − , (35)
i=1
σi2
for some positive constants M0 and Mi for i = 1, . . . , m.

The model (34) is an extension of the well-known Krogh cylinder model
Krogh (1919), which consists of one capillary of circular cross-section con-
centric with a circular cross-section of muscle tissue. For this concentric
annulus geometry ε < |x| < 1, the exact radially symmetric solution pe is
ε
M r2 − ε2 ε2 − 1
pe (r) = pc1 + + + log . (36)
2 2 κ1 r
This shows that pe = O(log ε) as ε → 0, as induced by the Neumann
boundary condition in (34b) on the boundary of the cross-section. In the
extended model (34), formulated originally in Titcombe and Ward (2000),
one allows for multiple capillaries of arbitrary location, of arbitrary cross-
sectional shape, and for the tissue domain to be arbitrary.
Most previous attempts to study the oxygenation of muscle tissue an-

alytically have assumed that the capillaries can be represented as point
sources for (34). However, by using the method of matched asymptotic
expansions, we show that this type of rough simplification represents only
the leading-order term in an infinite asymptotic expansion of the oxygen
partial pressure in powers of −1/ log ε, where ε is a measure of the capillary
cross-section. From a physiological viewpoint, this type of point-source ap-
proximation ignores the effect of the shape of the capillary cross-section and
the effect of the interaction between the capillaries. When many capillaries
are present, the effect of the capillary interaction should be significant.
Our goal here is to extend the hybrid method of §2 to calculate the
asymptotic solution to (34) with an error that is smaller than any power of
−1/ log ε. Such an approach, which effectively sums the infinite logarithmic
series, takes into account the effect of the capillary interaction.
In the outer region we expand the solution to (34) as
p(x; ε) = P0 (x; ν1 , . . . , νN ) + σ(ε)P1 (x; ν1 , . . . , νN ) + · · · . (37)
Here νj = O(1/ log ε) for j = 1, . . . , N are gauge functions to be chosen,

and we assume that σ νjk for any k > 0 as ε → 0. Thus, P0 contains all
of the logarithmic terms in the expansion. Substituting (37) into (34a) and
(34b), and letting Ωεj → xj as ε → 0, we get that P0 satisfies
P0 = M , x ∈ Ω\{x1 , . . . , xN } , (38a)
∂n P0 = 0 , x ∈ ∂Ω , (38b)
P0 is singular as x → xj . (38c)
The matching of the outer and inner expansions will determine singularity
behaviors for P0 as x → xj for j = 1, . . . , N .
In the inner region near the j th capillary Ωεj we introduce the inner
variables
y = ε−1 (x − xj ) , p(y; ε) = qj (xj + εy; ε) , (39)
together with the local expansion
qj = pcj + q0j (y; ν1 , . . . , νN ) + μq1j (y; ν1 , . . . , νN ) + · · · . (40)
Here we assume that μ νjk for any k > 0. We then write q0j in the form
q0j = Aj qcj (y) , (41)
where Aj = Aj (ν1 , . . . , νN ) is an unknown constant to be determined, and

qcj (y) ∼ log |y| as y → ∞. By substituting (39), (40), and (41), into (34a)
and (34c), we readily derive that qcj is the solution to
y qcj = 0 , y∈
/ Ωj ; ∂n qcj + κj qc = 0 , y ∈ ∂Ωj , (42a)
qcj ∼ log |y| , as |y| → ∞ , (42b)
where Ωj ≡ ε−1 Ωεj . The unique solution to (42) has the following far-field
asymptotic behavior:

1
qcj (y) ∼ log |y| − log dj + O , |y| 1 . (42c)
|y|
In comparing (42) with (7) for the pipe problem of §2, we observe that here
dj = dj (κj ). For a particular cross-sectional shape of the capillary and
for a given value of κj , one must compute dj = dj (κj ) numerically from a
boundary integral method applied to (42). For a circular capillary of radius
ε, for which qcj can be found analytically, we readily calculate that
dj = exp (−1/κj ) . (43)
Moreover, by comparing (6b) with (41) we observe that here we have intro-
duced a slight change in the definition of the inner solution. In the analysis
below, we will show that Aj = O(1) as ε → 0 in (41), which is a direct
consequence of the Neumann boundary condition in (34b).
By using (40) and (42c), we re-write the far-field form for |y| 1 of the
inner solution in terms of the outer variables as
Aj
qj ∼ pcj + Aj log |x − xj | + . (44a)
νj
Here we have introduced the logarithmic gauge function νj by

1
νj ≡ − . (44b)
log(εdj )
The matching condition is that the far-field form (44a) of the inner solution
must agree with the near-field behavior of the outer solution for p. Thus,
P0 satisfies (38) subject to the following singularity behavior
Aj
P0 ∼ pcj + Aj log |x − xj | + , as x → xj , j = 1, . . . , N . (45)
νj
As remarked in §2, we emphasize that the singularity behavior in (45)

specifies both the regular and singular part of a Coulomb singularity at
each xj . As such, the singularity behaviors (45) for j = 1, . . . , N will
provide N equations for the determination of the unknown constants Aj for

j = 1, . . . , N . By using the divergence theorem, it readily follows that (38),
together with (45), has a solution if and only if

N
1
Aj = − M(x) dx . (46)
2π Ω
j=1
This provides one equation for the determination of Aj for j = 1, . . . , N ,

and shows that Aj = O(1) as ε → 0.
Next, we decompose the solution to (38) and (45) in the form

N
P0 = PR (x) − 2π Ai GN (x; xi ) + χ . (47)
i=1
Here χ is an unknown constant, and PR (x) is the unique solution of

1
PR = M − M(x) dx , x ∈ Ω ; ∂n PR = 0 , x ∈ ∂Ω , (48)
|Ω| Ω

with Ω PR (x) dx = 0. Here |Ω| denotes the area of Ω. When M is a
spatially independent, then PR = 0 for each x ∈ Ω. In (47), GN (x; ξ) is the
Neumann Green’s function, defined as the solution to
1
ΔGN = − δ(x − ξ) , x ∈ Ω ; ∂n GN = 0 , x ∈ ∂Ω , (49a)
|Ω|
1
GN (x; ξ) ∼ − log |x − ξ| + RN (ξ; ξ) + o(1) , as x → ξ , (49b)
2π

with Ω GN (x; ξ) dx = 0. The constant RN (ξ; ξ) is called either the self-
interaction term or the regular part of GN . Since GN and PR have zero
spatial averages, then χ in (47) is the spatial average of P0 .
Finally, we expand the solution (47) as x → xj and we compare the
regular part of the resulting expression with the regular part of the required
singularity structure in (45). In this way, we obtain N algebraic equations
for the unknowns χ and A1 , . . . , AN :
⎡ ⎤
⎢ N
⎥ Aj
PR (xj )−2π ⎣Aj RN jj + Ai GN ji ⎦ +χ = +pcj , j = 1, . . . , N . (50)
i=1
νj
i=j
Here we have defined RN jj ≡ RN (xj ; xj ) and GN ji ≡ GN (xj ; xi ). The

remaining equation relating these unknowns is (46). We summarize our
asymptotic construction as follows:
Principal Result 2: For ε → 0, the asymptotic solution to (34) near the j th

capillary, is
p ∼ pcj + Aj qcj (y) , y = ε−1 (x − xj ) = O(1) , (51a)
where qcj satisfies (42). In the outer region, defined at O(1) distances from
the centers of the capillaries, we have

N
p ∼ PR (x) − 2π Ai GN (x; xi ) + χ . (51b)
i=1
Here PR (x) satisfies (48), and GN is the Neumann Green’s function, as

defined by (49). The constants Aj for j = 1, . . . , N and χ satisfy the N + 1
dimensional linear algebraic system defined by (50) and (46).
To implement the hybrid method, we must compute the Neumann Green’s
function GN and its regular part RN . This can be done explicitly for the
unit disk (see equation (4.3) of Kolokolnikov et al. (2005)) and for a rectan-
gle. In particular, upon representing points as complex numbers, we obtain
for the unit disk that

1 ξ 1 3
GN (x; ξ) =
− log |x − ξ| − log x|ξ| − 2 2
+ (|x| + |ξ| ) − ,
2π |ξ| 2 4
(52a)

1 ξ 3
RN (ξ; ξ) = − log ξ|ξ| − + |ξ|2 − . (52b)
2π |ξ| 4
For more general domains, GN and its regular part can be computed numer-
ically from a boundary integral method (see Pillay et al. (to appear, 2010)).
Then, after specifying M, we can compute the smooth function PR numer-
ically from (48), and evaluate it at each capillary location xj . Finally, the
effect of the cross-sectional shape of the capillary and the permeability of
the capillary wall enters only into the determination of the shape-dependent
parameters dj to be used in νj = −1/ log(εdj ) in (50). This information,
required in (50) and (46), represents the numerical part of the hybrid al-
gorithm. The numerical solution to the linear system (50) and (46) then
determines the strengths, Ai , of the singularities and the average pressure
χ as functions of ε.
As an illustration of the theory, we consider N = 4 capillaries of cir-
cular cross-section, each of radius ε, located inside a circular tissue do-
main Ω of unit radius. Therefore, di = 1 for i = 1, . . . , 4. For each
fixed j, with j = 1, 2, 3, the capillaries are centered at the locations xji =
j/4 (cos ((2i − 1)π/4) , sin ((2i − 1)π/4)) for i = 1, . . . , 4 (see Fig. 3(a) for
4.95
j=3
4.9
j=2
min
ε j=1 4.85
p
4.8 j=1 (+)
j=2 (x)
4.75 j=3 (*)
4.7
0.01 0.02 0.03 0.04 0.05 0.06 0.07 0.08 0.09 0.1
ε
(a) Three arrangements of (b) pmin vs. ε for the three

four traps in the unit disk arrangements
Figure 3. Left figure: Locations of four identical circular capillaries, each of

radius ε, centered at (±j/4 cos(π/4), ±j/4 sin(π/4)) for j = 1, 2, 3 inside the
unit disk. Right figure: Minimum oxygen partial pressure pmin versus the
capillary radius ε for the three arrangements shown in the left figure. The
parameter values are M = 0.3, with pci = 5 and κi = ∞ for i = 1, . . . , 4.
The solid curves for j = 1, 2, 3 are from the hybrid-method solution, while
the discrete points are the full numerical results.
the geometry). For simplicity we choose a constant oxygen consumption

rate M = 0.3, with capillary permeability coefficients κ i = ∞, and intra-
capillary oxygen partial pressure pci = 5, for i = 1, . . . , 4. In Fig. 3(b) we
plot the minimum oxygen partial pressure pmin = min p(x) versus ε for
x∈Ω\Ωp
each of the three arrangements of four traps as calculated from the hybrid
formulation (51), (50), and (46). In this figure we also show that the full
numerical results for pmin , as computed directly from (34) using the PDE
Toolbox of MATLAB (1996), agree very well with the hybrid results.
Additional illustrations of the asymptotic theory are given in Titcombe
and Ward (2000). It is an open problem to extend the asymptotic method-
ology to analyze more biologically realistic models of oxygen transport in
muscle tissue by considering the local fluid flow inside each capillary and to
allow for nonlinear saturation effects on the oxygen consumption rate. More
elaborate mathematical models addressing some of these issues are given in
Mikelić and Primicerio (2006).
3.2 A Nonlinear Elliptic Problem

Next, we show how the method of §2 can be extended to treat a nonlin-
ear elliptic second-order problem on a bounded two-dimensional domain Ω,
which contains a small hole Ωε , formulated as
w + F (w) = 0 , x ∈ Ω\Ωε , (53a)

∂n w + b(w − wb ) = 0 , x ∈ ∂Ω , (53b)
w = α, x ∈ ∂Ωε . (53c)
Here α is constant, F (w) is a smooth function of w, ∂n denotes the out-

ward normal derivative, b > 0, and Ωε is a small hole of radius O(ε) with
Ωε → x0 ∈ Ω uniformly as ε → 0. Nonlinear problems of this type arise in
many applications, including steady-state combustion theory where F (w)
is an exponential function (cf. Ward et al. (1993)). The primary difference
between the linear problem (1) and the unperturbed problem correspond-
ing to (53) is that, depending on the form of the nonlinearity F (w), the
unperturbed problem may have no solution, a unique solution, or multiple
solutions. We shall assume that the unperturbed problem has at least one
solution, and we will focus on determining how a specific solution to this
problem is perturbed by the presence of the subdomain Ωε .
In the outer region we expand w as in (3). The leading-order term
W0 (x; ν) in this expansion satisfies
W0 + F (W0 ) = 0 , x ∈ Ω\{x0 } , (54a)

∂n W0 + b(W0 − wb ) = 0 , x ∈ ∂Ω , (54b)
W0 is singular as x → x0 . (54c)
The analysis of the solution in the inner region is the same as that in §2
since the effect of the nonlinear term in the inner region is O(ε 2 ), which is
transcendentally small compared to the logarithmic terms. Hence, W0 must
have the following singular behavior (see equation (10)):
W0 = α + γ + γν log |x − x0 | + o(1) , as x → x0 . (55)
Here γ = γ(ν) is to be found and ν is defined in terms of the logarithmic

capacitance d of (7) by ν = −1/ log(εd).
At this stage, the analysis of (53) differs slightly from its linear counter-
part (1). We suppose that for some range of the parameter S we can find a
solution to (54) with the singular behavior
W0 ∼ S log |x − x0 | , as x → x0 . (56)
Then, in terms of this solution we define the regular part R = R(S; x0 ) of

this Coulomb singularity by
R(S; x0 ) = lim (W0 − S log |x − x0 |) . (57a)

x→x0
In general R is a nonlinear function of S at each x0 . Therefore, we have
W0 ∼ S log |x − x0 | + R(S; x0 ) + o(1) , as x → x0 . (57b)
By equating (57b) to (55) we get S = νγ and R = α + γ, where ν =

−1/ log(εd). For fixed εd and α, these relations are two nonlinear algebraic
equations for the two unknowns S and γ. Alternatively, we can view these
relations as providing a parametric representation of the desired curve γ =
γ(ν) in the form ν = ν(S) and γ = γ(S), where
S
γ = R(S; x0 ) − α , ν= . (58)
R(S; x0 ) − α
The equation for ν in (58) is an implicit equation determining S in terms

of ε from ν = −1/ log(εd). Therefore, we can analytically sum all of the
logarithmic terms in the expansion of the solution to (53) provided that we
compute the solution to (54), with singular behavior (56), and then identify
R(S; x0 ) from (57a). In general this must be done numerically. However, we
now illustrate the method with an example where R(S; x0 ) can be calculated
analytically.
Let Ω be the unit disk, and take b = ∞, wb = 0, F (w) = ew , and assume
that Ωε is an arbitrarily-shaped hole centered at the origin. Then, (54) and
(56) reduce to a radially symmetric problem for W0 (r), given by
1
W0 + W0 + eW0 = 0 , 0 < r ≤ 1 ; W0 = 0 , on r = 1 , (59a)
r
W0 ∼ S log r , as r → 0 , (59b)
where r = |x|. This problem (59) can be solved analytically by first intro-
ducing the new variables v and η defined by
v = W0 − S log r , η = r1+S/2 . (60)
When S > −2, we then obtain that v = v(η) is smooth and satisfies
−2
1 S
v + v + 1 + ev = 0 , 0 ≤ η ≤ 1; v = 0, on η = 1 . (61)
η 2
The well-known solution v = v(η) to (61) (see Ward et al. (1993)) can be
written in parametric form as
−2
1+ρ S 8ρ
v(η) = 2 log , 1+ = . (62)
1 + ρη 2 2 (1 + ρ)2
The maximum of the right-hand side of the implicit expression for ρ(S) in
(62) is 2 and occurs when ρ = 1. Therefore, for the existence√ of a solution
2
to (59)
√ we require that (1 + S/2) > 1/2, which yields S > 2 − 2. When
S > 2 − 2, then ρ(S) from (62) has two roots for ρ, and hence (59) has
two solutions. Let us consider the smaller root, labeled by ρ− (S), given by

1/2
ρ− (S) = (S + 1)(S + 3) − (S + 2) (S + 2)2 − 2 . (63)
Setting η = 0 in (62), and using (60), we compare with (57a) to obtain
v(0) = R(S; 0) = 2 log(1 + S/2) + log [8ρ− (S)] . (64)
Substituting (64) into (58) gives a parametric representation of the curve

γ = γ(ν) in the form ν = ν(S) and γ = γ(S).
4 Slow Viscous Flow Over a Cylinder

Next, we consider the classical problem of slow, steady, two-dimensional flow
of a viscous incompressible fluid around an infinitely long straight cylinder.
For simplicity, we assume that the cross-sectional shape of the cylinder is
symmetric about the direction of the oncoming stream, but otherwise is
arbitrary. By slow we mean that the Reynolds number ε ≡ U ∞ L/ν is small,
where U∞ is the velocity of the fluid at infinity, ν is the kinematic viscosity,
and 2L is the diameter of the cross-section of the cylinder.
For ε → 0, the method of matched asymptotic expansions was used
systematically in Kaplun (1957) and in Proudman and Pearson (1957) to
resolve the well-known Stokes paradox, and to calculate asymptotically the
stream function in both the Stokes region, which is near the body, and in
the Oseen region, which is far from the body. These pioneering studies
showed that, for ε → 0, the asymptotic expansion for the drag coefficient
CD of a circular cylindrical body starts with CD ∼ 4πε−1 S(ε), where S(ε)
is an infinite series in powers of 1/ log ε. The coefficients in this series are
determined in terms of the solutions to certain forced Oseen problems. For
a cylinder of arbitrary cross-section, it was shown in Kaplun (1957) that
CD ∼ 4πε−1 S(εdf ), where df is an ‘effective’ radius of the cylinder. This
result establishes a certain asymptotic equivalence for CD between cylin-

ders of various cross-sectional shapes and is known as Kaplun’s equivalence
principle.
In an effort to determine CD quantitatively, analytical formulae for the
first three coefficients in S(ε) were derived in Kaplun (1957). However, as a
result of the slow decay of 1/ log ε with decreasing values of ε, the resulting
three-term truncated series for CD agrees rather poorly with the experimen-
tal results of Tritton (1959) unless ε is very small. Owing to the complexity
of the calculations required, it is impractical to obtain a closer quantitative
determination of the drag coefficient by calculating further coefficients in
S(ε) analytically. As a result of these fundamental long-standing difficul-
ties, the problem of slow viscous flow around a cylinder has served as a
paradigm for problems where a matched asymptotic analysis fails to be of
much practical use, unless ε is very small. A comprehensive recent survey
of asymptotic and renormalization group methods applied to slow viscous
flow problems is given in Veysey and Goldenfeld (2007).
In Kropinski et al. (1995), this problem was re-examined and a hybrid
asymptotic-numerical method was formulated and implemented to effec-
tively sum the infinite logarithmic expansions that arise from the singular
perturbation analysis of slow viscous flow around a cylinder. Our approach
differs from the hybrid method employed in Lee and Leal (1986) in which
numerical methods are used within the framework of the method of matched
asymptotic expansions to calculate the first few coefficients in the logarith-
mic expansions of the flow field and the drag coefficient. Instead, we show
that these entire infinite logarithmic series are contained in the solution to
a certain related problem that does not involve the cross-sectional shape of
the cylinder. The overall framework of our approach is similar to that done
in §2 and §3, and is outlined below.
The model formulation is as follows. Consider steady, incompressible,
viscous flow around a cylindrical body with a uniform stream of velocity
U∞ in the x direction at large distances from the body. We assume that
the cross-section Ω of the cylinder is symmetric with respect to the oncom-
ing stream. Then, in terms of polar coordinates centered inside the body,
it follows from the Navier-Stokes equations that the dimensionless stream
function ψ satisfies
2 ψ + ε Jρ [ψ, ψ] = 0 , for ρ > ρb (θ) , (65a)

ψ = ∂n ψ = 0 , on ρ = ρb (θ) , (65b)
2 2 1/2
ψ ∼y, as ρ = (x + y ) → ∞. (65c)
Here ε ≡ U∞ L/ν 1 is the Reynolds number based on the radius L of

Ω, lengths are in units of L, ∂n denotes the normal derivative, and 2

denote the Laplacian and Biharmonic operators, respectively, and Jρ is the
Jacobian defined by Jρ [a, b] ≡ ρ−1 (∂ρ a ∂θ b − ∂θ a ∂ρ b). The boundary of
the scaled cross-section is denoted by ρ = ρb (θ) for −π ≤ θ ≤ π, and the
symmetry condition ρb (θ) = ρb (−θ) is assumed to hold. In terms of ψ, the
dimensionless negative vorticity ω is ω = ψ, and the x and y components
of the fluid velocity, u and v, are
cos θ sin θ
u = ∂y ψ = sin θ ∂ρ ψ + ∂θ ψ , v = −∂x ψ = − cos θ ∂ρ ψ + ∂θ ψ .
ρ ρ
(66)
We first outline the conventional singular perturbation analysis of (65)
for ε → 0 (cf. Kaplun (1957) and Proudman and Pearson (1957)). We then
formulate the hybrid method for summing the infinite-order logarithmic
expansions that arise from the analysis.
In the Stokes, or inner, region defined by ρ = O(1), the stream function
has an infinite logarithmic expansion of the form
∞

ψs (ρ, θ) = ν j ψj (ρ, θ) + · · · . (67)
j=1
$ %
Here, we define ν = ν(εdf ) ≡ −1/ log εdf e1/2 , where df is a shape-
parameter that is specified below in terms of the far-field behavior of a
Biharmonic problem. For a circular cylinder of radius one then df = 1.
Upon substituting (67) into (65a), we obtain that ψj = aj ψc , where the aj
for j ≥ 1 are undetermined constants and ψc ≡ ψc (ρ, θ) is the solution to
the following canonical inner or Stokes problem:
2 ψc = 0 , for ρ > ρb (θ) ; ψc (ρ, θ) = −ψc (ρ, −θ) , (68a)
ψc = 0 and ∂n ψc = 0 , on ρ = ρb (θ) . (68b)
The asymptotic form of ψc as ρ → ∞ involves linear combinations of
{ρ3 , ρ log ρ, ρ, ρ−1 } sin θ. However, to match ψs with the Oseen expansion
below we require that the coefficient of ρ3 must vanish. Then, to specify
ψc uniquely, we impose that the coefficient of ρ log ρ is unity. Therefore,
we define ψc as the unique solution to (68a) and (68b), with the far-field
asymptotic behavior

ψc ∼ ρ log ρ − ρ log df e1/2 sin θ , as ρ → ∞ . (68c)
The constant df , depending on the specific shape of the body, is determined

uniquely by the solution to (68). This exterior Biharmonic problem (68) is
analogous to the exterior Laplace problem (7) for the pipe problem of §2.
Upon substituting ψj = aj ψc into (67), the Stokes expansion becomes

∞

ψs (ρ, θ) = ν j aj ψc (ρ, θ) + · · · . (69a)
j=1
Then, by using (68c), we obtain the following far-field behavior of (69a):

∞

ψs (ρ, θ) ∼ ν j aj log ρ − log df e1/2 ρ sin θ , as ρ → ∞ . (69b)
j=1
Next, we consider the outer or Oseen region defined for ρ = O(ε −1 ). In

this region, we introduce the new Oseen, or outer, length-scale r by r = ερ
with r = O(1). We then re-write the far-field behavior of the Stokes solution
(69b) in terms of the outer Oseen variable r to obtain
⎛ ⎞
∞
1⎝
ψs ∼ a1 r sin θ + ν j [aj log r + aj+1 ] r sin θ⎠ . (69c)
ε j=1
This expression (69c) yields a singularity structure for the outer Oseen so-
lution as r → 0. This behavior suggests that we introduce the new variable
Ψ by Ψ(r, θ) = εψ(ε−1 r, θ), and that we expand Ψ as
∞

Ψ(r, θ) = r sin θ + νΨ1 (r, θ) + ν j Ψj (r, θ) + · · · , (70)
j=2
in order to satisfy the free-stream condition as r → ∞ in (65c). Upon

substituting (70) into (65a), and matching Ψ as r → 0 to the required
singular behavior (69c), we find that a1 = 1 and that Ψ1 and Ψj , for j ≥ 2,
satisfy the following Oseen problems on 0 < r < ∞:
$ %
L0s Ψ1 ≡ 2 Ψ1 + r−1 sin θ ∂θ − cos θ ∂r Ψ1 = 0 , (71a)
Ψ1 ∼ (log r + a2 ) r sin θ , as r → 0 ; ∂ r Ψ1 → 0 , as r → ∞ , (71b)

j−1
L0s Ψj = − Jr [Ψk , Ψj−k ] , (71c)
k=1
Ψj ∼ (aj log r + aj+1 ) r sin θ , as r → 0 ; ∂r Ψj → 0 , as r → ∞ .
(71d)
Here L0s is referred to as the linearized Oseen operator, and Ψ1 is the

linearized Oseen solution.
The limiting conditions (71b) and (71d) are the required singularity be-
haviors of Ψ1 and Ψj for j ≥ 2, respectively, as r → 0. For (71b) the
strength of the singular part r log r sin θ is set to unity. In terms of the solu-
tion to (71a) with Ψ1 ∼ r log r sin θ as r → 0, we then calculate the constant
a2 of the regular part of the singularity structure from the limiting process
Ψ1 − r log r sin θ ∼ a2 r sin θ as r → 0. Then, with a2 determined in this
way, we solve for Ψ2 from (71c) with singular behavior Ψ2 ∼ a2 r log r sin θ
as r → 0, The constant a3 in the regular part of (71d) is then determined
from the limiting process Ψ2 − a2 r log r sin θ ∼ a3 r sin θ as r → 0.
Hence, the coefficients aj for j = 2, 3, .., which are independent of ε and
of the shape of the body, are determined recursively from (71), in a similar
way as in §2. The first two coefficients are (cf. Kaplun (1957), Proudman
and Pearson (1957))
a2 = γe − log 4 − 1 ≈ −1.8091 , (72a)

∞
−1
a3 − a22 = − r I1 (2r) + 1 − 4K1 (r)I1 (r) K0 (r)K1 (r) dr ≈ −0.8669 .
0
(72b)
Here K1 , K0 , I0 and I1 are the usual modified Bessel functions, and γe is

Euler’s constant. This formula for a2 was obtained in Kaplun (1957) and
Proudman and Pearson (1957), while the expression for a3 was given in
Kaplun (1957). The expression for a2 was obtained in Proudman and Pear-
son (1957) in terms of the explicit solution to (71a) with singular behavior
Ψ1 ∼ r log r sin θ as r → 0 given by
∞
cn (r/2)
Ψ1 (r, θ) = − r sin(nθ) , cn (s) ≡ 2 [K1 (s)In (s) + K0 (s)In (s)] .
n=1
n
(73)
As r → 0, then cn (r/2) = O(rn−1 ) for n > 1, and c1 (r/2) ∼ 1−log(ρ/4)−γe ,
where γe is Euler’s constant. Therefore, we conclude that Ψ1 −r log r sin θ →
r (γe − log 4 − 1) sin θ as r → 0. Hence, from the regular part in (71b), we
obtain that a2 = γe − log 4 − 1. In contrast, the derivation in Kaplun (1957)
of the explicit formula for a3 given in (72b) is considerably more involved.
Explicit analytical formulae for aj when j ≥ 4 are not available.
A formula for the drag coefficient CD is given in Imai (1951) in terms of
an arbitrary closed contour around the body. From this formula, and from
the symmetry of the flow, it follows that

π
2 1 2 2
CD = ρ cos θ ψρ − 2 ψθ − sin θ ψρ ψθ dθ
0 ρ ρ
π π
−1
− 2ρ ω ψθ sin θ dθ − 2ε ρ (ρωρ − ω) sin θ dθ , (74)
0 0
where ψ satisfies (65) and ω = ψ. Here ρ, in terms of the Stokes length-

scale, is the radius of an arbitrary circular contour that encloses the body.
To derive an asymptotic formula for the drag coefficient, we substitute the
far-field form (69b) into (74) and evaluate the resulting expression on a
large circle ρ = ρ0 1. In this way, we obtain for ε → 0 that the drag
coefficient CD for a cylinder of arbitrary cross-section is given in terms of
the coefficients aj by
⎛ ⎞
∞
1
CD ∼ 4πε−1 ν(εdf ) ⎝ aj+1 ν j (εdf ) + · · · ⎠ , ν(εdf ) ≡ −
.
j=0
log εdf e1/2
(75)
Kaplun’s (see Kaplun (1957)) approximation for CD results from substi-
tuting a1 = 1 and (72) for a2 and a3 into (75). The resulting three-term
expansion, in equivalent asymptotic form, is
4π
−1
CD ∼ ν̂(εdf ) 1 − 0.8669 ν̂ 2(εdf ) , ν̂(z) ≡ [log (3.7027/z)] .
ε
(76)
For a circular cylinder, the explicit form (76) provides a rather poor
determination of the experimental drag coefficient unless ε is rather small
(cf. Dyke (1975)). One way to overcome this difficulty would be to com-
pute numerically further coefficients aj , for j ≥ 4, from the the infinite
sequence of PDE’s (71c) with singularity structures (71d). This would still
require truncating the series (75) at some finite j. As an alternative to series
truncation, we now follow Kropinski et al. (1995) and formulate a hybrid
asymptotic-numerical method that has the effect of summing all the terms
on the right-hand side of (75), but which avoids computing the coefficients
aj for j ≥ 1 individually.
To do so, we let A (z) denote a function that is asymptotic to the sum
of the terms written explicitly in the brackets on the right-hand side of (75):
∞

A (z) ∼

ν j−1 (z) aj , z ≡ εdf . (77)
j=1
Then, the Stokes expansion (69a) is asymptotic to

ψs (ρ, θ) = ν(z)A (z)ψc (ρ, θ) + · · · , z = εdf . (78)
Substituting (68c) into (78), and writing the resulting expression in terms of
the Oseen variable r = ερ, we obtain the far-field form in the Stokes region,
ψs ∼ ε−1 A (z) [1 + ν(z) log r] r sin θ . (79)
This expression yields the singularity structure for the outer solution.
In the Oseen, or outer region, we do not expand the solution in powers
of ν as in (70). Instead, we solve the full problem (65a) and (65c) for r > 0
subject to the singularity structure (79), which is to hold as r → 0. There-
fore, in analogy with the approach used in §3.2 (see equations (54)–(58) of
§3.2) to treat nonlinear elliptic problems in perforated domains, we intro-
duce the parameter-dependent auxiliary streamfunction ΨH ≡ ΨH (r, θ; S),
with ΨH (r, θ; S) = −ΨH (r, −θ; S), satisfying
2 ΨH + Jr [ΨH , ΨH ] = 0 , r > 0, (80a)
ΨH ∼ r sin θ , as r → ∞ , (80b)
ΨH ∼ Sr log r sin θ , as r → 0 . (80c)
In Kropinski et al. (1995) (see also Keller and Ward (1996)), this parameter-
dependent problem is solved numerically for a range of S values, and in
terms of this solution we identify the regular part R = R(S) of this singu-
larity structure by the following limiting process
ΨH − Sr log r sin θ = R(S)r sin θ + o(r) , as r → 0 . (81)
A plot of the numerically computed function R = R(S) is shown in Fig. 4.
Given that the hybrid problem (80) is posed on 0 < r < ∞, it is an inter-
esting open problem to investigate whether it is possible to find an exact
solution of (80), in a similar way as was found in (59)–(64) for the nonlinear
elliptic problem of §3.2. In this regard, the class of exact solutions to the full
2-D incompressible Navier-Stokes equations found in Ranger (1995) may be
useful.
Finally, since the required singularity behavior from (79) is that
ΨH = A [1 + ν(z) log r] r sin θ + o(r) , as r → 0 , (82)
we conclude that A (z) and ν(z), with z ≡ εdf , are given parametrically in
terms of the singularity strength S and its regular part R(S) by
1 S
ν(z) = −
= , A (z) = R(S) . (83)
log ze1/2 R(S)
1.0
0.5
R(S) 0.0
−0.5
−1.0
0.1 0.2 0.3 0.4 0.5 0.6 0.7
S
Figure 4. Plot of R = R(S) computed numerically from the hybrid formu-

lation (80) and (81).
The problem (80) is a hybrid asymptotic-numerical formulation of the

full problem (65). More specifically, the cylinder in (65) is replaced by the
singularity structure (82) that was derived by exploiting the far-field form
of the infinite-order logarithmic expansion in the Stokes region. Instead of
having to compute solutions to the infinite sequence of problems (71), the
hybrid method requires the solution of the parameter-dependent problem
(80), with singular behavior (80c) given in terms of the parameter S. Then,
the regular part R = R(S) of this singularity behavior is calculated from
(81). Finally, (83) determines A (z) in terms of z = εdf implicitly.
In terms of A and df , the asymptotic formula for the drag coefficient,
valid to within all logarithmic correction terms, is given by
4π −1
CD = [ν(z)A (z) + · · · ] , ν(z) =
, z = εdf . (84)
ε log z e1/2
Kaplun’s equivalence principle follows from the fact that the curve A (z)
versus z can be used for a cylinder of arbitrary cross-section. To determine
A (εdf ) for a body of a specific shape, we need only compute the single
constant df from the numerical solution to the canonical Stokes problem
(68). This feature provides a significant advantage over a direct numerical
approach on the full problem (65).
For a few simple cross-sectional shapes, the constant df can be deter-
mined analytically from (68). In particular, for a circular cross-section,
where ρb (θ) = 1, then the solution to (68) is

ρ 1
ψc = ρ log ρ − + sin θ , (85)
2 2ρ
so that df = 1. Next we consider an elliptical domain defined by (x/a)2 +
(y/b)2 = 1 where max(a, b) = 1. In the case where a = 1, for which the
major axis is aligned parallel to the oncoming stream, the solution to (68)
can be found by introducing elliptic cylinder coordinates (see Kropinski
et al. (1995)). In this way, we obtain that

a+b b−a
df = exp . (86)
2 2(b + a)
This formula for df also holds for the case when b = 1 (the major axis is
is aligned perpendicular to the oncoming stream). A plot of df for various
ellipses is shown in Fig. 5.
Figure 5. The shape-parameter df of (86) for an ellipse with a semi-major

axis of unity.
In Kropinski et al. (1995) a numerical conformal mapping method was

used to calculate df numerically from (68) for other simple cross-sectional
shapes that can be mapped to the unit disk. In particular, such an analytical
conformal mapping is known for the family of symmetric Karman-Trefftz
(KT) airfoils (cf. Milne-Thomson (1958)). The mapping function, z = z(σ),
for the boundary of these profiles is

(ξ + c)k + (ξ − c)k
z(σ) = β0 kc , ξ ≡ σ −1 + c − 1 , (87a)
(ξ + c)k − (ξ − c)k
Table 2. Numerical values for df corresponding to the KT airfoils (87).

The tail angle (in degrees) is θT , and the thickness ratio is δ. The last
column gives the value of b for an ellipse, with a = 1, that has the same
value of df as the corresponding airfoil.
δ θT k c df b
.050 0◦ 2.000 0.961 0.328 0.040
.080 5◦ 1.972 0.952 0.344 0.066
.100 13◦ 1.928 0.960 0.354 0.082
.120 16◦ 1.910 0.954 0.364 0.098
.120 20◦ 1.889 0.968 0.363 0.096
.200 25◦ 1.861 0.915 0.410 0.170
where σ = eiθ with 0 ≤ θ ≤ 2π. By fixing the length of the airfoil to be 2,

we find that the mapping constant β0 is given in terms of k and c by

1 − (1 − c)k
β0 = . (87b)
kc
A parametric representation for the airfoil profile is obtained by setting
σ = eiθ in (87a). In (87), the parameters k and c, where 1 < k < 2 and
0 < c < 1, determine the thickness ratio δ of the airfoil and the tail angle
θT , given by θT = (2 − k)π. Numerical values for df for some KT airfoils,
as computed in Kropinski et al. (1995), are given in Table 2.
For each of the KT airfoil examples given in Table 2 there is an equivalent
ellipse with a = 1 that has the same value of df . The values of b for
these equivalent ellipses, which are computed using (86), are given in the
last column of Table 2. By Kaplun’s equivalence principle each of these
equivalent ellipses has the same asymptotic drag coefficient, to within all
logarithmic correction terms, as the corresponding KT airfoil. However, it
is clear that the transcendentally small terms in the expansion of the drag
coefficient, which are smaller than any power of ν, will not satisfy the same
equivalence principle. Such terms are not accounted for in our analysis.
We now compare the hybrid results for the drag coefficient of a circular
cylindrical body for which df = 1 and z = ε. In Fig. 6(a) we plot the
hybrid drag coefficient CD versus ε, given by (84) with df = 1. In this
figure we compare our hybrid results with Kaplun’s three-term expansion
(76), with full numerical results computed directly from (65) (see Kropinski
et al. (1995)), and the experimental results of Tritton (1959). We observe
that the hybrid method provides a significantly better determination of CD

over the range 0.50 < ε < 2.0 than does the three-term expansion (76).
20.0
c
cc
15.0 cc
cc
c
c
cc
c c
c c
CD 10.0 c
cc c
c cc
c c
c c
c
c
c c c c
5.0 c
0.0
0.2 0.4 0.6 0.8 1.0 1.2 1.4 1.6 1.8 2.0

(a) CD vs. ε: Circular cylindrical (b) CD vs. ε: Two other shapes

cross-section
Figure 6. The drag coefficient CD versus the Reynolds number ε. Left

figure: for a circular cylinder the hybrid result (84) (solid curve), the full
numerical results (heavy solid curve), the three-term result (76) (dotted
curve), and the experimental results of Tritton (1959) (discrete points), are
compared. Right figure: the hybrid result (84) is compared with the three-
term result (76) for a cylindrical body of either an elliptical or a KT airfoil
cross-section.
Finally, we consider flow around other cylindrical bodies. In Fig. 6(b)

we plot the hybrid drag coefficient for flow around certain cylinders having
either elliptical or KT airfoil cross-sections. In this figure, we compare, for
an ellipse and an airfoil, the hybrid results for CD with Kaplun’s three-term
asymptotic result (76). These results were obtained from (83) and (84), and
by using the data from the plot of R = R(S) in Fig. 4. The value of df ,
needed in (83), is given in (86) for the ellipse, and in Table 2 for the KT
airfoil.
We now make several remarks concerning some extensions of the analy-
sis. We first remark that our hybrid method does not incorporate the effect
of the transcendentally small inertial terms arising from the Stokes region.
Therefore, the asymmetry in the flow field near the body, which becomes
more prominent as the Reynolds number is increased, is not captured by our
analysis. For a circular cylinder, the leading-order effects of these inertial

terms on the flow field and on the drag coefficient were analyzed in Skinner
(1975). In §3 of Keller and Ward (1996) an extension of the hybrid approach
of Kropinski et al. (1995) was used to calculate these transcendentally small
terms for a circular cylinder, and to predict the asymmetry in the flow field
near the body.
Secondly, we remark that a similar hybrid method was developed in Tit-
combe et al. (2000) to calculate the drag and lift coefficient for slow viscous
flow over a cylindrical body of arbitrary cross-sectional shape. For this
problem, the hybrid method solution involves a 2 × 2 matrix, depending on
the shape of the body, which replaces the single shape-dependent parameter
df for symmetric cylindrical bodies. In Titcombe et al. (2000) the hybrid
results were compared with those of Shintani et al. (1983).
Finally, we remark that our hybrid method can also be adapted to treat
some new microfluid flow problems. In Matthews and Hill (2009) (see also
Matthews and Hill (2006)) the drag coefficient for steady slow viscous flow
over an infinite nanocylinder was analyzed by asymptotically calculating two
terms in the infinite logarithmic series for the flow field and drag coefficient.
The novel feature of this microfluid flow problem is that, due to the small
scales involved, the usual no-slip boundary condition on the cylinder is
replaced by the Navier boundary condition, which takes into account the
effect of boundary surface roughness. An analysis of this related problem
by a hybrid asymptotic-numerical method will, essentially, only require the
modification of the boundary condition in (68b).
4.1 Summing Logarithmic Expansions: A Linear Biharmonic

Problem
In this subsection we consider a linear Biharmonic problem on a concen-

tric annular domain with a small inner radius ε, formulated as
2 u = 0 , ε < r < 1 , (88a)

u = sin θ , ur = 0 , on r = 1 , (88b)
u = ur = 0 , on r = ε . (88c)
We will calculate the exact solution to this problem, and then show how a
hybrid method similar to that used for the low Reynolds number flow prob-
lem can be readily formulated and implemented to calculate an approximate
solution to (88) that contains all logarithmic correction terms.
The exact solution of (88a), which satisfies (88b), is

1 B 1 B 1
u = Ar3 + Br log r + −2A + − r+ +A+ sin θ ,
2 2 2 2 r
(89)
for any constants A and B. Then, imposing that u = ur = 0 on r = ε, we
get two equations for A and B:

1 B 1 B
Aε3 + Bε log ε + −2A + − ε+ +A+ ε−1 = 0 , (90a)
2 2 2 2

1 B 1 B
3Aε2 + B + B log ε + −2A + − − +A+ ε−2 = 0 . (90b)
2 2 2 2
By comparing the O(ε−1 ) and O(ε−2 ) terms in (90), it follows that
1 B
+A+ = κε2 , (91)
2 2
where κ is an O(1) constant to be found. Substituting (91) into (90), and
neglecting the higher order Aε3 and 3Aε2 terms in (90), we obtain the
approximate system

1 B 1 B
B log ε + −2A + − ≈ −κ , B + B log ε + −2A + − ≈ κ.
2 2 2 2
(92)
By adding the two equations above to eliminate κ, we obtain that
B + 2B log ε + (−4A + 1 − B) = 0 . (93)
From (93), together with A ∼ −(1 + B)/2 from (91), we obtain that
3ν 3 −1
B∼ , A=1− , where ν ≡
. (94)
2−ν 2−ν log εe1/2
Finally, substituting (94) into (89), we obtain that the outer solution has
the asymptotics

u ∼ (1 − Ã)r3 + ν Ãr log r + Ãr sin θ , r O(ε) . (95a)
where Ã is defined by
3 −1
Ã ≡ , ν≡
. (95b)
2−ν log εe1/2
We remark that (95) is an infinite-order logarithmic series approximation

to the exact solution. However, it does not contain transcendentally small
terms of algebraic order in ε as ε → 0.
Next, we show how to derive (95) by employing the hybrid formulation
used in the low Reynolds number flow problem of §4. In order to sum the
infinite logarithmic series we formulate a hybrid method by following (77)–
(79). In the inner region, with inner variable ρ ≡ ε−1 r, we look for an inner
solution in the form (see (78) and (85))

ρ 1
v(ρ, θ) = u(ερ, θ) ∼ εν Ã(ν) ρ log ρ − + sin θ . (96)
2 2ρ

Here ν ≡ −1/ log εe1/2 and Ã ≡ Ã(ν) is a function of ν to be found. The
extra factor of ε in (96) is needed since the solution in the outer region is
not algebraically large as ε → 0. Now letting ρ → ∞, and writing (96) in
terms of the outer variable r = ερ, we obtain that the far-field form of (96)
is
v ∼ Ãνr log r + Ãr sin θ . (97)
Therefore, the approximate outer hybrid solution w H to (88) that sums all
the logarithmic terms must satisfy
2 wH = 0 , 0 < r < 1, (98a)
wH = sin θ , wHr = 0 , on r = 1 , (98b)

wH ∼ Ãνr log r + Ãr sin θ , as r → 0 . (98c)
The solution to (98a) and (98b) is given explicitly by

3 1 β 1 β 1
wH = αr + βr log r + −2α + − r+ +α+ sin θ .
2 2 2 2 r
(99)
The condition (98c) then yields the three equations
1 β 1 β
β = Ãν , −2α + − = Ã , +α + = 0, (100)
2 2 2 2
for α, β, and Ã. We solve this system to obtain
3
β = Ãν , Ã = , α = 1 − Ã . (101)
2−ν
Upon substituting (101) into (99), we obtain that the resulting expression
agrees exactly with the result (95) obtain from the exact solution.
4.2 A Convection-Diffusion Problem

Convection-diffusion problems in two dimensional regions with obstacles
in the low Peclet number limit can be analyzed in a similar way. A recent
analytical study of such problems in both the low and high Peclet number
limit using a different and highly innovative approach is given in Choi et al.
(2005). The following analysis is related to the work in Titcombe and Ward
(1997).
Consider the steady-state convection-diffusion equation for T (X), with
X = (X1 , X2 ) posed outside two circular disks Ωj for j = 1, 2 of a common
radius a, and with a center-to-center separation 2L between the two disks:
κT = U · ∇T , X ∈ R2 \ ∪j=1 Ωj , (102a)
T = Tj , X ∈ ∂Ωj , j = 1, 2 , (102b)
T = T∞ , |X| → ∞ . (102c)
Here κ > 0 is constant, Tj for j = 1, 2 and T∞ are constants, and U = U(X)

is a given bounded flow field with U(X) → (U∞ , 0) as |X| → ∞, where U∞
is constant. We introduce the dimensionless variables x, u(x), and w(x) by
x = X/γ , T = T∞ w , u(x) = U(γx)/U∞ , γ ≡ κ/U∞ . (103)
We also define the dimensionless centers of the two circular disks by xj for
j = 1, 2, and their constant boundary temperatures αj for j = 1, 2, by
xj = Xj /γ , αj = wj /T∞ , j = 1, 2 . (104)
Then, (102) transforms in dimensionless form to
w = u · ∇w , x ∈ R2 \ ∪2j=1 Dεj , (105a)

w = αj , x ∈ ∂Dεj , j = 1, 2 , (105b)
w ∼ 1, |x| → ∞ . (105c)
Here Dεj = {x | |x − xj | ≤ ε} is the circular disk of radius ε centered at xj .

The center-to-center separation is
|x2 − x1 | = 2lε , l ≡ L/a . (106)
The dimensionless flow has limiting behavior u ∼ (1, 0) as |x| → ∞. There

are two interesting limiting cases of (105), which can be analyzed.
Case 1: We assume that l = O(1) as ε → 0, so that |x2 − x1 | = O(ε). This
is the case where the bodies are close together. It leads below to a different
inner problem, not considered in §2.
We assume without loss of generality that x1 +x2 = 0. We then introduce

the inner variables y and v(y) by
y = ε−1 x , v(y) = w(εy) . (107)
Then, we obtain that (105a) and (105b) transform to
y v = εu0 · ∇y v , y ∈ R2 \ ∪2j=1 Dj , (108a)

v = αj , y ∈ ∂Dj , j = 1, 2 . (108b)
Here Dj = {y | |y − yj | ≤ 1} is the circular disk centered at yj = xj /ε of

radius one, and u0 ≡ u(0). The inter-disk separation is
|y2 − y1 | = 2l . (109)
We then look for a solution to (108) in the form
v = v0 + νAvc , (110)
where ν = O(−1/ log ε) and A = A(ν) is to be found. Here v0 solves
y v0 = 0 , y ∈ R2 \ ∪2j=1 Dj , (111a)
v0 = αj , y ∈ ∂Dj , j = 1, 2 , (111b)
v0 bounded as |y| → ∞ . (111c)
Moreover, vc (y) is the solution to
y vc = 0 , y ∈ R2 \ ∪2j=1 Dj , (112a)
vc = 0 , y ∈ ∂Dj , j = 1, 2 , (112b)
vc ∼ log |y| , as |y| → ∞ . (112c)
Since Dj for j = 1, 2 are non-overlapping circular disks, the problem

(111) can be solved explicitly using conformal mapping and the introduction
of symmetric points. In this way, we can derive that
v0 ∼ v0∞ + o(1) , as |y| → ∞ . (113)
The simple calculation of v0∞ is omitted. When α1 = α2 = αc , then

clearly v0∞ = α1 . Next, we can solve (112) exactly by introducing bipolar
coordinates as in Appendix B of Coombs et al. (2009). In this way, we
calculate that
vc (y) ∼ log |y| − log d + o(1) , |y| → ∞ , (114)

where d is given by
∞
ξc e−mξc
log d = log (2β) − + . (115)
2 m=1
m cosh(mξc )
Here β and ξc are determined in terms of l by

β = l2 − 1 ; ξc = log l + l2 − 1 . (116)
We remark that in this analysis we have neglected the transcendentally

small O(ε) term in (108), representing a weak drift in the inner region.
Upon substituting (113) and (114) into (110), and writing y = ε−1 x, we
obtain in terms of outer variables that the far-field behavior of v is
−1
v ∼ v0∞ + A + νA log |x| , ν≡ . (117)
log (εd)
The behavior (117) is the singularity behavior for the infinite-logarithmic
series approximation V0 (x; μ) to the outer solution as x → 0. This approx-
imation satisfies
V0 = u · ∇V0 , x ∈ R2 \{0} ; V0 ∼ 1 , |x| → ∞ , (118)
with singularity behavior (117) as x → 0. To solve (118), we introduce the

Green’s function G(x; ξ) satisfying
G = u · ∇G − δ(x − ξ) , x ∈ R2 , (119a)
1
G(x; ξ) ∼ − log |x − ξ| + R(ξ; ξ) + o(1) , x → ξ , (119b)
2π
with G(x; ξ) → 0 as |x| → ∞. Here R(ξ; ξ) is the regular part of this
Green’s function at x = ξ.
The solution to (118) with singular behavior V0 ∼ νA log |x| as x → 0 is
V0 = 1 − 2πνAG(x; 0) . (120)
By expanding (120) as x → 0, and equating the regular part of the result-

ing expression with that in (117), we get 1 − 2πνAR00 = A + v0∞ . This
determines A = A(ν) by
1 − v0∞ −1
A= , ν≡ , (121)
1 + 2πνR00 log(εd)
where R00 ≡ R(0; 0). The outer and inner solutions are then given in
terms of A. Finally, one can calculate the Nusselt number, representing
the average heat flux across the bodies, by using the divergence theorem
together with the form (117) of the far-field behavior in the inner region.
We leave this simple calculation to the reader.
Case 2: We assume that l = O(ε−1 ) as ε → 0, and define l = l0 /ε with

l0 = O(1), so that |x2 − x1 | = 2l0 . This is the case where the small disks of
radius ε are separated by O(1) distances in (105). In the analysis there are
two distinct inner regions; one near x1 and the other at an O(1) distance
away centered at x2 . Since each separated disk is a circle of radius ε, it has
a logarithmic capacitance d = 1. Therefore, the infinite-logarithmic series
approximation V0 (x; μ) to the outer solution satisfies
V0 = u · ∇V0 , x ∈ R2 \{0} ; V0 ∼ 1 , |x| → ∞ , (122a)

−1
V0 ∼ αj + Aj + νAj log |x − xj | , ν≡ . (122b)
log ε
The solution to (122) is given explicitly by
2

V0 = 1 − 2πν Ai G(x; xi ) . (123)
i=1
We then let x → xj for j = 1, 2 in (123) and equate the nonsingular part

of the resulting expression with the regular part of the singularity structure
in (122b). This yields that A1 and A2 satisfy the linear algebraic system
A1 (1 + 2πνR11 ) + 2πνA2 G12 = 1 − α1 , (124)

A2 (1 + 2πνR22 ) + 2πνA1 G21 = 1 − α2 . (125)
Here Gij = G(xj ; xi ) and Rjj = R(xj ; xj ) are the Green’s function and its
regular part as defined by (119).
Finally, we remark that for the case of a uniform flow where u = (1, 0),
then the explicit solution to (119) is

1 x1 − ξ1
G(x; ξ) = exp K0 (|x − ξ|) , (126a)
2π 2
where x = (x1 , x2 ) and ξ = (ξ1 , ξ2 ). By letting x → ξ, and using K0 (r) ∼
− log r + log 2 − γe , as r → 0+ , where γe is Euler’s constant, we readily
calculate that
1
R(ξ, ξ) = (log 2 − γe ) . (126b)
2π
These results for G and its regular part can be used in the results of either
(121) or (125) for Case I or Case II, respectively.
5 The Fundamental Neumann Eigenvalue in a Planar

Domain with Localized Traps
In this section we follow Kolokolnikov et al. (2005) and consider an op-
timization problem for the fundamental eigenvalue of the Laplacian in a
planar bounded two-dimensional domain with a reflecting boundary that is
perturbed by the presence of K small holes in the interior of the domain.
The perturbed eigenvalue problem is

Δu + λu = 0 , x ∈ Ω\Ωp ; u2 dx = 1 , (127a)
Ω\Ωp
∂n u = 0 , x ∈ ∂Ω ; u = 0, x ∈ ∂Ωp ≡ ∪K
i=1 ∂Ωεi . (127b)
Here Ω is the unperturbed domain, Ωp = ∪K i=1 Ωεi is a collection of K small

interior holes Ωεi , for i = 1, . . . , K, each of ‘radius’ O(ε), and ∂n u is the
outward normal derivative of u on ∂Ω. We assume that the small holes in
Ω are non-overlapping and that Ωεi → xi as ε → 0, for i = 1, . . . , K. A
schematic plot of the domain is shown in Fig. 7.
wandering particle
O(ε )
x2
N small absorbing holes
x1
n
reflecting
walls
Figure 7. A schematic plot of the perturbed domain for the eigenvalue

problem (127).
We let λ0 (ε) denote the first eigenvalue of (127), with corresponding

eigenfunction u(x, ε). Clearly, λ0 (ε) → 0 as ε → 0. Our objective is to
determine the locations, xi for i = 1, . . . , K, of the K holes of a given
shape that maximize this fundamental eigenvalue. Asymptotic expansions
for the fundamental eigenvalue of related eigenvalue problems in perfo-
rated multi-dimensional domains, with various boundary conditions on the
holes and outer boundary, are given in Ozawa (1981), Ward et al. (1993),
Ward and Keller (1993), Davis and Llewellyn-Smith (2007), and Lange and
Weinitschke (1994) (see also the references therein).
As an application of (127), consider the Brownian motion of a particle
in a two-dimensional domain Ω, with reflecting walls, that contains K small
traps Ωεi , for i = 1, . . . , K, each of ‘radius’ ε, for i = 1, . . . , K. The traps
are centered at xi , for i = 1, . . . , K. If the Brownian particle starts from
the point y ∈ Ω\Ωp at time t = 0, then the probability density v(x, y, t, ε)
that the particle is at point x at time t satisfies
vt = Δv , x ∈ Ω\Ωp ; ∂n v = 0 , x ∈ ∂Ω ; v = 0, x ∈ ∂Ωp , (128)
with v = δ(x − y) at time t = 0. By calculating the solution to (128) in

terms of an eigenfunction expansion, and by assuming that y is uniformly
distributed over Ω\Ωp , it is easy to show that the probability P0 (t, ε) that
the Brownian particle is in Ω\Ωp at time t is given by
P0 (t, ε) = e−λ0 (ε)t [1 + O(ν)] . (129)
Therefore, the expected lifetime of the Brownian particle is proportional to

1/λ0 (ε). In this context, our optimization problem is equivalent to choosing
the locations of K small traps to minimize this expected lifetime.
We first consider (127) for the case of one hole. In Ward et al. (1993)
(see also Ward and Keller (1993)) it was shown that as ε → 0 the first
eigenvalue λ0 of (127) has the asymptotic expansion:
λ0 (ε) = λ00 + ν(ε)λ01 + ν 2 (ε)λ02 + · · · .
Here, ν(ε) = −1/ log(εd) where d is the logarithmic capacitance of the hole.
For the unperturbed problem with ε = 0, we have λ00 = 0. In the O(ν)
term, λ01 is independent of the location of the hole at x = x0 (cf. Ward
et al. (1993)). Therefore, we need the higher-order coefficient λ02 in order
to determine the location of the hole that maximizes λ0 .
For the case of one hole, an infinite logarithmic expansion for λ0 (ε) has
the form
ε 1
λ0 (ε) = λ∗ (ν) + O , ν≡− .
log ε log(εd)
To calculate λ∗ (ν) we use the hybrid method of Kolokolnikov et al. (2005).
Near the hole, we identify an inner (local) region in terms of a local spatial
variable y = ε−1 (x−x0 ), and where the hole is rescaled so that Ω0 ≡ ε−1 Ωε .
Denoting the inner (local) solution by v(y, ε) = u(x0 + εy, ε), we then
expand v(y, ε) as
v(y, ε) = A ν vc (y) + · · · . (130)
Here, A = A(ν) ∼ O(1) as ε → 0, and vc (y) is the solution of the canonical

inner problem (7), re-written here as
Δy vc = 0 , y ∈ Ω0 ;
vc = 0 , y ∈ ∂Ω0 , (131a)
p·y
vc ∼ log |y| − log d + , as |y| → ∞ . (131b)
|y|2
In (131b), the logarithmic capacitance d and the dipole vector p = (p1 , p2 )

are determined from the shape of the hole.
We expand the eigenvalue λ0 and the outer (global) solution as
λ0 (ε) = λ∗ (ν) + μλ1 + · · · , u(x, ε) = u∗ (x, ν) + μu1 (x, ν) + · · · , (132)
where μ O(ν k ) for any k > 0. Substituting (132) into (127a) and the
boundary condition (127b) on ∂Ω, we obtain the full problem in a domain
punctured by the point x0 ,

Δu∗ + λ∗ u∗ = 0 , x ∈ Ω\{x0 } ; (u∗ )2 dx = 1 ; ∂n u∗ = 0 , x ∈ ∂Ω .
Ω
(133)
The singularity condition for (133) as x → x0 given below arises from
matching u∗ to the inner solution. Substituting (131b) into (130), and
expressing the result in global variables, we obtain
p · (x − x0 )
v(y, ε) ∼ A ν log |x−x0 |+A+εA ν +· · · , as y → ∞ . (134)
|x − x0 |2
Here, we have used ν ≡ −1/ log(εd). To match u ∗ to (134), we require that

u∗ has the singularity behavior
u∗ (x, ν) ∼ A ν log |x − x0 | + A , as x → x0 . (135)
Comparing the terms in (134) and (132) at the next order, we see that
μ = O(εν).
Next, we must determine u∗ (x, ν) and λ∗ (ν) satisfying (133) and (135).
To do so, we introduce the Helmholtz Green’s function, Gh (x; x0 , λ∗ ), and
its regular part, Rh (x0 ; x0 , λ∗ ), satisfying
ΔGh + λ∗ Gh = −δ(x − x0 ) , x ∈ Ω ; ∂n Gh = 0 , x ∈ ∂Ω , (136a)

1
Gh (x; x0 , λ∗ ) ∼ − log |x − x0 | + Rh (x0 ; x0 , λ∗ ) + o(1) , as x → x0 .
2π
(136b)
In terms of this Green’s function, u∗ (x, ν) is given by

u∗ (x, ν) = −2π A ν Gh (x; x0 , λ∗ ).
By using (136b), we expand u∗ as x → x0 to obtain
u∗ (x, ν) ∼ A ν log |x − x0 | − 2π A ν Rh (x0 ; x0 , λ∗ ) , as x → x0 . (137)
The matching condition is that the expressions in (135) and (137) agree.
The log |x − x0 | terms agree automatically, and from the remaining terms,
we obtain a transcendental equation for λ∗ (ν) given by
1
Rh (x0 ; x0 , λ∗ ) = − . (138)
2πν
To obtain the asymptotic behavior for λ0 , we need the solution λ∗ of (138)
that tends to zero as ν → 0.
Equation (138) can, in general, only be solved numerically as a function
of ν. Below, we only determine an expression for λ∗ that is correct to terms
of order O(ν 2 ). To obtain this expression, we expand the Helmholtz Green’s
function, Gh (x; x0 , λ∗ ), in terms of λ∗ 1, as
1
G(x; x0 , λ∗ ) = G0 (x; x0 ) + G1 (x; x0 ) + λ∗ G2 (x; x0 ) + · · · . (139)
λ∗
Substituting (139) into (136), we get a series of problems for the Gj (x; x0 ),
j = 0, 1, 2, . . .. At order O(1/λ∗ ), G0 satisfies ΔG0 = 0 in Ω and ∂n G0 = 0
on ∂Ω, from which we obtain that G0 is a constant. The higher-order
corrections Gj for j ≥ 1 are readily found to satisfy
&
−δ(x − x0 ) − G0 , j = 1,
ΔGj = x ∈ Ω ; ∂n Gj = 0, x ∈ ∂Ω ,
−Gj−1 , j > 1,
(140)
with Ω Gj dx = 0 for j ≥ 1. Applying the Divergence Theorem, we obtain
that G0 = −1/|Ω|, where |Ω| is the area of Ω. The function G1 (x; x0 )
(which we shall henceforth call GN ) is the Neumann Green’s function, with
regular part RN (x0 ; x0 ) defined by (49).
From (139) and (49b), we write the two-term expansion for G when
λ∗ 1 as
1
Gh (x; x0 , λ∗ ) = − + GN (x; x0 ) + O(λ∗ ) . (141)
|Ω|λ∗
A similar two-term expansion for the regular part Rh of the Helmholtz
Green’s function in terms of the regular part of the Neumann Green’s func-
tion is
1
Rh (x0 ; x0 , λ∗ ) = − + RN (x0 ; x0 ) + O(λ∗ ) . (142)
|Ω|λ∗
Substituting this expression into (138), we get the following two-term asymp-
totic result:
Principal Result 3:(One Hole) For ε → 0, the first eigenvalue λ0 of (127)
has the two-term asymptotic behavior
2πν 2πν 4π2 ν 2
λ0 (ε) = +O(ν 3 ) = − RN (x0 ; x0 )+O(ν 3 ) .
|Ω| (1 + 2πνRN (x0 ; x0 )) |Ω| |Ω|
(143)
Here ν = −1/ log(εd), and d is the logarithmic capacitance determined from
the inner problem (131). An infinite-order logarithmic expansion for λ0 is
given by λ0 ∼ λ∗ , where λ∗ is the first positive root of (138).
Next, we extend the asymptotic framework to the case of K holes. Much
of the analysis above remains the same, except that now the single hole x 0
is replaced by xi , for i = 1, . . . , K. The hybrid formulation for K holes is
Δu∗ + λ∗ u∗ = 0 , x ∈ Ω\{x1 , . . . , xK } ; ∂n u∗ = 0 , x ∈ ∂Ω , (144a)

u∗ ∼ Ai νi log |x − xi | + Ai , as x → xi , i = 1, . . . , K ,
(144b)

with normalization condition Ω (u∗ )2 dx = 1. Here, νi = −1/ log(εdi ),
where d is the logarithmic capacitance of the ith hole. In this formulation,
i
we have the K unknowns, Ai , for i = 1, . . . , K, and one normalization
condition for u∗ . The normalization condition effectively sets one relation
between the Ai , for i = 1, . . . , K.
We write u∗ in terms of the Helmholtz Green’s function defined in (136),
and then take the limit x → xi to get

K
u∗ = −2π Aj νj Gh (x; xj , λ∗ ) ∼ Ai νi (log |x − xi | − 2πνi Rh (xi ; xi , λ∗ ))
j=1

K
− 2π Aj νj Gh (xi ; xj , λ∗ ). (145)
j=1
j=i
The matching condition is that the expressions in (144b) and (145) agree.
The logarithmic terms agree, and from the remaining terms, we obtain a
K × K homogeneous linear system to solve for the Ai , given by

K
Ai (1 + 2πνi Rh (xi ; xi , λ∗ )) + 2π Aj νj Gh (xi ; xj , λ∗ ) = 0 , i = 1, . . . , K .
j=1
j=i
(146)
A solution to (146) exists only when the determinant associated with the
linear system (146) vanishes. This condition provides an expression for
λ∗ (ν1 , . . . , νK ) that sums all the logarithmic terms in the asymptotic ex-
pansion of λ0 (ε).
As with the case for one hole in the domain, we can derive an asymptotic
formula for λ∗ that has an error of O(ν 3 ). This formula is again determined
in terms of the Neumann Green’s function GN and its regular part RN ,
defined in (49). By using (141) and (142) in (146), we obtain a homogeneous
linear system for the Ai for i = 1, . . . , K, given by
K
2πνi 1
Ai 1 + 2πνi RN (xj ; xj ) − +2π A ν
j j − + G N (x j ; xi ) = 0.
|Ω|λ∗ j=1
|Ω|λ∗
j=i
(147)
It is convenient to write (147) in matrix form as
2π
Ca = BVa , C ≡ I + 2πGN V , (148a)
|Ω|λ∗
where
⎛ ⎞ ⎛ ⎞
ν1 0 ··· 0 1 1 ··· 1 ⎛ ⎞
⎜ .. ⎟ ⎜ .. ⎟ A1
⎜ ··· 0 ⎟ ⎜ ··· 1 ⎟
V ≡⎜
0 . ⎟, B≡⎜ 1 . ⎟, a≡⎜ . ⎟
⎝ .. ⎠ .
⎜ .. .. .. .. ⎟ ⎜ .. .. .. . ⎟
⎝ . . . . ⎠ ⎝ . . . .. ⎠
AK
0 0 · · · νK 1 1 ··· 1
(148b)
In (148a), the Neumann Green’s function matrix GN is the K ×K symmetric
matrix with entries
(GN )ij ≡ GN (xi ; xj ) , i = j ; (GN )jj = RN (xj ; xj ) . (148c)
Let νm = max νj . Then, for νm sufficiently small, we can invert C ap-
j=1,...,K
proximately, to obtain that λ∗ is an eigenvalue of the matrix eigenvalue
problem
2π −1
Aa = λ∗ a , A= C BV . (149)
|Ω|
By using this representation of λ∗ we obtain the following result:
Principal Result 4:(K Holes) For ε → 0, the first eigenvalue λ0 of (127) has
the explicit two-term asymptotic behavior
⎛ ⎞
2π K K K
λ0 (ε) ∼ λ∗ , λ∗ = ⎝ νj − 2π νj νi (G)N ij ⎠ + O(νm
3
).
|Ω| j=1 j=1 i=1
(150)
Here (G)N ij are the entries of the Neumann Green’s function matrix GN
defined in (148c).
Proof. We first notice that the matrix BV has rank one, since V is diagonal
and B = e0 et0 , where et0 = (1, 1, . . . , 1). This implies that A has rank one,
and so λ∗ is the unique nonzero eigenvalue of A. Hence, λ∗ = TraceA. By
using the structure of A in (149), we readily calculate that

2π $ %
K K
∗
λ = νj cij , cij ≡ C −1 ij . (151)
|Ω| j=1 i=1
Finally, we use the asymptotic inverse C −1 ∼ I − 2πGN V + · · · for νm 1

to calculate cij . Substituting this result into (151) we obtain (150).
As a Corollary to this result, we obtain the following simplification for

the case of K identical holes:
Corollary 5:(K Identical Holes) Suppose that the K holes are identical, in
the sense that εdj is independent of j. Then, (150) can be written as the
explicit two-term expansion
2πKν 4π 2 ν 2
λ0 (ε) ∼ λ∗ , λ∗ = − p(x1 , . . . , xK ) + O(ν 3 ) , (152)
|Ω| |Ω|
where ν ≡ −1/ log(εd), and the function p(x1 , . . . , xK ) is defined by

⎛ ⎞

K
K
K
⎜
K
⎟
p(x1 , . . . , xK ) = (G)N ij = ⎝RN (xi ; xi ) + GN (xj ; xi )⎠ .
j=1 i=1 i=1 j=1
j=i
(153)
Here (G)N ij are the entries in the matrix GN in (148c). For K circular
holes of radius ε, then dj = 1 for j = 1, . . . , K, and so ν = −1/ log ε.
When Ω is the unit disk, the optimal spatial configurations of the cen-
ters {x1 , . . . , xK } of K distinct traps of a common radius ε were com-
puted numerically in Kolokolnikov et al. (2005) by optimizing the function
p(x1 , . . . , xK ) in (153). For the unit disk, the Neumann Green’s function
GN (x; ξ) and its regular part RN (ξ; ξ) are explicitly available (see equation
(52)). By using this Green’s function, it is readily shown that the prob-
lem of minimizing the function p(x1 , . . . , xK ) is equivalent to the discrete
variational problem of minimizing the function F(x1 , . . . , xK ) defined by

K
K
K
K
K
F(x1 , . . . , xK ) = − log |xj −xk |− log |1−xj x̄k |+K |xj |2 ,
j=1 k=1 j=1 k=1 j=1
k=j
(154)
for |xij| = 1 and xj = xk when j = k. Here x̄k denotes the complex
conjugate of xk .
An interesting open problem is to determine the optimal arrangement of
K 1 traps in the dilute fraction limit Kε2 1. In particular, does the
optimal arrangement approach a hexagonal lattice structure with a bound-
ary layer near the rim of the unit disk?
6 Conclusion
In this article we have surveyed the development and application of a hybrid
asymptotic-numerical method for solving linear and nonlinear PDE mod-
els in two-dimensional domains that have small inclusions or obstructions.
Related theoretical approaches have also been developed to treat similar
strongly localized perturbation problems including, an eigenvalue perturba-
tion problem in a three-dimensional domain (cf. Ward and Keller (1993)),
cell-signaling problems in mathematical biology (cf. Bressloff et al. (2008),
Straube and Ward (2009)), the narrow escape problem from a sphere or a
disk that has small absorbing windows on its boundary (cf. Cheviakov et al.
(to appear, 2010), Pillay et al. (to appear, 2010)), and the mean first pas-
sage time in a three-dimensional domain with interior traps (cf. Cheviakov
and Ward (to appear, 2010)).
Acknowledgments
M. J. W. gratefully acknowledges the fundamental contributions of my other
co-authors D. Coombs (UBC), J. B. Keller (Stanford), T. Kolokolnikov
(Dalhousie), A. Peirce (UBC), S. Pillay (J.P. Morgan), R. Straube (Max
Planck Institute, Magdeburg), and M. Titcombe (Champlain College) to
some of the work surveyed herein. M. J. W. and M. C. K acknowledge the
grant support of NSERC (Canada).
Bibliography
C. Bandle and M. Flucher. Harmonic radius and concentration of energy;
hyperbolic radius and liouville’s equation. SIAM Review, 38(2):191–238,
1996.
P. C. Bressloff, D. Earnshaw, and M. J. Ward. Diffusion of protein recep-
tors on a cylindrical dendritic membrane with partially absorbing traps.
SIAM J. Appl. Math., 68(5):1223–1246, 2008.
A. Cheviakov and M. J. Ward. Optimizing the fundamental eigenvalue of
the laplacian in a sphere with interior traps. Mathematical and Computer
Modeling, to appear, 2010.
A. Cheviakov, M. J. Ward, and R. Straube. An asymptotic analysis of the
mean first passage time for narrow escape problems: Part ii: The sphere.
SIAM J. Multiscale Modeling, to appear, 2010.
J. Choi, D. Margetis, T. M. Squires, and M. Z. Bazant. Steady advection-
diffusion around finite absorbers in two-dimensional potential flows. J.
Fluid Mech., 536:155–184, 2005.
D. Coombs, R. Straube, and M. J. Ward. Diffusion on a sphere with local-
ized traps: Mean first passage time, eigenvalue asymptotics, and fekete
points. SIAM J. Appl. Math., 70(1):302–332, 2009.
A. M. Davis and S. G. Llewellyn-Smith. Perturbation of eigenvalues due
to gaps in two-dimensional boundaries. Proc. Roy. Soc. A, 463:759–786,
2007.
W. W. Dijkstra and M. E. Hochstenbach. Numerical approximation of
the logarithmic capacity. CASA Report 08-09, Technical U. Eindhoven,
2008.
M. Van Dyke. Perturbations Methods in Fluid Mechanics. Parabolic Press,
1975.
J. E. Fletcher. Mathematical modeling of the microcirculation. Math Bio-
sciences, 38:159–202, 1978.
I. Imai. On the asymptotic behavior of viscous fluid flow at a great distance
from a cylindrical body, with special reference to filon’s paradox. Proc.
Roy. Soc. A, 208:487–516, 1951.
S. Kaplun. Low reynolds number flow past a circular cylinder. J. Math.
Mech., 6(5):52–60, 1957.
J. B. Keller and M. J. Ward. Asymptotics beyond all orders for a low
reynolds number flow. J. Engrg. Math., 30(1-2):253–265, 1996.
J. Kevorkian and J. Cole. Multiple Scale and Singular Perturbation Methods.
Applied Mathematical Sciences, Vol. 114, Springer-Verlag, 1993.
T. Kolokolnikov, M. Titcombe, and M. J. Ward. Optimizing the fundamen-
tal neumann eigenvalue for the laplacian in a domain with small traps.
Europ. J. Appl. Math., 16(2):161–200, 2005.
A. Krogh. The number and distribution of capillaries in muscles with cal-
culations of the oxygen pressure head necessary for supplying the tissue.
J. Physiology (London), 52:409–415, 1919.
M. C. Kropinski, M. J. Ward, and J. B. Keller. A hybrid asymptotic-

numerical method for calculating low reynolds number flows past sym-
metric cylindrical bodies. SIAM J. Appl. Math., 55(6):1484–1510, 1995.
P. A. Lagerstrom. Matched Asymptotic Expansions. Applied Mathematical
Sciences, Vol. 76, Springer-Verlag, 1988.
C. Lange and H. Weinitschke. Singular perturbations of elliptic problems
on domains with small holes. Stud. Appl. Math., 92(1):55–93, 1994.
S. H. Lee and L. G. Leal. Low reynolds number flow past cylindrical bodies
of arbitrary cross-sectional shape. J. Fluid Mech., 164:401–427, 1986.
MATLAB. Partial Differential Equations Toolbox, User’s Guide. The Math-
Works, Inc., 1996.
M. Matthews and J. Hill. Flow around nanospheres and nanocylinders. Q.
J. Mech. Appl. Math., 59(2):191–210, 2006.
M. Matthews and J. Hill. Asymptotic analysis of the viscous micro/nano
pump at low reynolds number. J. Engrg. Math., 63(2-4):279–292, 2009.
A. Mikelić and M. Primicerio. A diffusion-consumption problem for oxygen
in a living tissue perfused by capillaries. Nonlinear Diff. Eq. Appl., 13(3):
349–367, 2006.
L. M. Milne-Thomson. Theoretical Aerodynamics. Dover Publications Inc.,
1958.
S. Ozawa. Singular variation of domains and eigenvalues of the laplacian.
Duke Math. J., 48(4):767–778, 1981.
S. Pillay, M. J. Ward, A. Peirce, and T. Kolokolnikov. An asymptotic
analysis of the mean first passage time for narrow escape problems: Part
i: Two-dimensional domains. SIAM J. Multiscale Modeling, to appear,
2010.
A. S. Popel. Theory of oxygen transport to tissue. Critical Reviews in
Biomedical Engrg., 17:257–321, 1989.
I. Proudman and J. Pearson. Expansions at small reynolds number for the
flow past a sphere and a circular cylinder. J. Fluid Mech., 2:237–262,
1957.
K. B. Ranger. Explicit solutions of the steady two-dimensional navier-stokes
equations. Stud. Appl. Math., 94(2):169–181, 1995.
T. Ransford. Potential Theory in the Complex Plane. London Mathematical
Society Student Texts Vol. 28, Cambridge University Press, 1995.
K. Shintani, A. Umemura, and A. Takano. Low reynolds number flow past
an elliptic cylinder. J. Fluid Mech., 136:277–289, 1983.
L. A. Skinner. Generalized expansions for slow flow past a cylinder. Q. J.
Mech. Appl. Math., 28(3):333–340, 1975.
R. Straube and M. J. Ward. Intraceulluar signalling gradients arising from

multiple compartments: A matched asymptotic expansion approach.
SIAM J. Appl. Math., 70(1):248–269, 2009.
M. Titcombe and M J. Ward. Convective heat transfer past small cylindrical
bodies. Stud. Appl. Math., 99(1):81–105, 1997.
M. Titcombe and M J. Ward. Summing logarithmic expansions for elliptic
equations in multiply-connected domains with small holes. Canadian
Appl. Math. Quart., 7(3):313–343, 1999.
M. Titcombe and M J. Ward. An asymptotic study of oxygen transport
from multiple capillaries to skeletal muscle tissue. SIAM J. Appl. Math.,
60(5):1767–1788, 2000.
M. Titcombe, M J. Ward, and M. C. Kropinski. A hybrid asymptotic-
numerical solution for low reynolds number flow past an asymmetric
cylindrical body. Stud. Appl. Math., 105(2):165–190, 2000.
D. J. Tritton. Experiments on the flow past a circular cylinder at low
reynolds numbers. J. Fluid Mech., 6:547–567, 1959.
J. Veysey and N. Goldenfeld. Simple viscous flows: From boundary layers
to the renormalization group. Reviews of Modern Physics, 79:883–927,
2007.
M. J. Ward, W. D. Henshaw, and J. B. Keller. Summing logarithmic ex-
pansions for singularly perturbed eigenvalue problems. SIAM J. Appl.
Math., 53(3):799–828, 1993.
M. J. Ward and J. B. Keller. Strong localized perturbations of eigenvalue
problems. SIAM J. Appl. Math., 53(3):770–798, 1993.
A. J. Ward-Smith. Internal Fluid Flow. Clarendon Press, Oxford, 1980.
Exponential Asymptotics and Generalized
Solitary Waves
*
Roger Grimshaw
*
Loughborough University, UK
Abstract Many problems in fluid mechanics involve asymptotic ex-

pansions in the form of a power series for a suitable small parameter.
Such expansions necessarily fail to find terms which are exponen-
tially small with respect to this parameter. Although small these
missing terms are often of physical importance. This chapter will
describe how to find such exponentially small terms, using as the
main tool matched asymptotic expansions in the complex plane and
Borel summation. The techniques are developed in the context of
model problems related primarily to the theory of weakly nonlocal
solitary waves (also called generalized solitary waves) which arise in
the study of gravity-capillary waves, internal waves and in several
other physical contexts. These waves have a central core of finite
amplitude, but are accompanied by co-propagating oscillatory tails
whose amplitude is exponentially small. Special interest lies in the
possibility that for certain parameter values, the amplitude of the
oscillatory tails may be zero, leading to the important concept of
embedded solitary waves.
1 Introduction
An inverse power series is asymptotic to a function f (x) if, for fixed N and
sufficiently large x > 0,
N
an 1
f (x) − n
= O( N +1 ) . (1)
n=0
x x
This formal definition, due to Poincare, describes the limit when x → ∞

for a fixed integer N . When (1) holds, we write
∞
an
f (x) ∼ . (2)
n=0
xn
© CISM, Udine 2010
72 R. Grimshaw
Importantly the series expansion in (2) need not converge as N → ∞ in

order for it to be asymptotic. Our concern here is that when the series is
asymptotic, but diverges, it fails to detect terms proportional to exp (−Kx),
where K > 0 is a constant. Such terms are smaller than x−N for any integer
N as x → ∞. The detection of such terms requires exponential asymptotics,
or asymptotics beyond all orders, see Bender and Orzsag (1978) or Boyd
(1998) for instance.
1.1 Stieltjes Integral

To begin, let us consider the function defined by the integral
∞
exp (−xt)
I(x) = dt . (3)
0 1+t
This function is well-defined for all x > 0. Now integrate by parts N times,
to get that
N
n!
I(x) = (−1)n n+1 + N (x) , (4)
n=0
x
∞
(N + 1)! exp (−xt)
N (x) = (−1)N +1 dt . (5)
xN +1 0 (1 + t)N +1
Since the integral term in (5) is bounded by 1/x, it is easily shown that
N (x) = O(x−N −2 ) for each fixed N , and so
∞
n!
I(x) ∼ (−1)n . (6)
n=0
xn+1
Note that the series diverges for all x > 0, but nonetheless is a useful
approximation to I(x) as x → ∞.
For each fixed x, as x → ∞, we can minimize |N (x)| with respect to
N . Using Stirling’s formula for large N , N ! ≈ (2π)1/2 exp (−N )N N +1/2 , we
find that the minimum occurs when N ≈ x, and then the error
N (x) = O(exp (−x))
is exponentially small. The plot shows x = 5, 10, 15 (red, blue, black)

and the exponential error (dots). This crucial observation leads to the
well-developed theories for optimal truncation, superasymptotics and hy-
perasymptotics, see Berry and Howls (1990), (Berry, 1992, 11-14), Boyd
(1998) or Boyd (1999) for instance.
Exponential Asymptotics and Generalized Solitary Waves 73
Figure 1. Plot of the error (5) as a function of N ; the dashed line is

proportional to exp (−x)
1.2 Fourier Integral

Suppose that the function F (x) is defined by
∞
F (x) = F̂ (k) exp (ikx) dk , (7)
−∞
where F̂ ∗ (−k) = F̂ (k) for all real k, in order that F (x) is real-valued (the
superscript ∗ denotes the complex conjugate). Suppose that the Fourier
transform F̂ (k) and all its derivatives vanish as k → ±∞. Then we can
estimate F (x) as x → ±∞ by integrating (7) by parts. The outcome is
F (x) = O(x−N ) , as x → ±∞ , (8)
for all integers N . The asymptotic series (2) has zero coefficients, and we
infer that F (x) is exponentially small. For x > 0, the contour of integration
is moved into the upper half of the complex k-plane, that is Im k > 0. Then
if the nearest singularity of F̂ (k) to the real k-axis is at k = a + ib, b > 0 we
can estimate that F (x) = O(exp (−bx)) as x → ∞. An analogous procedure
74 R. Grimshaw
is used if x < 0, when the contour is moved into the lower half of the complex
k-plane.
For example, suppose that F̂ (k) = sech(γk) so that
∞
F (x) = sech(γk) exp (ikx) dk . (9)
−∞
There are poles at k = ±iπ/2γ, ±3iπ/2γ, · · · . Evaluating the residue at the

nearest pole to the real axis yields
2π π|x|
F (x) ∼ exp (− ). (10)
γ 2γ
Indeed, in this case the integral (9) can be explicitly evaluated to give
π πx
F (x) = sech( ) ,
γ 2γ
from which the estimate (10) readily follows.
1.3 Airy Functions

Airy’s differential equation is
d2 w
= zw , (11)
dz 2
expressed here for a complex-valued function w(z) of the complex variable
z = x+iy. Note that on the real positive-axis, z = x > 0, we expect the two
linearly independent solutions to be either exponentiall growing or exponen-
tially decaying, but on the real negative-axis, z = x < 0, both solutions will
be oscillatory. The task is to connect these contrasting behaviours, and we
will show that this can be accomplished using ideas of exponential asymp-
totics in the complex z-plane. On the real axis z = x, the Airy function of
the first kind can be defined by

1 ∞ t3
Ai(x) = cos (xt + ) dt . (12)
π 0 3
This is real-valued on the real axis, and is extended to all complex z by

1 t3
Ai(z) = exp ( − zt) dt , (13)
2πi C 3
where the contour C runs from ∞e−iπ/3 to ∞eiπ/3 .

To obtain the second solution, define σ = e2iπ/3 ,Then if w(z) is any

solution of (11), so are w(σz) and w(σ −1 z). Any two of these three functions
are linearly independent, and further
w(z) + σw(σz) + σ −1 w(σ −1 ) = 0 . (14)
Hence we now define the Airy function of the second kind as
Bi(z) = e−iπ/6 Ai(σz) + eiπ/6 Ai(σ −1 z) . (15)
Note that, like Ai(z), Bi(z) (15) is real-valued on the real axis. Our interest
here is in the asymptotic expansions of Ai(z), Bi(z) in the complex z-plane.
These can be obtained from the integral (13) using the method of steepest
descent. Here we shall combine that with WKB asymptotic expansions
directly from Airy’s differential equation (11) in order to exhibit the Stokes
phenomenon. First, from the integral (13) we find that (using the saddle
point at t = z 1/2 )
1 2z 3/2
Ai(z) ∼ exp (− ), as |z| → ∞ , |arg(z)| < π . (16)
2π 1/2 z 1/4 3
1.4 WKB Solutions and Stokes Phenomenon

Although this asymptotic expression (16), combined with the relation
(14) can be used to get the desired connection formulas, it is instructive
to take a different approach based on finding WKB asymptotic solutions of
the differential equation (11) directly. Hence, seek a solution of the form
w(z) = exp (φ(z)) . (17)
Substitution into (11) (wzz = zw) yields
φzz + φ2z = z , (18)
which is a Riccati equation for φz . The WKB procedure for the limit |z| →
∞ yields the solutions
1
φz = ∓z 1/2 − + O(z −5/2 ) . (19)
4z
The outcome are the WKB asymptotic solutions,
1 2z 3/2
w1,2 = exp (∓ )(1 + O(z −3/2 )) . (20)
z 1/4 3
76 R. Grimshaw
It follows that we can write that
1 2z 3/2
Ai(z) = a1 w1 (z) + a2 w2 (z) , where w1,2 ∼ exp (∓ ). (21)
z 1/4 3
The issue is to find the constants a1,2 . The key requirement is to find
the regions where w1,2 are either exponentially decaying or growing. These
regions are defined by anti-Stokes lines where the exponential is purely
oscillatory, while the Stokes lines are where the exponent is real-valued.
Here the anti-Stokes lines are arg = ±π/3, ±π (shown in red), and the
Stokes lines are arg = 0, ±2π/3 (shown in blue).
Figure 2. Plot of the ant-Stokes lines (red) and the Stokes lines (blue).
Thus w1,2 (z) is exponentially small (large) ( subdominant, dominant) in

the sector |arg(z)| < π/3, but is exponentially large (small) in the sectors
π/3 < |arg(z)| < π. It follows that in the sector |arg(z)| < π/3, a2 = 0
since we require that Ai(z) is to be the solution which is exponentially
small for z = x > 0. But a2 = 0 in the sectors π/3 < |arg(z)| < π. This
discontinuity in the constant a2 is the Stokes phenomenon.
The constant a1 must be real-valued in order for Ai(z) to be real-valued
when z = x > 0; it is found by comparison with the asymptotic formula
(16) and so a1 = 1/2π 1/2 . The final task then is to find the value of a2
in terms of a1 . The fact that a2 = 0 when we evaluate the asymptotic
expression (21) for Ai(z) as z → x < 0 is crucial. The evaluation of a2 can
be achieved by recourse to the integral expression (13), or more simply by
the requirement that Ai(z) should be real-valued when z = x < 0. This
latter yields a2 = ±ia1 as z = x < 0, arg(z) = ±iπ. Substitution into (21)
yields the asymptotic expression
1 2(−x)3/2 π
Ai(x) ∼ sin ( + ), as x → −∞ . (22)
π 1/2 (−x)1/4 3 4
The corresponding expressions for Bi(z) can now be found from the formula
(15). Thus, the counterparts of (16, 22) for Ai(z) are
1 2z 3/2 π
Bi(z) ∼ exp ( ), as |z| → ∞ , |arg(z)| < , (23)
π 1/2 z 1/4 3 3
1 2(−x)3/2 π
Bi(x) ∼ cos ( + ), as x → −∞ . (24)
π 1/2 (−x)1/4 3 4
2 Wave Scattering and Reflection

2.1 Forced Harmonic Oscillator
The classical D’Alembert wave equation is
Utt − c2 Uxx = −c2 F (x, t) , (25)
where c is the constant wave speed, and F (x, t) is a specified forcing term.
Suppose F (x, t) = f (x) exp (−iωt) and seek a time-harmonic solution U (x, t) =
u(x) exp (−iωt). Then we get a forced harmonic oscillator equation for u(x),
uxx + k 2 u = f (x) , (26)
where k = ω/c. We shall suppose that f (x) is localized, so that f (x) and
all its derivatives vanish as x → ±∞. In the absence of any forcing, the
general solution is
u = a exp (ikx) + b exp (−ikx) , (27)
for some constants a, b, describing a wave to the right or left respectively.

This is also the general asymptotic solution at infinity, since f (x) → 0
there, where now in general a = a± , b = b± according as we take the limit
78 R. Grimshaw
x → ±∞. They correspond to incident, reflected or transmitted waves. The

aim now is to determine these constants, and in particular, which may be
taken equal to zero.
Although the forced equation (26) can be explicitly solved, it is instruc-
tive to consider the limit k → ∞. Then we can construct an asymptotic
expansion
f (x) fxx (x)
u(x) ∼ 2 − + O(k−6 ) . (28)
k k4
Clearly, this can be continued to all orders in k−2N , and no free waves are
obtained. We infer that the wave amplitudes at infinity are exponentially
small. The general solution of (26) is
∞
1
u(x) = a exp (ikx) + b exp (−ikx) + exp (ik|x − x̂|)f (x̂)dx̂ , (29)
2ik −∞
where a, b are arbitrary constants. We now impose the boundary condi-

tions all waves should be outgoing, so that as x ± ∞, u → a+ exp (ikx),
b− exp (−ikx) for some constants a+ , b− ; note this implies there is a Stokes
phenomenon. Now let x ± ∞ in (29),
u(x) → a exp (ikx) + b exp (−ikx) + α± exp (±ikx) ,

∞
1 (30)
α± = exp (∓ikx̂)f (x̂)dx̂ .
2ik −∞
Hence we must choose a = b = 0 in (29). It follows that the radiated

outgoing waves are given by
∞
1
u(x) → α± exp (±ikx) , α± = exp (∓ikx̂)f (x̂)dx̂ . (31)
2ik −∞
The radiation coefficients α± (k) are not zero, but are exponentially small
∗
as k → ∞. Note that α− = −α+ (the superscript ∗ denotes the complex
conjugate). Hence it is sufficient to evaluate α+ . The contour of integration
is moved into the lower half of the complex x̂-plane, that is Im x̂ < 0. Then
if the nearest singularity of f (x̂) to the real x̂-axis is at x̂ = K − iL, L > 0
we can estimate that α+ = O(exp(−kL)) as k → ∞. For example, fˆ(x̂) =
sech(γ x̂) has poles at x̂ = ±iπ/2γ, ±3iπ/2γ, · · · . Evaluating the residue at
the nearest pole to the real axis yields
π πk
α+ ∼ exp (− ) . (32)
ikγ 2γ
Indeed, in this case the integral (31) can be explicitly evaluated to give
π πk
α+ = sech( ) ,
2ikγ 2γ
from which the estimate (32) readily follows.

Suppose now that
f (x̂) = exp (−γ 2 x̂2 ) .
This has no singularities in the complex x̂-plane, and so the scattering
coefficients must be estimated in a different way. Indeed, this Gaussian
function, while decaying exponentially on the real axis, is only subdominant
in the sectors |argx̂| < π/4, 3π/4 < |argx̂| < π, and is dominant in π/4 <
|argx̂| < 3π/4. Thus the contour of integration in (31)
∞
1
α± = exp (∓ikx̂)f (x̂)dx̂
2ik −∞
cannot easily be moved from the real axis. Instead, in this case, we can
evaluate the integrals explicitly to get that
π 1/2 k2
α± = exp (− 2 ) .
2ikγ 4γ
2.2 Balanced Flow and Slow Manifold

Consider the system, introduced by Lorenz (1986), see also Lorenz and
Krishnamurthy (1987) and Vanneste (2003), as a model of large-scale at-
mospheric flow interacting with internal-inertial gravity waves,
ut = −vw + bvy , vt = uw − buy , wt = −uv , (33)
xt = −y , yt = x + buv . (34)
The variables (u, v, w) represent the large-scale flow, and the variables (x, y)
represent the internal gravity waves. The parameter b represents a Froude
number, and we shall consider the case when b << 1. When b = 0, the
system uncouples into a set of equations describing the slow manifold, a
resonant triad of large-scale planetary (Rossby) waves, and an equation for
fast oscillations, representing gravity waves. Hence, seek an asymptotic
expansion
∞
∞

(u, v, w) ∼ (un , vn , wn )bn , (x, y) ∼ (xn , yn )bn . (35)
n=0 n=1
80 R. Grimshaw
Then at the leading order, the system (33, 34) reduce to
u0t = −v0 w0 , v0t = u0 w0 , w0t = −u0 v0 , x1tt + x1 = −u0 v0 . (36)
We take a solution for the slow manifold as
u0 = F sech(F t) , v0 = F tanh (F t) , w0 = F sech(F t) . (37)
Then the equation for the leading-order fast oscillations is
x1tt + x1 = f1 = F 2 sech(F t) tanh(F t) , (38)
which has the general solution

∞
1
x1 = A exp (it) + B exp (−it) + exp (i|t − t̂|)f1 (t̂)dt̂ . (39)
2i −∞
The arbitrary constants A, B are now determined by the requirement that

there should be no oscillations as t → −∞. Then we can show that
∞
π
x1 ∼ −C1 cos t , C1 = sin (t̂)f1 (t̂)dt̂ = πsech( ) x → ∞. (40)
−∞ 2F
Thus there is no pure slow manifold, as gravity waves are inevitably gen-
erated as t → ∞. Note that the oscillations are exponentially small as the
parameter F → 0.
2.3 Waves in a Variable Medium

The classical D’Alembert wave equation in a variable medium is
Utt − {c2 (x)Ux }x = 0 , (41)
where the speed c = c(x) varies spatially. Now seek a time-harmonic solu-
tion U (x, t) = u(x) exp (−iωt) and then (41) becomes
{c2 (x)ux }x + ω 2 u = 0 . (42)
We shall seek high-frequency (ω → ∞) WKB solutions. For this purpose

we make the transformation
x
dx̂
X= , v(X) = c(x)1/2 u(x) , (43)
x0 c(x̂)
so that vXX + ω 2 v = Q(X)v , (44)

(c−1/2 cX )X cXX c2
Q(X) == 1/2
= − X2 . (45)
2c 2c 4c
In the limit ω → ∞, the right-hand side of (2.3) can be ignored at
the leading order, assuming that Q(X) is a smooth function. Then WKB
solutions are found by putting
v(X) = exp (iωΦ(X)) , (46)
iωΦXX − ω 2 Φ2X + ω 2 = Q(X) . (47)

At the leading order Φ ∼ ±X, corresponding to a wave to the right, or a
wave to the left respectively. The leading terms are
X
Q̃(X) iQ(X)
Φ ∼ ±X ∓ − + · · · , Q̃(X) = Q(X̂)dX̂ . (48)
2ω 2 4ω 3 −∞
Thus the WKB asymptotic solutions are
Q̃(X) Q
v± (X) = exp (±i[ωX − ]){1 + 2 + O(ω −3 )} . (49)
2ω 4ω
The general solution of (42) is
v(X) = A+ v+ (X) + A− v− (X) . (50)
Now suppose that Q(X) (45) is a localized function such that Q(X) → 0
as X ± ∞. That is, the variable speed c(x) is such that c → c± as x → ±∞,
where c± are constants. Then consider a wave incident on the variable
medium from the left. That is, we impose the boundary conditions that
v(X) → exp (iωX) + R exp (−iωX) as X → −∞ ,

(51)
v(X) → T exp (iωX) , as X → ∞ ,
where the reflection and transmission coefficients R, T are constants to be

determined; again this formulation implies the presence of a Stokes phe-
nomenon. Also note that in terms of u(x) the transmission coefficient is
T (c− /c+ )1/2 .
Then from the WKB solution (50) we infer that A+ = 1, A− = 0, noting
that as x → −∞, Φ ∼ ±X and as x → ∞, Φ ∼ ±X ∓ Q̃(∞)/2ω 2 . Hence
v(X) ∼ v+ (X) . (52)
Thus, T ∼ exp (iQ̃(∞)/2ω), |T | ∼ 1, R ∼ 0 in this WKB approximation,

and we infer that to all orders in ω −N , there is no reflected wave. The
82 R. Grimshaw
reflection coefficient R must be exponentially small as ω → ∞, and its

calculation needs exponential asymptotics, see Meyer (1975) and Grimshaw
(1976).
First note that from the exact equation for v(X) (44), we find that
v ∗ vX − vvX
∗
= constant , (53)
which is an expression of conservation of wave energy flux. Then, substitu-

tion of the boundary conditions at infinity (51) into (53) yields
|R|2 + |T |2 = 1 . (54)
Next, again from the the exact equation for v(X) (44) we can show that
v(X) = A exp (iωX) + B exp (−iωX)

∞
1 (55)
+ exp (iω|X − X̂|)Q(X̂)v(X̂)dX̂ ,
2iω −∞
The constants A, B are found by applying the boundary conditions at in-

finity (51), so that A = 1, B = 0, and then
∞
1
R, T − 1 = ∓ exp (±iω X̂)Q(X̂)v(X̂)dX̂ . (56)
2iω −∞
Equation (55) is an integral equation for v(X), while the expressions (56)
allow the determination of the reflection coefficients. Since v(X) ∼ v+ (X) to
all orders O(ω −N ) (52), it follows that the reflection coefficient R can now be
obtained from (56). Keeping just the leading term for v+ (X) ∼ exp (iωX),
∞
1
R∼− exp (2iω X̂)Q(X̂)dX̂ . (57)
2iω −∞
Since we are assuming that Q(X̂) is a smooth function of X we can infer that
R is O(ω −N ), and is exponentially small, determined by the singularities of
Q(X̂) in the complex X̂-plane. This if the nearest singularity to the real
axis is at X̂ = K + iL, L > 0, then R = O(exp (−2ωL) as ω → ∞.
Now recall that
x ˆ
(c−1/2 cX )X cXX c2X d(x)
Q(X) = 1/2
= − 2
, X = ,
2c 2c 4c x0 c(x̂)
c1/2 (c1/2 cx )x ccxx c2

or Q(x) = = + x.
2 2 4
As an example, suppose that

β
c(X) = α + β tanh (γX) , so that x = αX + ln (cosh (γX) ,
γ
where α > |β|, γ > 0. It follows that as X → ±∞, x ∼ (α ± β)x and
c± = α ± β, 2(α, β) = c+ ± c− . Evaluating Q(X) we find that the relevant
poles are double poles at X̂ = iπ/2, iπ/2 + ln (c+ /c−). Since,
∞
1
R∼− exp (2iω X̂)Q(X̂)dX̂ ,
2iω −∞
evaluation at these dominant poles gives

iπ πω c+
R∼− exp (− ){1 + exp [2iω ln ( )]} .
2 γ c−
3 Borel Summation: Forced Nonlinear Harmonic

Oscillator
We now consider the forced harmonic oscillator with a nonlinear term
μ2 uxx + u − u2 = f (x) , (58)
where , μ are small parameters. We shall suppose that f (x) is localized, so

that f (x) and all its derivatives vanish as x → ±∞. In this limit, f (x) → 0
and the general solution of (58) includes a family of periodic solutions, which
in the limit of small amplitude take the form
u ≈ a± cos (x/μ) + b± sin (x/μ) , (59)
for some constants a± , b± .

We shall now further suppose that f (x) = f (−x) is symmetric and an
analytic function of x in the complex x-plane. Then we seek the solutions
of (58) which are also symmetric, that is u(x) = u(−x). In particular, it
then follows that a+ = a− , b+ = −b− . To be explicit, we shall consider two
representative examples:
(a) : f (x) = sech2 (x) ,

(60)
(b) : f (x) = sech(x) ,
A balance is now required between the two small parameters, and this
depends on the form of f (x). Since we expect that the structure of the
84 R. Grimshaw
solution is related to the singularities of f (x) in the complex x-plane nearest

to the real axis. We note that for cases (a,b) these are poles at x = ±iπ/2.
Near each pole,
−1 −i
(a) : f (x) ≈ , (b) : f (x) ≈ , (61)
(x − iπ/2)2 (x − iπ/2)
respectively. Hence we write
iπ
x= + λq , v(q) = u(x) , (62)
2
so that equation (58) becomes
μ2 − −i
vqq + v − v 2 = 2 2 for case (a) , for case (b) , (63)
λ2 λ q λq
respectively. In this inner equation all terms must be in balance, and so
(a) : λ = μ = 1/2 , (b) : λ = μ = . (64)
Each of these cases will now be examined using exponential asymptotics

and Borel summation to calculate the small tails (59), see Pomeau et al
(1988), Kruskal and Segur (1991) and Grimshaw and Joshi (1995) for the
development and application of this approach. An alternative approach
using Fourier transforms was developed by Akylas and Yang (1995).
3.1 Case (a)

Equation (58) is now
uxx + u − u2 = sech2 (x) . (65)
First consider the outer expansion

∞
n

u ∼ us = un (x)n , u0 = sech2 (x) , un+1 = uj un−j − unxx . (66)
n=0 j=0
As expected, for real x this asymptotic solution is symmetric, and no

tail oscillations emerge to all orders in n . Instead, these are found from
the singularties of us in the complex x-plane, which are located at x =
±iπ/2, ±3iπ/2, · · · . It will be sufficient to consider just the singularity at
x = iπ/2 closest to the real axis in the upper half-plane, and write, see (62),
iπ
x= + 1/2 q . (67)
2
Substitution into (66) and evalaution in the limit 1/2 q → 0 yields
1 1 7 1 2
us ∼ (− + 4 + · · · ) + ( − 2 + · · · ) + O() . (68)
q2 q 3 3q
Next we consider the inner expansion, where we replace x with q (67),

and put v(q) = u(x) (62), so that
1
vqq + v − v 2 = − 2
1/2
∼ − 2 + + ··· . (69)
sinh ( q) q 3
The symmetry condition becomes
Im v(q) = 0 , on Re q = 0 . (70)
Equation (69) is exact, and we seek the expansion

∞

v(q) ∼ vn (q)n , (71)
n=0
1 1
so that v0qq + v0 − v02 = − .v1qq + v1 − 2v0 v1 = − . (72)
q2 3
Matching with the outer expansion (68) for us yields
1 7 1 2
v0 ∼ − + 4 + ··· , v1 ∼ − + ··· , (73)
q2 q 3 3q 2
taken in the limit q → ∞ for Re q > 0, Im q < 0.
1
v0qq + v0 − v02 = −
q2
Now seek an asymptotic solution
∞
bn
v0 ∼ , as q → ∞ , Re q > 0, Im q < 0 (74)
n=1
q 2n
Substitution into equation (72) yields b1 = −1, b2 = 7 as required by (73)

and then
n−1

bn+1 + 2n(n + 1)bn − bj+1 bn−j = 0 , n = 1, 2, · · · . (75)
j=0
86 R. Grimshaw
Figure 3. The complex s-plane for Req > 0, Im q < 0. The line Re (sq) = 0
is blue, and the contour Γ is red.
The aim now is to sum the series (74) using Borel summation We seek a
solution of equation (72) in the form of a Laplace transform

v0 = exp (−sq) V (s) ds , (76)
Γ
where the contour Γ runs from zero to infinity in the complex s-plane such
that Re(sq) > 0.
Substitution of the Laplace transform into equation (72) for v0 yields a
Fredholm integral equation for V (s),
s
(s2 + 1)V (s) − V (ŝ)V (s − ŝ) dŝ = −s . (77)
0
To solve this we seek a solution as a power series

∞

V (s) = an s2n+1 . (78)
n=0
Substitution of (78) into equation (77) shows that a0 = −1 and

n−1
(2j + 1)!(2n − 2j − 1)!
an + an−1 − aj an−j = 0, n = 1, 2, · · · . (79)
j=0
(2n + 1)!
Next, substitution of (78) into the Laplace transform (76) recovers the
asymptotic series (74) with
bn+1 = (2n + 1)!an . (80)
Thus solving the recurrence relation (79) for an effectively sums the asymp-
totic series (74) and yields the solution of equation (72) for v0 as a Laplace
transform. Examination of the recurrence relation (79)
n−1
(2j + 1)!(2n − 2j − 1)!
an + an−1 − aj an−j = 0, n = 1, 2, · · · .
j=0
(2n + 1)!
shows that as n → ∞ , an ∼ (−1)n K where K = 1.55... is a constant found

numerically. Hence the series (78)
∞

V (s) = an s2n+1 .
n=0
converges for |s| < 1. Analytic continuation into the complex s-plane, using
the integral equation (77), yields a complete solution for V (s) and hence
v0 (q) as the Laplace transform (76). )
Since an ∼ (−1)n K as n → ∞ we see that V (s) = ∞ n=0 an s
2n+1
has
singularities at s = ±i given by
Ks
V (s) ≈ , for|s| ≈ 1 . (81)
s2 + 1
There are similar poles at s = 2i, 3i, · · · , but these will be seen to generate
higher harmonics in the tail oscillations, and so are not our immediate
concern. We must know make a specific choice of the contour Γ in the
Laplace transform (76). Since we are seeking a symmeric solution, which
satisfies the condition (70), it is sufficient to suppose that at first Re q > 0
and Im q < 0. Then choose the contour Γ to lie initially in Re s > 0, Im s ≥
0, so that Re sq > 0 and the Laplace transform is well defined for the allowed
values of q. In particular, since V (s) can be represented by the power series
(78), which generates the asymptotic series (74), which in turn is equivalent
to the asymptotic series (66), we conclude that v0 ∼ vs = us .
88 R. Grimshaw
The next step is to deform the contour Γ onto the imaginary s-axis. in
this process we will need to deform around the poles at s = i, 2i, 3i, · · · and
collect the (half) residues. Hence we find that (76) becomes, on putting
s = iy on the deformed contour Γ,
∞
iπK
v0 = exp (−iyq) V (iy) i dy + exp (−iq) + · · · , (82)
0 2
where the dots denote terms proportional to exp (−2iq), exp (−3iq), · · · .
The integral is interpreted as a principal value integral at the singularities
at y = 1, 2, 3, · · · . This holds in Re q > 0, Im q < 0. To apply the symmetry
condition (70) that v0 should be real-valued on the imaginary q-axis, we
must now let Re q → 0, and put)q = −iQ, Q > 0 in the expression (82).
∞
From the series (78), V (iy) = i n=0 an (−1)n y 2n is pure imaginary since
the coefficients an are all real-valued. Hence the integral term is real-valued
as required. But the contributions from the poles are pure imaginary, and
hence the expression (82) cannot satisfy the symmetry condition. The rem-
edy is to note that the term exp (−iq) is exponentially small in the sector
Re q > 0, Im q < 0 and hence is subdominant, so that we are allowed to add
such terms asymptotically to (82).
Thus we replace (82) with
∞
iπK ib
v0 = exp (−iyq) V (iy) i dy + exp (−iq) + exp (−iq + iδ) + · · · ,
0 2 2
(83)
where b, δ are real constants. Now application of the symmetry condition
shows that
b cos δ = −πK . (84)
Thus the final solution for v0 in Re q > 0, Im q < 0 is

ib
v0 ∼ exp (−sq) V (s) ds + + exp (−iq + iδ) + · · · , (85)
Γ 2
where b is determined in terms of δ by (84). It remains to bring this solution

back to the real x-axis, for x > 0, using (67), x = iπ/2 + 1/2 q. Here we
must also collect a similar contribution from the singularities in the lower
half of the complex x-plane. Thus the full solution consists of a central core
and an exponentially small tail oscillation.
The final result is that, for x > 0 (the case x < 0 follows from the
imposed symmetry),
b π x
u ∼ us + exp (− 1/2 ) sin ( 1/2 − δ) , b cos δ = −πK , (86)
2
where us ∼ sech2 (x) + O() is given by (66). The tail oscillations form a
one-parameter family characterized by the phase shift δ, 0 ≤ δ < π/2, where
the minimum amplitude occurs for δ = 0.
Figure 4. Plot of the asymptotic solution (86) for δ = 0 and = 1/36.
Suppose that instead of the symmetry condition (70) we look for one-
sided solutions such that u(x) → 0 as x → ∞. The same procedure can
be followed, and again the solution for Re q > 0 is given by the Laplace
transform (76), with the same contour Γ, which ensures that there are no
oscillations as x → ∞. But now, to find the behaviour as x → −∞, the
contour Γ must be moved across the imaginary q-axis to Re q < 0, and in
doing so the solution collects the residue at the poles s = i, 2i, 3i, · · · . The
residue at s = i is iπK exp (−iq), exactly twice the contribution from the
half residue shown in (82). Then, bringing the solution back to the real
x-axis, in x < 0, we find that
2πK π x
u ∼ us + exp (− 1/2 ) sin ( 1/2 ) . (87)
2
For these one-sided solutions there are no free parameters.

90 R. Grimshaw
3.2 Case(b): Outer expansion

Equation (58) is now
2 uxx + u − u2 = sech(x) . (88)

)∞
The outer expansion is again (66), u ∼ us = n=0 un (x)n , but now
n+1

u0 = sech(x) , u1 = u20 , un+2 = uj un−j − unxx . (89)
j=0
Again, for real x this asymptotic solution is symmetric, and no tail oscil-
lations emerge to all orders in n . The singularties of us in the complex
x-plane, are again located at x = ±iπ/2, ±3iπ/2, · · · , and we now write in
place of (67), see (62),
iπ
x= + q . (90)
2
Substitution into (89) and evalaution in the limit q → 0 yields
1 i 1 iq 1
us ∼ (− − 2 + · · · ) + ( − + · · · ) + O(3 ) . (91)
q q 6 3q
Here we replace x with q (90), and put v(q) = u(x) (62), so that
i i iq
vqq + v − v 2 = − ∼− + + ··· . (92)
sinh (q) q 6
The symmetry condition (70) is unchanged. Instead of (71) we write

∞

v(q) ∼ vn (q)2n , (93)
n=0
i 1
so that v0qq + v0 − v02 = − .v1qq + v1 − 2v0 v1 = − . (94)
q 3
i 1 1 2
v0 ∼ − − 2 + · · · , v1 ∼ − + ··· , (95)
q q 3 3q 2
taken in the limit q → ∞ for Re q > 0, Im q < 0.

i
v0qq + v0 − v02 = −
q

∞
bn
v0 ∼ n
, as |q| → ∞ , Imq < 0 (96)
n=1
q
Substitution into equation (72) yields b1 = −i, b2 = −1 as required by (73)

and then
n+1

bn+2 + n(n + 1)bn − bj bn+2−j = 0 , n = 1, 2, · · · . (97)
j=1
The aim now is to sum the series (96) using Borel summation We again
seek a solution of equation (94) in the form of a Laplace transform (76).
As before, we substitute the Laplace transform into equation (94) to get
now the Fredholm integral equation
s
(s2 + 1)V (s) − V (ŝ)V (s − ŝ) dŝ = −i . (98)
0

∞

V (s) = (−i)n+1 an sn . (99)
n=0
Substitution of (99) into equation (98) shows that a0 = 1, a1 = 1 and

n
(j − 1)!(n − j)!
an − an−2 − aj−1 an−j = 0, n = 2, 3, · · · . (100)
j=1
n!
Next, again as before, substitution of (99) into the Laplace transform

(??) recovers the asymptotic series (96) with
bn+1 = n!(−i)n+1 an . (101)
Examination of the recurrence relation (100) shows that
an ∼ Cn + O(n−1 ) , as n → ∞,
where C = 0.94 is a constant found numerically. Hence the series (99)

converges for |s| < 1, and so as in case (a), we can obtain a complete
solution for V (s) and v0 (q). There is a singularity at s = i, given by
C iC
V (s) ≈ + , for s ≈ i . (102)
(s − i) (s − i)2
92 R. Grimshaw
The essential difference here from case (a) is the double pole at s = i. This
leads to a half-residue of
πC(q + i) exp (−iq) (103)
where the new feature is the non-uniform factor “q”.

To explain the origin of this new secularity, we return to the basic equa-
tion (94) and seek a solution of the form
v0 ∼ vs + vw , (104)
where as |q| → ∞, vs is given by the asymptotic expansion (96) and is the

dominant term. Substitution into (94) and linearization yields
vwqq + vw − 2vw vs ≈ 0 . (105)
Then, as |q| → ∞,
vw ∼ B(q + i) exp (−iq + iδ) , (106)
for some real constants B, δ. The secular term arises, because in contrast
to case (a), vs ∼ q −1 rather than q −2 .
Then we choose the same contour Γ as in case (a), move the contour to
the imaginary q-axis, collect the half-residues and add the allowed subdom-
inant terms. The outcome is that (83) is replaced by
v0 ∼ vint + πC(q + i) exp (−iq) + B(q + i) exp (−iq + iδ) + · · · . (107)

∞
vint = exp (−iyq) V (iy) i dy . (108)
0
To apply the symmetry condition

) (70), we let q = −iQ, Q > 0 in (107).
The series (99) is V (iy) = i ∞ n n
n=0 an (−1) y is pure imaginary since the
coefficients an are all real-valued. Hence vint is real-valued, and as in case
(a) we must choose the constants B, δ to ensure that v0 is real-valued on
the imaginary q-axis. The outcome is
B cos δ = −πC . (109)

v0 ∼ exp (−sq) V (s) ds + +B(q + i) exp (−iq + iδ) + · · · , (110)
Γ
where the integral term is just the asymptotic series vs = us .

It remains to bring this solution back to the real x-axis, for x > 0, using
(90), x = iπ/2 + q. Here we must also collect a similar contribution from
the singularities in the lower half of the complex x-plane. The outcome will
be
u ∼ us + uw , (111)
where us ∼ sech(x) + O() is given by (89). Substitution into (88) and
linearization yields
2 uwxx + uw − 2uw us ≈ 0 . (112)
Just as for the corresponding equation (105) for vw , a WKB analysis yields
uw ∼ A(x) exp (iφ(x)) + A(x) exp (−iφ(x)) , (113)
where both A(x), φ(x) are real-valued for real x. We find that
1
φx ∼ − us + O() , A ∼ A0 + O() , (114)

Using the leading expression for us ,
x
φ∼ − 2 tan−1 (exp (x)) + φ0 + O() . (115)

In order to find the constants A0 , φ0 , the WKB expression for uw (113)
should be matched with the corresponding expression for vw = uw (106)
when x = iπ/2 + q, for Re q > 0, Im q < 0. Thus, we find that
iπ q
φ∼ + q − π + i ln [tanh ( )] + φ0 + O() ,
2 2
so that, as |q| → ∞,
A0 q π
uw ∼ − exp ( − iq − iφ0 ) , (116)
2 2
noting that in Im q > 0, exp (−iq) is dominant over exp (iq). The matching
condition now shows that
2B π
A0 = − exp (− ) , φ0 = −δ . (117)
2 2
Finally we get that, as x → ∞, the tail oscillations are given by
4B π x
uw ∼ 2
exp (− ) cos ( − δ) , (118)
2
where B cos δ = −πC (109). Again, they form a one-parameter family.
94 R. Grimshaw
4 Generalized Solitary Waves

4.1 Korteweg-de Vries equation
Solitary waves are steady localized traveling waves of permanent form. In
the weakly nonlinear long wave regime, the canonical equation describing
solitary waves is the famous Korteweg-de Vries (KdV) equation, first derived
in the water-wave context in 1895, and subsequently found to hold in many
physical systems,
ut + c0 ux + μuux + δuxxx = 0 . (119)
The dominant term is ut + c0 ux ≈ 0, showing that the wave propagates
approximately with a speed c0 . At the next order weak dispersion due to
the term δuxxx is balanced by weak nonlinearity due to the term μuux , The
speed c0 (linear long wave phase speed), and the coefficients μ, δ depend on
the particular physical system being considered. For water waves
3c0 c0 h2
c0 = (gh)1/2 , μ= , δ= ,
2h 6
when u(x, t) is the elevation of the water surface above the undisturbed
depth h, and g is the acceleration due to gravity. The KdV equation is, in
the reference frame moving with speed c0 (transform x → x − c0 t),
ut + μuux + δuxxx = 0 . (120)
This is an integrable equation, a result first established in the 1960’s
by Kruskal and collaborators. Its principal solutions are solitons. A single
soliton is the solitary wave, given by
μa
u = a sech2 (γ(x − V t)) , V = = 4δγ 2 . (121)
3
Note that it solves the second order ordinary differential equation
μ
δuxx − V u + u2 = 0 . (122)
2
The solution (121) is a one-parameter family of solutions, parametrized by
the amplitude a say. The speed V is proportional to the amplitude a and
is positive (negative) as δ > 0(< 0), and is also proportional to the square
of the wavenumber γ; thus large waves are thinner and travel faster. They
are waves of elevation (depression) when μδ > 0(< 0). Integrability means
that the general initial-value problem for a localized initial condition can be
solved through the Inverse Scattering Transform, with the generic outcome
of a finite number of solitons propagating in the positive x-direction, and
some dispersing radiation, propagating in the negative x-direction (when
μδ > 0).
Figure 5. The generation of three solitons from a localized initial condition

for the KdV equation At + 6AAx + Axxx = 0.
4.2 Linear spectrum

The basic paradigms for solitary waves, at least for weakly nonlinear
long wave regime, is the central role played by model equations, such as the
KdV equation. Our aim now is to describe how small-amplitude solitary
waves can be found by an asymptotic perturbation procedure directly from
the full system, rather than from a model equation.
Because solitary waves are required to decay in their tail regions, some
information about their possible existence or otherwise, can be obtained by
an examination of these tail regions. Here, except in certain exceptional
cases, a linearized analysis is applicable. One-dimensional steady solitary
waves, propagating in the x-direction with speed c are functions of ξ = x−ct,
together with a set of other spatial transverse variables which define the
modal structure. For instance, for water waves, the dependence is on ξ
and z (x is horizontal and z is vertical), and there is no dependence on the
remaining horizontal variable y.
In the tail region, where we assume that a linearized analysis holds, we
seek solutions proportional to
Re {exp (ik(x − ct))} . (123)
96 R. Grimshaw
The linearized equations will then yield the linear dispersion relation
c = c(k) , (124)
written here for the phase speed c rather than the frequency ω = ck.
Whereas usually this dispersion relation (which may have several branches)
is considered as an equation for c given a real wavenumber k, for solitary
wave tails it needs to be considered as an equation for a complex-valued k
given a real speed c. Indeed, it is immediately clear that if there exist real-
valued solutions of (124) for the given value of c, then it is unlikely that the
solitary wave can decay to zero in its tail region. Instead, it will probably be
accompanied by a non-decaying co-propagating oscillatory wave field. This
consideration leads to the notion that solitary waves generally can only exist
in the gaps in the linear spectrum.
Figure 6. Plot of the water wave dispersion relation (125) for B = 0

(violet), B = 0.2 (red), B = 0.4 (blue).
For instance, for water waves, the dispersion relation in the presence of
surface tension is
c2 (1 + Bq 2 )
= tanh q , q = kh , (125)
gh q
where the Bond number B = σ/gh2 and ρσ is the coefficient of surface

tension (ρ is density), and has a value of 74 dynes/cm at 20o C. It may
then be shown that solitary waves of the KdV type can exist only when
either B = 0, c2 > gh or when B > 1/3, c2 < gh with a bifurcation from
wavenumber zero (k = 0) and c2 = gh in both cases. Otherwise when
0 < B < 1/3 solitary waves can exist for |c| < cm where c2m is the minimum
value that c2 can take in (125) as q takes all real values. In deep water,
|q| → ∞, c2m = 2(gσ)1/2 and occurs at |k| = km = (gσ)1/2 , a wavelength
2π/km = 0.23 cm. These solitary waves bifurcate at a finite wavenumber km
and from the speed cm , and have decaying oscillations in their tail regions.
They are envelope solitary waves, of a quite different kind from the afore-
mentioned KdV-type solitary waves, and closely related to the nonlinear
Schrodinger (NLS) equation.
4.3 Reformulation as a dynamical system

This approach has recently been developed into the basis of a rigorous
approach to finding solitary waves, often called the “dynamical-systems”
method, see Dias and Iooss (2003), Grimshaw and Iooss (2003) and Grimshaw
(2007) for recent reviews. Since here we are considering only solitary waves
which occur in conservative systems, which is the common and traditional
scenario for solitary waves, we shall suppose that the underlying physical
system is Hamiltonian (energy-conserving) and reversible (symmetry under
the transformation ξ → −ξ). In this case it can be shown the the solutions
k of the dispersion relation c = c(k) (124) for each real value of c have the
property that −k and k∗ (complex conjugate) are also solutions. It follows
that generically the solutions form a quartet (k, k∗ , −k, −k ∗ ), with an associ-
ated four-dimensional subspace for the corresponding wave mode. For soli-
tary waves we require solutions with Im(k) > 0(< 0) when ξ → ∞(→ −∞),
in order to ensure that the solution decays to zero in its tail region. In the
general case when Im(k) = 0 we see that there are generically two such
roots available as ξ → ∞ and, due to the reversible symmetry, two other
roots available as ξ → −∞. Thus, for the corresponding wave mode, as
ξ → ∞ two boundary conditions are needed at each of ±∞. This count is
consistent with the existence of a solitary wave solution, which from this
dynamical systems point of view, is a homoclinic orbit.
Next, consider how this quartet structure may change as some system
parameter is varied. Bifurcations arise when two solutions for k coalesce,
for which the necessary condition is that ∂c/∂k = 0 simultaneously with
the dispersion relation (124). When this occurs at a real value of k, it is
98 R. Grimshaw
equivalent to the condition that c = cg where
∂ω ∂(ck) ∂c
cg = = =c+k
∂k ∂k ∂k
is the group velocity. Generically, there are four possibilities:
(1) (0, 0, iγ, −iγ) where γ > 0 is real-valued .

(2) (0, 0, β, −β) where β > 0 is real-valued .
(3) (β, β, −β, −β) where β > 0 is real-valued .
(4) (iγ, iγ, −iγ, −iγ) where γ > 0 is real-valued .
Case (1) corresponds to a KdV-type solitary wave, and case (3) corresponds
to an envelope (NLS) solitary wave. Case (2) corresponds to a so-called
generalized solitary wave, which does not decay at infinity, but instead is
accompanied there be small-amplitude co-propagating oscillations. Case
(4) has only rarely been studied and corresponds to a transition from a
KdV-type solitary wave to an envelope solitary wave.
The full system is now projected onto the appropriate four-dimensional
subspace, and the resulting bifurcation analyzed within the framework of
this subspace. Of course, rigorous results require a delicate and sophisti-
cated justification of this process. Here we shall instead briefly describe the
structure of the subspace, which we suppose is represented by the 4-vector
W(ξ). It satisfies an equation of the form
Wξ = L( W;
) + N ( W) . (126)
Here L( W;
) is a linear operator and N ( W) contains all the nonlinear
terms. The bifurcation parameter is
, and is such that the spectrum of L
at
= 0 reproduces one of the cases (1) to (4) describe above. That is, the
eigenvalues λ = ik of the linear operator L( W; 0) are respectively :
(1) (0, 0, −γ, γ) .

(2) (0, 0, iβ, −iβ) .
(3) (iβ, iβ, −iβ, −iβ)
(4) (−γ, −γ, γ, γ) .
4.4 Case (1)

Let us first consider case (1). At the bifurcation point (
= 0) the
linearized system (126) has a double-zero eigenvalue. Then, without loss
of generality, we can transform the linear operator L( W;

) so that its
projection onto the two-dimensional is represented by the matrix

0 1
. (127)

0
Generically there is a corresponding single eigenvector V0 , and a single

generalized eigenvector V1 , chosen orthogonal to V0 . Small-amplitude
solutions are then sought in the form
W = A(ξ) V0 + B(ξ) V1 + W(2) . (128)
Here A, B are real variables of O(α), measuring wave amplitude, while

W(2) , is orthogonal to V0 , V1 and of O(α2 ). The two remaining eigen-
values ∓γ play no role at the leading order here, since they correspond to
strong exponential decay at infinity, and their effects are included in W(2) .
Projection onto the two-dimensional subspace yields
Aξ = B + a1 A2 + a2 AB + a3 B 2 + O(α3 ) , (129)
Bξ =
A + b1 A2 + b2 AB + b3 B 2 + O(α3 ) . (130)
The next step is to make a near-identity transformation
Ã = A + α1 A2 + β1 AB + γ1 B 2 , B̃ = B + α2 A2 + β2 AB + γ2 B 2 , (131)
Substitution into (129, 130) and choosing the coefficients so that
α1 = β2 = −a2 , β1 = 2a1 + b2 , γ1 = −a2 + b3 , α2 = a1 , γ2 = 2a1 + a3 + b2 ,
where we show only the leading order terms as

→ 0, enables the elimination
of all the quadratic nonlinear terms in (129, 130) except A2 in (130) with
coefficient b1 . Hence, omitting the superscript we find the normal form,
Aξ = B + · · · ,
Bξ =
A + μA2 + · · · , (132)
where μ = b1 is a real-valued coefficient, specific to the system being con-

sidered, and the omitted terms are O(α3 ). The coefficient
yields the per-
turbed eigenvalues ±
1/2 for
> 0, and ±i|
|1/2 for
< 0; the former case
yields the solitary wave solution. Comparison with the dispersion relation
(124) leads to the identification of
as
2(c − c(0))

=− . (133)
ckk (0)
100 R. Grimshaw
It follows that for solitary waves, c > (<)c(0) according as ckk (0) < (>)0,
as expected. When the error terms in (132) are omitted, it becomes the
steady-state KdV equation (122) with the “sech2 ” solution (121) It is then
a delicate and intricate task to establish that this solitary wave solution
persists when the error terms are restored.
4.5 Case (2)

Next consider case (2). At the bifurcation point (
= 0) the linearized
system (126) again has a double-zero eigenvalue, with a corresponding single
eigenvector V0 , and a single generalized eigenvector V1 . However, account
must now be taken of the other two eigenvalues ±iβ, with their associated
eigenvectors V2 , V∗2 , since they do not now lead to decaying solutions at
infinity. Small-amplitude solutions are sought in the form
W = A(ξ) V0 + B(ξ) V1 + C(ξ) V2 + C ∗ (ξ) V∗2 + W(2) . (134)
Here C is a complex-valued variable, and the leading terms form a four-

dimensional subspace (A, B, C), while W(2) is again a small error term.
Projection onto this four-dimensional subspace, and a normal form analysis
reveals that (A, B, C) satisfy the system
Aξ = B + · · · ,
Bξ =
A + μA2 + ν|C|2 · · · ,
Cξ = iγ(1 + δA)C + · · · . (135)
Here μ, ν, δ are real-valued coefficients specific to the system being consid-

ered, and the omitted terms are small error terms as before.
When the error terms are omitted the system is integrable. Indeed in
that limit, |C| = C0 is a constant, and after a change of origin, the sys-
tem reduces to the same form as (132) in case (1). Thus, for the case
> 0 (when case (1) is a KdV-type solitary wave), the solution is a one-
parameter family of homoclinic-to-periodic solutions, with |C| = C0 con-
stant and (A, B) → (A0 , ) as ξ → ±∞ where A0 is a real constant, given
by
A0 + μA20 + νC02 = 0. The solution is a generalized solitary wave which
typically has a “sech2 ” core, and decays at infinity to non-zero oscillations
of constant amplitude C0 and wavenumber γ, see Figure 4. A delicate anal-
ysis of the full system (126) with the the small error terms shows that at
least two of these solutions persist; the minimal amplitude C0 being expo-
nentially small, that is O(exp (−K/|
|1/2 )) where K is a positive real con-
stant. Although such waves are permissible as solutions of the steady-state
equations, they have infinite energy and their associated group velocity is
inevitably inward at one end and outward at the other end. Hence, they
cannot be realised in a physical system from any localized initial condition.
Instead localized initial conditions will typically generate a one-sided gener-
alized solitary wave, whose central core is accompanied by small-amplitude
outgoing waves on one side only. Such waves cannot be steady, and instead
will slowly decay with time.
4.6 Case (3)

Finally we consider case (3), when there is a double eigenvalue λ = iβ
with generically a corresponding single eigenvector V0 , and a single gener-
alized eigenvector V1 , while the complex conjugate double eigenvalue λ =
−iβ has corresponding complex conjugate eigenvectors. Small-amplitude
solutions are now sought in the form
W = A(ξ) V0 + B(ξ) V1 + A∗ (ξ) V∗0 + B ∗ (ξ) V∗1 + W(2) . (136)
Here A, B are complex-valued variables, forming a four-dimensional sub-

space while W(2) is again a small error term. Projection onto this subspace
and a normal form analysis reveals that
Aξ = iβA + B + iAP (
, |A|2 , K) + · · · ,
Bξ = iβB + iBP (
, |A|2 , K) + AQ(
, |A|2 , K) + · · · . (137)
where K = i(AB ∗ − A∗ B) , (138)
Here P, Q are real-valued polynomials of degree 1, that is we may write
P (
, |A|2 , K) =
+ ν1 |A|2 + ν2 K ,
Q(
, |A|2 , K) = 2
β + μ1 |A|2 + μ2 K (139)
where all coefficients are real-valued.

The truncated system, obtained when the error terms are omitted, is
integrable. There are two integrals, K, H both constants, where
μ1 4
H = |B|2 − (2
β|A|2 + |A| + μ2 K|A|2 ) . (140)
2
For a solitary wave solution we must have K = H = 0 and it then follows
that
μ1 4
|A|2ξ = 2
β|A|2 + |A| . (141)
2
Thus solitary wave solutions exist provided that
> 0, and that the non-
linear coefficient μ1 < 0. The condition
> 0 implies that the perturbed
102 R. Grimshaw
eigenvalues, λ ≈ iβ + (2
β)1/2 have split off the imaginary axis, and so pro-
vide the conditions needed for exponential decay at infinity; the condition
μ1 < 0 depends on the particular physical system being considered.
The solution of the truncated system is
4
β
A = a exp (i[β +
]ξ)sech(γξ) , where γ = (2
β)1/2 , |a|2 = − . (142)
μ1
This solution describes an envelope solitary wave, with a carrier wavenum-

ber β +
and an envelope described by the “sech”-function. These soli-
tary waves can also be obtained from the soliton solutions of the nonlin-
ear Schrodinger equation (NLS) equation, for that special case when the
phase velocity equals the group velocity, c = cg , or more precisely when
c + Ω/K = cg + V , where V is the soliton speed and Ω, K are the frequency
and wavenumber corrections. Note that the solution (142) contains an ar-
bitrary phase in the complex amplitude a, meaning that the location of the
crests of the carrier wave vis-a-vis the maximum of the envelope (here lo-
cated at ξ = 0) is arbitrary. However, restoration of the error terms leads to
the result that only two of these solutions persist, namely, those for which a
carrier wave crest or trough is placed exactly at ξ = 0, so that the resulting
solitary wave is either one of elevation or depression. This result requires
very delicate analysis, but could be anticipated by noting that these are
the only two solutions which persist under the symmetry transformation
ξ → −ξ.
For water waves, for which the dispersion relation is (125), these two
cases (1) and (2) imply that pure solitary waves of elevation exist for B =
0, and of depression for B > 1/3, while generalized solitary waves arise
whenever 0 < B < 1/3. Concerning case (3) the conditions are met for
capillary -gravity waves with 0 < B < 1/3, where it can be shown that the
coefficient μ1 < 0 as required. Hence we find envelope solitary waves.
For the case of generalized solitary waves, there is always the possibility
that the amplitude of the oscillations is zero, and the solution then reduces
to a pure solitary, called an “embedded” solitary wave. There are now
many examples of such embedded solitary waves arising in various physical
systems, notably for internal waves, but from various numerical and analyt-
ical studies, it would seem that they do not arise in the water wave context.
This “dynamical-systems” approach to finding solitary waves has also been
applied to interfacial waves, where again the linear dispersion relation holds
the key to where solitary waves can be found.
5 Fifth-order Korteweg-de Vries equation

5.1 Formulation and outer expansion
In the next two sections we shall look at how generalized solitary waves
arise in two illustrative model systems, one in this section representing
gravity-capillary waves, and one in the next section representing internal
waves. First, for gravity-caplillary waves, we shall consider a model equa-
tion, valid when B ≈ 1/3 and for small amplitudes. This is the usual KdV
equation supplemented by a fifth-order linear dispersive term, known as the
fifth-order Korteweg-de Vries (5KdV) equation,
ut + c0 ux + μuux + δuxxx + δ1 uxxxxx = 0 . (143)
where, for gravity-capillary waves

3c0 c0 h 2 τ 1 τ τ2
c0 = (gh)1/2 , μ= , δ= , δ1 = c0 h4 { + − },
2h 6 90 18 72
where τ = 1 − 3B. When B = 1/3, τ = 0, and then δ = 0, δ1 > 0. We are
concerned here with the case when 0 < τ << 1, when δ > 0, δ1 > 0. With
appropriate changes of variable, we replace (143) with
ut + 6uux + uxxx +
2 uxxxxx = 0 . (144)
When
= 0 this is the standard KdV equation. But when
= 0 it has a
linear dispersion relation
c = −k2 +
2 k 4 , (145)
and so there is a resonance at c = 0 between k = 0 and k = ±
−1 .
We seek solutions of the form
u = u(x − ct) , (146)
so that the 5KdV equation (144) becomes
−cu + 3u2 + uxx +

2 uxxxx = 0 , (147)
A constant of integration have been set to zero, essentially by imposing

solitary wave boundary conditions, or better, by translating u by a constant.
Equation (147) is a fourth order ordinary differential equation. We shall seek
a symmetric generalized solitary wave solution, that is u(x) = u(−x), with
a co-propagating oscillatory tail of small amplitude. This amplitude will be
found using exponential asymptotics and Borel summation, in an analogous
manner to that used in section 3, see Pomeau et al (1988) and Grimshaw
and Joshi (1995) for more details.
104 R. Grimshaw
First consider the outer expansion

∞
∞

u ∼ us = un (x)
2n , c∼
2n c2n , (148)
n=0 n=0
2 2 2
u0 = 2γ sech (γx) , c0 = 4γ , (149)
15
u1 = −10γ 2 u0 + u20 , c1 = c20 . (150)
2
For real x this asymptotic solution can be continued to all orders in
2n .
As expected, it is symmetric, and no tail oscillations emerge. Instead, these
are found from the singularties of us in the complex x-plane, which are
located at x = ±iπ/2γ, ±3iπ/2γ, · · · . It will be sufficient to consider just
the singularity at x = iπ/2γ closest to the real axis in the upper half-plane,
and write
iπ
x= +
q . (151)
2γ
Substitution into (148) and evalaution in the limit
q → 0 yields
1 2 30 −6 2 1 −6 2 2γ 2 q 2
us ∼ {− + + O(q )} + 2γ { + O(q )} +
{− }. (152)

2 q2 q4 3 15
5.2 Inner expansion and Borel summation

Next we replace x with q (151), and put v(q) =
2 u(x), so that (147)
becomes
vqqqq + vqq + 3v 2 −
2 cv = 0 . (153)
The symmetry condition becomes
Im v(q) = 0 , on Re q = 0 . (154)
Equation (153) is exact, and we seek the expansion

∞

v(q) ∼ vn (q)
2n , (155)
n=0
so that v0qqqq + v0qq + 3v02 = 0 . (156)

2 30
v0 ∼ − + 4 + O(q −6 ) , (157)
q2 q
taken in the limit q → ∞ for Req > 0, Im q < 0. Note that the speed c does
not appear at the leading order.
v0qq + v0 + 3v02 = 0
∞
bn
v0 ∼ 2n
, as q → ∞ , Re q > 0 Im q < 0 . (158)
n=1
q
Substitution into equation (156) yields b1 = −2, b2 = 30 as required by

(157) and then
(2n − 3)(2n + 4)bn + (2n − 2)(2n − 1)n(2n + 1)bn−1

n−1
(159)
+3 bj bn+1−j = 0 , n = 1, 2, · · · .
j=2
The aim now is to sum the series (158) using Borel summation We seek a
solution of equation (156) in the form of a Laplace transform

v0 = exp (−sq) V (s) ds , (160)
Γ
where the contour Γ runs from zero to infinity in the complex s-plane such
that Re(sq) > 0, see Figure 3.
Substitution of the Laplace transform into the equation (156) for v0
yields the Fredholm integral equation for V (s),
s
(s2 + 1)V (s) + 3 V (ŝ)V (s − ŝ) dŝ = 0 . (161)
0

∞

V (s) = an s2n+1 . (162)
n=0
Substitution of (162) into equation (161) shows that a0 = −2, a1 = 5 and
(2n − 1)(2n + 6)
an + an−1 +
(2n + 2)(2n + 3)
n−1
(163)
(2j + 1)!(2n − 2j + 1)!
3 aj an−j = 0, n = 1, 2, · · · .
j=1
(2n + 3)!
106 R. Grimshaw
Next, substitution of (162) into the Laplace transform (160) recovers the
asymptotic series (158) with
bn+1 = an (2n + 1)! , n = 0, 1, 2, · · · . (164)
Thus solving the recurrence relation (163) for an effectively sums the asymp-
totic series (158) and yields the solution of equation (156) for v0 as a
Laplace transform. Examination of the recurrence relation (163) shows
that as n → ∞, the nonlinear terms drop out, and so an ∼ (−1)n K where
K = −19.97 is a constant found numerically. Hence the series (162)
∞

V (s) = an s2n+1 .
n=0
converges for |s| < 1. Analytic continuation into the complex s-plane, using
the integral equation (161), yields a complete solution for V (s) and hence
v0 (q) as the Laplace transform (160). )
Since an ∼ (−1)n K as n → ∞ we see that V (s) = ∞ n=0 an s
2n+1
has
singularities at s = ±i given by
Ks
V (s) ≈ , for|s| ≈ 1 . (165)
s2 + 1
There are similar poles at s = 2i, 3i, · · · , but these generate higher harmon-
ics in the tail oscillations, and so are not our immediate concern. We must
know make a specific choice of the contour Γ in the Laplace transform (160)

v0 = exp (−sq) V (s) ds .
Γ
Since we are seeking a symmeric solution, which satisfies the condition (154),
it is sufficient to suppose that at first Re q > 0 and Im q < 0. Then choose
the contour Γ to lie initially in Re s > 0, Im s ≥ 0, so that Re sq > 0 and the
Laplace transform is well defined for the allowed values of q. In particular,
since V (s) can be represented by the power series (162), which generates
the asymptotic series (158), which in turn is equivalent to the asymptotic
series (148), we conclude that v0 ∼ vs =
2 us .
The next step is to deform the contour Γ onto the imaginary s-axis. In
this process we will need to deform around the poles at s = i, 2i, 3i, · · · and
collect the (half) residues. Hence we find that (160) becomes, on putting
s = iy on the deformed contour Γ,
∞
iπK
v0 = exp (−iyq) V (iy) i dy + exp (−iq) + · · · , (166)
0 2
where the dots denote terms proportional to exp (−2iq), exp (−3iq), · · · .
The integral is interpreted as a principal value integral at the singulari-
ties at y = 1, 2, 3, · · · . This holds in Re q > 0, Im q < 0. To apply the
symmetry condition (154) that v0 should be real-valued on the imaginary
q-axis, we must now let Re q → 0, and put q = )−iQ, Q > 0 in the ex-
∞
pression (166). From the series (162), V (iy) = i n=0 an (−1)n y 2n is pure
imaginary since the coefficients an are all real-valued. Hence the integral
term is real-valued as required. But the contributions from the poles are
pure imaginary, and hence the expression (166) cannot satisfy the symmetry
condition. The remedy is to note that the term exp (−iq) is exponentially
small in the sector Re q > 0, Im q < 0 and hence is subdominant, so that we
are allowed to add such terms asymptotically to (166).
Thus we replace (166) with
∞
iπK
v0 = exp (−iyq) V (iy) i dy + exp (−iq)
0 2
(167)
ib
+ exp (−iq + iδ) + · · · ,
2
where b, δ are real constants. Now application of the symmetry condition
shows that
b cos δ = −πK . (168)

ib
v0 ∼ exp (−sq) V (s) ds + + exp (−iq + iδ) + · · · , (169)
Γ 2
where b is determined in terms of δ by (168). It remains to bring this

solution back to the real x-axis, for x > 0, using (151), x = iπ/2 +
q.
Here we must also collect a similar contribution from the singularities in
the lower half of the complex x-plane. Thus the full solution consists of a
central core and an exponentially small tail oscillation.
The final result is that, for x > 0 (the case x < 0 follows from the
imposed symmetry),
b π x
u ∼ us + 2
exp (− ) sin ( − δ) , b cos δ = −πK , (170)

2
γ
where us ∼ 2γ 2 sech2 (γx)+O(

2 ) is given by (149). The tail oscillations form
a one-parameter family characterized by the phase shift δ, 0 ≤ δ < π/2.
The minimum amplitude occurs for δ = 0 and then the generalized solitary
wave has the structure shown in Figure 4. This result agrees with Pomeau
108 R. Grimshaw
et al (1988) (for the case δ = 0), and with the numerical solutions of Boyd
(1991). Amick and Toland (1992) have established theoretically that the
nonlocal solutions of the fourth order ordinary differential equation (147)
form a one-parameter family homoclinic to periodic solutions for (147).
5.3 One-sided oscillations

Suppose that instead of the symmetry condition (154) we look for one-
sided solutions such that u(x) → 0 as x → −∞. This choice is motivated
by the observation that for the fifth-order KdV equation (144) the group
velocity cg = 2
−2 > 0 at the resonant wavenumber k =
−2 and hence the
radiation is emitted into x > 0 The same procedure can be followed, and
again the solution for Re q > 0, Im q < 0 is given by the Laplace transform
(160), with the contour Γ now chosen to lie in Re s < 0, Im s > 0 (that is,
to the left of the imaginary axis in Figure 3), which ensures that there are
no oscillations as x → −∞. But now, to find the behaviour as x → ∞, the
contour Γ must be moved across the imaginary s-axis in order to evaluate
the solution when Re q > 0. In doing so the solution collects the residue at
the poles s = i, 2i, 3i, · · · . The residue at s = i is −iπK exp (−iq), exactly
twice the contribution and oppositely-signed from the half residue shown in
(166). Then, bringing the solution back to the real x-axis, in x > 0, we find
that
2πK π x
u ∼ us − 2 exp (− ) sin ( ) . (171)

2

For these one-sided solutions there are no free parameters. A schematic

example of a one-sided oscillation is shown in Figure 7. Further, these
waves are unsteady and slowly decay due to this radiation.
5.4 Higher-order terms

We now return to the inner expansion (155)
∞

v(q) ∼ vn (q)
2n ,
n=0
and examine the effect of the next term v1 , where we recall from (152) that
the matching condition is that v1 → 2γ 3 /3 as q → ∞ in Re q > 0, Imq < 0.
First, we re-examine the tail oscillations and seek a solution of (147) of the
form
u ∼ us + uw , (172)
Figure 7. A schematic plot of a one-sided oscillation.
where us is the outer expansion (148) and uw are the tail oscillations. Sub-
stitution into (147) and linearization about us yields
−cuw + 6us uw + uwxx +

2 uwxxxx = 0 . (173)
When x → ∞, the term 6us uw can be omitted, and we find that
kx
uw ∼ α sin ( − δ) , k 4 − k 2 =
2 c . (174)

for some constants α, δ. Since c = 4γ 2 + 16

2 γ 4 + O(
4 ), we find that
k = 1 + 2
2 γ 2 + O(
4 ).
To obtain the higher-order terms, we put
kx
uw ∼ α(1 +
2 R(x)) sin ( − δ +
φ(x)) . (175)

Substitution into (173) yields
R(x) = 15γ 2 sech2 (γx) + O(

2 ) , φ(x) = −6 tanh (γx) + O(
2 ) . (176)
110 R. Grimshaw
Then, to match with he inner solution, we use x = iπ/2γ +

q (151) to get
iα kπ 6i
uw ∼ exp ( ) exp (−iq + δ)[1 + + O(q −2 )] . (177)
2 2
γ q
Next, we re-examine the inner solution and replace (160) with

v0 ∼ vs + w0 , vs = exp (−sq) V (s)ds , (178)
Γ
where the contour Γ runs from zero to infinity in Re q > 0, Im q < 0, as

before. Substitution into equation (153) and linearization about vs gives
w0qqqq + w0qq + 6vs w0 = 0 . (179)
As q → ∞, the term 6vs w0 can be omitted, and we get that w0 ∝

exp (−iq) in agreement with (167). But a more careful analysis which takes
account of the term 6vs w0 yields
ib 6i
w0 ∼ exp (−iq + iδ){1 + + O(q −2 )} . (180)
2 q
The reaming analysis proceeds as before, with deformation of the contour
Γ onto the imaginary axis, and we again get (168) that b cos δ = −πK. The
expression (180) can now be matched successfully with uw (177) to leading
order in
2 .
Next, consider the term
2 v1 in the expansion (155) for v, which satisfies
the equation
v1qqqq + v1qq + 6v0 v1 − 4γ 2 v0 = 0 , (181)
where we have used the leading order expression (149) for c0 = 4γ 2 . From
the outer expansion (152) for us expressed in terms of q we find that the
matching condition is that
2γ 2
v1 ∼ , as q → ∞ , Re q > 0 , Im q < 0 . (182)
3
But now we see that v1 = 2γ 2 /3 is an exact solution of (181), and hence
there is no need for any Borel summation. But again, we need to insert a
subdominant term, and so write
2γ 2
v1 = + w1 , (183)
3
It is then readily shown that w1 satisfies the same equation (179 as that for
w0 , and hence can be absorbed into the same solution (180) by allowing the
constants b, δ to be expanded in powers of

2 . Matching between uw and
w0 then yields
b kπ
α = 2 exp (− ). (184)

2
γ
Here we recall that b cos δ = −πK, and that the wavenumber k is given by
(174), so that k = 1 + 2
2 γ 2 + O(
4 ), while uw is then given by (175). As
x → ∞, we get
kx
uw ∼ α sin ( − δ − 6
) . (185)

Hence the main effect of the higher-order terms is to replace k = 1 with the
full expression (174) for k.
6 Coupled Korteweg-de Vries Equations

6.1 Internal waves
Generalized solitary waves also occur for interfacial waves when there is a
free surface, and for all internal waves with mode numbers n ≥ 2, see Akylas
and Grimshaw (1992) and Boyd (1998). The underlying reason is that there
is a resonance between a long wave with wavenumber k ≈ 0 and a short wave
with a finite wavenumber k, see the dispersion curves shown in Figure 8. In
the absence of the surface mode, only mode 1 supports a genuine solitary
wave. In the presence of the surface mode, all internal modes support only
generalized solitary waves. When the amplitude of the central core is small,
O(
2 ), the amplitude of the oscillations with wavenumber k is exponentially
small, typically O(exp (−K/
)) where K is a positive constant. Hence they
cannot usually be found by conventional asymptotic expansions, and again
we need exponential asymptotics.
Steady generalized solitary waves are necessarily symmetric. But this
means they cannot be realized physically as then the group velocity of the
small oscillations is the same at both ends, which implies that energy sources
and sinks are needed. In practice, they are generated with a core and small
oscillations only on one side, determined by the group velocity, as shown
for gravity-capillary waves in Figure 7. However, for internal waves, unlike
gravity-capillary waves, the group velocity is less than the phase velocity,
and so the tail oscillations occur on the opposite side to those shown in
Figure 7. An example of an observation of a mode-2 internal wave with a
mode 1-tail is shown in Figure 2 of Akylas and Grimshaw (1992).
The technique we use to find the tail oscillations is again based on ex-
tending the usual asymptotic expansion into the complex plane, and using
Borel summation. It is similar to the techniques used by Pomeau et al
112 R. Grimshaw
Figure 8. Plot of a typical set dispersion curves for internal waves for the
phase speed c = c(k) in terms of wavenumber k: mode 1 (blue), mode 2
(red), mode 3 (green) and the surface mode (violet).
(1988) and Kruskal and Segur (1991). The full system was analyzed us-
ing this approach by Akylas and Grimshaw (1992). Here, for simplicity,
we consider instead two coupled Korteweg-de Vries (KdV) equations, which
can be shown to describe the interaction between two weakly nonlinear long
internal waves whose linear long wave speeds are nearly equal,
1
ut + 6uux + uxxx + (pvxx + quv + rv 2 )x = 0 , (186)
2
1
vt + Δvx + 6vvx + vxxx + λ(puxx + ruv + qu2 )x = 0 . (187)
2
This system is Hamiltonian, and conserves the ”mass” u, v, the “momen-

tum” λu2 + v 2 , and the Hamiltonian. For stability we choose the coupling
parameter λ > 0. Δ is the detuning parameter, proportional to the differ-
ence between the two linear long wave speeds; without loss of generality we
take Δ > 0.
First examine the linear spectrum, for waves of wavenumber k and phase
speed c = c(k),
Δ Δ2 1/2
c= − k2 ± {λp2 k 4 + } . (188)
2 4
If we let the coupling parameter λ → 0 these linear modes uncouple into a
u-mode with spectrum c = −k2 and a v-mode with spectrum c = Δ − k 2 .
This situation persists for λ > 0, and there is a resonance between the long
wave (u-mode, in red) and a short wave (v-mode, in blue), with a resonant
wavenumber k0 = (Δ/1 − λp2 )1/2 provided that λp2 < 1, see Figure 9.
Figure 9. A plot of the dispersion relation (188) for Δ = 1, p = 0.5, λ = 0.2
We seek solutions of the form

u = u(x − ct) , v = v(x − ct) , (189)
so that the coupled KdV system (186, 187) becomes
1
−cu + 3u2 + uxx + pvxx + quv + rv 2 = 0 , (190)
2
1
−cv + Δv + 3v2 + vxx + λ(puxx + ruv + qu2 ) = 0 . (191)
2
The two constants of integration have been set to zero, essentially by im-
posing solitary wave boundary conditions, or better, by translating u, v by
114 R. Grimshaw
constants. Equations (190, 191) form a fourth order ordinary differential

equations system. We shall show that they have symmetric generalized
solitary wave solutions, with co-propagating oscillatory tails of small ampli-
tude. This amplitude will be found using either exponential asymptotics,
or more directly by expanding in λ.
6.2 Outer expansion

First, expand around k = 0 for the long (u-mode) wave. Thus, we
introduce a small parameter
<< 1, and seek a solution as an asymptotic
expansion,
∞
∞
∞

2n 2n
us (
x) =
un , vs (
x) =
vn , c=
2n cn . (192)
n=1 n=1 n=1
Substitution into (190, 191 ) yields

u1 = 2γ 2 sech2 (
γx) , v1 = 0 , c1 = 4γ 2 , (193)
λ
u2 = {(20p2 + q 2 − 8pq)c1 u1 − (q − 6p)(q − 10p)u21 } , (194)
Δ
λ 1
v2 = − {pc1 u1 + (q − 6p)u21 } , (195)
Δ 2
λ
c2 = − p2 c21 . (196)
Δ
This expansion can be continued to all orders in
2 without any oscillatory
tail being detected.
To find the tail oscillations, we observe that un , vn are singular in the
complex plane at x = ±iπ/2
γ, ±3iπ/2
γ, · · · . This motivates us to examine
this singularity by the change of variables
iπ
x= +z, (197)
2
γ
Then as
z → 0, sech2 (
γx) ∼ −1/
2 γ 2 z 2 , and so
2 λ
us ∼ − 2
− (q − 6p)(q − 10p) + · · · + O(
2 ) , (198)
z 2Δz 4
2λ
vs ∼ − 4 (q − 6p) + · · · + O(
2 ) . (199)
Δz
6.3 Inner expansion and Borel summation

Next, we consider the inner problem in which we seek solutions of (190,
191) in the form u = u(z), v = v(z), for which the expressions (198, 199)
form an outer boundary condition. The outcome is just the same system
(190, 191) with x replaced by z. Note that c = O(
2 ), and can be omitted
at the leading order.
We seek a solution of this inner problem as a Laplace transform

[u, v] = exp (−zs)[U (s), V (s)] ds , (200)
Γ
where the contour Γ runs from 0 to ∞ in the half-plane Re(sz) > 0, see
Figure 3 and note that here z replaces q. We seek a power series solution
∞

[U (s), V (s)] = [an , bn ]s2n−1 , (201)
n=1
where a1 = −2 , b1 = 0, a2 = −λ(q −6p)(q −10p)/12Δ, b2 = −λ(q −6p)/3Δ.

In general, substitution of (201) into the Laplace transform (200) generates
the asymptotic series
∞

[u, v] ∼ [αn , βn ]z −2n , [αn , βn ] = (2n − 1)![an , bn ] . (202)
n=1
This agrees with the asymptotic series (198, 199), and in effect the Laplace
transform is a Borel summation of the asymptotic series.
Substitution of the Laplace transform (200) and the series (201) into
the differential equation system (190, 191) yields a recurrence relation for
[an , bn ]. Putting Δ[An , Bn )] = (−k02 )n [an , bn ], we get
(n + 1)(2n + 5) q
An−1 + {p − }Bn−1 = Fn , (203)
(n − 1)(2n − 1) (n − 1)(2n − 1)
rBn−1 + qAn−1
(1 − λp2 )Bn − Bn−1 − λpAn−1 + λ = Gn , (204)
(n − 1)(2n − 1)
where Fn , Gn are quadratic convolution expressions in A2 , · · · An−2 , B2 · · · ,
Bn−2 . As n → ∞, these nonlinear terms can be neglected, and we find that
[An , Bn ] → [−p, 1]K as n → ∞, (205)
where K is a constant whose value depends on λ, p, q, r. It now follows that

the series (201) converges for |s| < k0 , k02 = Δ/(1 − λp2 ). The result (205)
shows that as s → ik0 there is a pole singularity given by
[p, −1]K
[U (s), V (s)] ≈ Δ . (206)
2(s − ik0 )
116 R. Grimshaw
We have now established that the solution in the z-variable is given by

[u, v] = exp (−zs)[U (s), V (s)] ds ,
Γ
where [U (s), V (s)] has a pole singularity at s = ik0 , also at the complex
conjugate point s = −ik0 and at all their harmonics s = ±ink0 , n = 2, 3
etc. Hence the contour Γ should be chosen to avoid the imaginary s-axis,
and to be explicit we choose it to lie in Re s > 0. But we seek a symmetric
solution, which in the z-variable requires that Im [u.v] = 0 when Re z = 0.
But the presence of the pole prevents (200) from satisfying this condition,
and so, as in section 5, we must correct it by adding a subdominant term

ib
[u, v] = exp (−zs)[U (s), V (s)] ds + [p, −1] exp (−ik0 z + iδ). (207)
Γ 2
Here b, δ are real constants, and note that | exp (−ik0 z)| is smaller than of
z −n as z → ∞ in Re z > 0 , Im z < 0, recalling that x = (iπ/2
γ) + z. The
symmetry condition is now applied by bringing the contour Γ onto Re s = 0
and deforming around the pole at s = ik0 .
The outcome is
b cos δ = πK . (208)
which is substituted into

ib
[u, v] = exp (−zs)[U (s), V (s)] ds + [p, −1] exp (−ik0 z + iδ).
Γ 2
The final step is to bring this solution (207) back to the real axis, using x =
(iπ/2
γ) + z. Taking account of the corresponding singularity at s = −ik0 ,
we finally get that
[u, v] ∼ [us , vs ] + bΔ[−p, −1] exp(−πk0 /2
γ) sin(k0 |x| − δ) . (209)
Here we recall (193) that us ∼ 2
2 γ 2 sech2 (γ
x), vs ∼ O(
4 ). This is a two-
parameter family, the parameters being
γ, δ, 0 < δ < π/2. The minimum
tail amplitude occurs at δ = 0. Note that the constant in the exponential
term is determined by the location of the singularity, but the amplitude
needs the exponential asymptotics.
The constant K is determined by the recurrence relations (203, 204).
It is a function of the system parameters λ, p, q, r and in general is found
numerically. But K = 0 for q = 6p (see (193, 194)), and in general it
was found by Grimshaw and Cook (1996) that there are many parameter
combinations where K = 0, see their Figures 1-4. In particular
λ(6p − q)
K≈ as λ → 0. (210)
3Δ
These special values imply that the solitary wave decays to zero at infin-
ity, even although its speed lies inside the linear spectrum, at least in this
asymptotic limit. These are called embedded solitons. They are usually
not stable, but are instead metastable, or are said to exhibit semi-stability.
Nevertheless they are found useful in several applications, such as nonlin-
ear optics and solid state physics. For water waves with surface tension,
generalized solitary waves exist for Bond numbers 0 < B < 1/3, but from
numerical simulations it seems there are no embedded solitons.
These symmetric solitary waves cannot be realized in practice, since they
require an energy source and sink at infinity. Instead, they are replaced
by solitary waves with radiating tails on one side only, determined by the
group velocity. That is, in x > 0 for cg > c, or in x < 0 for cg < c,
where cg is the group velocity at the resonant wavenumber. For the present
case, the linear dispersion relation is (188) and so for the relevant u-mode,
cg = Δ − 3k2 < c = Δ − k 2 . Hence there are no oscillations in x > 0, but
they will appear in x < 0. Thus, in this case for one-sided oscillations, in
x > 0, or more generally in Re z > 0, the solution is completely defined by
the Laplace transform integral (200), with the contour Γ lying in Re s > 0.
Then for x < 0, or Re z < 0, the contour Γ must be moved to Re z < 0
across the axis Re s = 0. In this process the solution collects a contribution
from the pole at s = ik0 , which generates the tail oscillation. The final
outcome is that (209) is replaced by
[u, v] ∼ [us , vs ] − H(−x)2πKΔ[−p, −1] exp(−πk0 /2

γ) sin(k0 x) (211)
where H(·) is the Heaviside function. That is, in effect the phase shift δ = 0,
there are no oscillations in x > 0 and the amplitude in x < 0 is exactly twice
the amplitude of the symmetric solution.
6.4 Weak coupling approximation

In the system (190, 191) suppose that 0 < λ << 1 and expand,
∞
∞

[u, v] ∼ λn [un , vn ] , c∼ λn cn . (212)
n=0 n=0
u0 = 2β 2 sech2 (βx) , v0 = 0 , c0 = 4β 2 . (213)

This leading term is a u-mode solitary wave. Note that in comparison with
the previous expansion (192) β =
γ, but now the amplitude can be order
unity. At the next order
−c0 u1 + 6u0 u1 + u1xx + pv1xx + qu0 v1 − c1 u0 = 0 , (214)

118 R. Grimshaw
q
(Δ − c0 )v1 + v1xx + pu0xx + u20 = 0 . (215)
2
u20
(Δ − c0 )v1 + v1xx = f (x) = −pc0 u0 + (6p − q) . (216)
2
Note that in this limit λ → 0, the resonant wavenumber k0 ≈ (Δ−c0 )1/2 and
takes account of the finite speed of the wave. We must now take c0 < Δ to
get tail oscillations, and for c0 > Δ the expansion yields a genuine solitary
wave. The general solution of (216) is
∞
1
v1 = A sin k0 x + B cos k0 x + f (x ) sin (k0 |x − x |)dx . (217)
2k0 −∞
To determine the constants A, B we impose a symmetry condition on v1 , so

that A = 0, and then
v1 ∼ b1 sin (k0 |x| − δ) as |x| → ∞ , (218)

∞
1
b1 cos δ = L = f (x) cos (k0 x)dx . (219)
2k0 −∞
With v1 known, we can find u1 from (214), and
(Δ − c0 )
u1 ∼ −p b1 sin (k0 |x| − δ) , as |x| → ∞ , (220)
Δ
(Δ − c0 )
[u1 , v1 ] ∼ [−p , 1] b1 sin (k0 |x| − δ) , as |x| → ∞ ,
c0
v1 ∼ b1 sin (k0 |x| − δ) as|x| → ∞ ,
∞
1
b1 cos δ = L = f (x) cos (k0 x)dx ,
2k0 −∞
u20
f (x) = −pc0 u0 + (6p − q) .
2
We find that
∞
6k0 2
L=− {k (q − 6p) + 4β 2 q} sech2 (βx) cos (k0 x)dx . (221)
β2 0 −∞
Then as β =
γ → 0, this reduces to
πk02
L∼ (6p − q) exp (−πk0 /2
γ) , (222)
3
which agrees with the previous result (210) from the exponential asymp-
totics, since L = πK. The one-sided solutions are obtained by setting
δ = 0, and replacing b1 in (218, 220) by 0, 2b1 for x > 0, x < 0.
Bibliography
T. R. Akylas and R. H. J. Grimshaw Solitary internal waves with oscillatory
tails J.Fluid Mech., vol 242, 279-298, 1992.
T. R. Akylas and T-S. Yang On short-scale oscillatory tails of long-wave
disturbances Stud. Appl. Math., vol 94, 1-20, 1995.
C. Amick and J. Toland Solitary waves with surface tension I: trajecto-
ries homoclininic to periodic orbits in four dimensions Arch. Rat. Mech.
Anal., vol 118, 37-69, 1992.
C.M. Bender and S. A. Orzsag, Advanced mathematical methods for scien-
tists and engineers, McGraw-Hill, (ISBN 0-07-066173-1) 1978,
M.V. Berry and C. J. Howls, Hyperasymptotics, Proc. Roy. Soc., vol A 430,
653-668,1990.
M. V. Berry, Asymptotics, superasymptotics, hyperasymptotics, “Asymp-
totics beyond all orders” , ed. H Segur and S Tanveer, Plenum, 1-14,
1992.
J. Boyd Weakly nonlocal solitons for capillary-gravity waves: fifth-degree
Korteweg-de Vries equation Physica D, vol 48, 37-69, 1991.
J. P. Boyd Weakly nonlocal solitary waves and beyond-all-order asymptotics,
Kluwer, Amsterdamr, (ISBN 0-7923-5072-3), 1998.
J.P. Boyd The devils invention: asymptotic, superasymptotic and hyper-
asymptotic series Acta Applicandae Mathematicae, vol 56, 1-98, 1999.
F. Dias and G. Iooss Water waves as a spatial dynamical system Handbook of
Mathematical Fluid Dynamics, ed, S.Friedlander and D.Serre Elsevier,
North Holland, Chapter 10, 443 -499, 2003.
R. Grimshaw The reflection of internal gravity waves from a shear layer
Quart. J. Appl. Math., vol 29, 511-525, 1976.
R. Grimshaw Solitary Waves in Fluids, Advances in Fluid Mechanics, ed.
R. Grimshaw, vol 47, WIT press, UK, Ch. 1 1-17, Ch. 7 159-179, 2007.
R. Grimshaw and P. Cook Solitary waves with oscillatory tails Proceed-
ings of Second International Conference on Hydrodynamics, Hong Kong,
1996, ”Hydrodynamics: Theory and Applications”, Vol. 1, ed. A.T.
Chwang, J.H.W. Lee and D.Y.C. Leung, A.A. Balkema, Rotterdam, 327-
336, 1996.
R. Grimshaw and G. Iooss Solitary waves of a coupled Korteweg-de Vries
system Mathematics and Computers in Simulation, vol 62, 31 - 40. 2003.
120 R. Grimshaw
R. Grimshaw and N. Joshi Weakly nonlocal solitary waves in a singularly

perturbed Korteweg-de Vries equation SIAM J. Math., vol 55, 124-135,
1995.
M. D. Kruskal and H. Segur Asymptotics beyond all orders in a model of
crystal growth Stud. Appl. Math., vol 85, 129181, 1991.
E.N. Lorenz On the existence of a slow manifold J. Atmos. Sci., vol 43,
1547-1557, 1986.
E. N. Lorenz and V. Krishnamurthy On the nonexistence of a slow manifold
J. Atmos. Sci.,, vol 44, 2940-2950, 1987.
R. E. Meyer Gradual Reflection of Short Waves SIAM J. Appl. Math., vol
29, 481-492, 1975.
Y. Pomeau, A. Ramani and B. Grammaticos Structural stability of the
Korteweg-de Vries solitons under a singular perturbation Physica D, vol
31, 127-134, 1988.
J. Vanneste Inertia-gravity wave generation by balanced motion: revisiting
the Lorenz-Krishnamurthy model J. Atmos. Sci., vol 61, 224-23, 2003.
Exponential Asymptotics and Stokes Line
Smoothing for Generalized Solitary Waves
Philippe H. Trinh
University of Oxford, UK
Abstract In another paper of this volume, Grimshaw has demon-

strated how techniques of Borel summation can be used to elucidate
the exponentially small terms that lie hidden beyond all orders of
a divergent asymptotic expansion. Here, we provide an alterna-
tive derivation of the generalized solitary waves of the fifth-order
Korteweg-de Vries equation. We will first optimally truncate the
asymptotic series, and then smooth the Stokes line. Our method
provides an explicit view of the switching-on mechanism, and thus
increased understanding of the Stokes Phenomenon.
1 Introduction
The Stokes Phenomenon describes the puzzling event in which exponen-
tially small terms can suddenly appear or disappear when an asymptotic
expansion is analytically continued across key lines (Stokes lines) in the
Argand plane—“as it were into a mist,” Stokes once remarked in 1902.
Fortunately, much of the inherent vagueness of this phenomenon, as well
as its deep implications for the study of asymptotic approximations has been
examined since Stokes’ time (see Boyd (1999) for a comprehensive review).
In another paper of this volume by Grimshaw—henceforth referred to as
[Grimshaw]—it was shown how Borel summation can be used to reveal
the exponentially small waves found in the fifth-order Korteweg-de Vries
equation (5KdV).
In this review paper, we will show how the methodology outlined in
Olde Daalhuis et al. (1995) and Chapman et al. (1998) can be used as an
alternative treatment of the 5KdV equation. The procedure is as follows:
(1) Expand the solution as a typical asymptotic expansion, (2) find the
behaviour of the late-order terms (n → ∞), and (3) optimally truncate the
expansion and examine the remainder as the Stokes lines are crossed.
The location of the Stokes lines, as well as the details of the Stokes
Phenomenon and resultant exponentials are intrinsically linked to the late-
© CISM, Udine 2010
122 P. Trinh
order terms of the asymptotic approximation—thus, as we proceed through

Steps 1 to 3, we are effectively deriving the beyond-all-orders contributions
by decoding the divergent tails of the expansion. The novelty in this ap-
proach (in contrast to the one shown in [Grimshaw]) is that all the analysis is
done in the (complexified) physical space, rather than in Borel-transformed
space. This provides us with a special vantage point—to see the smooth
switching-on of the exponentially small terms as each Stokes line is crossed
(see Figure 1). Come, let us stare into Stokes’ mist.
u(x)
|S|
/()72) 5,*+72)

672.(6 672.(6
/,1( /,1(
(x)
- 10 -5 5 10

η

x = πi/2γ
672.(6/,1(
(x)

x = −πi/2γ
Figure 1. The analytic continuation of the traditional asymptotic solution

(the classical solitary wave) contains singularities up and down the imagi-
nary axis, with Stokes lines emanating from each of these singularities. By
re-scaling near the singularities and optimally truncating, we will be able to
observe the smooth switching-on of the exponentially small terms (top-left).
2 Generalized Solitary Waves and the 5KdV

We will consider the existence of solutions to the 5KdV equation,
2 uxxxx + uxx + 3u2 − cu = 0 (1)
with u → 0 as x → ±∞. Although the problem is for x ∈ R, it will be

important to consider the effects of allowing u and x to be complex.
Exponential Asymptotics and Stokes Line Smoothing 123
2.1 Initial Asymptotic Analysis and Late Terms

We begin as usual by substituting the asymptotic expansions,
∞
∞

u= 2n un and c= 2n cn
n=0 n=0
into Equation (1). This yields the first two orders as,
u0 = 2γ 2 sech2 (γx) c0 = 4γ 2 (2)

15
u1 = −10γ 2 u0 + u20 c1 = c20 (3)
2
while at O(2n ),
u(n−1)xxxx + unxx + 6u0 un − c0 un + . . . = 0. (4)

Here, the key observation is that there exists singularities in the analytic
continuation of the leading order solution, u0 (x) at x = ±πi/2γ, ±3πi/2γ, . . .
This use of ill-defined approximations in order to represent perfectly well-
defined phenomena is one of the caveats of singular asymptotics, but one
would feverishly hope that a singularity far from the region of interest (in
this case, x ∈ R) has little effect on the approximations!
Unfortunately this is not the case. We see from Equation (4) that at each
order, un is partly determined by differentiating un−1 twice and thus each
additional order adds to the power of the singularities in the early terms.
We would therefore expect the late terms of the asymptotic expansion to
exhibit factorial over power divergence of the form,
Q(z)Γ(2n + γ)
un ∼ , as n → ∞. (5)
[χ(z)]2n+γ
Here, γ is a constant, while Q(z) and χ(z) are functions to be determined.
Substituting this ansatz into Equation (4) yields at leading order,
4 2
dχ dχ
− + = 0, as n → ∞. (6)
dz dz
Now from the above discussion, we would expect that χ = 0 at the relevant
singularities, x = σi for some i; we then conclude that χ = ±1 and thus
without loss of generality, χ = x − σi . In general, un will be a sum of
terms of the form (5), one for each singularity. However along the real
axis, the behaviour of un will be dominated by those singularities closest to
124 P. Trinh
the axis and thus we need only concern ourselves with the singularities at
x = ±σ ≡ ±iπ/2γ. Finally, at next order as n → ∞, we find that Q(z) = Λ,
a constant.
The determination of γ, Λ, and in fact, the Stokes line smoothing in
the next section will require an analysis near each of the two singularities,
x = ±σ; for brevity, we will henceforth focus on the singularity at x = σ in
the upper-half plane.
First, since by Equation (2), u0 ∼ −2/(x − σ)2 as x → σ, we must
require that γ = 2. Second, in order to determine the final constant Λ, we
need to re-scale near the singularity x = σ, express the leading-order inner
solution as a power series (in inner coordinates) and match with the outer
solutions. In the end, however, Λ is determined by the numerical solution
of a canonical inner problem. As was shown in [Grimshaw], Λ ≈ −19.97.
Finally, let us discuss the significance of χ. Following Dingle (1973), we
expect there to be a Stokes line wherever un and un+1 have the same phase
as n → ∞, or in this case where,

−χ2 = 0 and −χ2 ≥ 0. (7)
Thus there exist Stokes lines from x = πi/2γ down the imaginary axis and
from x = −πi/2γ up the imaginary axis (as illustrated in Figure 1). In
the next section, we will optimally truncate the asymptotic expansion and
examine the switching-on of exponentially small terms as these two Stokes
lines are crossed.
2.2 Optimal Truncation and Stokes Smoothing

By now we have entirely determined the late terms of the asymptotic ex-
pansion. In order to identify the exponentially small waves, we truncate the
expansion and study its remainder,

N −1
u= 2n un + RN (x).
n=0
Substitution into Equation (1) yields the equation
2 R N

+ RN + 6u0 RN − c0 RN + . . . ∼ 2N uN , (8)
which, using Stirling’s formula, we can write the right-hand side as

Exponential Asymptotics and Stokes Line Smoothing 125
Λ(−1)N Γ(2N + γ + 2)χ

2N uN ∼ 2N
χ2N +γ+2
N
*√ +
2N Λ(−1) 2πe−(2N +γ+2) (2N + γ + 2)2N +γ+3/2
∼ (9)
χ2N +γ+2
We can see now that the remainder is only algebraically small unless N ∼
|χ|/2 (where the ratio of consecutive terms are equal) and thus we set
N = r/2 + ρ where ρ is bounded as → 0.
Although there are four homogeneous solutions to Equation (8) as → 0,
we will show that one in particular,
RN (x) ∼ S(x)e−i(x−σ)/ (10)

is switched on as the Stokes line is crossed. We will call the function S(x)
the Stokes multiplier, and we expect it to vary smoothly from one constant
to another across the Stokes line. We write
d ie−iθ d
χ = x − σ = reiθ and =− , (11)
dx r dθ
where now, since N is fixed (and thus also the modulus, r), we are only
interested in the “fast” variation in θ across the Stokes line. Then using
Equations (9) and (11) in (8) gives
√ r/+2ρ+γ+2 r/+2ρ
dS Λ rπ iθ
∼√ e−r/ eire / e−iθ e−πi/2 eiθ
dθ 2γ+1/2
√ ,
Λ rπ r iθ πi
=√ × exp − 1 − ie + iθ +
2γ+1/2 2
- π .
+i −2ρ θ + − θ(γ + 1) (12)
2
From the terms within the curly braces, we see that the change in S is
exponentially small, √except near the Stokes line θ = −π/2. Here, we will
re-scale θ = −π/2 + η and integrate Equation (12) from left (η → −∞)
to right to show that
√ √rη
Λ π πi(γ+1)/2 2
S ∼ const + √ e e−s /2 ds. (13)
2γ
−∞
This integral (the error function) precisely illustrates the smoothing of the
Stokes line in Figure 1. Thus the jump in the Stokes multiplier and conse-
quently, the remainder is
126 P. Trinh
Λπ 3πi/2 Λπ
S ∼ e =⇒ R N ∼ 2 e3πi/2 e−i(x−σ)/ . (14)
Stokes γ Stokes
We must remember that the analysis must be repeated for analytic continua-
tion into the lower-half x-plane and thus near the singularity at x = −πi/2γ.
The result is another exponentially small contribution which is the complex
conjugate of Equation (14) and thus along the real axis, the sum of contri-
butions from crossing the pair of Stokes lines is,
2Λπ −π/2γ
uexp ∼ − e sin (x/) . (15)
2
Let us recap our analysis: (1) The singular nature of the 5KdV equation
produces singularities in the early terms, (2) As more and more terms are
taken, the effects of the singularities grow, eventually producing factorial
over power divergence in the late terms, (3) Stokes lines emerge from each
of the singularities, and (4) By optimally truncation and examining the
jump in the remainder as the Stokes lines are crossed, we see the Stokes
Phenomenon and thus the appearance of exponentially small terms.
So finally, we are ready to answer the original question: Do there exist
classical solitary wave solutions of the 5KdV equation? No. For suppose
that we did impose the condition that only the base (non-oscillatory) asymp-
totic solution applies at x = −∞. Then as we pass through x = 0, the term
in Equation (15) necessarily switches on and u ∼ u0 + uexp for x > 0.
We have thus passed through Stokes’ mist and subsequently, realized
that there do not exist classical solitary wave solutions of the 5KdV.
Bibliography
J.P. Boyd. The Devil’s invention: Asymptotics, superasymptotics and hy-
perasymptotics. Acta Applicandae, 56:1–98, 1999.
S.J. Chapman, J.R. King, and K.L. Adams. Exponential asymptotics and
Stokes lines in nonlinear ordinary differential equations. Proc. R. Soc.
Lond. A, 454:2733–2755, 1998.
R.B. Dingle. Asymptotic Expansions: Their Derivation and Interpretation.
Academic Press, London, 1973.
A.B. Olde Daalhuis, S.J. Chapman, J.R. King, J.R. Ockendon, and R.H.
Tew. Stokes Phenomenon and matched asymptotic expansions. SIAM
J. Appl. Math., 55(6):1469–1483, 1995.
G.G. Stokes. On the discontinuity of arbitrary constants that appear as
multiples of semi-convergent series. Acta. Math. Stockholm, 26:393–397,
1902.
Multiple scales methods in meteorology
Rupert Klein ‡ , Stefan Vater ‡ , Eileen Paeschke ‡ ,

and Daniel Ruprecht ‡
‡
FB Mathematik & Informatik, Freie Universität Berlin, Germany
Abstract With emphasis on meteorological applications, we discuss

here the fluid dynamical fundamental governing equations, their
nondimensionalization including the identification of key nondimen-
sional parameters, and a general approach to meteorological mod-
elling based on multiple scales asymptotics.
1 Overview
In Chapter 2 we will derive the fluid mechanical conservation laws. We
explore the basic principles considering “pure” fluid mechanics, i.e., we ne-
glect the influences of gravity, Earth’s rotation (Coriolis force), molecular
transport, and of the so called “diabatic effects”. The latter subsume all
processes that involve external energy supply by radiation or conversion
of energy due to condensation, chemical reactions, etc.. Gravity and the
Coriolis force will be included in subsequet sections. The chapter concludes
with a summary of the governing equations, now extended to also include
a general set of species transport equations. These will be important, e.g.,
in describing (atmospheric) chemistry or moist processes.
Chapter 3 introduces the technique of multiple scales asymptotics using
the (almost trivial) example of a linear oscillator. After deriving analyti-
cal solutions, we will focus on a situation which, in many ways, resembles
situations arising frequently in geophysical problems: a slow background
motion caused by an external force is accompanied by rapid oscillations
around it, with the oscillation amplitudes generally not being small. To
give some meaning to the notions of “smallness” and “rapidity”, we will
first nondimensionalize the oscillator equations and identify small parame-
ters that lend themselves for comparison. By means of single and multiple
scales analyses we will then try to derive simplified approximate solutions
that become more and more accurate as the small parameters vanish.
One important aim of theoretical meteorology is the development of sim-
plified model equations that describe the large variety of scale-dependent
© CISM, Udine 2010
128 R. Klein et al.
phenomena observed in atmospheric flows. Chapter 4 summarizes the ba-

sic scaling arguments that justify a unified approach to the derivation of
such models based on multiple scales asymptotic techniques. We note that
Keller and Ting (1951) already anticipated the foundations of this approach
in an internal report of the Institute for Mathematics and Mechanics of
New York University. In particular, Chapter 4 non-dimensionalization to
the equations of compressible flows on a rotating sphere as a first step in
building the unified multiscale modelling framework. The subsequent steps
are the introduction of a quite generally applicable set of distinguished lim-
its, and multiple-scales asymptotics. For simplicity, diabatic effects, such as
radiation, water phase transitions, or turbulent transport are represented
as lumped terms in the governing equations to be specified later. For ex-
tensions see Klein and Majda (2006).
Chapter 5 employs the general asymptotics-based approach to rederive
the classical quasi-geostrophic model, see Pedlosky (1987).
The reader may want to consult Klein (2010) for further references.
2 Fluid mechanical conservation laws

In this chapter we will derive the fluid mechanical conservation laws. In sec-
tion 2.1 we explore the basic principles considering “pure” fluid mechanics,
i.e., we neglect the influences of gravity, Earth’s rotation (Coriolis force),
molecular transport, and of the so called “diabatic effects”. The latter
subsume all processes that involve external energy supply by radiation or
conversion of energy due to condensation, chemical reactions, etc.. Gravity
and the Coriolis force will be included later in Sections 2.3 and 2.4. Section
2.6 provides a summary of the governing equations, now extended to also in-
clude a general set of species transport equations. These will be important,
e.g., in describing (atmospheric) chemistry or moist processes.
Remark: In the present context, some quantity, say U , is conserved if the
total content of U within a given, fixed control volume in space can change
in time only by exchange of U across the control volume’s interface.
2.1 Pure fluid dynamics

Mass conservation During the motion of a mass parcel its mass is con-
served while, in general, the parcel’s volume can change. The change of
density (mass per unit volume) caused by this change of volume is expressed
in the law of mass conservation. The mass M of a fixed control volume Ω
Multiple Scales Methods in Meteorology 129
∂Ω
Δt v
Ω
Δσ
Ω n
v Δt (v · n)
n
Figure 1. Change of a volume’s mass with time
at time t can be expressed as an integral over the density,

M (t; Ω) = (t, x) dV . (1)
Ω
This mass will change during a time interval Δt if mass parcels cross in-
terface ∂Ω, being carried along by the flow velocity, v (see Fig. 1). The
change of mass, ΔM , associated with the passage of parcels across a con-
trol surface segment Δσ ⊂ ∂Ω, is then equal to − (Δt v) · n Δσ, where n is
the outward pointing normal on Δσ. The scalar multiplication v · n selects
the component of (Δt v) perpendicular to Δσ as the one relevant for mass
transport across the surface element. By summation (integration) along the
entire boundary of the control volume, and for time increments covering a
finite interval t ∈ [t1 , t2 ] we find
t2
M (t2 ; Ω) − M (t1 ; Ω) = − (v) · n dσ dt . (2)
t1 ∂Ω
This is the most general formulation of the law of mass conservation which
holds for arbitrary control volumes for which the integrals in (1), (2) are
meaningfully defined. Notice that the above definitions merely require suit-
able integrability for the mass and momentum densities, and v. These
quantities need not be differentiable in either space or time for the mass
balance in (2) to make sense!
If, however, we may assume differentiability of M (t; Ω) w.r.t. time, t, we
may let (t2 − t1 ) → 0 to find

dM
=− (v) · n dσ . (3)
dt ∂Ω
130 R. Klein et al.
If, in addition, v satisfies

the conditions
of Gauß’ integral theorem, (see
appendix 5.6), then ∂Ω (v) · n dσ = Ω ∇ · (v) dV , and

t + ∇ · (v) dV = 0 for arbitrary Gauß domains Ω. (4)
Ω
This equation can hold, for continuously differentiable fields , v and for ar-
bitrary control volumes Ω, only if pointwise the following partial differential
equation is satisfied:
t + ∇ · (v) = 0 . (5)
Adopting this differential form restricts solutions to the class of continu-
ously differentiable fields. Yet, in practice one uses (5) as a short-hand
for (2), thereby implying that whereever t and ∇ · (v) are singular, their
spacio-temporal integrals remain well defined. For discussions of such weak
solutions of conservation laws see, e.g., LeVeque (1990) and Kröner (1997);
for a measure theoretical approach to conservation laws see Temam and
Miranville (2000).
General conservation laws The considerations of the last section lead

us to the following general formal pattern of a conservation law: Let U de-
note an extensive conserved quantity. For extensive quantities, their “total
amount”, U (t, Ω), is well defined for arbitrary control volumes, Ω, that are
Gauß domains, and they are additive in that
U (t; [Ω1 ∪ Ω2 ]) = U (t; Ω1 ) + U (t; Ω2 ) ∀ Ω1 , Ω2 : Ω1 ∩ Ω2 = ∅ . (6)
Let u(t, x) denote the density field associated with U , so that

U (t; Ω) = u(t, x) dV , (7)
Ω
and f its flux density. Then conservation of U in time is expressed by the

Integral Conservation Law
t2
U (t2 ; Ω) − U (t1 ; Ω) = − f · n dσ dt . (8)
t1 ∂Ω
If, in addition, u and f are sufficiently smooth, then they satisfy the
Partial Differential Equation in Conservation Form
ut + ∇ · f = 0 . (9)
Remark: The integral form of the conservation law in (8) is the most general
basis for the formulation of numerical methods for problems in continuum
mechanics because, by construction, it allows for the correct representation
of non-smooth, e.g., discontinuous, solutions. See, e.g., LeVeque (1990),
Kröner (1997).
Remark: Partial differential equations which, in addition to divergence
terms, include other expressions that have no equivalent divergence form
cannot be cast in the more general integral form (8). Such equations do not
describe conservation in the present sense.
Energy conservation The conservation of energy follows the same pat-

tern just described, yet here the conserved quantity can not only be ex-
changed by the motion of fluid parcels. Rather, the energy contained in a
control volume is also changed by the work done by the pressure and other
stresses when the fluid moves across or along the boundary of the consid-
ered control volume, or when heat is added through thermal conduction,
radiation, and the like. Further effects, such as those due to changes in
potential energy in the Earth’s gravity field will be discussed in the next
chapter. The energy contained in some control volume Ω is

E(t) = e(t, x) dV , (10)
Ω
where e is the total energy per unit mass, or specific total energy, and is
again the mass density.
Energy is transported by the motion of fluid parcels in analogy with the
flux of mass considered in section 2.1. The associated contribution to the
energy flux density is e v. In addition, there are forces acting within a
fluid between adjacent fluid parcels. Those forces are represented by means
of a second order tensor field (p id + τ ), where p is the thermodynamic
pressure, and id, τ (t, x) ∈ IR3×3 are the unit tensor and the viscous stress
tensor, respectively. The interpretation of this tensor, (p id + τ ), and its
two contributions to the energy flux is as follows:
Consider the boundary, ∂Ω, of a control volume (or some similar surface
embedded in the flow domain). At any location x ∈ ∂Ω the vector (p id +
τ ) · n, with n the outward pointing normal on ∂Ω, denotes the force per
unit area, i.e., the stress, which the fluid within the control volume exerts
onto the fluid outside. If the fluid is in motion, with flow velocity v(t, x),
then v · (p id + τ ) · n is the work per unit time and unit area done by the
fluid inside the control volume on the fluid outside. This is the second
contribution to the energy flux density to be considered here. It consists
132 R. Klein et al.
of (i) the work done by the thermodynamic pressure forces, p v · n, and (ii)
the work done by the viscous stresses, v · τ · n.
We know from experience that two bodies of finite mass and different
temperature tend to exchange thermal energy so as to eventually approach
states of equal temperature. Let us denote the associated energy flux per
unit area by j.
Combining the three effects just discussed we obtain the energy conser-
vation law, written here in its differential form as
(e)t + ∇ · ([e + p] v + v · τ + j) = 0 . (11)
In much of the subsequent discussions we will neglect the terms, (v · τ + j),
which are associated with molecular transport processes, for simplicity of
exposition.
Momentum conservation The momentum of the fluid contained in a

control volume is defined as

I(t) = v(t, x) dV , (12)
Ω
so that v is the momentum density. Fluxes of momentum arise again

through advection, i.e., through transport by the fluid motion, and the
associated flux density across a surface with unit normal n is v(v · n).
Newton’s law of motion then states that the forces acting on some finite
mass equal the rate of change of its momentum. In the present continuum
mechanics we have seen that (p id + τ ) · n is the force per unit area which
some mass of fluid within our control volume Ω exerts onto the fluid outside
it (when n is the outward-pointing normal unit vector). Thus, (p id + τ ) · n
represents a flux of momentum from the control volume to its environment,
and the momentum conservation law reads
(v)t + ∇ · (v ◦ v + p id + τ ) = 0 , (13)
where ◦ denotes the tensorial product.
Remark: The divergence of (p id) equals the pressure gradient. In cartesian
co-ordinates we have
⎛ ⎞
∂ ∂ ∂ p 0 0
∂p ∂p ∂p

∇ · (p id) = , , · ⎝ 0 p 0 ⎠ = , , . (14)
∂x ∂y ∂z ∂x ∂y ∂z
0 0 p
In the notation of co-ordinate-free tensor analysis,
∇ · (p id) = ∇p . (15)
2.2 Equations of state

The system of mass, momentum, and energy conservation laws, (5),
(13), and (11), is not closed, as the pressure, p, the stress tensor, τ , and
the heat flux density, j, have not yet been related to the primary variables,
(, v, e). For the present purposes it suffices to take the model of an
ideal gas with constant specific heat capacities, endowed with Newtonian
friction and Fourier-type heat conduction as an example. For this case, we
introduce the temperature,
p
T = (16)
R
where R = R∗ /M , R∗ = 8.3141 J mol−1 K−1 is the idal gas constant, and
M the gas’ molecular weight. Then we express the total energy density, e,
as the sum of the internal (or thermal) and the kinetic energy via
v 2
e = (eth + ekin ) = cv T + . (17)
2
Here the coefficient cv is known as the specific heat capacity at constant
volume and it is assumed constant below.
From the ideal gas law (16) we obtain
cv 1 p 1
e = p + v 2 = + v 2 , (18)
R 2 γ−1 2
c
where γ ≡ 1 + cRv = cvp is the isentropic exponent of the gas (and cp is its
heat capacity at constant pressure). For atmospheric air a good estimate is
γ = 1.4 = const. with variations due to admixtures of water vapor and other
trace gases being of the order of a few percent at most (see the subsequent
remarks).
Remark: Generalization of these constitutive laws for mixtures of ideal gases
with molecular weights (M1 , M2 , . . . Mn ) and mass fractions (Y1 , Y2 , . . . Yn )
maintains (16) and (17) but replaces the gas constant and specific heat
capacity at constant volume with

n
(R, cv ) = Yi (Ri , cv,i ) .
i
See also section 2.6.

Remark: For an ideal gas with non-constant specific heat capacities,
T
eth = cv (T ) dT
0
134 R. Klein et al.
with the specific heat capacity at constant volume, cv (T ), being a known

function of temperature. The ideal gas law from (16) is maintained in this
case with R remaining independent of temperature.
For the stress tensor τ , we assume Newton’s law

τ = −μ ∇v + (∇v)T − μ̂ (∇ · v) id , (19)
where the dynamic viscosity, μ, and the dimensionless coefficient of bulk

viscosity, μ̂, depend on the fluid considered, and on the thermodynamic state
through the temperature (quite generally) and the pressure (for some fluids).
For air, μ ≈ 1.7 × 10−5 kg m−1 s−1 at typical atmospheric conditions, Gill
(1982). The coefficient μ̂ is very difficult to measure experimentally, and for
lack of more precise information one often assumes μ̂ = 23 , which simplifies
the equations to some extent by eliminiating trace(τ ).
For the thermal energy flux density we adopt Fourier’s law of heat con-
duction, so that
j = −λ · ∇T . (20)
Here λ is the thermal conductivity which again depends on the medium

as well as possibly on temperature and pressure. For instance, in water
we have λH2 O = 0.6 Wm−1 K−1 , while in air λair = 0.023 Wm−1 K−1 , Gill
(1982)).
The constitutive laws (16)–(20), when added to the conservation laws
for mass, momentum, and energy in (5), (13), and (11), yield the desired
closed set of partial differential equations for the flow of an ideal gas.
2.3 The influence of gravity

In the previous section we neglected Earth’s gravity. Gravity exerts a
bulk force, which cannot be directly expressed as a flux divergence (although
in a broad range of applications in meteorology it can!). The change of
momentum caused by this bulk force is proportional to the fluid density,
, and directed oppositely to some unit vector ĝ that points away from
the Earth’s center of mass. In the present notes, we will restrict to flows
covering sufficiently small domains, so that we may safely assume ĝ ≡ k,
where k is the “vertical” unit vector, perpendicular to a suitably chosen
tangent plane to the Earth’s surface. The factor of proportionality is the
acceleration of gravity,
m
g ≈ 9.81 2 , (21)
s
which may be considered constant here.
Remark: Because of the shape of the earth, g actually varies at sea level
around ±0.3 percent in north-south direction and around 0.3 percent with
a change of height of 10 km Gill (1982).
To account for the influence of gravity, we must endow the momentum
balance in (13) with a source term,
(v)t + ∇ · (v ◦ v + p id + τ ) = − g k . (22)
In the energy balance, (11), we must account for the potential energy as-
sociated with the position of the fluid mass in the Earths gravity field. This
is done by extending the constitutive law from (17), to include a potential
energy term, viz.,
v 2
e = (eth + ekin + epot ) = cv T + + Φ(x) . (23)
2
Here Φ(x) is the Earth’s geopotential, which in the present setting (tangent
plane approximation) we approximate by
Φ(x) = gz (24)
with z denoting height above sea level.

Remark: For the present setting of flows in a tangential plane with g,
k ≡ const., even the momentum equation is effectively in conservative form,
too. By introducing the hydrostatic pressure Phy , defined by
∞
∂Phy
Phy (z) = g (z ) dz and = −g (25)
z ∂z
we rewrite the right hand side of the momentum equation to become

⎛ ⎞
0 0 0
∂Phy
−( g k) = (0, 0, − g) = 0, 0, = ∇· ⎝ 0 0 0 ⎠, (26)
∂z
0 0 Phy
or
− g k = −∇ · Πhy . (27)
We thus obtain the momentum equation in conservative form
(v)t + ∇ · (v ◦ v + p id + τ + Πhy ) = 0 . (28)

136 R. Klein et al.
ΔX b
ΔΘ
X b (t + Δt)
|X b | sin((Ω, X b ))
X b (t)
(Ω, X b )
Figure 2. X b at times t and t + Δt
2.4 The effects of Earth’s rotation

Rotating frame of reference Up to now we have considered a non-
accelerating coordinate system and neglected the effects of Earth’s rotation.
Physical phenomena are independent of the choice of a coordinate system,
but their description depends on the observer and, in particular, on his
choice of a coordinate system. For obvious reasons, we are interested in
observers that follow Earth’s rotation.
Consider some point on the Earth’s surface with position vector X b =
X b (t) in an absolute, intertial frame of reference. Because of Earth’s rota-
tion, X b is rotating with angular velocity Ω. In a small time interval Δt
the vector X b turns by an angle Δθ = |Ω|Δt, where |Ω| is the absolute
value of Ω (figure 2, see also Pedlosky (1987)). This small change of X b is
described by

X b (t + Δt) − X b (t) ≡ ΔX b = n |X b | sin((Ω, X b )) Δθ + O (Δθ)2 (29)
with the unit vector

Ω × Xb
n= , (30)
|Ω × X b |
pointing in direction of the change of X b (perpendicular to X b and Ω). As
Δt → 0 we find
ΔX b dX b dθ Ω × X b
lim = = |X b | sin((Ω, X b )) , (31)
Δt→0 Δt dt dt |Ω × X b |
and, using |Ω × X b | = |Ω| |X b | sin((Ω, X b )),
Ẋ b = Ω × X b . (32)
Both observers see the same vector X b but their perception of how it
changes is completely different.
Remark: The length of X b is constant, independent of the used coordinate
system. Because of X b ⊥(Ω × X b ) it follows that
d|X b |2 dX b
= 2 Xb · = 2 X b · (Ω × X b ) = 0 . (33)
dt dt
To describe the time-dependent vector X b (t) in a non-rotating coor-

dinate system, the vector X b (t) is split into a vector that describes the
distance of the circle of latitude on which X b (t) moves to the equator and
two other vectors that define the position on this circle of latitude. Let
eΩ be the unit vector in direction of the earth rotation vector Ω, then the
vector describing its circle of latitude is
cos[(Ω, X b (0))] |X b (0)|eΩ = (eΩ · X b (0))eΩ . (34)
The position of X b (t) on the circle of latitude can be determined by the

linear combination of a distance vector X b (0), expressing the distance to
the axis of rotation and the vector perpendicular to it. The distance vector
can be computed by
X b (0) − (eΩ · X b (0))eΩ = (id − eΩ ◦ eΩ )X b (0) . (35)
The vector perpendicular to it with same length is (because of eΩ ×eΩ = 0)
[X b (0) − (eΩ · X b (0))eΩ ] × eΩ = X b (0) × eΩ . (36)
Thus we get
X b (t) = (eΩ · X b (0))eΩ +cos(|Ω|t)(id−eΩ ◦eΩ )X b (0)+sin(|Ω|t)(X b (0)×eΩ ) .

(37)
138 R. Klein et al.
Governing equations in a rotating frame of reference With help

of some tedious but straightforward computations we transform our con-
servation laws. In this section we use cartesian coordinates throughout,
representing vectors as 3-columns, tensors as 3×3-matrices, etc.. In par-
ticular, rotation by the Earth rotation vector Ω is represented by matrix-
multiplication with the skewsymmetric matrix Ω built from the components
(Ωx , Ωy , Ωz )t of the rotation vector:
⎛ ⎞⎛ ⎞ ⎛ ⎞
0 −Ωz Ωy ux Ωy uz − Ωz uy
Ω × u = ⎝ Ωz 0 −Ωx ⎠ ⎝ uy ⎠ = ⎝ Ωz ux − Ωx uz ⎠ . (38)
−Ωy Ωx 0 uz Ω x u y − Ω y ux
Here vectors and tensors are represented by their coordinate tupels and
matrices as indicated, and u is the column of cartesian coordinates of u.
With this notation, the mass balance in the intertial frame reads
t + ∇ · (v) = 0 . (39)
We transform the time derivative (∂/∂t) according to (226) as

∂(x, t) ∂ ˜ ˜ )T ,
= + x̃T Ω(∇˜ (40)
∂t ∂ t̃
and the divergence term, following (228), as

∇ · (v) = ∇ · [v rel + Ω x]
= (∇)v rel + (∇)(Ω x) + ∇ · v rel + ∇ · (Ω x) (41)
˜ ˜)ṽ rel + (∇˜

= (∇ ˜ · ṽ rel .
˜ )(Ω x̃) + ˜∇
˜ )T we have
For the scalar quantity x̃T Ω(∇˜
T
˜ )T = x̃T Ω(∇˜
x̃T Ω(∇˜ ˜ )T ˜ ˜)ΩT x̃
= (∇ (42)
and, as Ω is skew symmetric, with ΩT = −Ω = −Ω̃, the mass conservation

law in the rotating frame becomes
∂ ˜
t + ∇ · (v) = ˜ )ṽ rel + ˜∇
+ (∇˜ ˜ · ṽ rel + (∇˜ ˜ )(−Ω)x̃
˜ )Ω x̃ + (∇˜
∂ t̃
∂ ˜ ˜
= ṽ rel ) .
+ ∇ · (˜
∂ t̃
(43)
Thus the equation for mass conservation is invariant under the present co-
ordinate transformation into a rotating frame.
The momentum balance, in the present notation, reads
T
(v)t + ∇ · ( v v T ) + ∇p = −∇Φ − ∇ · τ . (44)
As only the first two terms do change under coordinate transformations
(the reader may want to verify this), we can neglect the others for the time
being. For the first two terms we use the product rule to obtain
(v)t = t v + v t and (45)
T
∇ · (v v T ) = (∇ · (v))v + (v T ∇T )v . (46)
Using mass conservation, we further have
T
(v)t + ∇ · (v v T ) = (t + ∇ · (v)) v + v t + (v T ∇T )v

= 0 + vt + (v T ∇T )v ,
(47)
and the transformation of these terms into the rotating coordinate system
yields

∂ ṽrel rel rel rel ˜ = −∇Φ− ˜ ˜ · τ̃ .
+ ((ṽ ) ∇ )ṽ + 2 (Ω ṽ ) + Ω(Ω x) +∇p
T T
∇
∂t
(48)
rel
Physically, the term 2 (Ω ṽ ) represents the Coriolis acceleration.
The term Ω(Ω x) expresses the centripetal acceleration due to the rota-
tion of the reference frame. As this term can be written as the density times
the gradient of a potential (namely which one?), it is often combined with
gravity term, thereby inducing a modified effective gravitational potential.
The order of magnitude of the centripetal inertia may be estimated by
m
(10−4 s−1 )2 · 6 · 106 m ≈ 10−2 2 . (49)
s
In contrast, the acceleration of gravity is of the order of g ≈ 10 m s−2 , and
the centripetal acceleration may be neglected for most practical purposes in
meteorology. Notice, however, that this may be a different issue in climate
models, because in long-time simulations, even small effects can accumulate
and eventually induce leading-order changes.
Like the equation for the mass conservation, the equation of energy con-
servation does not change when introducing a rotating coordinate system.
We leave the verification of this claim to the reader.
140 R. Klein et al.
2.5 Adiabatic motions and the concept of potential temperature

Consider a flow field that is sufficiently smooth so that the differential
form of the mass, momentum, and energy balances are valid, i.e.,
t + ∇ · (v) = 0
(v)t + ∇ · (v ◦ v + p id) + 2Ω × v = −∇ · τ − (g∇Φ + Ω × Ω × x)
(e)t + ∇ · ([e + p]v) = −∇ · (v · τ + j)

(50)
and they are closed by the equation of state connecting the pressure p with
the conserved quantities, (, v, e),
p v2
e = + + (Φ + ΦΩ ) , (51)
γ−1 2
with ΦΩ defined such that ∇ΦΩ = Ω × Ω × x.
In these equations, the (molecular) stress tensor, τ , and heat flux density
vector, j, represent transport processes of momentum and energy within the
fluid that occur without any mass being exchanged. When these terms are
absent, control volumes within the fluid are restricted to exchange mechan-
ical energy only, either by advection of total energy, as represented by the
flux term ev, or by mechanical work, as represented by the energy flux pv.
Flows of this kind are called adiabatic.
Remark: When considering mixtures of different fluid species, we will also
require diffusion, i.e., the molecular-level transport of the individual species
relative to the mean flow, to be zero for the notion of an adiabatic process
to apply.
For adiabatic, smooth flows, the conservation laws may be linearly com-
bined to yield an evolution equation for the pressure field. Together with
the equation for density, it reads
t + v · ∇ + ∇ · v = 0
(52)
pt + v · ∇p + γp∇ · v = 0.
By eliminating the velocity divergence, we obtain
1 D 1 Dp
− = 0, (53)
Dt γp Dt
where
D
= (∂t + v · ∇) (54)
Dt
is the time derivative which an observer moving with the fluid (at velocity
v) will measure.
The above is equivalent to

DΘ (p/p0 )1/γ
=0 where Θ = T0 (55)
Dt /0
is the potential temperature. Here 0 , T0 , and p0 = 0 RT0 are some arbi-

trary, yet for any given flow fixed, reference values. They are introduced
to render the expression of taking a non-integer power of some quantity
mathematically meaningful by first non-dimensionalizing it.
If we pick 0 , T0 , p0 to denote the standard reference values of typical
conditions at sea level, i.e., p0 = 105 N/m2 , T0 = 273 K, then Θ has a neat
interpretation: Take any parcel of air at thermodynamic conditions p, , and
let it undergo an adiabatic process that brings its pressure up or down to
the reference pressure p0 . Then Θ is the temperature the parcel will acquire
when that process is finished.
Remark: The potential temperature is closely related to thermodynamic
entropy. In the present case, one can directly be expressed as a function of
the other. Thus, (55) is where our coefficient γ received its name isentropic
exponent from.
Remark: Mixtures of different fluid species may undergo changes of com-
position when the pressure and temperature adjust during an adiabatic
process. In that case, the notion of a potential temperature with exactly
the same physical meaning as given above remains valid. Yet the formula
in (55) becomes more involved.
2.6 Summarizing the equations

In the sequel, we will study (i) the conservation of mass, momentum,
and energy, extented by a set of species balance equations. This extension
will allows us to later account for moist processes.
t + ∇ · (v) = 0 (56a)
(v )t + ∇ · (v ◦ v ) + 2(Ω × v) + ∇ p = −(∇ · τ ) (56b)
(w)t + ∇ · (vw) + 2(Ω × v)⊥ + pz = −(∇ · τ )⊥ − g (56c)

⎛ nsp
⎞

(e)t + ∇ · (v [e + p]) = −∇ · ⎝j + v · τ + hi di ⎠ . (56d)
i=1
142 R. Klein et al.
(Yi )t + ∇ · (Yi v) = ωi − ∇ · di . (i = 1, . . . , nsp − 1) (56e)

Here we have split the momentum equations into their horizontal “ ”
and vertical “⊥ ” components anticipating that the vertical direction in me-
teorological applications usually plays a mathematically special role.
Also, we have added a set of transport equations for nsp energy-carrying
(chemical) species. The composition of the gas is described by the species’
mass fractions Yi . Potential species conversion processes, such as chemical
reactions or the formation of cloud water from water vapor are represented
by the source terms ωi . The flux terms di cover the diffusion of species
relative to the mean flow. (Why does the counter in the last equation run
up to nsp − 1 only?)
The system is closed by adding the equations of state
T
nspec

1 2
e = cv (T )dT + Φ + v + Yi Qi , and p = RT (57)
0 2 i=1
where
nspec nspec

cv (T ) = Yi cv,i (T ) , R= Yi R i , (58)
i=1 i=1
and T
hi = cp,i (T )dT + Qi . (59)
0
Here the constants Ri , Qi are the gas constants and formation enthalpies
of the species, cv,i , cp,i are their specific heat capacities at constant volume
and at constant pressure, respectively.
In addition, we have to adopt appropriate expressions for the stress ten-
sor, heat flux density, and species diffusion fluxes, τ , j, and d, respec-
tively. For a Navier-Stokes fluid, the former two are given by (19) and
(20). However, in practical meteorological modelling applications involving
turbulence, one often replaces these fluxes with effective turbulent closure
schemes so as to describe not the transport due to molecular motions but
rather the transport due to turbulent fluctuations. In that case, the func-
tional form of these “subscale” momentum and heat flux terms may take a
wide variety of forms which we will not address here in detail.
3 Introduction to multiple scales asymptotics

To motivate the mathematical techniques we are going to apply to the at-
mospheric flow equations in later chapters, we will now analyze the simple
example of a linear oscillator. After deriving analytical solutions, we will

focus on a situation that in many ways resembles situations arising fre-
quently in geophysical problems: a slow background motion caused by an
external force is accompanied by rapid oscillations around it. The oscil-
lation amplitudes are generally not small! To give some meaning to the
notions of “smallness” and “rapidity”, we will first nondimensionalize the
oscillator equations and identify small parameters that lend themselves for
comparison. By means of single and multiple scales analyses we will then
try to derive simplified approximate solutions that become more and more
accurate as the small parameters vanish.
3.1 Exact solutions for the linear oscillator

A typical example of a linear oscillator is a (small) piece of material with
mass m attached to a spring with stiffness c. The stronger the spring, the
higher the spring constant c. At time t, the mass is located at position
x = x(t) (see Fig. 3). If we displace it away from its equilibrium position,
at which the spring’s force just balances the weight of the piece, and then
let the system evolve freely, it will oscillate around its equilibrium with
constant amplitude. If we let the mass dive into some not-too-viscous fluid,
the system will perform a damped oscillation. If the viscosity of the fluid
is sufficiently high, the mass will not oscillate anymore but just move back
monotonically to its equilibrium position.
FF = −cx
mẍ
FD (t)
(Demon) FR = −k ẋ
Figure 3. Damped spring–mass system

144 R. Klein et al.
There are two forces acting within the system: the restoring force of the
spring in the direction opposite to the displacement of the mass, and the
frictional force FR = −k ẋ of the viscous fluid which acts in the direction
opposite to the motion of the mass. Newton’s law, which says that the
temporal change of momentum equals the sum of all acting forces, yields
an equation of motion for the system,
mẍ = −cx − k ẋ + FD (t) . (60)
Here we have included a general external force FD = FD (t) which some
demon may exert on the mass.
The solution to this second-order ordinary differential equation (ODE)
is uniquely determined once initial conditions x(0) = x0 and ẋ(0) = ẋ0
are prescribed. Following the theory of linear ODEs [Walter (1996)] we
can construct all solutions x(t) as a superposition of the general solution
of the homogenous problem and one so-called particular solution of the
inhomogenous equation, i.e.,
x(t) = xh (t) + xp (t) . (61)
Free oscillations We will now derive the general solution of the spring-
mass system’s homogenous ODE ((60) with FD ≡ 0),
mẍ + k ẋ + cx = 0 . (62)
To solve this equation, we choose an exponential ansatz,
x(t) = exp(ωt) . (63)
Inserting, we have

m ω2 exp(ωt) + k (ω exp(ωt)) + c (exp(ωt)) = 0 , (64)
and after division by exp(ωt) = 0 we find
mω 2 + kω + c = 0 , (65)
which is the system’s characteristic equation. The solutions are
1k √ k2 c
ω1/2 = − ± D with D := − . (66)
2m 4m2 m
If ω1 = ω2 , i.e., if the discriminant D = 0, the solutions differ qualitatively
from those obtained when ω1 = ω2 . Generally, if ω is a k-fold multiple solu-
tion of the characteristic equation, it corresponds to k linearly independent
solutions of the form
eωt , teωt , . . . , tk−1 eωt (67)
of the associated differential equation (Walter, 1996, S.173ff). One impor-

tant property of linear homogenous differential equations is that the sum of
two solutions is again a solution. Using this for the case D = 0 we obtain
the general solution to (62),
x(t) = A exp(ω1 t) + B exp(ω2 t) , (68)
where A and B are constants that remain to be determined. For the case
D = 0 we have
x(t) = A exp(ωt) + Bt exp(ωt) = (A + Bt) exp(ωt) (69)
again with yet unknown constants A und B.
Next we will distinguish several cases that differ w.r.t. the relative mag-
nitudes of the spring constant, c, the damping coefficient, k and the mass,
m:
1st case: k = 0; c > 0.

Thus we consider an inviscid oscillator for which

c
ω1,2 = ±iω0 with ω0 = , (70)
m
and
x(t) = A exp(iω0 t) + B exp(−iω0 t) . (71)
The solution x(t) is now complex-valued, but since we seek real-valued so-
lutions, we allow A, B ∈ C to be complex as well,
A = Ar + iAi , B = Br + iBi . (72)
Later we will choose A and B in such a way that the final result is again
real valued and phyiscally meaningful.
Using Euler’s formula, eix = cos x + i sin x, respectively, e−ix = cos x −
i sin x, and splitting A and B into their real and an imaginary parts, we
transform (71) into

x(t) = i (Ar + Br ) cos(ω0 t) + (−Ai + Bi ) sin(ω0 t)
(73)
i (Ar − Br ) sin(ω0 t) + (Ai + Bi ) cos(ω0 t) .
Because sine and cosine are linearly independent, the solution is real valued
if the coefficients (Ar − Br ) and (Ai + Bi ) satisfy
(Ar − Br ) = 0 ⇔ Ar = Br
(74)
(Ai + Bi ) = 0 ⇔ Ai = −Bi .
146 R. Klein et al.

x [m]
ï
ï
ï

time [s]
Figure 4. Exact solution for the oscillator without friction (k = 0) with

C = 2 and ω0 = 1.
Now (73) can be reduced to
x(t) = 2Br cos(ω0 t) + 2Bi sin(ω0 t)

(75)
= a cos(ω0 t) + b sin(ω0 t)
The constants a and b in (75) can be determined from the given initial
conditions and we obtain a particular solution for the considered system.
Example: Let us chose x0 = x(0) = 0, so that the mass is at its equilibrium
at time t = 0, and ẋ0 = dx
dt (0) = Cω0 , so that it has initial velocity Cω0 .
Inserting, we find from (75)
ẋ0
a=0 und b = =C, (76)
ω0
and the solution reads

x(t) = C sin(ω0 t) . (77)
See Fig. 4.
2nd case: k > 0; c > 0.

In this case with non-zero friction, the discriminant, D, in the general so-
lution for ω in (66) has to be examined in more detail. It may be greater
than, equal, or less than zero.

k2 k2
a) D<0 4m2
< c
m
⇒ ω1,2 = − 2m
k
±i c
m
− 4m2
Inserting ω1,2 into (68) and using Euler’s formula provides

k
x(t) = exp − t ã cos(ω̃a t) + b̃ sin(ω̃a t) , (78)
2m
where
c k2
ω̃a = − . (79)
m 4m2
The coefficients ã and b̃ again have to be determined so as to satisfy the
required initial conditions. The mass now performs a damped oscillation
around its equilibrium position with a modified frequency compared to first
case. See Fig. 5.

k2 k2
b) D > 0 4m2
> c
m
⇒ ω1,2 = − 2m
k
± 4m2
− c
m
Again, we insert into (68), and A and B are constants that have to be
computed using given initial conditions. We find

k
x(t) = exp − t A exp(ω̃b t) + B exp(−ω̃b t) (80)
2m
where
k2 c
ω̃b = − . (81)
4m2 m
Comparison of the solutions in (78) and (80) shows that both cases describe
a damped motion, the term exp(ω̃b t) “losing” against exp(− 2mk
t) for t → ∞,
but there is a fundamental difference: in case a), the system with the ex-
k
ponential function exp(− 2m t) performs a damped harmonic oscillation and
passes the origin several times. In case b), however, the damping is so strong
that the mass is not oscillating at all and just moves back monotonically to
its equilibirium (creeping case).

k2
c) D=0 4m2 = c
m ⇒ ω1,2 = − 2m
k
The solution in this case is in line with the one in the case already mentioned
above with general solution (69). We find

k
x(t) = exp − t (A + Bt) , (82)
2m
148 R. Klein et al.

x [m]
ï
ï

time [s]
Figure 5. Exact solution for the oscillator with friction (k > 0): a) D < 0
(blue, line), b) D > 0 (green,dashed), c) D = 0 (red, point-dashed).
where again the constants A and B have to be determind from the initial
conditions.
Thus, case c) is exactly “between” cases a) and b). It is called aperiodic
limiting case. In case b) the damping is still stronger and faster than in
case c) but in both cases there is no oscillation in the system.
Forced oscillations Now we account for an external force FD = FD (t)

acting on the linear oscillator by computing the particular part xp (t) of the
solution of the inhomogeneous differential equation (60). We will restrict to
the case of a periodic force
FD (t) = F0 cos(Ωt) . (83)
(Why is this not a severe restriction?). Experiments show that an oscillator
exposed to such a force performs a forced harmonic oscillation after some
adjustment time has elapsed. The frequency of this oscillation equals that
of the driving force, FD (t). Let us verify that this will in fact be a valid
particular solution. We let
xp (t) = Ap sin(Ωt) + Bp cos(Ωt) (84)
with coefficients Ap and Bp remaining to be determined. Inserting the
corresponding time derivatives
ẋp (t) = −Ap Ω cos(Ωt) − Bp Ω sin(Ωt)
(85)
ẍp (t) = −Ap Ω2 sin(Ωt) − Bp Ω2 cos(Ωt)

x [m]
ï
ï

time [s]
Figure 6. Exact solution for the oscillator driven by an external force.

a) Superposition of fast oscillations and slow background movement (blue
line) b) resonance (green dashed line) c) driving external force (red dashed-
pointed line)
into equation (60) we obtain

F0 cos(Ωt) = −m(Ap Ω2 sin(Ωt) + Bp Ω2 cos(Ωt)) +
k(−Bp Ω sin(Ωt) + Ap Ω cos(Ωt)) + (86)
c(Ap sin(Ωt) + Bp cos(Ωt)) .
Linear independence of the sine and cosine functions allows us to split this
equation, such that
(c − mΩ2 )Ap − kΩBp = 0
(87)
(c − mΩ2 )Bp + kΩAp = F0 .
This is a system of two equations with two unknowns Ap and Bp that is
easily solved. If k = 0 (2nd case), then
c − mΩ2
Bp = Ap
kΩ
2 −1 (88)
F0 c − mΩ2
Ap = 1+ .
kΩ kΩ
If k = 0 and (c − mΩ2 ) = 0, then

F0
Ap = 0 and Bp = (89)
c − mΩ2
150 R. Klein et al.
and the case k = 0 and (c − mΩ2 ) = 0 yields the resonant solution (see for
example Walter (1996)) with time-dependent coefficients
F0
Ap (t) = t and Bp = 0 . (90)
2mΩ
We have found particular solutions that satisfy the inhomogenous equa-
tion for each of the parameter regimes. The sum of the homogenous and
the particular solution is the general solution of the oscillator equation with
external periodic forcing. For example, in case 2a) this solution reads

k
x(t) = exp − t ã cos(ω̃a t) + b̃ sin(ω̃a t) + Ap sin(Ωt) + Bp cos(Ωt)
2m
(91)
The constants Ap and Bp have to be defined according to equation (88) and
only ã and b̃ are to be computed from the intial conditions. The solutions
for the other cases can be derived analogously.
3.2 Dimensionless representation and small parameters

In the preceding section we were able to derive the general solution for
the linear oscillator with the equation of motion
mẍ + k ẋ + cx = F0 cos(Ωt) (92)
and initial conditions x(0) = x0 and ẋ(0) = ẋ0 . This solution consists
of a homogeneous and an inhomogenous part. As we have seen in the
second case, the homogenous solution is decaying exponentially when there
is non-zero damping, so that in the longtime motion only the inhomogenous
(particular) part of the solution prevails.
In the present section, we will continue our analysis by studying situ-
ations in which the free oscillation, damping, and forcing act on very dif-
ferent characteristic time scales. (The notion of a characteristic scale will
hopefully be reasonably clear by the end of the section. The reader may
trust her or his intuition for the time being.) An important tool for deriv-
ing concise mathematical descriptions of such scale-separated processes is
Multiple-Scales Asymptotics, the main ideas and techniques of which will
be explained here using the linear oscillator as an example.
To determine the different time scales of the system we will nondimen-
sionalize the general equation of motion (92) in the first step.
Remarks on dimensional analysis Within the governing equation for

the linear oscillator, we identify three fundamental physical dimensions
{Xi }3i=1 : Length L, Time T , and Mass M. Each physical quantity φj

that appears in this governing equation has a physical dimension that is a
product of these fundamental ones, so that
3
/ bij
Dim(φj ) = (Xi ) . (93)
i=1
For the linear oscillator, we have
quantity φ physical dimension Dim(φ)
dependent and independent variables

x L
t T
parameters of governing equation

m M
k M/T (94)
c M/T 2
F0 ML/T 2
Ω 1/T
initial data
x0 L
x˙0 L/T
Once a system of concrete units is chosen based on which these fundamental

dimensions shall be measured, each of the physical quantities and coefficients
in the governing equations can be quantified by a sole number. The familiar
SI-system is one example, where (T , L, M) are measured in terms of (Second
s, Meter m, Kilogram kg).
Knowing a quantity’s physical dimension and the underlying system of
units one can always transform these non-dimensional numbers back into
measurable physical values. Obviously, there is a one-to-one map between
any two different systems of units, so that the exact solutions of the gov-
erning equations will not depend on which system is chosen.
As it stands, the oscillator equation in (60) does not reveal anything
besides what was built into it to begin with: Newton’s law of motion for
the particular case of the mechanical system in Fig. 3. To obtain a some-
what improved intuition about possible solutions one may study classes of
solutions distinguished by some particular global mathematical characteri-
zation.
152 R. Klein et al.
For any given solution of the equations one can identify “characteristic
values” [φj,ref ]N
j=1 of the total of N physical quantities in the system which
roughly describe their orders of magnitude throughout the solution or at
least during a certain time interval (and within a selected region in space
for pdes). These dimensional characteristic quantities may be combined
into non-dimensional characteristic numbers
/
N
ajk
Πk = (φj,ref ) , (95)
j=1
with the exponents ajk chosen so as to guarantee that the Πk do not have a
physical dimension as will be explained shortly.
These numbers are extremely useful as they provide a comparison be-
tween various quantities that may have the same physical dimension but
very different physical origin. An example is the ratio of the oscillator’s
frequency of free, undamped oscillations, c/m, and that of its harmonic
excitation,
c/m
Π∗ = . (96)
Ω
For the non-dimensional Π’s to be actually non-dimensional, all the phys-
ical dimensions have to cancel exactly in the product. Using (93), we may
rephrase this statement as
3 ajk ⎡ ⎤ " #
3 3 P
N
/N / i / /
N
i j / bij ajk
Dim(Πk ) = b
(Xi ) j = ⎣ (Xi ) j k ⎦ =
b a
(Xi ) j=1 ≡ 1.
j=1 i=1 i=1 j=1 i=1
(97)
For this equation to hold, the respective powers of each of the fundamental
dimensions Xi must vanish independently, so that

N
bij ajk ≡ 0 (i = 1 . . . 3, k arbitrary) . (98)
j=1
These are 3 linear constraints on the N -tuples ak = (a1k , . . . , aN

k ), which
therefore span a total space of dimension N − 3. This, in turn, is equiv-
alent to the existence of a set of N − 3 independent characteristic num-
−3
bers {Πk }Nk=1 , which is the key statement of the famous Buckingham’s
π-theorem.
Remark: Often this theorem is quoted from E. Buckingham (1914) as Buck-
ingham’s Theorem. Yet, Barenblatt (1996) acknowledges a A. Federmann
(St. Petersburg 1911) for the first proof, and Görtler (1975) for a concise
formulation.
Remark: For further aspects of dimensional analysis, the reader may want
to consult Barenblatt (1996).
Dimensionless representation of the oscillator problem With the

nine quantities from (94) we have 9-3=6 linear independent dimensionless
combinations. Two of these are our new dimensionless dependent and in-
dependent variables, y and τ . Four of them are dimensionless real numbers
which relate and characterize all those quantities that influence the solution
in one way or the other. A possible specific choice of these quantities is
Dependent and independent variables
x T2 M
y = ; Dim (y) = L· · = 1
F0 /c ML T 2
1
τ = Ωt ; Dim (τ ) = ·T = 1
T
Characteristic numbers
(99)
Ω2 1 T2
μ = m ; Dim (μ) = M· 2 · = 1
c T M
Ω M 1 T2
κ = k ; Dim (κ) = · · = 1
c T T M
x0 T2 M
y0 = ; Dim (y0 ) = L· · = 1
F0 /c ML T 2
ẋ0 L T2 M
y0 = ; Dim (y0 ) = ·T · · = 1
ΩF0 /c T ML T 2
Here μ characterizes the system’s inertia, κ its damping, and y0 , y0 the
initial data that allow us to select a specific solution. In interpreting these
quantities, notice that F0 /c is the static displacement of a spring with stiff-
ness c under the effect of a (constant) force F0 .
Notice that y : IR+ → IR, is a function, not just a number, and τ varies
all over IR+ . The original unknowns x(t), t and the new ones, y(τ ), τ , must
154 R. Klein et al.
,,,
,,
Figure 7. Different possibilities of performing a limit in an asymptotic

system with small parameters μ and κ
satisfy
x(t)
= y(Ωt) and Ωt = τ . (100)
F0 /c
Using this identity, we have
F0 dy dτ F0 Ω
ẋ(t) = = y (τ ) (101)
c dτ dt c
2
F0 d2 y dτ F0 Ω2
ẍ(t) = = y (τ ) , (102)
c dτ 2 dt c
and (92) then allows us to specify a differential equation for y(τ ),
mΩ2 kΩ
y + y + y = cos τ (103)
c c
or
μy + κy + y = cos τ . (104)
The appropriate initial conditions read
x0 c ẋ0 c
y(0) = y0 = and y (0) = y0 = . (105)
F0 F0 Ω
Oscillators with small mass and small damping – exact solutions

Here we examine the behavior of the (dimensionless) oscillator system
y = y(τ ; μ, κ, y0 , y0 ) (106)
for fixed initial data, but for smaller and smaller values of the inertia and
damping parameters μ, κ (i.e., μ, κ ∈ IR+ ; μ, κ 1).
Distinguished limits Letting μ, κ → 0 it turns out to be important just

how these values tend to zero: For example, if we let μ vanish first, following
path I in Fig. 7, we formally get rid of the second derivative in (104) and
obtain a first-order ordinary differential equation. But if we let κ → 0
first (path II), then the system formally reduces to the undamped oscillator
equation. Along the first path, we expect non-oscillatory, purely damped
motions, whereas along the second, oscillations of constant amplitude about
some slow mean harmonic motion should arise. To verify or reject the stated
hypotheses regarding paths I and II, or to decide what behavior will emerge
for some path in between, such as III, requires some deeper analyses.
The first ansatz that may come to mind when looking for solutions in
the vicinity of the location μ = 0; κ = 0 in parameter space might be a
Taylor expansion with respect to the μ-κ-dependence of the solution. This
would promise to provide the best-possible linear, quadratic, or higher-order
polynomial approximations as long as the dependence of the solution on
these parameters is sufficiently smooth.
In fact, when such an expansion exists at all, it reads

∂y ∂y
y = y|μ,κ=0 + μ +κ + O(μ, κ) (μ, κ → 0) , (107)
∂μ μ,κ=0 ∂κ μ,κ=0
where
∂y ∂y
, = grad(μ,κ) y (108)
∂μ ∂κ
is the gradient (or (Fréchet-) derivative) of the solution with respect to our
two small parameters. (For the definition of a Fréchet-derivative see, e.g.,
(Werner, 2000, S. 113)).
However, from our previous considerations regarding the path-dependence
of the solution behavior as μ → 0 and κ → 0 we conclude that even though
there is a limit solution, y|μ,κ=0 (τ ) = cos(τ ), it is not the limit of either
solution found along paths I or II in the parameter space of Fig. 7. Thus we
cannot decide whether the “proper behaviour” for very small μ, κ should be
purely oscillatory, purely damped, or something in between. We are led to
conclude that the gradient in (108) does not exist!
Fortunately, not all is lost. Analysis has it that even if a Fréchet-
derivative does not exist, linear approximations to the solution along straight
lines, i.e., directional derivatives, may still be well-defined. This would lead
156 R. Klein et al.
us to consider coupling μ, κ in such a way that κ = αμ as μ → 0 for some

fixed α.
We allow here for more general μ-κ–relationships which include this for-
mer one by letting (μ, κ) approach the origin of their parameter space along
some parameterized trajectory. To this end, we introduce a new expansion
parameter ε 1, and two functions μ̂(ε), κ̂(ε) that should satisfy
(μ̂(ε), κ̂(ε)) = o(1) (ε → 0) . (109)
Then, dropping the solution’s explicit dependence on the initial data in the
notation for the moment, we Taylor-expand w.r.t. ε,
∂ ŷ
y(τ ; μ, κ) = ŷ(τ ; ε) = ŷ(τ ; μ̂(ε), κ̂(ε)) = ŷ(τ ; 0)+ε (τ ; 0)+O(ε) (ε → 0) .
∂ε
(110)
The mappings
(μ̂, κ̂) : IR → IR2
(111)
ε → (μ̂(ε), κ̂(ε))
comprise a distinguished limit.
Remark: Here ∂ ŷ/∂ε(τ ; 0) is a generalization of the directional or Gâteaux
derivative
∂
(y(τ ; αε, βε))ε=0 α, β = const. (112)
∂ε
In general we know from functional analysis that for some mapping f
⇒
f Fréchet-differentiable f Gâteaux-differentiable . (113)

We conclude that asymptotic expansions based on an approximation in

ε with respect to appropriate distinguished limits are more general than a
multi-parameter expansions, because–as pointed out above–the latter corre-
spond to classical Taylor-expansions in μ, κ and would require the existence
of the Fréchet-derivative gradμ,κ y at μ = κ = 0.
Time scales In analyzing distinguished limits below, we will pay special

attention to the timescales on which oscillation, damping,and background
forcing will act as ε vanishes. The expressions m/k and m/c determine
the timescales of damping and free oscillation of the system: In our solution
k
from (91), the damping is described by exp(− 2m t). Thus, t has to change
by O(m/k) to change the argument of the exponential function by O(1).
For
an undamped, free oscillation, t has to change by O(1/ω0 ) = O m/c
to change the arguments of sine and cosine by O(1). The dimensionless
quantities

m m √
KD = Ω = μ/κ and KO = Ω = μ (114)
k c
allow us to compare the internal damping and oscillation time scales of the
oscillator with that of the external forcing. Notice that both are functions
of our dimensionless mass and damping parameters, (μ, κ).
Three examples shall underline the consequences of picking particular
distinguished limits in the oscillator problem:
1st case: μ∼κ (μ = ε and κ = εκ̂ with κ̂ = const. as ε → 0)

√
Here we have KD = 1/κ̂ = const. and KO = ε. Thus, as ε → 0, the
timescale of free oscillation becomes much shorter than that of the external
forcing, while the damping timescale remains comparable to it. We verify
this considering the upper graph in Fig. 8. The frequency of oscillation
increases as we reduce ε from 10−2 to 5 · 10−4 , whereas the time which the
oscillation needs to decay remains the same.
√
2nd case: μ ∼ κ2 (μ = ε and κ = εκ̂ with κ̂ = const. as ε → 0)
√ √
Here, since KO = ε and KD = ε/κ̂, the oscillation and damping time-
scales remain comparable in the limit. As a consequence, the number of
oscillations which the system performs before the oscillation is essentially
damped away remains nearly independent of ε. We corroborate this by
inspecting the centred graph in Fig. 8.
1
3rd case: μ ∼ κ3 (μ = ε and κ = ε 3 κ̂ with κ̂ = const. as ε → 0)
For this case, Fig. 8 reveals that the system does no longer oscillate at all.
2
For small ε, the damping timescale KD√ = ε 3 /κ̂ is always much smaller
than the timescale of oscillation KO = ε. Therefore, the inertial motion
of the oscillator is already damped before the first overshoot due to the
incipient oscillation can take place. After a short initial transient, the mass
is essentially “slaved” by the external forcing.
It turns out that the regime μ ∼ κ2 is precisely the threshold that
separates the regions in μ-κ–space in which, as ε → 0, either an oscillatory
or a purely damped motion prevails. If μ vanishes slower than κ2 , the system
will oscillate, otherwise it is strongly damped as summarized in figure 9.
158 R. Klein et al.

\τ
ï
ï

τ

\τ
ï
ï

τ

\τ
ï
ï

τ
Figure 8. Impact of different choices for the distinguished limit. Exact

solutions for ε = 0.01 (blue dashed lines), ε = 0.0005 (green lines) and
background oscillation (red dashed-pointed lines). μ = κ = ε (upper),
μ = κ2 = ε (centre), and μ = κ3 = ε (lower graph).
κ 1
κ ∼ μ3
√
κ = κ̂ μ
GDPSHG
RVFLOODWRU\
κ∼μ
Figure 9. Ratio between mass and damping in the problem and the result-
ing behaviour of the solution.
If there exist solutions for which y and y remain bounded while μ → 0

and κ → 0, equation (104) reduces to
y = cos τ . (115)
in the limit. The solution written in dimensional terms becomes
F0 F0
x(t) = y(τ ) = cos(Ωt) . (116)
c c
This is in line with “experimental observations” (and with the exact solu-
tion, of course). After a certain initial period of adjustment, the oscillator
with external force FD (t) = F0 cos(Ωt) performs a periodic oscillation with
angular frequency Ω. This is a limiting long-time behavior to be expected
from any approximate (asymptotic) solution below, as long as κ > 0.
3.3 Regular perturbation analysis for small mass and damping

In the sequel we will derive approximate solutions to the dimensionless
equation of motion (104) using techniques of asymptotic analysis. To do so,
we choose case 1 from the last section, i.e., we let
μ = ε, κ = εκ̂ with κ̂ = const. as ε → 0. (117)
Remark: Choosing a different distinguished limit would yield different asymp-
totic results!
With this distinguished limit fixed, the oscillator’s governing equation re-
duces to
εy + εκ̂y + y = cos τ . (118)
160 R. Klein et al.
We consider the solution to explicitly depend only on τ and ε and denote
it by y(τ ; ε), dropping the “hat” notation used earlier in (110).
Slow-time asymptotic expansion For the solution y = y(τ ; ε) of (118),

we choose here an asymptotic ansatz of the form

y(τ ; ε) = y (0) (τ ) + εy (1) (τ ) + ε2 y (2) (τ ) + O ε2 . (119)
Remark: Generally, a series of the form

N
φn (ε) un (x) (120)
n=1
with φn+1 (ε) = O(φn (ε)) for ε → 0 is called an asymptotic N-term expansion
of the function u if

N
u(x; ε) − φn (ε) un (x) = O(φN (ε)) (121)
n=1
for ε → 0 (see for example Kevorkian and Cole (1996) and Schneider (1978)).
Notice that the result of the analysis will depend on the choice of the asymp-
totic sequence {φn (ε)}n∈IN .
The chosen ansatz is a Taylor expansion in ε = 0 of the desired solution
y = y(τ ; ε), i.e., we look for the coefficients of ε, ε2 etc. in

N
1 n ∂ny
y(τ ; ε) = ε n
(τ ; 0) + O εN
n=0
n! ∂ε (122)
ε2
= y(τ ; 0) + ε(∂ε y)(τ ; 0) + (∂ε2 y)(τ ; 0) + O ε2
2
Letting y (0) (τ ) := y(τ ; 0), y (1) (τ ) := (∂ε y)(τ ; 0) etc. naturally leads to (119).
We proceed to determine the asymptotic behaviour of the solution y for a
fixed τ and an arbitrary but small ε. Depending on the power of ε consid-
ered, one speaks of the behaviour of the solution at a certain order. Thus,
y (0) (τ ) describes the behaviour at leading or zeroeth order, y (1) (τ ) the be-
haviour at first order, etc.
Inserting the time derivatives from (119) (time means the dimensionless
time τ )

y (τ ; ε) = y (0) (τ ) + εy (1) (τ ) + ε2 y (2) (τ ) + O ε2
(123)

y (τ ; ε) = y (0) (τ ) + εy (1) (τ ) + ε2 y (2) (τ ) + O ε2
into (118) we find

0 = εκ̂ y (0) + εy (1) + ε2 y (2) + O ε2 +

εκ̂ y (0) + εy (1) + ε2 y (2) + O ε2 + (124)

εκ̂ y (0) (τ ) + εy (1) + ε2 y (2) + O ε2 − cos τ
or

y (0) − cos τ +ε y (0) + κ̂y (0) + y (1) +ε2 y (1) + κ̂y (1) + y (2) = O ε2 .
(125)
If this equation is to hold for arbitrary (but small) ε, each of the coeffi-
cients of εi for (i = 1, 2, . . . ) (the expressions in brackets) has to be zero
individually. Therefore,
y (0) = cos τ

y (1) = −y (0) − κ̂y (0) = cos τ + κ̂ sin τ (126)

y (2) = −y (1) − κ̂y (1) = (1 − κ̂2 ) cos τ + 2κ̂ sin τ .
Our second-order accurate asymptotic solutions thus reads

y(τ ; ε) = cos τ + ε (cos τ + κ̂ sin τ ) + ε2 (1 − κ̂2 ) cos τ + 2κ̂ sin τ + O ε2 ,
(127)
and this is compared with the exact solution in Fig. 10.
Discussion For ε → 0, the approximate asymptotic solution in (127)

reduces to
y(τ ; 0) = y (0) (τ ) = cos τ , (128)
and the asymptotic solution in the limit coincides with (115). Yet, there is
a severe catch:
We have no degrees of freedom to meet the initial data!
Instead, the initial displacement and velocity from the leading-order asymp-

totics are y (0) (τ ) = 1 and y (0) (0) = 0. Any different values for y0 and y0 can
nowhere be accounted for. The reason is that with the present ansatz we can
only see the solution after any initial transient, which would be determined
by the initial data, has already decayed.
162 R. Klein et al.

\τ ε
ï
ï
ï
ï

τ
Figure 10. Asymptotic expansion of the solution in τ with y0 = 2, y0 = 0,

ε = 0.2 and κ̂ = 0.2. Exact solution: black line, y(τ ; ε) = y (0) : red line,
y(τ ; ε) = y(0) + εy (1) : green dashed, y(τ ; ε) = y (0) + εy (1) + ε2 y (2) : blue
dash-pointed
This can be verified in Fig. 10, even when the next two higher-order
terms are included. Adding the first- and second-order contributions, makes
merely a slight difference for the asymptotic approximation. It can do no
more than reproduce with better accuracy the background oscillation as
more and more expansion terms are included.
Fast-time asymptotic expansion Using the time coordinate τ = Ωt

we miss
the fast oscillatory motions associated with the free oscillation time
scale m/c. The rescaled time variable
t τ τ
ϑ= = =√ (129)
m/c 2
mΩ /c ε
would remedy this problem. We will try out a new expansion scheme in
which the unknowns will depend on ϑ instead of on τ . Since each time
derivative d/dτ that√ appears in the governing equation in that case will
produce a √factor 1/ ε by the chain rule, we should expand the solution in
powers of ε instead of in powers of ε. (What happens if we don’t but use
ϑ as the independent variable while expanding in powers of ε?) Thus we
choose the new asymptotic expansion scheme
√
y(τ ; ε) =: y (0) (ϑ) + εy (1) (ϑ) + εy (2) (ϑ) + O(ε) , (130)
which is equivalent to

(0) τ √ (1) τ (2) τ
y(τ ; ε) = y √ + εy √ + εy √ + O (ε) . (131)
ε ε ε
We will need to be aware of the latter form of writing the expansion when
we insert into the governing equation in (118), which is written in terms of
τ . In preparation, we compute the τ -derivatives of (131) for fixed ε,

∂y dy (0) dϑ √ dy (1) dϑ dy (2) dϑ dϑ
(ϑ; ε) = + ε +ε +O ε
∂τ ε dϑ dτ dϑ dτ dϑ dτ dτ
1 dy (0)
dy (1) √ dy (2) √
= √ + + ε +O ε (132)
ε dϑ dϑ dϑ

∂ 2 y 1 d2 y (0) 1 d2 y (1) d2 y (2)
(ϑ; ε) = + √ + + O(1) . (133)
∂τ 2 ε ε dϑ2 ε dϑ2 dϑ2

∂y
The notation ∂τ ε
shall underline that, for any solution of the oscillator
problem, ε is a fixed parameter and thus held constant when differentiating.
If, instead, we were to consider ε as varying in time, then the time scales, and
thus the mass, spring stiffness, and the damping coefficient, would change
in time, too.
We express√the forcing term in (118) in terms of ϑ, and then Taylor-
expand w.r.t. ε,
√ ε
cos τ = cos( εϑ) = 1 − ϑ2 + O ε2 . (134)
2
Inserting the appropriate derivatives and expansions into (118) we find
d2 y (0) √ d2 y (1) d2 y (2)

+ ε + ε
dϑ2 dϑ2 dϑ2
√ dy (0)
dy (1)
+ κ̂ ε + κ̂ε (135)
dϑ dϑ
√ ϑ2
+ y (0) + εy (1) + εy (2) = 1−ε + O(ε) .
2
This yields, after collection of like powers of ε, the following hierarchy of

164 R. Klein et al.
perturbation equations,
d2 y (0)
O(1) : + y (0) = 1
dϑ2
√ d2 y (1) dy (0)
O ε : + y (1) = −κ̂ (136)
dϑ2 dϑ
d2 y (2) dy (1) ϑ2
O(ε) : + y (2) = −κ̂ − .
dϑ2 dϑ 2
At leading order (i.e., at O(1)) we find the equation for an undamped os-
cillator with time-independent forcing. The solution is
y (0) = a0 sin ϑ + b0 cos ϑ + 1 . (137)
Here a0 and b0 are constants that depend on the √ initial conditions.

Next we solve the first-order equation (at O( ε)) in the expansion, which
becomes
d2 y (1)
+ y (1) = −κ̂ (a0 cos ϑ − b0 sin ϑ) (138)
dϑ2
(1)
The solution is a superposition of the homogeneous solution, yh = a1 sin ϑ+
b1 cos ϑ, and a particular solution that takes care of the right-hand side. The
coefficients a1 and b1 will have to be computed from the initial conditions
as before. To derive a particular solution, we use the technique of variation
of coefficients, i.e.,
yp(1) = f (ϑ) sin ϑ + g(ϑ) cos ϑ , (139)
where f and g remain to be determined. Inserting this ansatz into the

differential equation (138) yields
(f¨ − 2ġ) sin ϑ + (g̈ + 2f˙) cos ϑ = −κ̂ (a0 cos ϑ − b0 sin ϑ) . (140)
Comparing coefficients, we find the constraints
f¨ − 2ġ = κ̂b0 and g̈ + 2f˙ = −κ̂a0 . (141)
The desired solutions are polynomials in ϑ of degree less than two, and we
let f (ϑ) = Af ϑ + Bf and g(ϑ) = Ag ϑ + Bg . Without loss of generality, we
may also assume Bf ≡ 0 and Bg ≡ 0, as these terms can be covered by the
homogenous part of the solution. Solving (141), yields
κ̂
yp(1) = −ϑ (a0 sin ϑ + b0 cos ϑ) (142)
2
so that
(1) κ̂
y (1) = yh + yp(1) = a1 sin ϑ + b1 cos ϑ − ϑ (a0 sin ϑ + b0 cos ϑ) . (143)
2
In the same way one may compute higher-order solutions.
It turns out that this ansatz based on the short-time variable is still
not satisfactory. As seen in figure 11, although the solution improves with
increasing order, this is true only for the early stages of the evolution. With
increasing time the solution deteriorates dramatically, and this gets worse
the higher the approximation order!

\τ ε
ï
ï
ï
ï

τ
Figure 11. Asymptotic expansion of the solution in ϑ with y0 = 2, y0 = 0,

ε = 0.001
√ and κ̂ = 0.8. y(τ ; ε) (exact solution): line; y (0) : red line;
√ (1)black(2)
(0) (1) (0)
y + εy : green dashed line; y + εy + εy : blue dash-pointed
line
Remark: For the present short-time expansion we could have chosen an

ansatz similar to (119) (i.e. with φn (ε) = εn ). It is easy to verifiy that this
yields only the trivial solution y(τ ; ε) ≡ 0. This shows that the choice of the
asymptotic sequence {φn (ε)}n∈IN is crucial when looking for an asymptotic
solution.
Remark: As we have seen, the differential equation √ (1)for y (1) has a resonant,
amplifying solution. After some time, the term εy becomes comparable
to the previous one, y (0) , and the whole idea of building a series with smaller
and smaller corrections as the order of approximation increases fails badly.
In fact, the present asymptotic √ approximate solution is valid only for ϑ =
O(1) or, equivalently, τ = O( ε), i.e., for asymptotically short times on the
time scale of the background forcing.
166 R. Klein et al.
3.4 Multiple scales analysis

√
The analyses based on single timescale representations, with τ = O( ε)
and τ = O(1), respectively, were at best partially successful. They did allow
us to cover the early, respectively, late development of the solution, but
were definitely not valid uniformly in time. Here we consider an asymptotic
expansion scheme that accounts for both of these timescales in a single
sweep,

τ √ τ τ
y(τ ; ε) = y (0) √ , τ + εy (1) √ , τ + εy (2) √ , τ + O(ε)
ε ε ε
√ (1)
= y (0) (ϑ, τ ) + εy (ϑ, τ ) + εy (2) (ϑ, τ ) + O(ε) .
(144)
Including both the time variables considered in our previous expansions, we
expect this scheme to allow us to fulfill the initial conditions, y(0; ε) = y0
and y (0; ε) = y0 , but to also capture the long-time behavior of the solution
without resonant growth.
Remark: The key challenge, of course, will be to find the dependencies of
the expansion functions, y (i) (ϑ, τ ), on two time variables, even though our
orginal govering equation involves merely a single independent variable, τ .
In fact, we will have to determine the variation of the expansion functions
within the entire (ϑ, τ )–plane, because we want to obtain solutions that are
valid for small, but otherwise arbitrary values of ε. Varying ε, the ratio
between ϑ and τ changes, as indicated in Fig. 12. Thus, for any fixed value
of τ we may be interested in y (i) (ϑ, τ ) within a range of values of ϑ, and
vice versa, the range being determined by the range of realistic values of ε.
To proceed, we need to work out the time derivatives needed in (144)
taking into account that the y (i) (ϑ, τ ) are functions of two independent
variables. Using the chain rule, we find

∂y (0) √ (1) √ (2)
= yϑ ϑτ + yτ(0) + εyϑ ϑτ + εyτ(1) + εyϑ ϑτ + εyτ(2) + O(εϑτ )
∂τ ε
1 (0) √ √
(1) (2)
= √ yϑ + yτ(0) + yϑ + ε yτ(1) + yϑ + O ε
ε
∂ 2 y 1 (0) 1 (0) (1)

(0) (1) (2)

= y + √ 2y + y + y τ τ + 2y + y + O(1)
∂τ 2 ε ε ϑϑ
ε ϑτ ϑϑ ϑτ ϑϑ
(145)
Inserting into (118) one has
√ √
(0) (0) (1) (0)
yϑϑ + y (0) − cos τ + ε 2yϑτ + yϑϑ + κ̂yϑ + y (1) +O ε = 0 . (146)
ϑ
ϑ1
ϑ2
ϑ3
ϑ∗
τ∗ τ
Figure 12. Relation between the time coordinates τ and ϑ for chosen ε,
√
where ϑi = τ / εi (ε1 < ε2 < ε3 ). For every τ ∗ there is exactly one ϑ∗ .
For this equation to hold for arbitrary ε 1 it is sufficient that the dif-
ferent terms in brackets vanish.
√ Considering that, through the dependence
of the y (i) on τ and ϑ = τ / ε, these brackets implicitly do depend on ε, it
is not immediately clear that the vanishing of the brackets is also necessary.
But, if we can make all the coefficients disappear for arbitrary (ϑ, τ ), then
our equation is in any case satisfied.
√ Let’s give it a try!
For the different orders in ε this leads to
(0)
O(1) : y (0) + yϑϑ = cos τ
√ (147)
(1) (0) (0)
O ε : y (1) + yϑϑ = −2yϑτ − κ̂yϑ .
It is of crucial importance that the first of these equations is an ODE for

the ϑ–depencence of y (0) , and that any τ –derivative of this leading-order
function appears only at the next order. Thus, if the first equation is solved
for the ϑ–dependence, τ remains as a parameter the influence of which will
remain to be determined.
Following the same procedures as earlier in this section, we find the
general solution in terms of ϑ for (147)1 to be
y (0) (ϑ, τ ) = A(τ ) cos ϑ + B(τ ) sin ϑ + cos τ . (148)
This is the superposition of a slow background motion and an oscillation

with increasing frequency as ε → 0. In addition, when A(τ ) ≡ B(τ ) ≡ 0 we
just retrieve the leading-order solution y (0) from the single-scale analysis.
Thus, the present result includes the previous solutions, and has extended
168 R. Klein et al.
the description of the fast oscillations to arbitrary τ .√(Remember that the

fast-time analysis restricted us to considering τ = O( ε) only!)
Considering (147)2 and inserting the partial derivatives of (148) we have,
(1)
yϑϑ + y (1) = (2A + κ̂A)(τ ) sin ϑ − (2B + κ̂B)(τ ) cos ϑ . (149)
This equation has a resonant solution (compare the remarks in the section
3.1):

(1) κ̂ κ̂ (1)
y (1) = yh −ϑ A + A (τ ) cos ϑ + B + B (τ ) sin ϑ =: yh +ϑỹp(1) .
2 2
(150)
(1) √ √ (1) (1)
If ỹp = 0, the term εy (1) (ϑ, τ ) = εyh + τ ỹp in the asymptotic ansatz
(144) is no longer negligible compared to the term y (0) (ϑ, τ ) if τ = O(1).
To exclude resonant solutions like this and thus to make sure that succes-
sive terms in our approximation
√ yield systematically smaller and smaller
corrections even if τ = εϑ = O(1), we demand that, for fixed τ

√ (i) τ
εy √ ,τ → 0 as ε → 0. (151)
ε
Remark: The condition (151) is known as sub-linear growth condition. In
general we impose the condition

φn (ε) (n) x1
u , x1 , x2 , . . . , xm → 0 as ε → 0, xi , i = 1, . . . , m fixed
φn−1 (ε) ψ(ε)
(152)
on the coefficients u(n) (η, x1 , x2 , . . . , xm ) of an asymptotic expansion of
(121) (with η = x1 /ψ(ε) → ∞ for ε → 0 at fixed x1 ). The name “sub-
linear growth condition” is motivated by the above condition for η = x1 /ε
and φn (ε) = εn . In this case,
u(n) (η, x1 , x2 , . . . , xm )
lim =0 , xi , i = 1, . . . , m fixed. (153)
η→∞ η
and this means the u(n) grow slower (sub-linear) than η for η → ∞, or for
ε → 0 at fixed x1 .
(1)
Demanding that y (1) contain no “resonant terms” of the type ϑỹp leads to
κ̂ κ̂
A (τ ) +
A(τ ) = 0 and B (τ ) + B(τ ) = 0 . (154)
2 2
These equations are solved by

κ̂ κ̂
A(τ ) = A(0) exp − τ and B(τ ) = B(0) exp − τ . (155)
2 2
At this point it is reasonable to continue with the expansion of the initial

conditions y(0) = y0 , y (0) = y0 . This will provide us with some additional
information that will be needed to determine A(0) and B(0). The expansion
of the initial conditions is done by inserting τ = 0 and ϑ = 0 into (144).
This results in
√
y0 = y (0) (0, 0) + εy (1) (0, 0) + εy (2) (0, 0) + O(ε) (156)

1 √ √
y0 = √ yϑ(0) + (yτ(0) + yϑ(1) ) + ε(yτ(1) + yϑ(2) ) (0, 0) + O ε (157)
ε
Assuming initial conditions of type y0 = O(1) and y0 = O(1) and an allow-
ing for small but otherwise arbitrary ε, we conclude that
y (0) (0, 0) = y0
y (i) (0, 0) = 0 (i = 1, 2, ...)
(0) . (158)
yϑ (0, 0) = 0
(0) (1)
(yτ + yϑ )(0, 0) = y0
(i) (i+1)
(yτ + yϑ )(0, 0) = 0 (i = 1, 2, ...)
Inserting these results into the leading-order solution from (148), and using
(0)
yϑ = −A(τ ) sin ϑ + B(τ ) cos ϑ , (159)
we derive initial data for A and B,
y (0) (0, 0) = A(0) + 1 ⇒ A(0) = y0 − 1

(0)
(160)
yϑ (0, 0) = B(0) ⇒ B(0) = 0.
Thus, with (155) and (148) we obtain the leading-order multiple-scales so-
lution,
(0) κ̂
y (ϑ, τ ) = (y0 − 1) exp − τ cos ϑ + cos τ . (161)
2
Remark: We have used the initial conditions exclusively at ϑ = τ = 0, but
not for ϑ = 0 and arbitrary τ . The latter would be incorrect, because in the
ϑ, τ –plane, for fixed system parameters μ, κ, i.e., for fixed ε, the solution
evolves along a straight line through the origin in the ϑ, τ –plane as shown
170 R. Klein et al.

\τ ε
ï
ï
ï
ï
ï

τ
Figure 13. Asymptotic multiscale expansion of the solution in τ and ϑ

with ε = 0.001, κ̂ = 0.8 and initial conditions y0 = 2 and y0 = 10. Exact
solution: black line, y (0) (τ ; ε): red dashed line
in Fig. 12. Obviously, at physical time t = 0, we access the asymptotic

solutions y (i) (ϑ, τ ) at ϑ = τ = 0.
Figure 13 displays the leading-order and exact solutions for y0 = 2; y0 =
0, i.e., for a setting in which y0 = OS (1) and y0 = OS (1). The agreement
is convincing, even after many of the fast oscillation cycles, and for times
τ = O(1).
Remark: We observe that the leading-order multiple scales solution from
Fig. 13 nowhere makes explicit use of the velocity initial datum, y(0) = y0 .
This raises several questions:
1. Why is that?
2. How will we account for this second initial condition in the present
expansion scheme?
3. What kind of initial data would be required to make the initial velocity
show up in the leading-order solution?
3.5 Some comments and a question

Multiple-Scales Asymptotics is
• a direct, constructive approach to approximate model reduction,
• first of all a formal approach; rigorous justification of any specific
asymptotic expansion must generally be handled on a case-by-case
basis,
• a means to systematically describe scale interactions,
Multiple-Scales Asymptotics is not
• a techique for deriving “the” general solution to any given pde- or
related problem,
• a cure to all multiscale problems; problems involving a continuous
range of interacting scales as observed, e.g., in turbulent flows, cannot
be handled by multiple scales asymptotics (at least not in a straight-
forward fashion.
Having read this section, how would you describe a “scale”, and how would
you mathematically describe scale-separation in a class of problems?
4 Universal parameters, distinguished limits, and

non-dimensionalization
One important aim of theoretical meteorology is the development of sim-
plified model equations that describe the large variety of scale-dependent
phenomena observed in atmospheric flows. Here we summarize the basic
scaling arguments that justify a unified approach to the derivation of such
models based on multiple scales asymptotic techniques. The approach was
proposed by Klein (2004) and has led to or been an important part of sev-
eral recent new developments, Majda and Klein (2003), Klein et al. (2004),
Majda and Biello (2004), Mikusky et al. (2005), Biello and Majda (2006),
Klein and Majda (2006), and references therein. Remarkably, Keller and
Ting (1951) already anticipated the foundations of this approach in an in-
ternal report of the Institute for Mathematics and Mechanics of New York
University.
To elucidate our main points, we restrict the discussion here to inviscid
compressible flows on a rotating sphere. Diabatic effects, such as radia-
tion, water phase transitions, or turbulent transport will be represented as
lumped terms in the governing equations to be specified later. Extensions
of the framework to include moist processes have been developed by Klein
and Majda (2006).
4.1 Universal parameters and distinguished limits

Table 1 displays several physical variables that are characteristic of at-
mospheric flows, and that are valid independently of the typical length and
time scales of any specific flow phenomenon: The mean sea level pressure pref
is set by the requirement that it balance the weight of a vertical column of
air. Thus, it is directly given by the total mass of the atmosphere which, to a
very good approximation, is evenly distributed over the sphere. A reference
temperature Tref is set roughly by the global radiation balance which, even
172 R. Klein et al.
Earth’s radius a = 6 · 106 m

Earth’s rotation rate Ω ∼ 10−4 s−1
Acceleration of gravity g = 9.81 ms−2
Sea level pressure pref = 105 Pa
Water freezing temperature Tref ∼ 273 K
Equator–pole temperature difference ΔT |peq ∼ 50 K
Dry air gas constant R = 287 ms−2 /K
Dry air isentropic exponent γ = 1.4
Table 1. Universal characteristics of atmospheric motions.
without greenhouse gases, would render the mean near-surface air temper-
ature near 250K. The actual value in Table 1 is the freezing temperature of
water under standard conditions, i.e., Tref ∼ 273K, which is about midway
between the observed maximal and minimal near-surface air temperatures.
The equator-to-pole air temperature difference near the surface, ΔT |peq , is
a consequence of the latitudinal variation of the sun’s irradiation. The dry
air gas constant, R, and isentropic exponent, γ, as thermodynamic proper-
ties are also universally characteristic for atmospheric flows, because their
variations due to admixtures of water vapor, trace gases, and the like are
no larger than a few percent in general.
Based on these eight basic reference quantities, four independent di-
mensionless combinations can be composed. To define combinations with
intuitive interpretations, we introduce as auxiliary quantities the pressure
scale height, hsc , the characteristic speed cref of barotropic1 gravity waves,
and a reference density, ref , via
hsc = pref/(g ρref ) ∼ 10 km

√
cref = ghsc ∼ 300 ms−1 (162)
−3
ρref = pref /(R Tref ) ∼ 1 kgm .
1
Atmospheric flow modes are called “barotropic” if their structure is homogeneous in
the vertical direction.
Then we let
hsc
Π1 = ∼ 1.67 · 10−3 ,
a
ΔT |peq (163)
Π2 = ∼ 0.18 ,
Tref
cref
Π3 = ∼ 0.5 .
Ωa
The interpretations of Π1 and Π2 should be obvious, while the parameter

Π3 compares a typical barotropic gravity wave speed with the tangential
speed of points on the equator as induced by Earth’s rotation.

Remark: The sound √ speed, γpref /ρref is comparable to the barotropic
wave speed, cref = ghsc according to (162).
The parameter Π1 is definitely quite small. Π2 is not extremely small,
yet, many successful developments in theoretical meteorology have relied on
scale analysis (asymptotics) in terms of, e.g., Rossby numbers or internal
wave Froude numbers with values in a similar range. Finally, for Π3 one
may argue that, even though it is less than unity, one may be hard pressed
to consider it “asymptotically small”. Deviating somewhat from our earlier
work cited above, we will consider Π3 1 in the present notes.
There is little hope for success with asymptotic expansions that would
allow Π1 , Π2 , and Π3 to vary independently in a limit process: even for the
simple example of a linear oscillator such expansions in two independent
parameters were found in section 3 not to exist! Thus, for the present
parameters we need to adopt a distinguished limit, and we investigate the
following scaling relationships below,
√
Π1 ∼ ε3 , Π2 ∼ ε , Π3 ∼ ε , as ε → 0. (164)
These limits are compatible with the estimates in (163) for actual values
of ε ∈ [1/7 . . . 1/8]. We will adopt ε as the reference expansion parame-
ter for asymptotic analyses below, and any additional small or large non-
dimensional parameter that may be associated with singular perturbations
in the governing equations is subsequently tied to ε through suitable further
distinguished limits.
Remark: Before we proceed to do so, we notice that Keller and Ting
(1951) already proposed to use the acceleration ratio, ε ∼ (aΩ2 /g)1/3 =
(Π1 /Π23 )1/3 , as a basic expansion parameter for meteorological modelling.
When Π3 = O(1), this is equivalent to (164) above.
174 R. Klein et al.
Remark: In contrast to (164), in my earlier work I have usually let Π3 =

O(1). The present, slightly modified limit I introduce because it appears to
unify current developments of planetary balanced models by my colleague
Stamen Dolaptchiev with Pedlosky’s derivations of the quasi-geostrophic
theory in (Pedlosky (1987)).
4.2 Nondimensionalization and general multiple scales ansatz

With pref and Tref , and through the ideal gas equation of state, ρ =
p/RT , Table 1 immediately suggests reference values for the nondimension-
alization of pressure, temperature, and density. But what about velocity,
length, and time?
Hydrostatic–geostrophic velocity scale Most theories for atmospheric

flows rely on the assumption that typical flow speeds are small compared
with the speed of barotropic gravity waves cref which, except for a factor
√
of γ, matches the speed of sound. Here and in the rest of this section we
make this assumption explicit by introducing a reference speed
ghsc ΔT |peq 3
uref = = cref Π2 Π3 ∼ ε 2 cref (165)
Ωa Tref
for the nondimensionalization of the flow velocity. (What is a typical value

for uref ?) The reader may verify that the above expression has in fact the
dimension of a velocity, but what is the motivation for this choice? We will
resolve this question later when we (re-)derive the quasi-geostrophic theory
(QG). (See also the Remark at the end of this section!)
The choice of a velocity scale in (165) allows us to express two classical
non-dimensional parameters of theoretical meteorology (and fluid dynam-
ics), the (barotropic) Froude and Mach numbers, in terms of our small
parameter,
M uref 3
Fr = √ = ∼ ε2 as ε → 0. (166)
γ cref
Scaling of space and time As we are interested in multiple scales

problems, and will consistently take into account different characteristic
lengths in our analyses, the specific choice of reference length and time
scales for non-dimensionalization should not make much of a difference.
We opt here to use, hsc , i.e., the smallest length scale that suggests itself
just from the fundamental parameters in Table 1 via equation (162), to
non-dimensionalize all lengths. The associated advection time serves as a
reference time. Thus,

hsc
ref = hsc and tref = . (167)
uref
Questions
• What would be the scaling in terms of ε of the Froude number based
on a typical internal gravity wave speed∗ ?
• What would be the scaling of the Rossby number based on our refer-
ence length, hsc ∗ ?
• By what power of 1/ε is the internal Rossby radius larger than hsc ∗ ?
• By what power of 1/ε is the Obhoukhov or external Rossby radius
larger than hsc ?
• By what power of 1/ε is the Obhoukhov or external Rossby radius
larger than the internal Rossby radius? — Compare your result with
a related remark in Pedlosky (1987).
• Does the Oboukhov or external Rossby radius come out larger, com-
parable or smaller than the Earth radius, a, which is representative of
the planetary scale?
∗
It is safe to assume that the variation of potential temperature across
the troposphere is comparable to the equator-pole temperature difference,
ΔT |peq .
4.3 Scaled governing equations and general multiple scales ex-

pansion scheme
With these scalings, the nondimensional governing equations in the ro-
tating earth system may be written as
1 1
v ,t + v · ∇ v + wv ,z + ε 2(Ω × v) +
∇ p = Qv ,
ε3 ρ
1 1 1
wt + v · ∇ w + wwz + ε 2(Ω × v)⊥ + 3 pz = Qw − ,
ε ρ ε3
pt + v · ∇ p + wpz + γp ∇ · v = Qp ,
Θt + v · ∇ Θ + wΘz = QΘ ,
(168)
where
p1/γ
Θ= (169)
ρ
is the dimensionless potential temperature, k is the local vertical unit vector
indicating the direction of the acceleration of gravity. The terms Qv , Qp ,
176 R. Klein et al.
and QΘ represent additional effects which in a concrete application may

stem from turbulence closures or similar models for the net influence of
non-resolved scales.
Klein (2004) suggested to consider the small parameter ε as introduced
above as the general singular asymptotic expansion parameter for theoret-
ical developments in meteorology (although suggesting a slightly different
distinguished limit for Π3 from (163)). To this end, the solution vector
U = (p, Θ, v) is expanded in powers of ε (or some fractional power thereof),
and all expansion functions would depend on a series of space-time coor-
dinates that are scaled again by powers of ε. The most straightforward
version of such a scheme reads
t x z
U(u, z, t, ; ε) = εi U (i) (. . . , , t, εt, . . . , , x, εx, . . . , z, . . . ) . (170)
i
ε ε ε
In practical applications it might be necessary to work with fractional pow-

ers of ε for the scaling of the coordinates, or more general asymptotic se-
quences, φ(i) (ε), as explained in the context of (121) in section 3.
In a number of publications, e.g., Majda and Klein (2003); Klein (2004);
Klein et al. (2004); Klein (2005), we have demonstrated that a wide range
of known simplified model equations of theoretical meteorology can be red-
erived in a unified fashion starting from the full compressible flow equations
in (168) and suitable specializations of the multi-scale ansatz in (170). To
derive a typical existing model, one would maintain one scaled time, one
scaled horizontal coordinate, and one pair of scaled vertical coordinates,
respectively, and this we consider a welcome “validation” of the approach.
We will demonstrate the procedure in the next chapter.
Of course, such “validation studies” are but a first step, as (170) strongly
suggests itself as the basis for systematic studies of multiple scales interac-
tions. See Majda and Klein (2003), Majda and Biello (2004), Klein et al.
(2004), Biello and Majda (2006), Klein and Majda (2006) for related devel-
opments.
Remark: The particular choice of a reference velocity in (165) does in no way
restrict our degrees of freedom in constructing simplified asymptotic models.
If, for example, we were to consider flows that are inherently compressible,
so that systematically |v| ∼ cref , then the asymptotic expansion scheme for
the (dimensionless) flow velocity would simply have to read

1 u z t
i (i)
v= 3 ε v (. . . , , u, εu, . . . , z, . . . , t, εt, . . . ) . (171)
ε2 i
ε ε ε
5 (Re-)derivation of the quasi-geostrophic (QG)

theory
In this chapter, we employ the general asymptotics-based approach from
section 4 to rederive the quasi-geostrophic model, see Pedlosky (1987).
5.1 Asymptotic expansion scheme

For the derivation of this theory, we will take the dimensionless form of
the compressible flow equations from (168) as our point of departure. For
simplicity of the exposition we assume adiabatic flow, dropping the source
and transport terms in the governing equations. Then (168) simplifies to
t + v · ∇ + wz + (∇ · v + wz ) = 0,
1 1
v ,t + v · ∇ v + wv ,z + ε 2(Ω × v) + ∇ p = 0,
ε3 ρ (172)
1 1 1
wt + v · ∇ w + wwz + ε 2(Ω × v)⊥ + 3 pz = − ,
ε ρ ε3
Θt + v · ∇ Θ + wΘz = 0,
and
p1/γ
Θ= . (173)
ρ
The quasi-geostrophic theory is designed to address flows on length scales
comparable to the internal Rossby radius, and on time scales corresponding
to horizontal advection across such distances. How can we access these
length and time scales within our multiple scales asymptotic scheme?
The internal Rossby radius is defined as the distance which an internal
gravity wave would travel during a characteristic Earth rotation time. This
is equivalent to requiring
N hsc
L Ro = , (174)
Ω
where
g ∂Θ
N= (175)
Θ ∂z
is the so-called Brunt-Väisälä or buoyancy frequency, and

hsc ∂Θ hsc ∂Θ
N hsc = ghsc = cref (176)
Θ ∂z Θ ∂z
is a typical travelling speed of internal gravity waves. The reader may want
to consult the established literature for corroboration.
178 R. Klein et al.
Remark: Here and below, primes mark dimensional variables!

Remark: Another interpretation of the internal Rossby radius considers it
the characteristic distance which an internal gravity wave would have to
travel to become affected by the Coriolis effect.
Non-dimensionalizing LRo by our reference length, ref = hsc , and using
the above we find
0
LRo N cref hsc ∂Θ cref a ΔT |peq
= = ∼ = Π 3 Π 1 Π2 (177)
hsc Ω Ωhsc Θ ∂z Ωa hsc Tref
or, using the distinguished limits introduced earlier for the Πi ,

LRo 1 1 1 1
=O ε 3ε 2 2 =O 2 , (178)
hsc ε ε
Here we have used the scalings of our fundamental parameters from Table 1
as discussed in (162)–(164), and the well established observation, [Schneider
(2006); Frierson (2008)], that the equator-to-pole temperature differences
are comparable to the vertical potential temperature variations across the
troposphere, so that
hsc ∂Θ ΔT |peq
∼ . (179)
Θ ∂z Tref
With the estimate in (178), if we want to describe horizontal variations
on scales comparable to LRo , we should use the dimensionless horizontal
coordinate
x hsc x
ξ= = = ε2 x . (180)
LRo LRo hsc
We will be interested here in phenomena associated with advection over
distances of LRo , so we will use the time variable
t hsc t
τ= = = ε2 t . (181)
LRo /uref LRo hsc /uref
Finally, in order to study phenomena which occupy the full depth of the
troposphere, we will use a vertical coordinate non-dimensionalized by the
pressure scale height, hsc , i.e., we use our original dimensionless coordinate
z
z= . (182)
hsc
These scalings are summarized in Fig. 14.
1
LRo ∼ h
ε2 sc
hsc
t ∼ LRo
vref
∼ hsc
ε2 vref
Figure 14. Length and time scales addressed by quasi-geostrophic theory.
Our asymptotic expansion scheme for the solution written in terms of

non-dimensional variables will thus read

U(t, x, z; ε) = εi U (i) (ε2 t, ε2 x, z) , U = (p, Θ, v , w)t , (183)
i
which is the announced specialization of the general multiple scales expan-

sion scheme in (170) adapted to resolve advection phenomena on the length
scale of the internal Rossby radius.
5.2 Some preliminaries

When inserting this expansion into the governing equations in (172) we
will have to account for the following transformation rules for the partial
derivatives,

∂ 2 ∂

=ε , ∇x |t,z;ε = ε2 ∇ξ |τ,z;ε . (184)
∂t x,z;ε ∂τ ξ,z;ε
Here the subscripts indicate which variables are to be held constant when
carrying out the partial differentiations.
We also anticipate the following properties of the background stratifica-
tion of the atmosphere in order to save us some tedious calculations:

(t, x, z; ε) = 0 (z) + ε1 (z) + ε2 (2) (τ, ξ, z) + O ε2 ,

p(t, x, z; ε) = p0 (z) + εp1 (z) + ε2 p(2) (τ, ξ, z) + O ε2 ,
(185)
Θ(t, x, z; ε) = 1 + εΘ1 (z) + ε2 Θ(2) (τ, ξ, z) + O ε2 ,

w(t, x, z; ε) = ε3 w(3) (τ, ξ, z) + O ε3 .
Remark: That the leading-order thermodynamic variables are independent
of time and do not vary horizontally can actually be derived within the
180 R. Klein et al.
present framework rather than having to be anticipated. The same is true

for the vanishing of the leading two orders of vertical velocity, w(0) , w(1) .
Remark: The leading-order potential temperature must be a constant, be-
cause of the order-of-magnitude analyses of the previous subsection which
restrict variations of potential temperature to ΔΘ/Tref = O(ε). We may set
this leading-order constant to Θ(0) ≡ 1 by choosing an appropriate reference
temperature.
5.3 Expansions of the governing equations

The next steps are standard procedure. We insert the expansion scheme,
collect like powers of ε, and separately equate the sum of these terms to zero,
so as to create the usual hierarchy of perturbation equations.
Mass conservation Expanding the mass conservation law, (172)1 , we

find
(0)
O ε0 : 0 ∇ξ · v = 0,
(186)
(1) ∂
O(ε) : 0 ∇ξ · v + 0 w(3) = 0.
∂z
(0)
In writing down the terms of O(ε) we have already neglected (1) ∇ξ · v on
account of (186).
Horizontal momentum balance
Splitting the Coriolis term Before expanding the momentum balances,

we need to explicitly split the Coriolis term into its horizontal a vertical
components.
Ω×v = (Ω + kΩ⊥ ) × (v + kw)
= (Ω × v ) + (Ω⊥ k × v + wΩ × k) + (Ω⊥ wk × k) . (187)

1 23 4 1 23 4 1 23 4
=(Ω×v)⊥ =(Ω×v) =0, as k×k=0
We will need the vertical component of Ω (see Fig. 15), which we expand
Pol
Ω
k
Ω0
eΩ ϑ
eeq Äquator
Figure 15. Splitting of the coriolis term into a horizontal and a vertical
component.
as
Ω⊥ = k·Ω
= (eeq cos ϑ + eΩ sin ϑ) · eΩ |Ω|
= |Ω| sin ϑ

y
= |Ω| sin ϑ0 + (188)
a
= |Ω| sin(ϑ0 + ε ξ2 )
= |Ω| sin(ϑ0 ) +ε |Ω| cos(ϑ0 ) ξ2 + O(ε)
1 23 4 1 23 4
=:Ω0 =:β
= Ω0 + εβξ2 + O(ε) .
Here we have taken into account that ϑ is the arclength along a longitudinal
circle divided by the radius of the reference sphere, a, introduced deviations
from a reference latitude, so that ϑ = ϑ0 +y /a, and recalled that hsc /a = ε3
and ξ2 = ε2 y /hsc . The rest is Taylor expansion of the sine function about
the reference latitude.
Since w(0) ≡ w(1) ≡ 0 we also know that wΩ × k = O(ε) and, in
182 R. Klein et al.
summary, we find
(Ω × v) = (Ω0 + εβξ2 )k × v + O(ε) ,
(189)
(Ω × v)⊥ = Ω × v .
Expansion of the horizontal momentum balance Consider now the

horizontal momentum balance, written in terms of the new variables, (τ, ξ, z),
1 1 ∇ξ p 1
v τ + (v · ∇ξ )v + 2
wv z + 3 + (Ω̂ × v) = 0 . (190)
ε ε ε
Using w(0) ≡ w(1) ≡ 0, and that p(0) ≡ p0 (z) and p(1) ≡ p1 (z), we im-
mediately move to the equation at O ε−1 where we find the geostrophic
balance,
(0) p(2)
Ω0 k × v + ∇ξ π (2) = 0 , where π (2) = , (191)
0
i.e., the balance of the horizontal Coriolis and pressure gradient forces.
Geostrophic balance implies that, at leading order, the horizontal flow di-
rection is perpendicular to the horizontal pressure gradient.
We verify for later purposes that
(0) (0) 1
∇ξ · v =0 and v = k × ∇ξ π (2) (192)
Ω0
by, respectively, applying (k · (∇ξ × [ · ])) and (k × [ · ]) to (191)1 .
Remark: With (191) we have found a time-independent constraint on the
leading order velocity and second order pressure fields. Such constraints
did not exist for the original system of the compressible flow equations!
The constraint implies that only if the initial data for a given flow problem
satisfy the constraint, at least at the given orders, can we hope that the
approximate asymptotic solution will remain close to the exact solution.
This kind of “change of type” of the asymptotic limit problem relative to
the original one is typical of singular perturbation problems. Also, we
recall that we encountered a similar issue with the slow-time expansion for
the linear oscillator in chapter 3.
Vertical momentum balance From the vertical momentum balance we

obtain at orders ε−5 to ε−2
∂p(i)
= −(i) (i = 0, 1, 2, 3) . (193)
∂z
Expanding the defining equation for the potential temperature, i.e., Θ =

1
p γ into
1
O ε0 : 0 = p0 γ
1 p1
O(ε) : 1 + 0 Θ1 = p0 γ (194)
γp0
(2)
1 p (1 − γ)p1 2
O ε2 : (2) + 1 Θ1 + 0 Θ(2) = p0 γ +
γp0 2γ 2 p0 2
we obtain from (193)

− γ1 ∂p(0)
p(0) = −1 (195)
∂z
with the exact solution
γ−1
γ
γ−1
p0 (z) = 1− z . (196)
γ
In a similar way one solves the first order equation explicitly for given Θ1 (z).
We leave this as an exercise.
Evolution of the potential temperature The first non-trivial asymp-

totic equation is extracted from the potential temperature transport equa-
tion at O(ε), and it reads

∂ (0) dΘ1
+ v · ∇ξ Θ(2) + w(3) = 0. (197)
∂τ dz
5.4 Summary of the leading-order balances

Using the expansion scheme in (185), we have found first that the back-
ground structure is in hydrostatic balance, i.e.,
dpi
= −i (i = 0, 1) . (198)
dz
The remaining primary unknowns for description of the flow field are
then
(0)
π (2) , v , w(3) , Θ(2) (τ, ξ, z) , (199)
where π (2) = p(2) /0 , and they satisfy the following balance and transport
equations:
184 R. Klein et al.
Hydrostatic Balance
∂π (2)
= Θ(2) (200)
∂z
Geostrophic Balance
(0)
Ω0 k × v + ∇ξ π (2) = 0 (201)
Anelastic Constraint
∂
(1)
0 ∇ξ · v + 0 w(3) = 0 (202)
∂z
Potential Temperature Transport

∂ (0) dΘ1
+ v · ∇ξ Θ(2) + w(3) =0 (203)
∂τ dz
(1)
If it were not for the appearance of the first-order divergence, ∇ξ · v in
(202), we would have the same number of equations as we have unknowns.
As it is, the system is as yet unclosed. We will extract additional information
(1)
on ∇ξ · v from the next higher order horizontal momentum equation in the
next section in the form of a solvability condition that may be interpreted
as a vorticity transport equation,
First-Order Solvability Condition / Vorticity Transport Equation

(0) (1)
∂τ + v · ∇ξ (ζ (0) + βξ2 ) + Ω0 ∇ξ · v = 0 . (204)
where,
(0)
ζ (0) = k · (∇ξ × v ) , (205)
is the vorticity of the leading-order velocity field.
This completes the summary of the quasi-geostrophic model equations.
Remark: We have given the QG equations here in a somewhat unusual
form, sticking as closely as possible to the original equations. In this way, it
remains transparent that (200) and (201) are direct consequences of the
vertical and horizontal momentum balances, respectively, (202) emerges
from mass conservation, and (203) from the potential temperature transport
equation. These equations can all directly be read off the original equations
at the appropriate orders in the asymptotic expansion.
(1)
closure for ∇ξ · v in (204) emerges as a solvability condition at
The
0
O ε in the horizontal momentum balance as will be shown in the next
section.
Remark: As can see clearly in the present summary of our asymptotic limit
equations, considering large spacial and long time scales only implies strong
constraints on the solutions. Instead of evolution equations for the primary
unknowns in the compressible flow equations (the densities of mass, momen-
tum, and energy) we find three time independent constraints or balances!
Only the potential temperature evolution equation in (203) and the vortic-
ity transport equation in (204) have maintained the original “prognostic”
(time evolution) character.
5.5 First-order solvability condition / existence of ∇ξ p(3)

Consider the scaled horizontal momentum balance from (190), which we
had already written in terms of our new coordinates, (τ, ξ, z), at O ε0 ,

(3)
(0) (0) (0) ∇ξ p (1) (0)
(v )τ + (v · ∇ξ )v + + Ω0 k × v + βξ2 k × v = 0 . (206)

Using the fact that 0 , 1 depend on z only, so that ∇ξ 0 = ∇ξ 1 = 0, and
π (i) = p(i) /0 we rewrite the pressure gradient term as

(3)
∇ξ p 1 1 1 (2)
= ∇ξ p(3) − 2 ∇ξ p(2) = ∇ξ π (3) − ∇ξ π . (207)
0 0 0
Next we regroup (206) into first-order “geostrophic terms” on left-hand side

and terms that distort the geostrophic balance, i.e., “ageostrophic terms”,
on the right,
(1)
∇ξ π (3) + Ω0 k × v =
(208)
(0) (0) (0) (0) 1 (2)
− (v )τ + (v · ∇ξ )v + βξ2 k × v − ∇ξ π .
0
We know that any gradient of a scalar is curl-free. In particular,
k · (∇ξ × ∇ξ φ) ≡ 0 (209)
for any scalar function φ(ξ) that is sufficiently smooth. By applying the
operator k · (∇ξ × [ · ]) to (208) we thus eliminate the terms involving π (2)
186 R. Klein et al.
and π (3) . The remaining terms become

(1) (1)
k · ∇ξ × Ω0 k × v = Ω0 k · k ∇ξ · v
(1)
= Ω0 ∇ξ · v ,

(0)
k · ∇ξ × (v )τ = ζτ(0) ,

(0) (0) (0)
k · ∇ξ × (v · ∇ξ )v = v · ∇ξ ζ (0) , (210)

(0) (0)
k · ∇ξ × βξ2 k × v = β k · k ∇ξ ξ2 · v
= βv (0)

∂ (0)
= β + v · ∇ξ ξ2 .
∂τ
Collecting, we find the vorticity transport equation,

(0) (1)
∂τ + v · ∇ξ (ζ (0) + βξ2 ) + Ω0 ∇ξ · v = 0 , (211)
as announced in (204).
5.6 Classical formulation of the QG-theory and PV transport

In (201)–(204) we have taken care to display the quasi-geostrophic bal-
ance equations in a form that reveals their close connection to the mass,
momentum, and potential temperature evolution equations. This may not
be the most practicable description in many applications, and it hides the
central role of potential vorticity (PV) in the quasi-geostrophic regime.
In fact, one can rewrite (211) as a transport equation for the QG-
potential vorticity,

Ω0 ∂ 0 Θ(2)
q = ζ (0) + β ξ2 + with Θ1 = dΘ1 /dz , (212)
0 ∂z Θ1
which then reads

(0)
∂τ + v · ∇ξ q = 0 . (213)
Equipped with the additional constitutive relations
(0) 1
v = k × ∇ξ p(2) ,
Ω0
∂π (2)
Θ(2) = , (214)
∂z
(0) 1 2 (2)
ζ (0) = k · ∇ξ × v = ∇ π
Ω0 ξ
we have the QG theory in its classical form (Pedlosky (1987)): Equation

(213) describes advection of potential vorticity by the leading order ve-
(0)
locity field v , which can be expressed in terms of the pressure gradient
(2)
∇ξ π according to (214)1 . Given the (advected) PV-field, one can retrieve
the pressure field solving the ellliptic equation that results from inserting
(214)2,3 into (212), viz.

Ω2 ∂ 0 ∂π (2)
2 (2)
∇ξ π + 0 = q − β ξ2 . (215)
0 ∂z Θ1 ∂z
Appendix
Gauß’ Integral Theorem
Let Ω ⊂ IR be a compact subset with a smooth boundary, n : ∂Ω → IRn
the field of outer unit normal vectors and U ⊃ Ω an open subset of IRn . Then
for every continuous differentiable vector field F : U → IRn the following is
true
∇ · F (x) dV = F (x) · n dσ
Ω ∂Ω
Proof see (Forster, 1984, S. 155)
The symbols O() and O() (Landau’s symbols)

The symbol O() is used in this text in two ways. One formulation is
based on the so called Landau symbol. Here, for functions f and g the
equity holds
f (x) = O(g(x)) as x → a
if and only if f (x)/g(x) → const. for x → a (in the asymptotic sense). On
the other hand the symbol is often used in the way that the statement the
”
quantity X is O()“ means that X is of the same order of magnitude as .
The particular meaning of O( · ) will be clear from the context.
188 R. Klein et al.
The second Landau symbol O( · ) (“little oh”). For functions f and g we

have
f (x) = O(g(x)) as x → a ,
if and only if f (x)/g(x) → 0 for x → a.

If for some function f we not only have f (x) = O(g(x)) but also g(x) =
O(f (x)) as x → a, then f (x) = OS (g(x)) as x → a.
Vector identities
For vectors A, B, C ∈ IRn and scalars ϕ, ψ ∈ IR the following general
identities are true :
A × (B × C) = (C × B) × A = B(A · C) − C(A · B) (216)
∇ · (ϕ A) = ϕ∇ · A + A · ∇ϕ (217)
∇ × (ϕ A) = ϕ∇ × A + ∇ϕ × A (218)
∇ · (A × B) = B ·∇× A − A·∇× B (219)
∇ × (A × B) = A(∇ · B) − B(∇ · A) + (B · ∇)A

− (A · ∇)B (220)
∇(A · B) = A × (∇ × B) + B × (∇ × A)
+ (A · ∇)B + (B · ∇)A (221)
∇ × ∇ϕ = 0 (222)
∇·∇ × A = div(rotA) = 0 (223)
∇ · (A ◦ B) = (∇ · A)B + (A · ∇)B (224)
Pressure scale height

The atmosphere is in hydrostatic balance if the vertical pressure gradient
is equal to the gravity acceleration, i.e.,
∂p
= −g
∂z
Starting from a reference pressure pref (e.g. a mean pressure on sea level)
we can define the pressure scale height by the height difference at which the
pressure in an atmosphere in hydrostatic balance and with constant density
changes of an order of magnitude of pref
pref
hsc := .
ref g
The thermal wind

The equation of the thermal wind
(0)
∂v 1
−Ω0 k × = ∇ Θ(3)
∂z Θ∞
does not give any information about the geostrophic wind itself, but only
about it’s vertical variation. The thermal wind denotes the velocity dif-
ferences that result from geostrophic balance across some vertical distance,
say, z : v T = v = v(z1 ) − v(z2 ).
Some details of the transformations into a rotating reference frame

We switch from IR3 to a Cartesian system of coordinates Z = IR × IR × IR
in which Ω and X b (t) are defined by
⎛ ⎞ ⎛ ⎞
Ω1 Xb1
Ω := ⎝ Ω2 ⎠ , X b (t) := ⎝ Xb2 ⎠ .
Ω3 Xb3
With the additional definition of
⎛ ⎞
0 −Ω3 Ω2
Ω := ⎝ Ω3 0 −Ω1 ⎠
−Ω2 Ω1 0
the change with time of X b (t) can also be written as a matrix vector prod-
uct, namely
⎛ ⎞
−Ω3 Xb2 (t) + Ω2 Xb3 (t)
Ẋ b (t) = Ω × X b (t) = Ω X b (t) = ⎝ Ω3 Xb1 (t) − Ω1 Xb3 (t) ⎠ . (225)
−Ω2 Xb1 (t) + Ω1 Xb2 (t)
With the knowledge from the theory of ordinary differential equations
we can derive equation (37) in a different way. We know that the solution
of the initial value problem
Ẋ b (t) = Ω X b (t) , X b (0) = X b0
190 R. Klein et al.
is
∞ ν
t
X b (t) = exp(Ωt) X b (0) , wobei exp(Ωt) := (Ω)ν .
ν=0
ν!
Because of
⎛ ⎞
−Ω22 Ω23 Ω21 Ω22 Ω21 Ω23
2
Ω = ⎝ Ω21 Ω22 −Ω21 Ω23 Ω22 Ω23 ⎠ = |Ω|2 (eΩ eTΩ − 1) ,
−Ω21 Ω23 Ω22 Ω23 −Ω21 Ω22
Ω3 = −|Ω|2 Ω , Ω4 = −|Ω|4 (eΩ eTΩ −1) , Ω5 = −|Ω|2 Ω3 = |Ω|4 Ω etc.

Ω

With eΩ = |Ω|
and |Ω| = Ω21 + Ω22 + Ω23 we then have

t2 t3
X b (t) = 1+t Ω+ |Ω|2 (eΩ eTΩ − 1) − |Ω|2 Ω
2! 3!

t4 4 t 5
4
− |Ω| (eΩ eΩ − 1) + |Ω| Ω + ... X b (0)
T
4! 5!
2 2

|Ω| t |Ω|4 t4
= (eΩ eΩ ) + 1 −
T
+ − ... (1 − eΩ eTΩ )
2! 4!

|Ω|3 t3 |Ω|5 t5 Ω
+ |Ω|t − + − ... X b (0)
3! 5! |Ω|

Ω
= (eΩ eTΩ ) + cos(|Ω|t)(1 − eΩ eTΩ ) + sin(|Ω|t) X b (0)
|Ω|
and this, because of the representation in coordinates, is just the same as

the previously derived result from (37). As (225), Ω X b can be represented
as Ω × X b and we obtain
X b (t) = (eTΩ X b (0))eΩ +cos(|Ω|t)(1−eΩ eTΩ )X b (0)+sin(|Ω|t)(X b (0)×eΩ ) .
Defining R(t) := exp(Ωt), R(t) has the following properties:

• R(−t) = RT (t)
As Ω is skew symmetric, i.e., it is ΩT = −Ω, and as the cos( · ) is an
even while the sin( · ) is an uneven function, we have
ΩT
RT (t) = (eΩ eTΩ )T + cos(|Ω|t)(1 − eΩ eTΩ )T + sin(|Ω|t)
|Ω|
−Ω
= (eΩ eTΩ ) + cos(−|Ω|t)(1 − eΩ eTΩ ) − sin(−|Ω|t) = R(−t)
|Ω|
• R(−t) = R−1 (t)

We will now analyze how the different terms of our conservation laws
change when we switch to a rotating coordinate system. To this end, we rep-
resent an arbitrary fixed vector with respect to the inertial basis {e1 , e2 , e3 }
as
⎛ ⎞ ⎞
⎛ ⎛ ⎞
1 0 0
e1 = ⎝ 0 ⎠ , e2 = ⎝ 1 ⎠ , e3 = ⎝ 0 ⎠
0 0 1
in the basis {ẽ1 , ẽ2 , ẽ3 } of the rotating coordinate system. The interrelation
between the coordinate systems is given in the following form :
ẽi = R(t)ei , i = 1, 2, 3 .
Then x can be represented by the two bases as
3
3
3

x= xi ei = x̃i ẽi = x̃i R(t)ei .
i=1 i=1 i=1
Multiplication from left with eTk yields
3
3

xk = x̃i (eTk R(t))ei = x̃i (R(t))ki .
i=1 i=1
Hence,
x = R(t)x̃ und x̃ = R−1 (t)x .
Furthermore we analyze the differential operators under the present a trans-

formation of coordinates. For the transformation (t, x) → (t̃, x̃) with t̃ = t
it holds for a function f that
f (t, x) = f˜(t̃, x̃) = f˜(t̃(t, x), x̃(t, x)) .

192 R. Klein et al.
This yields
˜ 3

∂f ∂ t̃ ∂f ∂ x̃i ∂ f˜
= +
∂t x ∂t x ∂ t̃ ∂t x ∂ x̃i
x̃ i=1 t̃,xj (j=i)
3
∂ f˜ ∂ f˜
= + (−Ω x̃)i
∂ t̃ i=1
∂ x̃i
3
∂ f˜ ∂ f˜ (226)
= + (−Ωij x̃j )
∂ t̃ i,j=1
∂ x̃i
3

∂ f˜ ∂ f˜
= + (x̃j Ωji )
∂ t̃ i,j=1
∂ x̃i
∂ f˜ ˜ f˜)T
= + x̃T Ω(∇
∂ t̃
and

∂f ∂ t̃ ∂ f˜
= +
∂xi t̃,xk (k=i) ∂xi t,xk (k=i) ∂ t̃
x̃
3
∂ x̃j ∂ f˜
+
∂xi t,xk (k=i) ∂ x̃j
j=1 t̃,x̃k (k=j)
3
∂ f˜
= (R−1 (t))ji
j=1
∂ x̃j
3
∂ f˜
= (R(t))ij
j=1
∂ x̃j
or
˜ f˜)T
(∇f )T = R(t)(∇ .
Consider now the transformation of the velocity of a particle: The position

of a particle in both coordinate systems can be expressed by
3
3

xp (t) = xpi (t)ei = x̃pj (t)ẽj (t)
i=1 j=1
and its velocity is
3 3

dxp
= ẋpi (t)ei = (x̃˙ pj (t)ẽj (t) + x̃pj (t)ẽ˙ j (t))
dt i=1 j=1
3

= ẋrel
p (t) + x̃pj (t)Ω ẽj (t)
j=1
= ẋrel
p (t) + Ω xp (t) .
Thus, if p denotes a particle of the fluid we look at, then ẋrel

p = vp
rel is the
relevant wind speed and the local flow velocity v(x, t) can be splitted
v = ẋ = v rel + Ω x
3
(227)
= ṽirel ẽi + Ω x
i=1
into a part arising from earth’s rotation and the relative wind speed. When
transforming the velocity divergence we find
∇ · v = ∇ · (v rel + Ω x) = ∇ · v rel + ∇ · (Ω x) ,
whereas because of the zeros on the diagonals of the rotation matrix Ω
3
3

∂
∇ · (Ω x) = (Ωij xj ) = Ωij δij = 0 ,
∂xi
i,j=1 i,j=1
(the symbol δij = 1 if i = j, 0 otherwise, denotes the Kronecker-Symbol).

∂ )3
As also eTi ẽj = eTi R ej = Rij and ∂x i
= j=1 Rij ∂∂x̃i , one has
3
3 3
∂v rel ∂
∇ · v rel = i
= T
(ei ṽjrel ẽj )
i=1
∂xi i=1
∂x i j=1
3
∂
= ṽjrel(t̃(t, x), x̃(t, x)) eTi ẽj
i,j=1
∂xi
3 (228)
∂ rel
= Rik ṽ Rij
∂ x̃k j
i,j,k=1
3 3
∂ rel ∂ṽkrel ˜ · ṽ rel
= δik ṽj = =∇ .
∂ x̃k ∂ x̃k
j,k=1 k=1
194 R. Klein et al.
Furthermore, because of R−1 (t) = RT (t) = R(−t) we have (RT R)kj =

)3
i=1 Rik Rij = δkj .
For momentum conservation the following auxiliary calculations might
be useful: Transforming (v t + (v T ∇T )v) into the rotating system of coor-
dinates and using (226) and (227) yields for the derivative with respect to
time

∂
vt = (v rel + Ω x)
∂t x
3

∂ rel ∂
= ṽi ẽi + Ωx
∂t x i=1 ∂t x
3
3
∂ṽ rel
= i ˜ rel )T ẽ +
+ x̃T Ω(∇ṽ ṽirel Ω ẽi + 0
i i
i=1
∂ t̃ i=1
3
∂ṽirel
= ẽi + x̃T Ω∇ ˜ T ṽ rel + Ω ṽ rel
i=1
∂t
Because of
3

T
(v rel ∇T )v rel = ((ṽ rel )T ∇
˜ T )ṽirel ẽi ,
i=1

˜ T )ṽ rel = −(x̃T Ω∇
(Ω x)T ∇T v rel = (x̃T ΩT ∇ ˜ T )ṽ rel ,
3
3
3
T ∂ ∂
(v rel ∇T )Ω x = virel Ω xi ei = ṽjrel (Ω ei x̃i )
i=1
∂xi i=1 i,j=1
∂ x̃j
3

= Ωṽjrel δij ei = Ω ṽ rel
i,j=1
and

(Ω x)T ∇T (Ω x) = −(xT Ω∇T )Ω x
3
3

∂
= − xi Ωik Ωjl xl ej = xi Ωki Ωjk ej
∂xk
i,j,k,l=1 i,j,k=1
T
= xT ΩT ΩT = Ω(Ω x) = Ω × (Ω × x)
we obtain

v t + (v T ∇T )v = v t + (v rel + Ω x)T ∇T v rel + Ω x
T
= v t + (v rel ∇T )v rel + (Ω x)T ∇T v rel
T
+(v rel ∇T )(Ω x) + (Ω x)T ∇T (Ω x)
3
3

∂ṽ rel
= i
ẽi + ((ṽ rel )T ∇
˜ T )ṽirel ẽi
i=1
∂t i=1
+2 Ω ṽ rel + Ω(Ω x)
Multiplication with ẽTk yields
∂ṽkrel
v t + (v T ∇)v k
= + ((ṽ rel )T ∇T )ṽkrel + 2 (Ω ṽ rel )k + Ω(Ω x) k
∂t
Bibliography
G. I. Barenblatt. Scaling, self-similarity, and intermediate asymptotics.
Cambridge Texts in Applied Mathematics. Cambridge University Press,
Cambridge, New York, Melbourne, 1996.
J.A. Biello and A.J. Majda. Transformations for temperature flux in mul-
tiscale models of the tropics. Theor. Comp. Fluid Dyn., under revision,
2006.
O. Forster. Analysis 3. Vieweg & Sohn Verlagsgesellschaft, Braunschweig,
Wiesbaden, 3. edition, 1984.
D. M. W. Frierson. Midlatitude static stability in simple and comprehensive
general circulation models. J. Atmos. Sci., 65:1049–1062, 2008.
A.E. Gill. Atmosphere-Ocean Dynamics, volume 30 of International geo-
physics series. Academic Press, New York, 1982.
H. Görtler. Dimensionsanalyse: Theorie der physikalischen Dimensio-
nen mit Anwendungen. Ingenieurwissenschaftliche Bibliothek. Springer,
Berlin, Heidelberg, New York, 1975.
J.B. Keller and L. Ting. Approximate equations for large scale atmospheric
motions. Internal Report, http://www.arxiv.org/abs/physics/0606114,
Inst. for Mathematics & Mechanics (renamed to Courant Institute of
Mathematical Sciences in 1962), NYU, 1951.
J. Kevorkian and J.D. Cole. Multiple Scale and Singular Perturbation Meth-
ods, volume 114 of Applied Mathematical Sciences. Springer-Verlag, New
York, 1996.
196 R. Klein et al.
R. Klein. An applied mathematical view of meteorological modelling. In

Applied Mathematics Entering the 21st century; Invited talks from the
ICIAM 2003 Congress, volume 116. SIAM Proceedings in Applied Math-
ematics, 2004.
R. Klein. Multiple spacial scales in engineering and atmospheric low mach
number flows. ESAIM: Math. Mod. & Num. Analysis (M2AN), 39:537–
559, 2005.
R. Klein. Scale-dependent models for atmospheric flows. Ann. Rev. Fluid
Mech., 42:249–274, 2010.
R. Klein and A.J. Majda. Systematic multiscale models for deep convection
on mesoscales. Theor. Comp. Fluid Dyn., accepted, 2006.
R. Klein, E. Mikusky, and A. Owinoh. Multiple scales asymptotics for
atmospheric flows. In A. Laptev, editor, 4th European Conference of
Mathematics, Stockholm, Sweden, pages 201–220. European Mathemat-
ical Society Publishing House, 2004.
D. Kröner. Numerical Schemes for Conservation Laws. J. Wiley & Sons
und B.G. Teubner-Verlag, Chichester, Stuttgart, 1997.
R.J. LeVeque. Numerical Methods for Conservation Laws. Lectures in
Mathematics, ETH Zürich. Birkhäuser Verlag, Basel, Boston, Berlin,
1990.
A. J. Majda and R. Klein. Systematic multiscale models for the tropics.
Journal of the Atmospheric Sciences, 60(2):393–408, 2003.
A.J. Majda and J.A. Biello. A multiscale model for tropical intraseasonal
oscillations. Proc. Natl. Acad. Sci. USA, 101:4736–4741, 2004.
E. Mikusky, A. Owinoh, and R. Klein. On the influence of diabatic effects
on the motion of 3D-mesoscale vortices within a baroclinic shear flow. In
Computational Fluid and Solid Mechanics 2005, pages 766–768. Elsevier,
Amsterdam, 2005.
J. Pedlosky. Geophysical Fluid Dynamics. Springer-Verlag, New York, 2.
edition, 1987.
T. Schneider. The general circulation of the atmosphere. Ann. Rev. Earth
Planet. Sci., 34:655–688, 2006.
W. Schneider. Mathematische Methoden der Strömungsmechanik. Vieweg
& Sohn Verlagsgesellschaft, Braunschweig, 1978.
R. Temam and A. Miranville. Mathematical Modeling in Continuum Me-
chanics. Cambridge University Press, Cambridge, UK, 2000.
W. Walter. Gewöhnliche Differentialgleichungen. Springer-Verlag, Berlin,
Heidelberg, New York, 6. edition, 1996.
D. Werner. Funktionalanalysis. Springer-Verlag, Berlin, Heidelberg, New
York, 3rd edition, 2000.
Multiple Scales Analysis of the Turbulent
Undular Hydraulic Jump
*
Abstract The undular hydraulic jump in turbulent open chan-
nel flow is considered in the double limit of very large Reynolds
numbers, and Froude numbers approaching the critical value, i.e.
F r = 1 + 32 ε with ε → 0+.
The undular jump is associated with a distinguished limit, which
is characterized by the similarity
√ parameters A and a. The square
root of the first parameter A is essentially the ratio of the dimen-
sionless friction velocity and the difference of the Froude number to
its critical value. The second parameter a is a scaled measure of
the difference of the incident turbulent flow to the fully developed
turbulent flow.
Since a wavy solution with a slowly varying amplitude is ex-
pected, a multiple scales expansion is performed. A new indepen-
dent variable is introduced such that the wave length becomes con-
stant and normalized to one. The perturbation equations of the
orders ε, ε3/2 , ε2 , and ε5/2 have to be considered in order to obtain
a complete first-order solution.
In case of fully developed incident flow analytical results for the
amplitude and wave length of the first wave are obtained. They
are compared with measured data and reasonable agreement is ob-
served.
1 Introduction
The undular hydraulic jump is a peculiar change of state that can be
observed in steady open channel flows if the upstream Froude number is
slightly above the critical value 1, cf. Böß (1927), Chow (1959), Hager and
Hutter (1984), Hager (1992), Henderson (1966), Chanson (1993), Chanson
and Montes (1995), Reinauer and Hager (1995), Ohtsu et al. (2001).
As far as the observations are concerned, the undular hydraulic jump is
characterized by a wavy shape of the free surface, with wave lengths that
© CISM, Udine 2010
198 H. Steinrück
are much larger than the depth of the liquid and amplitudes that decay
slowly in main flow direction.
From a theoretical point of view there is the difficulty that, though the
viscosity effects are very weak, an inviscid-flow solution does not exist, cf.
Benjamin and Lighthill (1954).
For laminar flow, Johnson (1972) was able to cope with these difficulties
and provide an asymptotic analysis in terms of Froude numbers near 1 and
large Reynolds numbers. His main result is a steady-state version of the
Korteweg-de Vries-Burgers equation that governs the perturbations of the
surface elevation.
For turbulent flow, attempts have been made to analyze the undular
hydraulic jump on the basis of various ad-hoc approximations, e.g. An-
dersen (1978), Hager and Hutter (1984), Kaufmann (1934), Lauffer (1935),
Chanson and Montes (1995). More detailed comments on those previous
investigations can be found in Grillhofer (2002).
In Grillhofer (2002), Grilllhofer and Schneider (2003) and Steinrück et al.
(2003) an asymptotic analysis of the undular jump in turbulent flow was
given. Plane flow over a bottom of constant slope was considered in the
double limit of very large Reynolds numbers, i.e. Reτ → ∞, and Froude
numbers approaching the critical value, i.e. F r = 1 + 32 ε with ε → 0.
Here we will follow the analysis in Steinrück et al. (2003) to study the
influence of small viscosity effects on near critical open channel flow. Based
on the shallow water approximation we will employ the method of multiple
scales to derive equations for the change of amplitude and wave length of
the undulated surface.
2 Governing equations and scaling

We consider a plane open channel flow of a fluid with constant density ρ and
constant kinematic viscosity ν over a plane bottom with constant angle of
inclination α, cf. Figure 1. The coordinate system is such that the x-axis is
in the bottom plane and the y-axis is perpendicular to it. The corresponding
velocity components are u and v, respectively.
As reference state we chose the volumetric mean velocity ūr , the height
of the free surface h̄r at some appropriate location xr and and wall shear
stress τwr of a locally fully developed open channel flow with volumetric
mean velocity ūr and height of the free surface h̄r .
Having a shallow water approximation in mind we choose a typical wave
length l to nondimensionalize the x-coordinate. The velocity components
are scaled accordingly. The pressure p is referred to the hydrostatic pressure
ρgh r at the bottom of the channel.
Multiple Scales Analysis of the Turbulent Hydraulic Jump 199
y g
ur l
hr
u(x r , y) h C
h(x)
x
α
Figure 1. The undular hydraulic jump in turbulent open channel flow
The Reynolds stresses (u 2 ), (u v ) and (v 2 ) are scaled with the square
of the reference friction velocity, which is given by:
τwr
uτ2 r = . (1)
ρ
We introduce the following dimensionless variables:
x y ū v̄ p̄ F r 2 h̄
X= , Y = , Ū = , V̄ = , P̄ = , H̄ = , (2a)
l hr ur δu r ρu2r hr
(u 2 ) (u v ) (v 2 ) uτ
(U 2 ) = , (U V ) = , (V 2 ) = , Uτ = , (2b)
uτ2 r uτ2 r uτ2 r uτ r
where the dimensionless parameters
ur hr uτ r
Fr = √ , δ= , α, γ= (3)
gh r l ūr
are the Froude number, the shallow water parameter, the angle of inclina-
tion and the ratio of the reference wall friction velocity and the volumetric
mean velocity, respectively. The later is a function of the Reynolds num-
ber Re = u r h r /ν only. If the fully developed flow in an open channel
with the inclination α is chosen as reference state the force balance yields
sin α = γ 2 F r 2 .
For high Reynolds numbers the flow field can be decomposed into a
defect layer and a viscous sub-layer. Since a universal solution exists for the
viscous sub-layer, cf. Kluwick (1998), Schlichting and Gersten (2000), it is
sufficient to consider the defect layer only.
200 H. Steinrück
Using the scalings (2a) and (2b), the continuity equation for the dimen-
sionless Reynolds averaged velocities becomes:
ŪX + V̄Y = 0. (4)
The momentum equations for the defect layer in dimensionless form are:
sin α γ2
F r2 Ū ŪX + V̄ ŪY = −P̄X + − F r2 δ(U 2 )X + (U V )Y , (5a)
δ δ

δ 2 F r 2 Ū V̄X + V̄ V̄Y = −P̄Y −cos α− F r2 γ 2 δ(U V )X + (V 2 )Y . (5b)
We define the averaged free surface Y = H̄(X) by a particular streamline

of the Reynolds averaged velocity field (ū, v̄). This particular streamline
corresponds to the averaged volume flow rate
H(X,Z,t) H̄(X)
V̇ := U (X, Y, t) dY = Ū(X, Y ) dY. (6)
0 0
Using the above definition of the mean free surface, the common kinematic
boundary condition for the averaged velocities holds:
V̄(X, H̄(X)) = Ū(X, H̄(X)) H̄X (X). (7)
In the absence of a general theory of turbulent free surfaces, the dynamic

boundary conditions are also formulated following common approaches, cf.
Rodi (1993). Thus

2 2 (U 2 ) (U V ) −δ H̄X δHX
γ Fr · = P̄ at Y = H̄.
(U V ) (V 2 ) 1 −1
(8)
2.1 The friction law

Matching with respect to the viscous sub-layer is accomplished by re-
quiring
V̄(X, 0) = 0 (9)
and, furthermore, making use of the logarithmic friction law (cf. Schlichting
and Gersten (2000), p. 524, eq. (17.54)), which may be written as follows:

1
ŪS (X) = γ Ūτ ln Reγ Ūτ H̄ + C + + C̄(X) , (10)
κ
where ŪS = Ū(X, H̄(X)) is the surface velocity, Ūτ2 = −(U V )Y =0 is the
dimensionless wall shear stress and the function C̄(X) is given by
H̄
1 dŪ 1
C̄(X) = − dY,
0 γUτ dY κY
where κ is von Karman’s constant. C + is another empirical constant, which,
as it will turn out, does not appear in the final results.
Let us denote ΔŪS = ŪS − ŪS r , ΔŪτ = Ūτ − 1, and ΔH̄ = H̄ − 1 the
difference of the dimensionless surface velocity, friction velocity and surface
height to its reference values at X = X r , respectively. Using the friction
law (10) the surface velocity ŪS r at the reference state, X = X r , is given
by
1
ŪS r = γ ln(Reγ) + C + + C̄(X r ) (11)
κ
and we can rewrite the friction law as

1
ΔŪS = ŪS r ΔŪτ + (1 + ΔŪτ )γ ln(1 + ΔH̄) + C̄(X) − C̄(X r ) . (12)
κ
Expanding (12) for small changes we obtain
−(U V )Y =0 = (1 + ΔŪτ )2 ∼ 1 + 2ΔŪS . (13)
2.2 External forces

In order to motivate the scaling assumptions we consider the external
forces: the component of the gravity force in flow direction integrated over
the fluid height H̄ sin α/δ and the shear stress F r2 γ 2 (U V )Y =0 /δ at the
bottom of the channel. Using the expansion (13) of the friction law the
external net force on a cross section of the flow can be written as
sin α F r2γ 2
Fext = H̄ + (U V )Y =0
δ δ
(14)
1 F r2 γ 2
∼ − ( F r 2 γ 2 − sin α) + 3ΔH̄.
δ δ
Here we have already used ΔŪS = −ΔH̄ + o(ΔH̄) which follows from the
shallow water approximation and will be shown later.
The difference γ 2 F r 2 −sin α measures the deviation of the reference state
from the fully developed turbulent flow. The second parameter γ 2 F r2 /δ
measures the change of the wall shear stress due to changes of the fluid
depth H̄.
202 H. Steinrück
2.3 Scaling assumptions

Our intention is to study the flow for slightly supercritical upstream
flow conditions. Thus we introduce the small perturbation parameter ε
according to the relation
3
F r = 1 + ε, (15)
2
where the coefficient 32 serves for later convenience.
We perform an asymptotic analysis with respect to the small parameters
α, δ, ε and γ. The relative sizes of these parameters have to be chosen
such that one can use the shallow water approximation and take turbulence
effects into account, but limit their magnitude in such a way that the well-
known results for fully developed flow can be applied locally to the disturbed
flow. Thus no turbulence model will be required to derive the results for
the free surface.
We choose the shallow water parameter δ to be
δ = 3ε1/2 , (16)
where the coefficient 3 is introduced for the sake of simplifying the final
equations.
In order to define a distinguished limit we have to couple the angle of
inclination α and the wall shear stress γ to the perturbation parameter ε
in such a way that to the leading order the shallow water equation will be
recovered. Therefore we introduce the two similarity parameters a and A,
which are assumed to be of order one as
γ 2 F r 2 − sin α 3γ 2 F r2
a= , A= . (17)
δε5/2 δε3/2
As mentioned above the parameter a measures the deviation of the reference
state from the fully developed flow. With other words if the flow in the
reference state is fully developed the parameter a vanishes. This assumption
has been made in Grilllhofer and Schneider (2003) and Steinrück et al.
(2003). The other parameter A is a measure for the change of the wall
shear stress due to a change of the fluid height.
The reference state We assume that the reference state state is attained
at Xr and the flow is a locally fully developed turbulent open channel flow.
Thus in the defect layer the velocity profile can be written as
√ 1
Ū(Xr ) = 1 + γ ŪD = 1 + ε AŪD with ŪD (Y ) dY = 0, (18)
0
where ŪD is the velocity defect.
3 Asymptotic Analysis
3.1 Shallow water approximation
In a first attempt we try to use a regular expansion of the form
Ū = 1 + εŪ1 + ε2 Ū2 + · · · , (19a)
P̄ = 1 − y + εP̄1 + ε2 P̄2 + · · · , (19b)

2
H̄ = 1 + εH̄1 + ε H̄2 + · · · , (19c)
(U V ) = (U V ) 0 + ε(U V )1 ··· . (19d)
Inserting into the Y -momentum equation and the dynamic boundary
condition yields
P̄1 = H̄1 . (20)
Using the X-momentum equation and the continuity equation we obtain
√
Ū1 = −H̄1 + AŪD , V̄1 = H̄1,X Y . (21)
However the perturbation H̄1 of the fluid height remains undetermined from
the analysis of the O(ε)-terms.
The component of the gravity term in flow direction is of order O(ε3/2 ).
Due to the scaling assumptions the derivative of the Reynolds’ shear stress
is of the same order. Using the dynamic boundary condition we obtain
(U V )0 = 1 − Y. (22)
Considering the O(ε)2 -terms the Y -momentum equation yields
9
P̄2 = H̄2 + H̄1,XX (1 − Y 2 ) − A(V 2 )0 (23)
2
Using the X-momentum equation and integration of the continuity equation
yields the second order term of the vertical velocity component.
V̄2 (X, 1) = H̄2,X + 3H̄1,XXX +

√ √ 1 (24)
+ −3 + H̄1 + AŪD − 2 A ŪD dY H̄1,X .
0
Employing the kinematic boundary condition a second expression for V̄2 (X, 1)
can be obtained
√
V̄2 (X, 1) + V̄1,Y H̄1 = H̄2,x + (−H̄1 + AΔŪD )H̄1,x . (25)
204 H. Steinrück
Comparing both expressions for V̄2 (X, 1) we obtain the third order differ-
ential equation for H̄1
H̄1,XXX + H̄1,X (H̄1 − 1) = 0. (26)
Integrating (26) once we obtain
1
H̄1,XX + H̄12 − H̄1 = R. (27)
2
Multiplying (27) by H̄1,X and integration yields
2
−3H̄1,X = p(H̄1 , R, S) := H̄13 − 3H̄12 − 6RH̄1 − 6S, (28)
where R and S are constants of integration. With the help of a phase

plane analysis it can be easily verified that (28) has periodic solutions if
the polynomial p(H̄1 ) has three different real roots. That is the case if
the discriminant D = 8R3 + 3R2 − 18RS − 9S 2 − 6S is positive. Now let
h1 = h1 (R, S), h2 = h2 (R, S) and h3 = h3 (R, S) be the three roots of
p(H̄1 , R, S) with h1 ≤ h2 ≤ h3 . Then the solution (28) can be written as
⎧ √
⎨ h2 + (h2 − h3 ) cn2 X h√2 −h1 |ν for ν < 1
H̄1 (X) = 2√3 (29)
⎩ h2 + (h2 − h3 ) sech2 X h√2 −h1 for ν = 1
2 3
with ν = hh32 −h2

−h1
.
If D < 0 the polynomial p(H̄1 ) has only one real root and the solution of
(28) is unbounded. Of special interest is the case when the reference state
at X = −∞ is a fully developed turbulent flow. Then we seek for a solution
which connects the fully developed flow with a flow with an undulated
surface. A candidate might be the solution of (28) for the case that the
p(H̄1 ) has a double root (D=0 or ν = 1). Then H̄1 has a single hump and
for X → ∞ the flow will return to the fully developed state again. Thus
none of the described solutions of (28) can be used to describe an undular
hydraulic jump.
3.2 Multiple Scales Expansion

Considering our goal to describe the undulated surface of the fluid in an
open channel we look for a solution which is almost periodic. The amplitude
and wave length of the surface are expected to change slowly by virtue of
dissipative effects. Of course such a solution structure is not possible with a
regular expansion discussed in the preceeding section. However, if we allow
the constants of integration R and S in (28) to vary slowly we can achieve

our goal.
Thus we expect to have two longitudinal length scales. The wave length
and a scale on which the wave length changes. Thus we have a typical
multiple scales problem.
Our strategy will be the following: First we define two length scales.
The original length scale X and a slow variable X = ε1/2 (X − X0 (ε)). Here
X0 is an appropriate shift of the origin of the coordinate system. Thus we
will allow all functions to depend on both length scales independently. For
example
H̄ = H̄(X, X ). (30)
Thus a derivative with respect to X becomes a sum of two partial derivatives
d ∂ ∂ dX ∂ ∂
H̄ = H̄ + H̄ = H̄ + ε1/2 H̄. (31)
dX ∂X ∂X dX ∂X ∂X
The derivatives with respect to the slow variable of terms of order O(ε)
are of order O(ε3/2 ) and the derivatives of terms of the order O(ε2 ) are
of the order O(ε5/2 ). Equations for the constants of integration R and S
as functions of the slow variable X will be obtained from the requirement
that all functions are periodic with respect to the fast variable X. Thus we
have to include terms of the order ε3/2 and ε5/2 additionally to the terms of
order ε and ε2 in the expansion of the dependent variables (19). However,
to facilitate the formulation of the periodicity condition with respect to the
fast variable X we introduce a transformed fast variable ξ
1
ξ= Ω(X ), with X = ε1/2 (X − X0 (ε)), (32)
ε1/2
such that all function have period 1 with respect to the new fast variable ξ.
We have
dξ = ω(Ω) dX, with ω = Ω . (33)
Thus 1/ω is the wave length in terms of the original fast variable X and
part of the solution.
Representative for all dependent state variables the expansion of the
pressure is given by:
P̄ ∼ P̄0 (Y ) + εP̄1 (ξ, Ω, Y ) + ε3/2 P̄3/2 (ξ, Ω, Y )+

(34)
2 5/2
ε P̄2 (ξ, Ω, Y ) + ε P̄5/2 (ξ, Ω, Y ) + · · · .
206 H. Steinrück
Terms of order O(ε) Using the Y -momentum equation (5b) and the
dynamic boundary condition (8) we obtain
P̄1 (ξ, Ω, Y ) = H̄1 (ξ, Ω). (35)
The X-momentum equation (5a) yields
Ū1,ξ = −P̄1,ξ , (36)
with the solution
√
Ū1 (ξ, Ω, Y ) = −H̄1 (ξ, Ω) + AŪD (Y, Ω). (37)
Note that the asymptotic analysis allows the velocity ’defect’ to depend on
the slowly varying variable Ω. From the continuity equation (4) we get the
Y -component of the velocity:
V̄1 (ξ, Ω, Y ) = ω H̄1,ξ (ξ, Ω)Y. (38)
From the dynamic boundary condition (8) and the relation for the surface
velocity (13) the perturbation of the Reynolds shear stress is obtained:
(U V )1 (ξ, Ω, 1) = −H̄1 (ξ, Ω), (U V )1 (ξ, Ω, 0) = 2H̄1 (ξ, Ω). (39)
The comparison of terms of order O(ε) leaves, however, the perturbation
H̄1 of the height of the free surface undetermined. An equation for H̄1 has
to be derived from equations for the higher–order terms.
Terms of order O(ε3/2 ) The analysis of the O(ε3/2 ) terms follows the
same lines as the analysis of the O(ε) terms. Formally in equations (35)-
(39) the subscript ’1’ has to be replaced by the subscript ’3/2’. We note
that in order to avoid secular terms ΔŪ1,Ω has to vanish. Thus we have
ŪD (Y, Ω) = ŪD (Y ).
Up to now we have not used the kinematic boundary condition at all.
Inspection shows that it is already satisfied.
In what follows the discussion of the terms of order O(ε2 ) and O(ε5/2 ),
respectively, will run along the same lines, with the exception that the
kinematic boundary condition will provide solvability conditions that will
serve as equations to determine the perturbation quantities H̄1 and H̄3/2 ,
respectively.
Terms of order O(ε2 ) Using the Y -momentum equation (5b) the O(ε2 )
pressure perturbation P̄2 can be expressed in terms of H̄2 as
9
P̄2 (ξ, Ω, Y ) = H̄2 + H̄1,ξξ (1 − Y 2 )ω 2 − A(V 2 )0 . (40)
2
Inserting P̄2 into the X-momentum equation (5a), using the continuity equa-
tion (4) and integrating with respect to Y gives:

√ √ 1
V̄2 (X, Ω, 1) = −3 + H̄1 + AŪD − 2 A ŪD dY ω H̄1,ξ
0
(41)
+3ω 3 H̄1,ξξξ + ω H̄2,ξ + ω H̄3/2,Ω
However, V̄2 has to satisfy the kinematic boundary condition as well, i.e.
V̄2 + V̄1,Y H̄1 = (H̄2,ξ + Ū1 H̄1,ξ + H̄3/2Ω )ω. (42)
Thus we obtain the solvability condition

ω 3 H̄1,ξξξ + ω H̄1,ξ H̄1 − 1 = 0, (43)
which is a third order differential equation for H̄1 as a function of the fast
variable ξ. It is essentially the same differential equation (26) which we have
obtained in our first approach. Integrating (43) with respect to ξ we obtain
1
ω 2 H̄1,ξξ + H̄12 − H̄1 = R, (44)
2
where R = (Ω) is a constant of integration which may depend on the slowly

varying variable Ω. Multiplying it with H̄1,ξ and integrating with respect
to ξ again yields
−3ω 2H̄1,ξ
2
= p(H̄1 , R(Ω), S(Ω)), (45)
where S = S(Ω) is another slowly varying constant of integration. We have

recovered equation (28) again. The only difference is that the constants of
integration are allowed to vary on the slow variable Ω. Differential equations
for both R and S as functions of the slow variable Ω will be obtained by
avoiding secular terms in the equation for the O(ε5/2 ) terms.
Terms of order O(ε5/2 ) Using the Y -momentum equation (5b) and the
dynamic boundary condition (8) we obtain the O(ε5/2 ) pressure perturba-
tion P̄5/2 :

9 2 2 dω
P̄5/2 = H̄5/2 + ω H̄3/2,ξξ + ω H̄1,ξΩ + H̄1,ξ ω 1−Y2 . (46)
2 dΩ
208 H. Steinrück
In the balance of the O(ε5/2 ) terms of the X-momentum equation the

Reynolds shear stress has to be taken into account, i.e.
(Ū5/2,ξ + Ū2,Ω )ω + 3(Ū3/2,ξ + Ū1,Ω )ω =

1
−(P̄5/2,ξ + P̄2,Ω )ω − a(U V )0,Y − A(U V )1,Y (47)
3
−Ū1 Ū1,Ω ω − Ū1 Ū3/2,ξ ω − Ū3/2 Ū1,ξ ω − V̄3/2 Ū1,Y − V̄1 Ū3/2,Y .
Using the continuity equation (4), the dynamic boundary condition (8) and
equation (39) we obtain
−V̄5/2 − 3(H̄3/2,ξ + H̄1,Ω )ω =

= −H̄5/2,ξ ω − 3H̄3/2,ξξξ ω 3 − H̄2,Ω ω−
dω 1
−6H̄1,ξξΩω 3 − 9ω 2 H̄1,ξξ − (H̄12 )Ω ω + (H̄1 H̄3/2 )ξ ω−
dΩ 2 (48)
1
√ √
−(H̄1,Ω + H̄3/2,ξ ) AŪD + 2 H̄3/2,ξ + H̄1,Ω ω A ŪD dy+
0
A
+a(U V )0 |Y =0 − (U V )1 |Y =1 − (U V )1 |Y =0 .
3
Finally, with the help of the kinematic boundary condition for the O(ε5/2 )
terms, i.e.
V̄5/2 + V̄1,Y H̄3/2 + V̄3/2,Y H̄1 =

(49)
= H̄5/2,ξ ω + H̄2,Ω ω + Ū1 H̄3/2,ξ ω + Ū3/2 H̄1,ξ ω + Ū1 H̄1,Ω ω,
we obtain the following equation for H̄3/2 :
ω 2 H̄3/2,ξξ + H̄3/2 H̄1 − H̄3/2 = r (50)
with
A a dω
ωrξ = H̄1 − − H̄1,ξξΩ ω 3 − H̄1,ξξ ω 2 − RΩ . (51)
3 3 dΩ
where we have used (U V )0,Y =0 = 1 and (U V )1,Y =0 −(U V )1,Y =1 = 3H̄1 .
In the following we derive differential equations for the slowly varying
’constants’ R and S. Integrating (51) with respect to ξ over one period
gives the following differential equation for R:
1
a A
ω RΩ = − + H̄1 (ξ, Ω) dξ. (52)
3 3 0
Multiplying equation (44) with H̄3/2,ξ and adding (50) multiplied by H̄1,ξ
we get 2
ω H̄3/2,ξξ + H̄3/2 H̄1 − H̄3/2− r H̄1,ξ +
1 (53)
+ ω H̄1,ξξ + H̄12 − H̄1 − R H̄3/2,ξ = 0.
2
Integrating (53) with respect to ξ over one period, and integrating the result
by parts yields
1
rξ H̄1 dξ = 0. (54)
0
Taking the derivative of (44) with respect to Ω and using (51) and (54) we
obtain the following differential equation for S:

a 1 A 1 2
ω SΩ = H̄1 dξ − H̄ dξ. (55)
3 0 3 0 1
Thus we have derived a set of differential equations for the slowly varying
constants of integration R and S. As we will discuss later the wave length
and amplitude of the undulated surface can be determined from R and S.
The main result of the multiple scales analysis can now be summarized
as follows: Let H(X; R, S) be the solution of
2 1
HX + H3 − H2 = 2R H + 2S, HX (0, R, S) = 0, HXX (0, R, S) > 0,
3
(56)
where R = R(Ω) and S = S(Ω) are the solutions of (52), (55). If H̄1 (0),
H̄1 (0) and H̄1 (0) is given initial conditions for R and S are obtained from
(27) and (28). Then the first-order perturbation of the free surface, i.e. H̄1 ,
is given by:
Ω
ξ − 12 dΩ Ω
H̄1 (X) = H ; R, S , X = , ξ=√ , (57)
ω 0 ω ε
with
1
X(ξ, ε) = √ X + X0 , (58)
ε
where X0 is chosen such that H(−X0 , R(0), S(0)) = H̄1 (0).
3.3 Uniformly valid differential equation

In Jurisits et al. (2007) and Grilllhofer and Schneider (2003) the following
third order differential equation
√
ε
ĤXXX + ĤX (Ĥ − 1) = −a + AĤ (59)
3
210 H. Steinrück
has been derived to describe the perturbation of the free surface of an open
channel flow near critical flow conditions. The equations is written here
in a form such that the parameters a, A and ε are the same as defined in
(15), (16, (17), respectively. In the derivation of (59) terms resulting from
the unbalanced external forces, which are of the order O(ε5/2 ) have been
included in the balance of the order O(ε2 ). As a result the perturbation
parameter ε appears in the equation for the leading order terms. However,
equation (59) can be taken as a the starting point of a multiple scales
analysis. We want to show here that the leading order terms for the fast
and slowly varying quantities agree with the results presents in the previous
section. Since equation (59) contains all the information of the multiple
scales analysis we call it a uniformly valid differential equation.
Equation (59) is equivalent to the first order system of differential equa-
tion
2
3ĤX = p(Ĥ, R̂, Ŝ), (60a)
√ √
ε ε
R̂X = −a + AĤ , ŜX = − −a + AĤ Ĥ, (60b)
3 3
where p(Ĥ, R̂, Ŝ) is the third order polynomial defined in (28). In order to
study the asymptotic behavior of the solution (59) in the limit ε → 0 we
can perform a multiple scale expansion according to the same lines as in
preceeding subsection
ˆ Ω̂) + . . . ,
Ĥ ∼ Ĥ1 (ξ, ˆ Ω̂) + . . . ,
R̂ ∼ R̂1 (ξ, ˆ Ω̂) + . . . ,
Ŝ ∼ Ŝ1 (ξ, (61)
where Ĥ1 , R̂1 , Ŝ1 are assumed to be periodic function with period 1 of the
fast variable ξ. ˆ The slow variable Ω̂ and the fast variable ξˆ are defined
analogously to (32). Now it is easy to see that the leading order terms Ĥ1 ,
R̂1 , Ŝ1 satisfy exactly the same differential equations as H̄, R and S. In
particular R̂1 and Ŝ1 are functions of the slow variable Ω̂ only.
We remark that one of the two perturbations terms on the right hand
side of (59) can be eliminated by setting

a √
Ĥ(X) = 1 + b (G(X/c) − 1) , b = 1 − c = 1/ b. (62)
A
We obtain √
ε c
G ± G (G − 1) = ÃG, Ã = , (63)
3 b
provided that A = a and A = 0. The plus sign holds for A > a and the
minus sign for A < a. √
If A = a we set b = A2/3 , c = 1/ b and obtain
√
ε c
G ±G G= ÃG, Ã = . (64)
3 b
3.4 The equations for the slowly varying variables

We have already derived the differential equations for the slowly varying
quantities R and S. Here we want to indicate how the integrals on the right
hand side can be considered as functions of R and S. It is obvious that H̄1
is a periodic function of the fast variable if the polynomial p(H̄1 , R, S) has
three real roots, h1 (R, S) ≤ h2 (R, S) ≤ h3 (R, S). Now let
h3 (R,S)
hj dh
Ij (R, S) = , (65)
h2 (R,S) (h − h1 (R, S))(h − h2 (R, S))(h3 (R, S) − h)
with j = 0, 1, 2. Then the equations for R, S and ω are

a 2A
RΩ = − + √ I1 (R, S), (66a)
3ω 3
2a 2A
SΩ = √ I1 (R, S) − √ I2 (R, S), (66b)
3 3
1
ω= √ . (66c)
2 3I0 (R, S)
We can state a sufficient condition for the existence of a solution of (66)

for all Ω > 0: Let A > a, and R(0) > 0 and S(0) < 0 such that the
polynomial p(H, R(0), S(0)) has three real roots then the solution of the
initial value problem (66) for the slow variables exists for all Ω > 0.
Proof
(i) It is sufficient to consider the case A > 0. If a = 0 the transform
a √
H̄1 (ξ, Ω), = 1 + b Ḡ1 (ξ/c, Ω/c), b=1− , c = 1/ b (67)
A
can be used to eliminate a.
(ii) We consider the polynomial p(H) √ defined in (28). If R > 0 it has a
∗
local maximum√ at H max = 1 − 1 + 2R < 0 and a local minimum at
∗ ∗
H min = 1 − 1 + 2R > 0. √ If p(h, R, S) has three roots: h1 < H max <
∗
0 and h3 > H min = 1 − 1 + 2R > 0. Thus h2 = 6S/(h1 h3 ) > 0.
(iii) Since h2 and h3 are positive the integrals Ij , j = 0, 1, 2 are also posi-
tive. With A > 0 it follows that RΩ > 0 and SΩ < 0. Thus zero is a
lower bound for R and an upper bound for S. It can be shown that
the integrals Ij can be estimated by
√
0 < Ij < CR (1 + R)j , j = 0, 1, 2, (68)
212 H. Steinrück
for some constant CR . Thus we obtain R < C̃Ω2 for an appropriate

constant C̃. Furthermore, SΩ can be estimated from below indepen-
dently of S. Integration yields a lower bound for S. It remains to show
that the right hand side of the differential equation is well defined for
Ω > 0. Thus we have to show that h2 and h3 never coalesce.
(iv) The differential equations for R and S can be reformulated as differ-
ential equations for
h2 + h3 h3 − h2
hm = , Δh = . (69)
2 2
(v) The solution of the differential equations for hm , Δh can be discussed
in the hm , Δh-plane. It turns out that on line Δh = 0, ΔhΩ = 0, but
hm,Ω > 0 holds. Thus Δh cannot become zero at a finite Ω and this
completes the proof.
Asymptotic behavior far downstream (Ω, X → ∞) To discuss the

behavior for X → ∞ it is appropriate to consider hm and Δh defined in
(69) instead of R, S. Expanding the integrals Ij for Δh/hm 1 we obtain
the following differential equations for Δh and hm :

Δh A 1
ΔhX ∼ −A , hm,X ∼ 1+ (70)
4hm 3 hm
with the solution

A A
hm ∼ X + ln 1 + X , Δh = O(X −3/4 ). (71)
3 3
Thus the amplitude of the oscillation decays and the mean value over an
oscillation period increases linearly with increasing X .
Breakdown of the multiple scales approximation If a > A the roots

h1 and h2 may coalesce and thus the multiple scale solution may terminate
at a finite value of Ω. This has been observed by Jurisits et al. (2007). They
solved the uniformly valid differential equation (59) numerically for the case
a = 3A and compared the solution with the multiple scales approximation,
see figure 2. The solution of the uniformly valid differential equation behaves
at first almost periodically and is in excellent agreement with multiple scales
approximation. Representative for the slowly varying quantities the zeros
h2 and h3 of the polynomial p(H̄1 ) are shown.
All of a sudden the solution of the uniformly valid differential equation
terminates in a singularity. At first glance h2 and h3 are almost constant.
PXOWLSOHVFDOHVDSSUR[LPDWLRQ
QXPHULFDOVROXWLRQ
h2 h3 X
4
ï
3 h2
2 ï
H̃1
1
0
−1 EUHDNGRZQRI
PXOWLSOHVFDOHV
DSSUR[LPDWLRQ
0 20 40 60 80 100 120
X
Figure 2. Breakdown of solution: Comparison of a numerical solution of
the uniformly valid differential equation (59) with the multiple scales ap-
proximation. Parameter values: ε3 × 10−12 , A = 1, a = 3, Inital conditions
H̄1 (0) = 0, H̄1 (0) = 0.3, H̄1 (0) = 0.314. The zero h2 of the ploynomial p is
shown in the insert to the right, see Jurisits et al. (2007).
But a detailed view of h2 shows that it terminates in a square-root singu-

larity shortly before the breakdown of the numerical solution. Considering
the multiple scales approximation it turns out that the two zeros h1 and
h2 coalesce. Thus the multiple scales approximation terminates at a final
value of Ω.
3.5 Initially fully developed flow

Here we focus on a flow which is fully developed far upstream. Thus the
external forces are in equilibrium, i.e. a = 0.
It remains to determine appropriate initial conditions for R and S. As
the origin of the X-axis we choose the location of the first crest of the free
surface, i.e. the first maximum of H̄1 . Since H̄1 has to decay to zero for
X → −∞, the period becomes infinite as X → −∞.
This gives the following initial conditions for R and S:
R(0) = S(0) = 0. (72)
In this case the shift of the origin X0 has to depend on ε. It is chosen such
that the first wave crest is at X = 0. We have
1/2
1 1
X(ξ, ε) = √ X + X0 (ε), X0 (ε) = − √ dξ. (73)
ε 0 ω( εξ)
214 H. Steinrück

h3
multiple
scales
sol.

H̄1

h2

homo-clinic orbit

ï ï
X
Figure 3. Multiple scales solution (57) for A = 3, ε = 10−4 and the homo-
clinic orbit (A = 0) according to eq. (74). The solution for the slowly
varying variables is represented by h2 and h3 , i.e. the local minima and
maxima, respectively.
We will see later that X (0) = 0 and X0 is finite. Thus the multiple scales
approximation (57) fails for X → −∞. However, an asymptotic analysis
using the original coordinate X yields the inviscid solution
H̄1 (X) = H(X; 0, 0), (74)
which is a homo-clinic orbit in the phase plane. It is the limiting solution
for the first elevation and describes the behavior for X → −∞ correctly,
but fails to approximate the undular behavior.
In Figure 3 the multiples scales solution (57), evaluated for ε = 10−4 ,
A = 3 and the homo-clinic orbit are plotted. Moreover the solution of the
equations for the slowly varying variables represented by the zeros h2 and
h3 of the polynomial p(h; R, S) are also shown in Figure 3. In Figure 4 the
multiple scales solution (57) and the homo-clinic orbit (74) are shown in the
phase plane.
Behavior near the origin (Ω 1) Although the polynomial p(h),

which occurs in the denominator of I1 and I2 , respectively, has a double
root at h2 = h1 = 0 for R = S = 0, these integrals exist for R = S = 0:
√ √
I1 (0, 0) = 2 3, I2 (0, 0) = 4 3. (75)

multiple scales solution

dH̄1
dX
ï
ï
homo-clinic orbit
ï

H̄1
Figure 4. Phase portrait of the multiple scales solution (57) for A = 3,
ε = 10−4 and the homo-clinic orbit (A = 0) according to eq. (74).
Using
R ∼ RΩ (0)Ω = 4A Ω, S ∼ SΩ (0)Ω = −8A Ω (76)
we obtain for the zeros of the polynomial p(h):
√ √ 8
h1 ∼ −4 AΩ, h2 ∼ 4 AΩ, h3 ∼ 3 + AΩ + ... . (77)
3
Expanding I0 for Ω 1 gives
1 1
∼ ln + 2 ln 6. (78)
ω AΩ
With
Ω
1 1
X (Ω) = dΩ ∼ Ω ln + 1 + 2 ln 6 (79)
0 ω AΩ
we get X (0) = 0 and obtain the shift

√ 1/2
1 ε 1 1 √ 3 1
X0 (ε) = − √ X ( )=− √ dξ ∼ ln A ε − ln2 − ln3 − .
ε 2 0 ω( εξ) 2 2 2
(80)
216 H. Steinrück

h̄C /h̄r #3
#1 #4
#5
data as in table 1
#2 0.95 ... 1.05 V̇

asymptotic result, eq. (81)

Fr
Figure 5. Comparison of calculated and measured amplitudes h C /h r of
the first elevation. The dotted lines indicate how the non-dimensionalized
height of the first crest changes for a variation of the volume flow rate within
±5% around the value given in table 1.

Λ = λ/l data as in table 1

0.95 ... 1.05 V̇
#3 #4
#1

calculated result (82)
#5
#2

asymp. approximation for β 1, eq. (83)

A√ α
ε = 3/2 β=
3 3ε
Figure 6. Comparison of calculated and measured distances λ/l be-
tween the first two maxima. The dotted lines indicate how the non-
dimensionalized first wave length changes for a variation of the volume flow
rate within ±5% around the value given in table 1.
# Exp. V̇ α h̄r w
[ m2 /s] [rad] [ m] [ m]
1 HCUJ8e 0.028 4.10×10−3 0.041 0.25
2 HCUJ10c 0.080 3.99×10−3 0.084 0.25
3 HCUJ4b 0.119 4.99×10−3 0.105 0.25
4 HCUH1b 0.0416 3.7×10−3 0.053 0.25
5 HCUJ4a 0.120 4.33×10−3 0.109 0.25
λas
# Exp. λ h̄C Fr β
λex
[ m] [ m] [−] [−] [−]
1 HCUJ8e 0.310 0.052 1.077 0.115 0.94
2 HCUJ10c 0.600 0.100 1.049 0.225 1.07
3 HCUJ4b 0.750 0.144 1.119 0.075 0.87
4 HCUH1b 0.450 0.068 1.088 0.086 0.83
5 HCUJ4a 0.660 0.138 1.062 0.171 1.20
Table 1. Experimental data, selected from Chanson (1993)
Comparison with experiments In the literature there is a large amount

of experimental data for the undular hydraulic jump, but to obtain a mean-
ingful comparison with the present analysis the experiments have to satisfy
the following criteria:
(i) A fully developed flow far upstream is required. This is only possible
in an inclined channel.
(ii) The theory is based on an asymptotic expansion in terms of the small
parameter ε. Thus we consider only experiments with 0 < ε < 0.1,
i.e. 1 < F r < 1.15.
(iii) The relative error of neglected terms has to be small. The smallest
terms that influence the leading
√ order result for the wave length are
of order O(Aε5/2 ) = O(α ε), whereas terms of order O(ε3 ) are ne-
glected. Thus the relative truncation error is of O(ε5/2 /α), and only
experiments with ε5/2 /α being sufficiently small are considered.
(iv) The height of the first crest and the first wave-length, i.e. the dis-
tance between two successive crests, must be among the measured
quantities.
In Table 1 experiments selected to meet these criteria are listed. The ex-
periments have been performed in a channel of width w.
In the analysis the maxima of the surface elevation occur at ξ = 12 , 32 , ...,
√ √
and so on. Thus the first maximum is H̄1,max = h3 ( 2ε ) = 3 + 4 A3 ε , cf.
(77), and for the height h̄C of the first crest we obtain:
2 8
h̄C /h r ∼ 1 + ( F r − 1)H̄1,max = 1 + 2( F r − 1)(1 + β). (81)
3 3
218 H. Steinrück
√
The parameter β = A 3 ε has been introduced in Grillhofer (2002) and Gril-
llhofer and Schneider (2003) and is used also here for convenience.
Let
√ Λ be the scaled first wave lengths. It is referred to reference lh̄r /δ =
h̄r /3 ε, see (2a), (3). Thus we obtain from
√ √ 3/2
1 3 ε ε 1
Λ(ε) = √ X( ) − X( ) = √ dξ. (82)
ε 2 2 1/2 ω( εξ)
In the limit ε → 0 eq. (82) becomes
√
8 3 8
Λ(ε) ∼ ln √ + 1 = ln √ + 1. (83)
A ε 3β
In Figures 5 and 6 a comparison of measured data with the results (81),
(82) and (83), respectively, is given. The volume flux in the experiments is
measured with an accuracy of 5% (cf Chanson (1993), p. 2-1). Therefore
the effect of a variation of the volume flux by ±5% is also shown in the
figures. It can be seen that the results are very sensitive with respect to
uncertainties in the volume flux. In Figure 6 results are given for both
a numerical evaluation
√ of Λ and the asymptotic approximation for small
values of β = A ε/3.
Taking into account that the aspect ration w/h r , where w is the width
of the channel, is larger than 0.16 in all experiments, so that side wall effects
may influence the results considerably, the agreement between the measured
data and the theoretical results for plane flow is reasonable.
4 Conclusions
The multiple scaling approach as proposed in this paper permits a self-
consistent asymptotic analysis of the undular hydraulic jump in turbulent
flow. Results for the perturbation of the height of the fluid are obtained
without making use of a turbulence model. Ordinary differential equations
have been derived for the slowly varying quantities, i.e. the amplitude and
the wave-length, as well as for the rapidly varying height of the free surface.
The price for the rigorous analysis is a limited range of parameters where
the requirements for the analysis are satisfied. In view of the high sensitivity
of the the wave length and the amplitude, the agreement with experimental
data appears reasonable.
Bibliography
V.M. Andersen. Undular hydraulic jump. J. Hydraulics Div., 104(HY8):
1185–1188, 1978.
T.B. Benjamin and M.J. Lighthill. On cnoidal waves and bores. Proc. Roy.
Soc., A224:448–460, 1954.
P. Böß. Berechnung der Wasserspiegellage. Forschungsarbeiten VDI, 284,
1927.
H. Chanson. Characteristics of undular hydraulic jumps. University of
Queensland, Department of Civil Engineering Research Report Series,
CE146, 1993.
H. Chanson and J.S. Montes. Characteristics of undular hydraulic jumps:
Experimental apparatus and flow patterns. J. Hydraulic Eng., 121:129–
144, 1995.
V.T. Chow. Open-Channel Hydraulics. McGraw-Hill, 1959.
W. Grillhofer. Der wellige Wassersprung in einer turbulenten Kanal-
strömung mit freier Oberfläche. Dissertation, Technische Universität
Wien, 2002.
W. Grilllhofer and W. Schneider. The undular hydraulic jump in turbulent
open channel flow at large reynolds numbers. Phys. Fluids, 15:730–735,
2003.
W. Hager. Energy Dissipators and Hydraulic Jump. Kluwer, 1992.
W.H. Hager and K. Hutter. On pseudo-uniform flow in open channel hy-
draulics. Acta Mechanica, 53:183–200, 1984.
F.M. Henderson. Open Channel Flow. Macmillan Company, 1966.
R.S. Johnson. Shallow water waves on a viscous fluid - the undular bore.
Phys. Fluids, 15:1693–1699, 1972.
R. Jurisits, W. Schneider, and Y.S. Bae. A multiple scales solution to the
undular hydraulic jump problem. PAMM, 7:4120007 – 4120008, 2007.
K. Kaufmann. Hydromechanik II. Springer, 1934.
H. Lauffer. Wassersprung bei kleinen Sprunghöhen. Wasserwirtschaft und
Technik, pages 137–140, 1935.
I. Ohtsu, Y. Yasuda, H. Gotoh, and M. Iahr. Hydraulic condition for
undular-jump formations. J. Hydraulic Res., 39:203–209, 2001.
R. Reinauer and W. Hager. Non-breaking undular hydraulic jump. J.
Hydraulic Res., 33:683–698, 1995.
H. Steinrück, W. Schneider, and W. Grillhofer. A multiple scales analysis of
the undular hydraulic jump in turbulent open channel flow. Fluid Dyn.
Res., 33:41–55, 2003.
Modern Aspects of High-Reynolds-Number
Asymptotics of Turbulent Boundary Layers
From Fully Attached to Marginally Separated Flows
Bernhard Scheichl†‡∗
†
Institute of Fluid Mechanics and Heat Transfer, Vienna University of
Technology, Vienna, Austria
‡
AC2 T research, Austrian Center of Competence for Tribology,
Wiener Neustadt, Austria
Abstract This contribution reports on recent efforts with the ul-

timate goal to obtain a fully self-consistent picture of turbulent
boundary layer separation. To this end, it is shown first how the
classical theory of turbulent boundary layers having an asymptoti-
cally small streamwise velocity deficit can be generalised rigorously
to boundary layers with a slightly larger, i.e. moderately large, ve-
locity defect and, finally, to situations where the velocity defect is
of O(1). In the latter case, the formation of short recirculation zones
describing marginally separated flows is found possible, as described
in a rational manner.
1 Introduction
Despite the rapid increase of computer power in the recent past, the calcula-
tion of turbulent wall-bounded flows still represents an extremely challeng-
ing and sometimes insolvable task. Direct-Numerical-Simulation computa-
tions based on the full Navier–Stokes equations are feasible for moderately
large Reynolds numbers only. Flows characterised by much higher Reynolds
numbers can be investigated if one resorts to turbulence models for the
small scales, as accomplished by the method of Large Eddy Simulation, or
for all scales, as in computer codes designed to solve the Reynolds-averaged
Navier–Stokes equations. Even then, however, the numerical efforts rapidly
∗
This material represents in some places a slight modification and in some an extension
of its original form published by Kluwick and Scheichl (2009) and is by courtesy of
these authors. Special thanks are due to AIC Androsch International Management
Consulting GmbH for generous financial support.
© CISM, Udine 2010
222 B. Scheichl
increase with increasing Reynolds number. This strongly contrasts the use
of asymptotic theories, the performance of which improves as the values of
the Reynolds number become larger and, therefore, may be considered to
complement purely numerically based work.
With a few exceptions (e.g. Deriat, 1986; Walker, 1998; Scheichl and
Kluwick, 2008a), studies dealing with the high-Reynolds-number properties
of turbulent boundary layers start from the time- or, equivalently, Reynolds-
averaged equations. By defining non-dimensional variables in terms of a
representative length L̃ and flow speed Ũ and assuming incompressible nom-
inally steady two-dimensional flow they take on the form
∂u ∂v
+ =0, (1a)
∂x ∂y
∂u ∂u ∂p 1 2 ∂u 2 ∂u v
u +v =− + ∇ u− − , (1b)
∂x ∂y ∂x Re ∂x ∂y
∂v ∂v ∂p 1 2 ∂u v ∂v 2
u +v =− + ∇ v− − . (1c)
∂x ∂y ∂y Re ∂x ∂y
Herein ∇2 = ∂ 2 /∂x2 + ∂ 2 /∂y 2 , and (x, y), (u, v), (u , v ), −u2 , −u v , −v 2 ,
and p are Cartesian coordinates measuring the distance along and perpen-
dicular to the wall, the corresponding time mean velocity components, the
corresponding velocity fluctuations, the components of the Reynolds stress
tensor, and the pressure, respectively. The Reynolds number is defined by
Re := Ũ L̃/ν̃, where ν̃ is the (constant) kinematic viscosity. Equations (1)
describe flows past flat impermeable walls when supplemented with the
usual no-slip and no-penetration conditions u = v = u = v = 0. Effects of
wall curvature can be incorporated without difficulty but are beyond the
scope of the present analysis.
When it comes down to the solution of the simplified version of these
equations provided by asymptotic theory in the limit Re → ∞, one is, of
course, again faced with the closure problem. The point, however, is that
these equations and the underlying structure represent closure-independent
basic physical mechanisms characterising various flow regions, identified by
asymptotic reasoning. This has been shown first in the seminal studies
by Yajnik (1970), Bush (1972), Fendell (1972), Mellor (1972), and more
recently and in considerable more depth and breath by Walker (1998) and
Schlichting and Gersten (2000) for the case of small-defect boundary layers,
which are subject of Section 2. Those exhibiting a slightly larger, i.e. a
moderately large, velocity defect are treated in Section 3. Finally, Section 4
deals with situations where the velocity defect is of O(1) rather than small.
High-Reynolds-Number Asymptotics of Turbulent Boundary Layers 223
2 Classical Theory of (Small-Defect) Turbulent

Boundary Layers
We first outline the basic ideas underlying an asymptotic description of
turbulent boundary layers.
2.1 Preliminaries
Based on dimensional reasoning put forward by L. Prandtl and Th. von
Kármán, a self-consistent time-mean description of firmly attached fully
developed turbulent boundary layers holding in the limit of large Reynolds
numbers Re, i.e. for Re → ∞, has been proposed first in the aforementioned
early asymptotic work (Yajnik, 1970; Bush, 1972; Fendell, 1972; Mellor,
1972). One of the main goals of the present investigation is to show that
this rational formulation can be derived from a minimum of assumptions:
(A) all the velocity fluctuations are of the same order of magnitude in
the limit Re → ∞, so that all Reynolds stress components are equally
scaled by a single velocity scale uref , non-dimensional with a global
reference velocity Ũ (hypothesis of locally isotropic turbulence);
(B) the pressure gradient does not enter the flow description of the viscous
wall layer to leading order (assumption of firmly attached flow);
(C) the results for the outer predominantly inviscid flow region can be
matched directly with those obtained for the viscous wall layer (as-
sumption of “simplest possible” flow structure).
In those outstanding papers (Yajnik, 1970; Bush, 1972; Fendell, 1972; Mel-
lor, 1972), uref is taken to be of the same order of magnitude in the fully
turbulent main portion of the boundary layer and in the viscous wall layer
and, hence, equal to the skin-friction velocity
1/2
uτ := Re −1 (∂u/∂y)y=0 . (2)
This in turn leads to the classical picture, according to which (i) the stream-
wise velocity defect with respect to the external impressed flow is small and
of O(uτ ) across most of the boundary layer, while (ii) the streamwise veloc-
ity is itself small and of O(uτ ) inside the (exponentially thin) wall layer, and,
finally, (iii) uτ /Ue = O(1/ ln Re). Furthermore, then (iv) the celebrated uni-
versal logarithmic velocity distribution
u/uτ ∼ κ−1 ln y + + C + , y + := y uτ Re → ∞ . (3)
is found to hold in the overlap of the outer (small-defect) and inner (viscous
wall) layer. Here κ denotes the von Kármán constant; in this connection
we note the currently accepted empirical values κ ≈ 0.384, C + ≈ 4.1, which
224 B. Scheichl
refer to the case of a perfectly smooth surface (see Österlund et al., 2000)
and have been obtained for a zero pressure gradient.
This might be considered to yield a stringent derivation of the logarith-
mic law of the wall (3), anticipating the existence of an asymptotic state
and universality of the wall layer flow as Re → ∞; a view which, however,
has been increasingly challenged in more recent publications (e.g. Baren-
blatt and Goldenfeld, 1995; Barenblatt et al., 2000; Barenblatt and Chorin,
2000). It thus appears that – as expressed by Walker (1998) – “. . . although
many arguments have been put forward over the years to justify the log-
arithmic behaviour, non are entirely satisfactory as a proof, . . . ”. As a
result, one has to accept that matching (of inner and outer expansions),
while ensuring self-consistency, is not sufficient to uniquely determine (3).
In the following, from the viewpoint of the time-averaged flow description
the logarithmic behaviour (3), therefore, will be taken to represent an exper-
imentally rather than strictly theoretically based result holding in situations
where the assumption (B) applies. The description of the boundary layer
in the limit Re → ∞ can then readily be formalised.
In passing, we mention that in the classical derivations (see Yajnik,
1970; Bush, 1972; Fendell, 1972; Mellor, 1972) the assumption (B) is not
adopted, in favour of rather heuristic dimensional arguments already put
forward by (among other authors) Clauser (1956) that constitute the scal-
ings of the thicknesses of both the entire boundary and the wall layer. In
view of assumption (A) and dimensional notation, these are given by L̃ uref
and ν̃/(Ũ uref ), respectively. As demonstrated in detail by Mellor (1972),
then (3) results from matching, which decisively contrasts the present study
where it is imposed. However, this procedure additionally requires to an-
ticipate a priori that in the outer layer u ∼ u0 + O(uref ), u0 = O(1). On
balance, we thus render that classical approach less generic than that pro-
posed here and elucidated in the following.
2.2 Leading-Order Approximation

Inside the wall layer where y + = y uτ Re = O(1) the streamwise velocity
component u, the Reynolds shear stress τ := −u v and the pressure p are
expanded in the form
u ∼ uτ (x; Re) u+ (y + ) + · · · , (4a)
τ∼ u2τ (x; Re) t+ (y + ) + ··· , (4b)
p ∼ p0 (x) + · · · , (4c)
+
where u exhibits the limiting behaviour implied by (3):
u+ (y + ) ∼ κ−1 ln y + + C + , y+ → ∞ . (5)
Assumption (C), quoted in Subsection 2.1, then uniquely determines the

asymptotic expansions of, respectively, u, τ , and p further away from the
wall where the Reynolds stress τ is predominant. Let δo (x; Re) characterise
the thickness of this outer main layer, i.e. of the overall boundary layer. In
turn, one obtains
u ∼ Ue (x) − uτ (x; Re) F1 (x, η) + · · · , (6a)

τ∼ u2τ (x; Re) T1 (x, η) + ··· , (6b)
p ∼ pe (x) + · · · , (6c)
where η := y/δo and F1 , T1 represent a perturbation stream function and

a shape function for the Reynolds shear stress, respectively. Here and in
the following primes denote differentiation with respect to η. Furthermore,
Ue and pe stand for the velocity and the pressure, respectively, at the outer
edge η = 1 of the boundary layer (here taken as a sharp line with suffi-
cient asymptotic accuracy) imposed by the external irrotational flow. The
asymptotic structure of the boundary layer is sketched in Figure 1.
y
δo
5
outer (predominantly
⎫
inviscid) region ⎪
⎪ overlap region :
⎪
⎪
O(uτ ) ⎪
⎬ u ∼ 1 ln y + + C + ,
uτ κ
⎪
⎪ +
⎪y →∞
⎪
⎪
u viscous wall layer⎭
Ue
Figure 1: Structure of classical two-tiered boundary layer.
Matching of the expansions (4) and (6) by taking into account (5) forces
a logarithmic behaviour of F1 ,
F1 ∼ −κ−1 ln η + C0 (x) , η→0, (7)
yields p0 (x) = pe (x), and is achieved provided γ := uτ /Ue satisfies the skin-
friction relationship
κ/γ ∼ ln(Reγδo Ue ) + κ(C + + C0 ) + O(γ) . (8)
From substituting (4) into the x-component (1b) of the Reynolds equa-
tions (1) one obtains the well-known result that the total stress inside the
226 B. Scheichl
wall layer is constant to leading order,
du+ /dy + + t+ = 1 . (9)
Moreover, the expansions (6) lead to a linearisation of the convective terms

in the outer layer, so that there Bernoulli’s law holds to leading order,
dpe /dx = −Ue dUe /dx . (10)
The necessary balance with the gradient of the Reynolds shear stress then
determines the magnitude of the thickness of the outer layer, i.e. δo = O(uτ ).
As a consequence, the expansions (6) are supplemented with
δo ∼ γ Δ1 (x) + · · · , (11)
which in turn gives rise to the leading-order outer-layer streamwise mo-

mentum equation. After integration with respect to η and employing the
matching condition T1 (x, 0) = 1, the latter is conveniently written as
(E + 2β0 )ηF1 − EF1 − Δ1 F1,e ∂F1 /∂x = F1,e (T1 − 1) , (12a)

F1,e := F1 (x, 1) , β0 := −Δ1 F1,e (dUe /dx)/Ue , (12b)
E := (F1,e /Ue )d(Δ1 Ue3 )/dx = 1 − Δ1 dF1,e /dx .
3
(12c)
The last equality is due to the usual boundary or patching conditions im-
posed, T1 (x, 1) = F1 (x, 1) = 0. The boundary layer equation (12a) is un-
closed, and in order to complete the flow description to leading order, tur-
bulence models for t+ and T1 have to be adopted. Integration of (12) then
provides the velocity distribution in the outer layer and determines the yet
unknown function C0 (x) entering (7) and the skin-friction law (8).
As a main result, inversion of (8) with the aid of (11) yields
γ ∼ κσ[1 − 2σ ln σ + O(σ)], σ := 1/ ln Re , ∂γ/∂x = O(γ 2 ) . (13)
Finally, the skin-friction law (13) implies the scaling law (iii), already men-
tioned in Subsection 2.1, which is characteristic of classical small-defect
flows.
2.3 Second-Order Outer Problem

Similar to the description of the leading-order boundary layer behaviour,
the investigation of higher-order effects is started by considering the wall
layer first. By substituting (4a), (4b), (8) into (1b) and taking into account
the slow variation of γ, cf. (13), one obtains upon integration (cf. Walker,
1998),
y+
1 ∂u dUe /dx + γ dUe /dx 2
+ τ ∼ γ 2 Ue2 − y + u+ dy + · · · . (14)
Re ∂y γ Re Re 0
Herein the second and the third term on the right-hand side account for,
respectively, the effects of the (imposed) pressure gradient, cf. (10), and the
convective terms, which have been neglected so far. By using (5) and (12),
one can easily derive the asymptotic behaviour of τ as y + → ∞ (e.g. Walker,
1998). When rewritten in terms of the outer-layer variable η, this is found
to be described by

τ δo dUe /dx Δ1 dUe /dx 2

∼1+2 η ln η + · · · + γ η(ln η) + · · · + · · · ,
(γUe )2 γκUe κ2 U e
(15)
as η → 0 and Re → ∞, which immediately suggests the appropriate gener-
alisation of the small-defect expansions (6a), (6b), (11):
u/Ue ∼ 1 − γF1 − γ 2 F2 + · · · , (16a)

2
τ /(γUe ) ∼ T1 + γT2 + · · · , (16b)
δo /γ ∼ Δ1 (x) + γΔ2 (x) + · · · . (16c)
On the basis of these second order results, here matching with the wall
layer is achieved if
F1 ∼ −κ−1 ln η + C0 (x) , F2 ∼ C1 (x) , (17a)

2Δ1 dUe dUe /dx Δ1

T1 ∼ 1 + η ln η , T2 ∼ η ln η ln η + 2Δ2 , (17b)
κUe dx κUe κ
as η → 0, provided that the skin-friction relationship (8) is modified to

explicitly include an additional term of O(γ),
κ/γ ∼ ln(Reγδo Ue ) + κ(C + + C0 + γC1 ) + · · · . (18)
Similar to C0 (x), the function C1 (x) depends on the specific turbulence

model adopted, as well as the upstream history of the boundary layer.
2.4 Can Classical Small-Defect Theory Describe Boundary Layer

Separation?
An estimate of the thickness δw of the viscous wall layer is readily ob-
tained from the definition of y + , see (3), and the (inverted) skin-friction
228 B. Scheichl
relationship (13): δw = O[γ −1 exp(−κ/γ)]. In the limit Re → ∞, therefore,

the low-momentum region close to the wall is exponentially thin as compared
to the outer layer, where Reynolds stresses cause a small O(γ)-reduction of
the fluid velocity with respect to the mainstream velocity Ue (x). This the-
oretical picture of a fully attached turbulent small-defect boundary layer
has been confirmed by numerous comparisons with experimental data for
flows of this type (e.g. Walker, 1998; Afzal, 1995; Monkewitz et al., 2007).
However, it also indicates that attempts based on this picture to describe
the phenomenon of boundary layer separation, frequently encountered in
engineering applications, will face serious difficulties. Since the momentum
flux in the outer layer, which comprises most of the boundary layer, dif-
fers only slightly from that in the external flow region, an O(1)-pressure
rise almost large enough to cause flow reversal even there appears to be
required to generate negative wall shear, which hardly can be considered as
flow separation. This crude estimate is confirmed by a more detailed anal-
ysis dealing with the response of a turbulent small-defect boundary layer
to a surface-mounted obstacle, carried out, among others, by Sykes (1980).
Moreover, to date no self-consistent theory of flow separation compatible
with the classical concept of a turbulent small-defect boundary layer has
been formulated.
The above considerations strikingly contrast the case of laminar bound-
ary layer separation, where the velocity defect is of O(1) across the whole
boundary layer and the associated pressure increase tends to zero as Re → ∞.
It, however, also indicates that a turbulent boundary layer may become
more prone to separation by increasing the velocity defect. That this is
indeed a realistic scenario can be inferred by seeking self-preserving solu-
tions of (12), i.e. by investigating equilibrium boundary layers. Such so-
lutions, where the functions F 1 , T1 , characterising the velocity deficit and
the Reynolds shear stress in the outer layer, respectively, solely depend on
η, exist if the parameter β0 in the outer-layer momentum equation (12a) is
constant, i.e. independent of x. Equation (12a) then assumes the form
(1 + 2β0 )ηF1 − F1 = F1,e (T1 − 1) , (19)
where
Ue ∝ (x−xv )m , m = −β0 /(1+3β0 ) , Δ1 F1,e = (1+3β0 )(x−xv ) . (20)
Herein x = xv denotes the virtual origin of the boundary layer flow. In the
present context flows associated with large values of β0 are of most interest.
1/2
By introducing suitably (re)scaled quantities in the form F1 = β0 F̂ (η̂),
1/2
T1 = β0 T̂ (η̂), η = β0 η̂, the momentum equation (19) reduces to
2η̂ F̂ = F̂e T̂ , F̂e := F̂ (1) (21)
in the limit β0 → ∞. Solutions of (21) describing turbulent boundary lay-

1/2
ers having a velocity deficit measured by uref := β0 uτ uτ have been
obtained first in Mellor and Gibson (1966). Unfortunately, however, it was
not realised that this increase of the velocity defect no longer allows for a
direct match of the flow quantities in the outer and inner layer, which has
significant consequences, to be elucidated below.
We note that in general β0 (x) can be regarded as the leading-order con-
tribution to the so-called Rotta–Clauser pressure-gradient parameter (e.g.
Schlichting and Gersten, 2000), formed with the pressure gradient given
by (10) and the displacement thickness δ ∗ ,
∞
dUe δ ∗ u
β := −Ue , δ ∗ := δo 1− dη . (22)
dx u2τ 0 Ue
As already mentioned by Mellor and Gibson (1966), the quantity β allows

for the appealing physical interpretation that uref is independent of the wall
shear stress u2τ for β0 1.
3 Moderately Large Velocity Defect

Following the considerations summarised in the preceding section, we now
seek solutions of (1) describing a relative velocity defect of O(), where
the newly introduced perturbation parameter is large compared to γ but
still small compared to one: γ 1. From assumption (A), see Subsec-
tion 2.1, we then have −u v ∼ 2 , and the linearised x-momentum equation
immediately yields the estimate δo = Δ for the boundary layer thickness,
where Δ is of O(1) and accounts for its streamwise variation. However, since
−u v ∼ 2 with u2τ , here the solution describing the flow behaviour in
the outer velocity defect region no longer matches with the solution for the
universal wall layer as in the classical case. As a consequence, the leading-
order approximation to the Reynolds shear stress must vanish in the limit
η = y/δo → 0. This indicates that the flow having a velocity defect of O() in
the outer main part of the boundary layer exhibits a wake-type behaviour,
leading to a finite wall slip velocity at its base and, therefore, forces the
emergence of a sublayer, termed intermediate layer, where the magnitude
of −u v reduces to O(u2τ ), being compatible with the wall layer scaling.
230 B. Scheichl
3.1 Intermediate Layer

Here the streamwise velocity component u is expanded about its value
at the base η = 0 of the outer defect region: u/Ue ∼ 1 − W − γUi + · · · ,
so that the quantities W , Ui , assumed to be of O(1), account, respectively,
for the wall slip velocity, given by u = Ue (1 − W ) with W > 0, and the
dominant contribution to u that varies with distance y from the wall. In-
tegration of the x-momentum balance then shows that the Reynolds shear
stress increases linearly with distance y for y/δo 1:
τ ∼ τw − y d(Ue2 W )/dx , y/δi = O(1) . (23)
Herein δi denotes the thickness of the intermediate layer and τ assumes

its near-wall value τw as y/δi → 0. Matching with the wall layer then re-
quires that τw ∼ u2τ , which immediately yields δi = O(u2τ /). By taking
into account (22), we then infer that δi /δo = O(β −1 ) and recover the re-
lationship = O(uτ β 1/2 ), already suggested by the final considerations of
Subsection 2.4. Formal expansions of u and −u v in the intermediate layer,
therefore, are written as
u/Ue ∼ 1 − W (x; , γ) − γUi (x, ζ) , (24a)

2
− u v /(γUe ) ∼ Ti (x, ζ; , γ) ∼ 1 + λζ , (24b)
where ζ := y/δi = y/(Δγ 2 ) and λ := −(Δ/Ue2 ) d(Ue2 W )/dx.

To close the problem for Ui , we adopt the common mixing length concept,

2 ∂u ∂u

−u v := , (25)
∂y ∂y
by assuming that the mixing length behaves as ∼ κy for y = O(δi ), which

is the simplest form allowing for a match with the adjacent layers. Integra-
tion of (24b), supplemented with (25), then yields

κUi = − ln ζ + 2 ln (1 + λζ)1/2 + 1 − 2(1 + λζ)1/2 , (26)
from which the limiting forms
κUi ∼ −2(λζ)1/2 + (λζ)−1/2 + O(ζ −3/2 ) , ζ→∞, (27a)

2
κUi ∼ − ln(λζ/4) − 2 − λζ/2 + O(ζ ) , ζ →0, (27b)
can readily be inferred.

The behaviour (27a) holding at the base of the outer defect layer is recog-
nised as the square-root law deduced first by Townsend (1961) in his study
of turbulent boundary layers exhibiting vanishingly small wall shear stress;

the outermost layer so to speak “anticipates” the approach to separation
as the velocity defect increases to a level larger than uτ . We remark that
Townsend (1961) identified the intermediate region as the so-called “equi-
librium layer”, where convective terms in (1b) are (erroneously within the
framework of asymptotic high-Reynolds-number theory) considered to be
negligibly small. Equation (27b) provides the logarithmic variation of Ui as
ζ → 0, required by the match with the wall layer, which gives rise to the
generalised skin-friction relationship

κ Reγ 2 Ue3
∼ ln 1/2
+ β0 κW + O(γβ0 ) ∼ (1 + W ) ln Re . (28)
γ β0
Note that (28) reduces to (8) when β0 = O(1).

Having demonstrated that the classical theory of turbulent boundary
layers in the limit of large Reynolds number can – in a self-consistent man-
ner – be extended to situations where the velocity defect is asymptotically
large as compared to uτ but still o(1), we now consider the flow behaviour
in the outer wake-type region in more detail.
3.2 Outer Defect Region: Quasi-Equilibrium Flows and Non-

Uniqueness
We first introduce the stream function ψ by u = ∂ψ/∂y, v = −∂ψ/∂x,
so that the continuity equation (1a) is satisfied identically. Following the
arguments put forward at the beginning of Section 3, we write the flow
quantities in the outer layer in the form
p ∼ pe (x) + 2 P (x, η; , γ) , (29a)

ψ/Ue ∼ y − δo F (x, η; , γ) , (29b)

−u2 , −v2 , −u v ∼ Ue2 2 [Rx , Ry , T ](x, η; , γ) . (29c)
As before, here η = y/δo and we accordingly expand
Q ∼ Q1 + Q2 + · · · , Q := F, P, Rx , Ry , T, W , (30a)
2
δo ∼ Δ1 + Δ2 + · · · , (30b)
β/βv ∼ B0 (x) + B1 (x) + · · · , βv → ∞ , (30c)
where we require (without any loss of generality) that βv equals β0 at x = xv ,

so that β0 = βv B0 and B0 (xv ) = 1, Bi (xv ) = 0, i = 1, 2, . . .. In analogy
232 B. Scheichl
to (12), the first-order problem then reads
[Ue2 d(Ue Δ1 )/dx]ηF1 − ∂(Ue3 Δ1 F1 )/∂x = Ue3 T1 , (31a)

F1 (x, 0) = F1 (x, 1)= F1 (x, 1) = T1 (x, 1) = 0 , (31b)
η→0: T ∼ (κηF1 )2 , F1 ∼ W1 (x) − (2/κ)(λη)1/2 . (31c)
In the following we concentrate on solutions which are self-similar up

to second order, i.e. ∂F1 /∂x ≡ ∂T1 /∂x ≡ 0 and ∂F2 /∂x ≡ ∂T2 /∂x ≡ 0. By
again adopting the notations F1 = F̂ (η), T1 = T̂ (η), and setting Δ1 = Δ̂(x),
Ue = Û (x), we recover the requirements (20), (21) for self-similarity at first
order resulting from classical small-defect theory in the limit of large values
of βv , in agreement with (30b) and the definition of β provided by (22):
B0 ≡ 1 , Δ̂F̂e = 3(x − xv ) , Û = (C/3)1/3 (x − xv )−1/3 , (32)
with a constant C, and
2η F̂ = F̂e T̂ , F̂ (0) = T̂ (0) = F̂ (1) = F̂ (1) = T̂ (1) = 0 . (33)
If, as in the discussion of the flow behaviour in the intermediate layer, a

mixing length model T̂ = [l(η)F̂ (η)]2 , l := /δo , in accordance with (25) is
chosen to close the problem, integration of (33) yields the analytical expres-
sions
1 1/2
2 1 1 1/2
2 ,1/2
1 z 1 z
F̂ (η) = dz , F̂e = dz dη . (34)
2F̂e η l(z) 2 0 η l(z)
Equations (34) have been evaluated numerically by using a slightly gener-

alised version of the mixing length closure originally suggested by Michel
et al. (1969),
l = /δo = c I(η)1/2 tanh(κη/c ) , I = 1/(1 + 5.5η6) , c = 0.085 . (35)
Herein I(η) represents the well-known intermittency factor proposed by

. .
Klebanoff (1955). One then obtains W1 = F̂ (0) = 13.868, F̂e = 5.682, and
.
dΔ̂/dx = 0.528, cf. (32). As seen in Figure 2(a), both F̂ and T̂ vanish
quadratically as η → 1 as a result of the boundary conditions expressing
vanishing vorticity, T̂ (1) = T̂ (1) = 0, cf. (33). Also, note that F̂ exhibits
the square-root behaviour required from the match with the intermediate
layer as η → 0.
Turning now to the second-order problem, we consider the most general
case that the wall shear enters the description of the flow in the outer layer
at this level of approximation (principle of least degeneracy). Therefore, we
require 3 T2 (0) ∼ γ 2 , which finally determines the yet unknown magnitude

of relative to γ, namely that = O(γ 2/3 ). Since, as pointed out before
1/2
and indicated by (22) and (32), here setting B0 ≡ 1 yields ∼ γβ0 , this
estimate for implies that β0 = Γ := 1/T2 (0) = O(1). Inspection of the re-
sulting second-order problem then indicates that self similar solutions exist
only if the external velocity distribution (32) predicted by classical theory
is slightly modified in the form
Û (x) = (C/3)1/3 (x − xv )−1/3+μ , μ ∼ γ 2/3 μ1 + · · · , (36)
where the O(1)-parameter μ1 satisfies a solvability condition that represents

the integral momentum balance obtained from integrating the second-order
x-momentum equation from η = 0 to η = 1. It can be cast into the canonical
form
9D̂ 2 μ̂ = 1 + D̂3 , (37a)

1
D̂ = r1/3 Γ 1/3 , μ̂ = r−2/3 μ1 , r = F̂e−1 (F̂ 2 − R̂x + R̂y ) dη . (37b)
0
A graph of the relationship (37a) which represents the key result of the
analysis dealing with boundary layers that are in quasi-equilibrium (i.e. self-
similar to first and second order) having a moderately large velocity defect
is shown in Figure 2(b). Most interesting, it is found that solutions describ-
ing flows of this type exist for μ̂ ≥ μ̂∗ = 21/3 /6 only and form two branches,
associated with non-uniqueness of the quantity D̂, which serves as a mea-
sure of velocity defect, for a specific value of the pressure gradient. Along
the lower branch, D̂ ≤ D̂ ∗ = 21/3 and decreases with increasing values of μ̂,
so that the classical small-defect theory is recovered in the limit μ̂ → ∞,
where D̂ ∼ (9μ̂)−1/2 . In contrast, this limit leads to an unbounded growth
of values D̂ ≥ D̂∗ associated with the upper branch: D̂ ∼ 9μ̂ as μ̂ → ∞.
This immediately raises the question if it is possible to formulate a gen-
eral necessarily nonlinear theory which describes turbulent boundary layers
having a finite velocity defect in the limit of infinite Reynolds number. We
also note that the early experimental observations made by Clauser (1956)
seem to strongly point to this type of non-uniqueness.
The non-uniqueness is intrinsically tied to the nonlinearity of the inertia
terms in (1b). For small-defect boundary layers, these come into play as
inhomogeneities in the second-order flow description, as reflected by the
term D̂3 in (37a). It is, therefore, instructive to seek for double-valued self-
preserving flows for various values of m by starting from an ad-hoc boundary
layer approximation of the governing equations (1). We hence assume that
234 B. Scheichl
T̂ D̂

η
T̂
large defect

D̂ ∗
F̂
small defect

W1

F̂ μ̂∗ μ̂
(a) (b)

Ĝ 107
D̂ 108
109

6
Re = 106
Re = 10

107
8
10
9
10

ï ï ï ï ï ï m μ̂
(c) (d)
Figure 2: Non-unique quasi-equilibrium flows; (a) F̂ (η), T̂ (η),
dashed : asymptotes found from (31b), (31c); (b) canonical representa-
tion (37a), dashed : asymptotes (see text) and osculating parabola in the
point (μ∗ , D̂∗ ); (c), (d) solid : solutions of (38) for various values of Re;
(c) circles: experimental data; (d) dashed : asymptote (37a) for Re → ∞.
ψ = δo Ue g(η), τ = Ue2 t(η), with Ue given by (20), and δo = a(x − xv ), to

end up with the boundary-value problem
t /a = m(g 2 − 1) − (m + 1)gg , (38a)

2
η→0: g→0, γηg → κ , t→γ , (38b)

η=1: g =1, g =t=0. (38c)
The boundary conditions (38b) and (38c) account for the behaviour of F1 ,
F2 , T1 , T2 entering the expansions (16), even in case of a moderately small
defect still formally valid in the entire small-defect region, near the base
and the outer edge of the latter, respectively. When supplemented with an
asymptotically correct closure for t, i.e. one consistent with these conditions,
for sufficiently small values of γ and external flows represented by m the
quantities g, t, a represent an asymptotically accurate description of small-
defect flows that are in equilibrium up to second order with respect to the
expansions (16).
We again adopt (25) by setting t = (lg )2 , with l satisfying (35). In
order to detect non-uniqueness, (38) was solved numerically for g(η), t(η),
and m by prescribing values of Re and a. Then γ is computed from (8),
and increasing the value of a means increasing the boundary layer thickness
at some position x. The approximation β0 ∼ −ma[1 − g(1)]/γ 2 + O(γ 2 ) in-
1/2 1/2
ferred from (22) is used to evaluate ∼ γβ0 , Γ 1/3 = γ 1/3 β0 , and, in
turn, μ̂ and D̂ from (37b). Here we set R̂x = R̂y = 0, owing to the lack
of reliable closures for these stress components but in agreement with the
boundary layer approximation adopted, so that r is identified with the so-
called shape factor Ĝ (cf. Schlichting and Gersten, 2000). Also, the esti-
mates F̂ ∼ (1 − g )/ + O(2 ), F̂e ∼ [1 − g(1)]/ + O(2 ) are employed.
Figure 2(c) shows the comparison of the resulting relationship m(Ĝ; Re)
with data extracted from the measurements by Simpson et al. (1981) of a
massively separating boundary layer under the action of an adverse pres-
sure gradient (Schlichting and Gersten, 2000, p. 590): here Ue is found to
obey a power law according to (20) with m being a slowly varying function
of x, and the value of Re is estimated roughly as 3 × 106 . Notwithstand-
ing this uncertainty, the neglect of effects due to Reynolds normal stresses,
and the fact that the experimental flow is in equilibrium only locally, the
agreement with the results that are consistent with the asymptotic theory
is encouraging. Figure 2(d) uncovers the rather slow convergence of the
function μ̂(D̂; Re) towards the canonical relationship (37), cf. Figure 2(b),
attained for Re → ∞, which originates from the logarithmic dependence
of γ on Re.
4 Large Velocity Deficit

As in the cases of small and moderately small velocity defect we require
the boundary layer to be slender. However, in contrast to the considera-
tions of Sections 2 and 3, the validity of this requirement can no longer
be inferred from assumption (A) and the balance between convective and
Reynolds stress gradient terms in the outer predominantly inviscid region of
the boundary layer which now yields ∂τ /∂y = O(1), rather than ∂τ /∂y 1
as earlier. A hint how this difficulty can be overcome is provided by the
observation that the transition from a small to a moderately large velocity
defect is accompanied with the emergence of a wake-type flow in this outer
236 B. Scheichl
layer. One expects that this effect will become more pronounced as the
velocity defect increases further, suggesting in turn that the outer part of
the boundary layer, having a velocity defect of O(1), essentially behaves
as a turbulent free shear layer. An attractive strategy then is to combine
the asymptotic treatment of such flows, (see Schneider, 1991) in which the
experimentally observed slenderness is enforced through the introduction
of a Reynolds-number-independent parameter α 1 with the asymptotic
theory of turbulent wall bounded flows.
4.1 Outer Wake Region

Let the parameter α 1 measure the lateral extent of the outer wake
region, so that ȳ := y/α = O(1). Appropriate expansions of the various field
quantities then are
p ∼ pe (x) + O(α) , q ∼ α q0 (x, ȳ) + o(α) , (39)
where q stands for Δ, ψ, τ = −u v (as well as the remaining components

−u2 , −v 2 of the Reynolds stress tensor). From substitution into (1b–1c)
the leading order outer wake problem is found to be
∂ψ0 ∂ 2 ψ0 ∂ψ0 ∂ 2 ψ0 dUe ∂τ0

− 2
= −Ue + , (40a)
∂ ȳ ∂ ȳ∂x ∂x ∂ ȳ dx ∂ ȳ
ȳ = 0 : ψ0 = τ0 = 0 , (40b)
ȳ = Δ0 (x) : ∂ψ0 /∂ ȳ = Ue , τ0 = 0 . (40c)
As in the case of a moderately large velocity defect, we expect a finite

wall slip Us (x) := ∂ψ0 /∂ ȳ at the base ȳ = 0 of this outer layer, which yields
the limiting behaviour
∂ψ0 /∂ ȳ ∼ Us (x) + O(ȳ 3/2 ) , τ0 ∼ Λ0 ȳ + O(ȳ 3/2 ) , (41)
with Λ0 := Us dUs /dx − Ue dUe /dx > 0.

It is easily verified that the various layers introduced so far in the de-
scription of turbulent boundary layers share the property that their lateral
extent is of the order of the mixing length , defined by (25), characteristic
of the respective layer. In contrast, the scalings given by (39) imply that
is much smaller than the thickness of the outer wake region: ∼ α3/2 α.
This is a typical feature of free shear layers, of course, but also indicates
that the outer wake region “starts to feel” the presence of the confining wall
at distances y ∼ α3/2 , which in turn causes the emergence of an inner wake
region.
4.2 Inner Wake Region

By introducing the stretched wall distance Y = y/α3/2 = O(1), inspec-
tion of (41) suggests the expansions
ψ ∼ α3/2 Us (x) + α9/4 ψ̄(x, Y ) + · · · , (42a)

3/2 3/2
τ ∼α T̄ (x, Y ) + · · · , ∼α L̄(x, Y ) + · · · , (42b)
which leads to
T̄ = Λ0 Y . (43)
Furthermore, T̄ and ψ̄ are subject to the boundary conditions
T (x, 0) = ψ̄(x, 0) = 0 , (44a)

1/2
2 Λ0
ψ̄Y ∼ Y 3/2 , Y →∞, L̄0 = limY →∞ L̄ . (44b)
3 L̄0
The solution of the inner wake problem posed by (43), (44) can be obtained
in closed form. It exhibits the expected square-root behaviour of ψ̄Y ,
(Λ0 Y )1/2
ψ̄Y ∼ Ūs (x) + 2 , L̄ ∼ χ(x)Ȳ , Y →0. (45)
χ(x)
Here Ūs (x) denotes the correction of the slip velocity Us (x) caused by the
inner wake region,
us ∼ Us (x) + α3/4 Ūs (x) + · · · , (46a)

∞
1 1
Ūs (x) = − − (Λ0 Y )1/2 dY (< 0) . (46b)
0 L̄ L̄0
At this point it is important to recall the basic assumption made at

the beginning of this section, namely, that the slenderness parameter α
is independent of Re, or more generally, asymptotes to a small but finite
value as Re → ∞. As a consequence, the outer and inner wake regions
provide a complete description of the boundary layer flow in the formal
limit Re −1 = 0. If, however, 0 < 1/Re 1 an additional sublayer forms at
the base of the inner wake region. This sublayer plays a similar role as the
intermediate layer discussed in Subsection 3.1: there the magnitude of the
Reynolds shear stress, still varying linearly with distance from the wall, is
reduced to O(u2τ ), which is necessary to provide the square-root behaviour
expressed in (45) and, finally, to allow for the match with the universal wall
layer (see Scheichl and Kluwick, 2007b).
238 B. Scheichl
4.3 Numerical Solution of the Leading-Order Outer-Wake Prob-
lem
As in Subsection 3.2, here again (25) and (35) shall be adopted to close
the outer wake problem posed by (40). Numerical calculations were carried
out for a family of retarded external flows controlled by two parameters ms ,
k, with ms < 0, 0 ≤ k < 1:
Ue (x; ms , k) = (1 + x)m(x;ms ,k) , (47a)

m k 3
=1+ Θ(2 − x) 1 − (1 − x)2 . (47b)
ms 1−k
Herein Θ denotes the Heaviside step function. Self-similar solutions of the

form ψ0 = Δ0 f (ξ), ξ := Y /Δ0 , Δ0 = b(1 + x), where b is a constant and
the position x = −1 defines the virtual origin of the flow, exist for k = 0 if
ms > −1/3 and are used to provide initial conditions at x = 0 for the down-
stream integration of (40) with Ue given by (47). As a specific example, we
consider the case F (0) = 0.95 of a relatively small velocity defect, imposed
at x = 0, for which the requirement of self-similarity for −1 < x < 0 yields
. .
b = 0.3656 and ms = −0.3292. The key results which are representative
for the responding boundary layer and, most important, indicate that the
present theory is capable of describing the approach to separation are dis-
played in Figure 3. If k is sufficiently small, the distribution of the wall slip
velocity Us is smooth and Us > 0 throughout. However, when k reaches
.
a critical value kM = 0.84258, the slip velocity Us is found to vanish at a
single location x = xM , but is positive elsewhere. A further increase of k
provokes a breakdown of the calculations, accompanied with the formation
of a weak singularity slightly upstream of xM at x = xG . A similar be-
haviour is observed for the scaled boundary layer thickness Δ0 , which is
smooth in the subcritical case k < kM , exhibits a rather sharp peak Δ0,M
for k = kM at x = xM , and approaches a finite limit Δ0,G in an apparently
singular manner in the supercritical case k > kM .
Following the qualitatively similar behaviour of the wall shear stress
that replaces Us in the case of laminar boundary layers (see Ruban, 1981,
1982; Stewartson et al., 1982) the critical solution with k = kM is termed a
marginally separating boundary layer solution. However, in vivid contrast
to its laminar counterpart, is is clearly seen to be locally asymmetric with
respect to x = xM where it is singular. This numerical finding is supported
by a local analysis of the flow behaviour near x = xM , carried out by Scheichl
and Kluwick (2007a): it indicates that Us decreases linearly with x upstream
Us Δ0
Δ0,M
Δ0,G

Us

Δ0

xG xM

x
Figure 3: Solutions of (40) for |x − xM | 1, |k − kM | 1, dashed : asymp-
totes expressed by (48b) and (49).
of x = xM but exhibits a square-root singularity as x − xM → 0+ ,

1/2
Us /P00 ∼ −B(x − xM ) , x − xM → 0− , (48a)
1/2
Us /P00 ∼ U+ (x − xM )1/2 ,
x − xM → 0+ , (48b)
.
where P00 = (dpe /dx)(xM ). It is found that P00 = 0.02272 in case of the
specific choice (47) of the external-flow speed Ue . Also, U+ takes on an
.
universal value, U+ = 1.1835, whereas the constant B remains arbitrary in
the local investigation and has to be determined by comparison with the
numerical results for x ≤ xM .
This local analysis also shows that a square-root singularity forms at a
position x = xG < xM for k > kM ,
1/2
Us /P00 ∼ U− (xG − x)1/2 , x − xG → 0 − , (49)
with some U− to be determined numerically, and that the solution cannot
be extended further downstream. This behaviour, which has been described
first by Melnik (1989), is reminiscent of the Goldstein singularity well-known
from the theory of laminar boundary layers and, therefore, will be termed
the turbulent Goldstein singularity. As shown in the next section, the bi-
furcating behaviour of the solutions for k − kM → 0 is associated with the
occurrence of marginally separating flow.
4.4 Marginal Separation

According to the original boundary layer concept, pressure disturbances
caused by the displacement of the external inviscid flow due to the mo-
mentum deficit, which is associated with the reduced velocities close to the
240 B. Scheichl
wall, represent a higher order effect. Accordingly, higher-order corrections

to the leading-oder approximation of the flow quantities inside and out-
side the boundary layer can be calculated in subsequent steps. However,
as found first for laminar flows, this so-called hierarchical structure of the
perturbation scheme breaks down in regions where the displacement thick-
ness changes so rapidly that the resulting pressure response is large enough
to affect the lowest-order boundary layer approximation (e.g. Stewartson,
1974). A similar situation is encountered for the type of turbulent flows dis-
cussed in the preceding section. Indeed, the slope discontinuity of Δ0 and,
in turn, of the displacement thickness forces a singularity in the response
pressure, indicating a breakdown of the hierarchical approach to boundary
layer theory. As for laminar flows (see Ruban, 1981, 1982; Stewartson et al.,
1982) this deficiency can be overcome by adopting a local interaction strat-
egy, so that the induced pressure disturbances enter the description of the
flow in leading rather than higher order (see Scheichl and Kluwick, 2007a).
We subsequently present the essential results of this theory.
Again, similar to laminar flows, three layers (decks) characterising re-
gions of different flow behaviour have to be distinguished inside the local
interaction region, see Figure 4. Effects of Reynolds stresses are found to
be confined to the lower deck region (LD), having a streamwise and lateral
extent of O(α3/5 ) and O(α6/5 ), respectively. Here the flow is governed by
equations of the form (40). The majority of the boundary layer, i.e. the
main deck (MD), behaves passively in the sense that it transfers displace-
ment effects caused by the lower deck region unchanged to the external flow
region taking part in the interaction process, the so-called upper deck (UD),
and transfers the resulting pressure response unchanged to the lower deck.
Solutions to the leading-order main and upper deck problems can be ob-
tained in closed form which finally leads to the fundamental lower deck
problem. By using suitably stretched variables, it can be written in terms
of a stream function ψ̂(X̂, Ŷ ) as
∂ ψ̂ ∂ 2 ψ̂ ∂ ψ̂ ∂ 2 ψ̂ dP̂ ∂ T̂
− = −1 − Λ̂(Γ̂ ) + , (50a)
∂ Ŷ ∂ Ŷ ∂ X̂ ∂ X̂ ∂ Ŷ 2 dX̂ ∂ Ŷ

∂ 2 ψ̂ ∂ 2ψ̂
T̂ = , (50b)
∂ Ŷ 2 ∂ Ŷ 2

1 ∞ Â (Ŝ)
P̂ (X̂) = − dŜ , (50c)
π −∞ X̂ − Ŝ
Ŷ = 0 : ψ̂ = T̂ = 0 , (50d)
Ŷ → ∞ : T̂ − Ŷ → Â(X̂) , (50e)
X̂ → −∞ : ψ̂ → (4/15)Ŷ 5/2 + Γ̂ Ŷ , 0 ≤ Γ̂ ≤ 1 , (50f)

X̂ → ∞ : ψ̂ → X̂ 5/6 F+ (η̂) , η̂ := Ŷ /X̂ 1/3 . (50g)
The first and the second term on the right-hand side of (50a) account for
the imposed and the induced pressure, respectively. The latter is given
by the Hilbert integral (50c), where Â characterises the displacement ef-
fect exerted by the lower deck region. The far-field condition (50e) ex-
presses the passive character of the main deck mentioned before, whereas
the conditions (50f), (50g) follow from the match with regions LD− , LD+
immediately upstream and downstream of the local interaction zone. The
analysis of region LD+ determines the function F+ (η̂). Finally, the parame-
ter Γ̂ measures the intensity of the interaction process as the monotonically
increasing but otherwise arbitrary function Λ̂(Γ̂ ) expresses the magnitude
of the induced pressure gradient. The distinguished limit underlying the
interacting-flow description is given by α3/10 ln |k − kM | = O(1) as k → kM ,
α → 0.
UD y
y ∼ αΔ0 (x) O(α3/5 )
OW MD−
MD O(α6/5 ) MD+
3/2
O(α )
IW LD− LD+
LD
x
3/5
O(α )
Figure 4: Triple-deck structure, for captions see text, subscripts “−” and
“+” refer to the continuation of flow regions up- and downstream of the
local interaction zone, dashed line indicates inner wake.
As a representative example of flows encountering separation, the distri-

butions of Â, P̂ , and the wall slip Ûs := (∂ ψ̂/∂ Ŷ )(X̂, Ŷ = 0), obtained by
numerical solution of the triple-deck problem (50) for Γ̂ = 0.019, Λ̂ = 3, are
depicted in Figure 5(a). Here the upstream and downstream asymptotes
for Â, P̂ , Ûs have been obtained from the analysis of the flow behaviour in
the pre- and post-interaction regions (subscripts “−” and “+” in Figure 4),
and X̂D and X̂R denote the X̂-positions of, respectively, detachment and
242 B. Scheichl
reattachment. For example, under the assumption Λ̂(Γ̂ ) ≡ 3, reversed-flow

. .
regions are observed for Γ̂min ≤ Γ̂ ≤ Γ̂max , Γ̂min = 0.017, Γ̂max = 0.205. It
is interesting to note that the passage of Ûs into the reversed-flow region
where Ûs < 0 causes the interaction pressure P̂ to drop initially before it
rises sharply, overshoots and finally tends to zero in the limit X̂ → ∞. This
is in striking contrast to laminar flows, where flow separation always is trig-
gered by an initial pressure rise, and reflects the fact that – in the case
of turbulent flows considered here – the streamwise velocity component at
the base Ŷ = 0 of the lower deck region is allowed to take on finite values,
whereas the no-slip condition is enforced in its laminar counterpart.
ï ï X̂

Â,
Ûs Ŷ
P̂
Ûs
P̂ ï

Γ̂ Â
X̂D X̂R ï
ï
ï ï ï X̂ ï X̂Dï ï ï X̂ ï ï X̂Rï
Figure 5: Specific solution of (50); (a) key quantities, bottom abscissa: X̂-
values for Â(X̂), Ûs (X̂), top abscissa: X̂-values for P̂ (X̂), dashed : asymp-
totes; (b) streamlines, bold : separating streamline ψ̂ = 0.
The streamlines inside the lower-deck region (constant values of ψ̂) are
displayed in Figure 5(b), which clearly shows the formation of a recirculat-
ing eddy. Also, we draw attention to the increasing density of streamlines
further away from the wall and downstream of reattachment, associated
with the strong acceleration of the fluid there as evident from the rapid
increase of Ûs .
The interaction process outlined so far describes the (local) behaviour of
marginally separated turbulent flows in the formal limit 1/Re = 0. As in the
case of conventional, i.e. hierarchical, boundary layers having a velocity of
defect of O(1), additional sublayers form closer to the wall if 1/Re 1 but
finite. Their analysis, outlined in detail by Scheichl and Kluwick (2007b),
provides the skin-friction relationship in generalised form to include the
effects of vanishing and negative wall shear – treated first in a systematic
way by Schlichting and Gersten (2000) – but also shows that these layers
behave passively insofar as the lower deck problem (50) remains intact.
5 Conclusions and Outlook

In this presentation an attempt has been made to derive the classical two-
layer structure of a turbulent small-defect boundary layer from a minimum
of assumptions. As in the overview given by Walker (1998), but in contrast
to earlier investigations (e.g. Mellor, 1972), the (logarithmic) law of the wall
is taken basically as an empirical observation rather than a consequence of
matching inner and outer layers, as the latter is not felt rich enough to
provide a stringent foundation of this important relationship reflecting the
dynamics of the flow close to the wall, which is not understood in full at
present. Probably the first successful model that describes essential aspects
of this dynamics is provided by Prandtl’s mixing length concept, proposed
more than 50 years before the advent of asymptotic theories in fluid me-
chanics. Significant progress has been achieved in more recent years and,
in particular, by the pioneering work of Walker (e.g. Walker, 1998), whose
untimely death ended a line of thought which certainly ought to be taken
up again.
Following the brief outline of the classical small-defect theory, it is shown
how a description of turbulent boundary layers having a slightly larger (i.e.
moderately large) velocity defect, where the outer predominantly inviscid
layer starts to develop a wake-type behaviour, can be formulated. Further
increase of the velocity defect to values of O(1) causes the wake region to
become even more pronounced and is seen to allow for the occurrence of
reversed-flow regions close to the wall, resulting in what we believe to be
the first fully self-consistent theory of marginally separated turbulent flows.
Unfortunately, however, this success seemingly does not shed light on
the phenomenon of global or gross separation associated with flows past
(more-or-less) blunt bodies or, to put it more precisely, flows which start
at a stagnation point rather than a sharp leading edge. Indeed, a recent
careful numerical investigation for the canonical case of a circular cylinder,
presented, among others, by Scheichl et al. (2008b), Scheichl and Kluwick
(2008c), undoubtedly indicates that the boundary layer approaching sepa-
ration exhibits a small rather than a large velocity defect, leading in turn
to the dilemma addressed in Subsection 2.4. The accompanying asymptotic
analysis based on the turbulence intensity gauge model introduced by Neish
and Smith (1992), however, strongly suggests that a boundary layer form-
ing on a body of finite extent and originating in a front stagnation point
does not reach a fully developed turbulent state, even in the limit Re → ∞.
Specifically, it is found that the boundary layer thickness δ and the Reynolds
shear stress are slightly smaller than predicted by classical small-defect the-
ory, while, most important, the thickness of the wall layer is slightly larger.
244 B. Scheichl
In fact, δ varies predominantly algebraically with Re, whereas the velocity

defect measure γ in the outer region is still of O(1/ ln Re), conforming to the
matching condition (8) decisive for the classical flow description. As a con-
sequence, the outer large-momentum region does not penetrate to distances
from the wall which are transcendentally small. In turn, this situation opens
the possibility to formulate a local interaction mechanism that describes the
detachment of the boundary layer from the solid wall within the framework
of free-streamline theory at pressure levels which are compatible with ex-
perimental observation. This is a topic of intense current investigations; for
preliminary results see Scheichl et al. (2009) or Scheichl et al. (2010).
Bibliography
N. Afzal. Scaling of Power Law Velocity Profile in Wall-Bounded Turbulent
Shear Flows. Meeting paper 2005-109, AIAA, 2005.
G. I. Barenblatt and A. J. Chorin. A note on the intermediate region in
turbulent boundary layers. Phys. Fluids, 12(9):2159–2161, 2000. Letter.
G. I. Barenblatt, A. J. Chorin, and V. M. Prototoskin. Self-similar interme-
diate structures in turbulent boundary layers at large Reynolds numbers.
J. Fluid Mech., 410:263–283, 2000.
G. I. Barenblatt and N. Goldenfeld. Does fully developed turbulence exist?
Reynolds number independence versus asymptotic covariance. Phys. Flu-
ids, 7(12):3078–3082, 1995.
W. B. Bush and F. E. Fendell. Asymptotic analysis of turbulent channel
and boundary-layer flow. J. Fluid Mech., 56(4):657–681, 1972.
F. H. Clauser. The Turbulent Boundary Layer. In H. L. Dryden and
Th. von Kármán, editors, Advances in Applied Mechanics, volume 4,
pages 1–51. Academic Press, New York, 1956.
E. Deriat and J.-P. Guiraud. On the asymptotic description of turbulent
boundary layers. J. Theor. Appl. Mech., 109–140, 1986. Special Issue on
Asymptotic Modelling of Fluid Flows. Original French article in J. Méc.
Théor. Appl.
F. E. Fendell. Singular Perturbation and Turbulent Shear Flow near Walls.
J. Astronaut. Sci., 20(11):129–165, 1972.
P. S. Klebanoff. Characteristics of turbulence in a boundary layer with
zero pressure gradient. NACA Report 1247, NASA – Langley Research
Center, Hampton, VA, 1955. See also NACA Technical Note 3178, 1954.
A. Kluwick and B. Scheichl. High-Reynolds-Number Asymptotics of Tur-
bulent Boundary Layers: From Fully Attached to Marginally Separated
Flows. Invited Paper. In A. F. Hegarty, N. Kopteva, E. O’Riordan,
and M. Stynes, editors, BAIL 2008 – Boundary and Interior Layers.
Proceedings of the International Conference on Boundary and Interior

Layers – Computational and Asymptotic Methods, Limerick, July 2008,
volume 69 of Lecture Notes in Computational Science and Engineering,
pages 3–22. Springer-Verlag, Berlin, Heidelberg, 2009.
G. L. Mellor. The Large Reynolds Number, Asymptotic Theory of Turbulent
Boundary Layers. Int. J. Engng Sci., 10(10):851–873, 1972.
G. L. Mellor and D. M. Gibson. Equilibrium turbulent boundary layers.
J. Fluid Mech., 24(2):225–253, 1966.
R. E. Melnik. An Asymptotic Theory of Turbulent Separation. Computers
& Fluids, 17(1):165–184, 1989.
R. Michel, C. Quémard, and R. Durant. Hypothesis on the mixing length
and application to the calculation of the turbulent boundary layers. In
S. J. Kline, M. V. Morkovin, G. Sovran, and D. J. Cockrell, editors, Pro-
ceedings of Computation of Turbulent Boundary Layers – 1968 AFOSR-
IFP-Stanford Conference, volume 1, pages 195–207. Stanford University,
CA, 1969.
P. A. Monkewitz, K. A. Chauhan, and H. M. Nagib. Self-consistent
high-Reynolds-number asymptotics for zero-pressure-gradient turbulent
boundary layers. Phys. Fluids, 19(11):115101-1–115101-12, 2007.
A. Neish and F. T. Smith. On turbulent separation in the flow past a bluff
body. J. Fluid Mech., 241:443–467, August 1992.
J. M. Österlund, A. V. Johansson, H. M. Nagib, and M. Hites. A note on
the overlap region in turbulent boundary layers. Phys. Fluids, 12(1):1–4,
2000. Letter.
A. I. Ruban. Singular Solutions of the Boundary Layer Equations Which
Can Be Extended Continuously Through the Point of Zero Surface Fric-
tion. Fluid Dyn., 16(6):835–843, 1981. Original Russian article in
Izvestiya Akademii Nauk SSSR, Mekhanika Zhidkosti i Gaza (6), 42–52,
1981.
A. I. Ruban. Asymptotic Theory of Short Separation Regions on the Lead-
ing Edge of a Slender Airfoil. Fluid Dyn., 17(1):33–41, 1982. Original
Russian article in Izvestiya Akademii Nauk SSSR, Mekhanika Zhidkosti
i Gaza (1), 42–51, 1982.
B. Scheichl and A. Kluwick. Turbulent Marginal Separation and the Tur-
bulent Goldstein Problem. AIAA J., 45(1):20–36, 2007a. See also
AIAA Meeting paper 2005-4936.
B. Scheichl and A. Kluwick. On turbulent marginal boundary layer separa-
tion: how the half-power law supersedes the logarithmic law of the wall.
Int. J. Computing Science and Mathematics (IJCSM), 1(2/3/4):343–359,
2007b. Special Issue on Problems Exhibiting Boundary and Interior Lay-
ers.
246 B. Scheichl
B. Scheichl and A. Kluwick. Asymptotic theory of bluff-body separation: a

novel shear-layer scaling deduced from an investigation of the unsteady
motion. J. Fluids Struct., 24(8):1326–1338, December 2008. Special Issue
on the IUTAM Symposium on Separated Flows and their Control.
B. Scheichl, A. Kluwick, and M. Alletto. “How turbulent” is the boundary
layer separating from a bluff body for arbitrarily large Reynolds num-
bers? Acta Mech., 201(1–4):131–151, 2008b. Special Issue dedicated to
Professor Wilhelm Schneider on the occasion of his 70th birthday.
B. Scheichl and A. Kluwick. Level of Turbulence Intensities Associated
with Bluff-Body Separation for Large Values of the Reynolds number.
Meeting paper 2008-4348, AIAA, 2008c.
B. Scheichl, A. Kluwick, and F. T. Smith. Time-mean description of tur-
bulent bluff-body separation in the high-Reynolds-number limit. In
B. Eckhardt, editor, Advances in Turbulence XII. Proceedings of the 12th
EUROMECH European Turbulence Conference, September 7-10, 2009,
Marburg, Germany, volume 132 of Springer Proceedings in Physics,
pages 577–580. Springer-Verlag, Berlin, Heidelberg, New York, 2009.
B. Scheichl, A. Kluwick, and F. T. Smith. Break away separation for high
turbulence intensity and large Reynolds number. J. Fluid Mech., sub-
mitted in revised form, 2010.
H. Schlichting and K. Gersten. Boundary-Layer Theory. Springer-Verlag,
Berlin, Heidelberg, New York, 8th edition, 2000.
W. Schneider. Boundary-layer theory of free turbulent shear flows. Z. Flug-
wiss. Weltraumforsch. (J. Flight Sci. Space Res.), 15(3):143–158, 1991.
R. L. Simpson, Y.-T. Chew, and B. G. Shivaprasad. The structure of a
separating turbulent boundary layer. Part 1. Mean flow and Reynolds
stresses. J. Fluid Mech., 113(2):23–51, 1981.
K. Stewartson. Multistructured Boundary Layers on Flat Plates and Re-
lated Bodies. In Y. Chia-Sun, editor, Advances in Applied Mechanics,
volume 14, pages 145–239. Academic Press, New York, San Francisco,
London, 1974.
K. Stewartson, F. T. Smith, and K. Kaups. Marginal Separation. Studies
in Applied Mathematics, 67(1):45–61, 1982.
R. I. Sykes. An asymptotic theory of incompressible turbulent boundary-
layer flow over a small hump. J. Fluid Mech., 101(3):631–646, 1980.
A. A. Townsend. Equilibrium layers and wall turbulence. J. Fluid Mech.,
11(1):97–120, 1961.
J. D. A. Walker. Turbulent Boundary Layers II: Further Developments.
In A. Kluwick, editor, Recent Advances in Boundary Layer Theory, vol-
ume 390 of CISM Courses and Lectures, pages 145–230. Springer-Verlag,
Vienna, New York, 1998.
K. S. Yajnik. Asymptotic theory of turbulent shear flows. J. Fluid Mech.,
42(2):411–427, 1970.
Interactive Boundary Layer (IBL)
P.-Y. Lagrée
CNRS & UPMC Univ Paris 06, UMR 7190,
Institut Jean Le Rond d’Alembert, Boı̂te 162, F-75005 Paris, France
[email protected] ; www.lmm.jussieu.fr/∼lagree
Abstract We present here the Interacting Boundary Layer Equa-
tions. It is called Inviscid-Viscous Interactions as well. This is a way
to solve an approximation of the Navier Stokes equations at large
Reynolds number using the Ideal Fluid / Boundary Layer decom-
position. But, instead of solving first the ideal Fluid and second the
Boundary Layer, both are solved together. This ”strong coupling”
(or this Viscous-Inviscid Interaction) allows to compute separated
flows. This was impossible in the classical Boundary Layer frame-
work, because in this framework, the boundary layer is constrained
by the Ideal Fluid which imposes its slip velocity at the wall. This
coupling is justified in the Triple Deck theory which is the rational
explanation of IBL. We present some numerical experiments show-
ing some simple academic examples of interactions such as flows
over bumps or wedges in subsonic, supersonic, subcritical and su-
percritical external flows and in pipes. Some examples from the
literature are then presented.
1 Introduction
The concept of Ideal Fluid/ Boundary Layer decomposition is classical (see
Schlichting books (42), (41) (17) or Prandtl (37)). The rational technique
of expansion has been presented by Van Dyke (48) and is explained in
its book (51). The procedure is as follows, starting from Navier Stokes
equations we put first 1/Re = 0; this gives the Euler description (called
”outer problem” see (51)). In this non viscous description, the flow slips
at the wall. This gives an ”outer velocity” at the wall, parallel to the wall.
This singular behavior √ is removed by the introduction of a thin layer of
relative thickness 1/ Re, the boundary layer (called ”inner problem” see
(51)). The velocity at the upper bound of this layer (at infinity in local
inner boundary layer variables) is by matching the ideal fluid velocity at
the wall. In this boundary layer, viscous effects act in order to decrease this
slip velocity to full fit the no slip condition.
© CISM, Udine 2010
248 P.-Y. Lagrée
)

7@501;-=2-5?

:@/417585?1
Figure 1. The typical problem a plane plate (neglect curvature) with a small
bump on it. The position L is used to scale x and y, the velocity U∞ is used to
scale the velocities
From a practical aeronautical point of view, the ideal fluid description

gives the lift of the wing, the viscous layer gives the drag (we are aware of
the induced drag on finite span wings which is a ideal fluid effect).
But, everything is not so simple as just described here and in section
2 were we present classical notations and solutions (Falkner Skan, Von
Kármán...). There are problems when computing the boundary layer. The
first most important problem is the boundary layer separation problem. It
is introduced in section 3. We introduce in this same section an other im-
portant problem which is the ”upstream influence problem”. We will show
in the section 4 that to solve those two problems, the good strategy is a
strategy of ”strong interaction” between the boundary layer and the ideal
fluid, furthermore, to obtain boundary layer separation one has to solve the
boundary layer in an inverse way. So in section 5 is presented what is called
”Interactive Boundary Layer” or ”Viscous Inviscid Interaction”, or ”Viscid-
Inviscid coupling” or ”Inviscid Viscous Interaction”, see (3), (7), (12), (28),
(9) and (44)). The strategy of coupling is presented. Some practical exam-
ples with boundary layer separation from literature and for various flows
régimes are then presented in the final section 6.
2 Classical Boundary Layer Equations

2.1 Prandtl Boundary Layer equations
Let us consider a simple semi infinite flat plate with a kind of bump on
it (see figure 1). Using L and U∞ as scales (x = Lx̄, y = Lȳ, u = U∞ ū,
and v = U∞ v̄), first, we compute the ideal fluid solution. We obtain the
”slip velocity”, written ūe the value of the ideal fluid velocity at the wall
(the transverse velocity is zero). Near the wall, the ideal fluid solution is no
more valid as the velocity should be zero at a wall. We have to introduce a
”Boundary Layer”. To obtain this we use the ”least degeneracy principle”
Interactive Boundary Layers 249
or ”dominant balance” (Van Dyke (51), Darrozès & François (11)): we want
the convective terms and at least re hook one diffusive term (as ȳ = ỹδ/L):
∂ ũ 1 ∂ 2 ũ
ũ ∝ ,
∂ x̄ Re(δ/L) ∂ ỹ 2
2
we then say that the boundary layer is of relative order Re−1/2 . So we

classically define
x = Lx̄, y = Re−1/2 Lỹ, u = U∞ ũ, v = U∞ Re−1/2 ṽ, p = ρU∞
2
p̃,
in those new scales, the Navier Stokes equation are the Prandtl equations:
∂ ũ ∂ṽ ∂ ũ ∂ ũ dūe ∂ 2 ũ
+ = 0, ũ + ṽ = ūe + 2, (1)
∂ x̄ ∂ ỹ ∂ x̄ ∂ ỹ dx̄ ∂ ỹ
with no slip boundary conditions ũ = ṽ = 0 on the wall, and the matching
ũ(x̄, ỹ → ∞) → ūe (x̄, ȳ → 0);
at the first position in x̄, a initial profile of ũ should be given to start the
resolution which is marching in x̄. With a given external velocity, those
equations are in principle solvable.
2.2 Blowing Velocity Induced by the Boundary Layer

But, before going further, let us present a useful relation for the sequel.
Let us look at the transverse velocity in the Boundary Layer, in fact, we do
not match the transverse velocity (at initial order). The reason is that it is
of order Re−1/2 , which is negligible for the Ideal Fluid. Starting from the
incompressibility equation and adding and substracting the same derivative
of the velocity ∂∂ūx̄e (in the spirit of Von Kármán integral equations see
thereafter), we obtain, after integration up to an ỹ (x̄ and ỹ are independent
variables), that the transverse velocity is:
ỹ
∂ ∂ ūe
ṽ(ỹ) − ṽ(0) = − (ũ − ūe )dỹ − ỹ .
∂ x̄ 0 ∂ x̄
So, if ỹ is large enough and as ṽ(0) = 0, we obtain the behavior for large
enough ỹ, with the help of the displacement thickness δ̃1 :
∞
d ∂ ūe ũ
ṽ(ỹ) (ūe δ̃1 ) − ỹ with δ̃1 = (1 − )dỹ. (2)
dx̄ ∂ x̄ 0 ū e
This velocity must be expressed in outer variables, so it is multiplied by

Re−1/2 , and ȳ = Re−1/2 ỹ is used. Now, we write the outer velocity in the
250 P.-Y. Lagrée
ideal fluid as a Taylor expansion near the wall for small ȳ (and taken into
account the incompressibility of the ideal fluid):
∂v̄ ∂ ūe
v̄ = v̄(x̄, 0) + ȳ + ... = v̄(x̄, 0) − ȳ + ...
∂ ȳ ∂ x̄
matching this velocity and the boundary layer velocity 2 shows that:
d
v̄(x̄, 0) = Re−1/2 (ūe δ̃1 ).
dx
Hence the boundary layer disturbs the ideal fluid at order Re−1/2 . It is
called the ”blowing velocity”. So the velocity in the ideal fluid (called
transpiration boundary condition as well) induces perturbations at the order
Re−1/2 in the fluid:
ū = ū1 + Re−1/2 ū2 , v̄ = v̄1 + Re−1/2 v̄2 , p̄ = p̄1 + Re−1/2 p̄2 ...
with ū1 (x, 0) = ūe (x). The classical asymptotic sequence (as described by
Van Dyke (51)) is then: the ideal fluid at order O(1) drives the boundary
layer at order O(1). In turn the boundary layer disturbs the ideal fluid at
order O(Re−1/2 ), then this perturbation creates a corrective boundary layer
at this O(Re−1/2 ) order, and so on. There is a cascade of disturbances at
increasing order (see figure from Van Dyke 10 left for this classical sequence).
The analysis was presented for a flat plate, but we will present in section
6 examples of flows over a flat plate with a small bump. The bump is
defined by the function f¯(x̄). Before going further, we have to present the
”Prandtl transform” which consists to change the transverse velocity and
variable: ṽ → ṽ − f¯ (x̄)ũ and ỹ → ỹ − f¯(x̄) and keep x̄ and ũ. With this
transformation, the Prandtl system is invariant, and now the position on
the wall is 0 again. The blowing velocity is then corrected by f¯ (x̄)ūe . This
is a trick which allows to change the bumpy wall in a flat one.
2.3 Self Similar Solutions of Prandtl Equations

There is a simple class of solution of the boundary layer equations
which correspond to the flow along a wedge of half angle βπ/2 (see fig-
ure 6 left). The ideal fluid slip velocity is then a power law x̄, the suitably
non-dimensional longitudinal velocity is:
β 2n
ūe = x̄n with n = , β= .
2−β n+1
This is the so called Falkner Skan problem (41). There is a self similar
n
in the boundary layer, the longitudinal velocity is f (η)x̄ with
solution
ỹ
η = ( n+1
2 ) x̄(1−n)/2 . The selfsimilar stream function equation f (η) is from
Prandt Equation 1:
f + f f + β(1 − f 2 ) = 0, f (0) = f (0) = 0 and f (∞) = 1. (3)
The Blasius solution is the solution of the flow √ over a∞flat plate, it corre-
sponds to n = 0 and β = 0, and f (0) = 0.332 2 and 0 (1−f )dη = 1.732.
The Hiemenz stagnation point solution is the solution of the flow against
a ∞flat plate, it corresponds to n = 1 and β = 1 and f (0) = 0.92 and

0
(1 − f )dη = 0.8. In principle, for a given geometry (β), one solves the
ideal fluid and obtains n. Then the equation is solved, one finds the value of
f (0) which allows to suit all the boundary conditions. See figure 6 where
f (0) is plotted as a function of β (we will see section 4.1 that the f (0)
function of β is in fact ∞ multivalued in β < 0. One other result is the dis-
placement thickness 0 (1 − f )dη. In fact, this naive direct resolution gives
a very stiff problem, and in practice, for a given β and a negative f (0) it
is impossible to solve the system (eq. 3) in the case of separated flows.
2.4 Von Kármán Equation Integral Relation

Boundary layer equations are a 2D PDE which is not so simple to solve.
Nevertheless, the velocity profile is sometimes self similar (as just seen). It
means that there is a kind of unique profile and that all the profiles are
deduced by stretching it (Pohlhausen introduced simple poynomia). The
velocity of the ideal fluid at the wall and the thickness of the profile are two
fundamental parameters which stretch the fundamental profile.
In this part we present the Von Kármán-Pohlhausen (1921) equation
which consists in writing only the global dependance between ūe and the
displacement thickness δ1 supposing that in fact all the profiles are nearly
similar. An other equivalent interpretation of δ1 (linked closely to the ”blow-
ing veloicity” just defined) is that the flux of mass trough an enough large
y is the same than the flux of a constant velocity across a smaller section
reduced by the amount of δ1 so that:
∞
δ1 ue = (ue − u)dy.
0
This gives the physical definition of the displacement thickness, it is the

distance by which the external stream lines are shifted due to the boundary
layer development. To define the Von Kármán equation, we write the total
252 P.-Y. Lagrée
derivative in conservative form and in a ”defect formulation” (Le Balleur

(28)), so the momentum equation is:
∂ ∂ ūe ∂ ∂ 2 ũ
(ũūe − ũ2 ) + (ūe − ũ) − (ṽ(ũ − ūe )) = − 2 .
∂ x̄ ∂ x̄ ∂ ỹ ∂ ỹ
Using the displacement thickness (eq. 2), and defining the momentum thick-
ness δ̃2 and the shape factor H:
∞
ũ ũ δ̃1
δ̃2 = (1 − )dỹ and H = ,
0 ūe ūe δ̃2
and defining a function f2 linked to the skin friction as: ∂∂ũỹ = f2 Hδū1 e gives
the following equation where the ideal fluid promotes the boundary layer:
d δ̃1 δ̃1 2 dūe f2 H

( ) + (1 + ) = , i.e. δ̃1 = F (ūe ). (4)
dx̄ H ūe H dx̄ δ̃1 ūe
Initial condition is for example δ̃1 (0) = 0 (but the Hiemenz value may be
a good first guess) and ūe (0) = 1. In the classical approach, δ̃1 is obtained
through the knowledge of ūe , which we write formaly δ̃1 = F (ūe ).
To solve this boundary layer equation, a closure relationship linking H
and f2 to the velocity and the displacement thickness is needed. This is of
course a strong hypothesis. Defining Λ1 = δ̃12 dū e
dx̄ , the system is closed from
the resolution of the Falkner Skan system using the following fit (fig. 2):
,
2.5905e−0.37098Λ1 if Λ1 < 0.6 1 4
H= , f2 = 1.05(− + 2 ). (5)
2.074 if Λ1 > 0.6 H H
It means that we suppose that each profile remains a Falkner Skan one in
the boundary layer. We used this crude solution in exponential with the
value of the sink solution H = 2.074 as a limiting value (Lorthois & Lagrée
(32)). We tested it to be enough good, other closures may be found in the
literature. Some closures use the concept of entrainment. The closure may
be done with other families of profiles, and Pohlhausen profiles are good
candidates (the solution is part of a polynomia). With those profiles the
reverse flow is over estimated compared to Falkner Skan.
In general, the Von Kármán equation is written with the momentum

thickness δ̃2 :
d dūe ∂ ũ
(δ̃2 ū2e ) + δ̃1 ūe =
dx̄ dx̄ ∂ ỹ ỹ=0

I
+

+
Figure 2. An example of closure of the integral relations. The dots are

the Falkner Skan values and the line the proposed function. Anticipating
problems of boundary layer separation f2 may be negative.
(often the symbol θ̃ is taken), we prefer to write it with δ̃1 as we we have

seen that this value has a real physical interpretation: the displacement
thickness and the ”blowing velocity”. The reason why mainly δ̃2 is used is
that its derivative is clearly linked to the skin friction (this gives a technique
to deduce the skin friction from even crude measurements of the boundary
layer profile.
Up to now, we have all the classical ingredients of the Boundary Layer

obtained from the ideal fluid solution. Let us examine some of the problems
of this theory.
3 Problems Associated with the Boundary Layer

3.1 Separation
The first problem of the boundary layer is the the problem of separation.
The point of separation is the point defined by ∂∂ũỹ = 0. After this point,
the flow is reversed, see an example of representation on figure 3 taken from
Prandtl himself (37).
On figure 4 is presented an example of boundary layer computation with
an external flow ūe = sin(x̄) corresponding to the flow on a cylinder of unit
radius. An Integral resolution of the equations is compared with a com-
plete boundary layer resolution showing that the Von Kármán approach is
enough precise.
The flow is accelerated from x̄ = 0 to π/2, near x = 0 we have an

Hiemenz linear flow. The flow is decelerated for x̄ > π/2, this deceleration
promotes an increase of the boundary layer thickness and a decrease of the
skin friction. At the point where ∂∂ũỹ = 0, the boundary layer is singular, we
254 P.-Y. Lagrée
Figure 3. A sketch from Prandtl (37) of the flow near the point of vanishing
shear stress.
can not compute numerically (here by finite difference) the boundary layer.
Using the Von Kármán equation (Eq. 4 and closure 5) gives the same be-
havior. It fails nearly at the same point (not exactly, but not so bad).
A simple way to see the problem is to observe the Von Kármán equation
linearised around a decelerating flow. Linearizing the velocity near the point
of separation is say ūe = sin(x̄s )−a(x̄− x̄s ) with a = cos(x̄s ) and linearizing
around small Λ1 (which is not true but is a enough good approximation)
gives H = H0 − Hp Λ1 where H0 = 2.59 and Hp −0.96, so the variation
of δ1 /H with respect to x̄ is:
d δ̃1 1 d δ̃1 dH dΛ1 1 d 2Hp aδ̃12

( ) = ( δ̃1 )(1 − )( δ̃1 )(1 − ),
dx̄ H H dx̄ H dΛ1 dδ̃1 H0 dx̄ H0
d
therefore dx̄ δ̃1 is infinite, we can not march in x̄ anymore. This crude
estimation shows that the separation point is impossible to cross, but direct
numerical finite difference solution of the boundary layer equations gives
the same result (figure 4).
This difficult problem has been examined, among others by Landau (in
the classical text book (25)) and by Goldstein (18). Landau (25) noticed
that as in the boundary layer v u, so the transverse velocity must increase
a lot to be as large as the longitudinal one. It is apparently the case when the
flow is separated (stream lines are ejected from the wall). In boundary layer
variables Landau infer that v = ∞ and ∂v/∂y = ∞ so that ∂u/∂x = −∞.
The velocity is strongly decelerated near the point of separation xs . So
∂x
he proposes to work with the inverse of the function ( ∂u ) which is zero at
separation and proposes a reciprocal expansion of x in u near xs as (where
δ1 τ
x x
Figure 4. Boundary layer separation on a cylinder, the outer velocity is
ūe = sin(x̄), points are numerical finite difference solution of the Boundary
Layer equations, line is the integration of Von Kármán equation with the
proposed closure (Eq. 4 and 5). Separation occurs for an angle of 104o .
β is an ad hoc function linked to ∂x2 u):
∂x 1 ∂2x 1
x−xs = (u−us )+ (u−us )2 +... = 0(u−us )+ 2 (u−us )2 +...
∂u 2 ∂u2 4β (y)
so that one may write the velocity u and by the continuity equation v as:
√ β(y)
u = us (y) + 2β (y) xs − x + ... and v = √ + ...
xs − x
Injecting it in the momentum equation, neglecting the viscosity and writing
the total derivative as u2 (∂y ( uv )) shows that v/u does not depend on y, an
hint for the profile near separation may be deduced as :
∂us ∂A √
u = us (y) + A(x) v = − us with A(x) ∝ xs − x.
∂y ∂x
Unfortunately this description does not fit the good boundary conditions at
the wall....
Goldstein showed that for a given external flow, one can not compute the
boundary layer if the skin friction vanishes (which is consistent to Landau
approach), the skin friction behaves as:

τp = 48a4 (xs − x)),
one of the problem being an impossible match of the constant a4 before

and after the point of separation. He obtained the same square root sin-
gularity. This is called Goldstein singularity, close to the point of separation.
256 P.-Y. Lagrée
So, for a given external decreasing velocity, there is a possibility of sep-

aration with a singularity. The computation can not pass the separation.
Most of classical text book of fluid mechanics do the same and end their
course on boundary layers by this dead end, for example one can read in
Kundu (19): ”the boundary layer equations are valid only far downstream as
the point of separation. Beyond it the boundary layer becomes so thick that
the basic underlying assumptions become invalid. Moreover, the parabolic
character of the boundary layer equations requires that a numerical integra-
tion is possible only in the direction of advection (along which information
are propagated), which is upstream within the reversed flow region. A for-
ward (downstream) integration of the boundary layer equations therefore
breaks down after the separation point. Last, we can no longer apply poten-
tial theory to find the pressure distribution in the separation region, as the
effective boundary of the irrotational flow is no longer the solid surface but
some unknown shape encompassing part of the body plus the separation
region.”
3.2 The Problem of the Influence of Downstream on Upstream
One other strange problem appeared in the 50’ at the time of the super-
sonic conquest: the problem of ”Upstream Influence”. A model configu-
ration for a supersonic wing was the aligned flat plate in a compressible
supersonic flow. In various experiments in supersonic flows (Ackeret Chap-
man and others, see (45)), it was observed that an impinging shock wave on
a boundary layer produces perturbations far upstream the point of reflex-
ion of the wave. The boundary layer deviates from its basic state upstream
of the impinging shock, see photos on figure 5 from Stewartson 64 book
(45). On this figure we even see that three different accidents (an imping-
ing shock, a forward facing step and a wedge) produce the same upstream
flow. The deviation occurs far away (in boundary layer thickness units)
from the accident.
In the classical supersonic framework this is impossible (figure 5 lower
left). First the ideal fluid is supersonic (hyperbolic) so perturbations move
downstream in the Mach cone. Second, the boundary layer is parabolic,
so perturbations move downstream and across the boundary layer. No dis-
turbance can theoretically move upstream against the flow. This is the
upstream influence paradox.
Is it a dead end? No!

Figure 5. The ”upstream influence” in supersonic flows. Figures from Stewart-

son book (45): upper left, a sketch of the shock wave boundary layer interaction.
Right, the three different accidents (an impinging shock, a forward facing step and
a wedge) produce the same upstream flow. Lower left: upstream influence para-
dox: the ideal fluid is supersonic (hyperbolic, perturbations move downstream in
the Mach cone), the boundary layer is parabolic (perturbations move downstream
and across the boundary layer).
4 Solutions of the paradoxes

4.1 Inverse Boundary Layer
In fact computing the reverse flow within a boundary layer is possible
with the Prandtl equations. The good idea is: impose the displacement
thickness and solve for pressure gradient. This was the idea of Catherall
and Mangler (6) in 66, and they were the first to succeeded to pass the
point of flow separation while solving the steady boundary-layer equations
with a prescribed displacement thickness (a kind of parabolic shape). With
an imposed ∞
ũ
δ̃1 = (1 − )dỹ
0 ūe
it is possible to compute ūe and to obtain a negative skin friction.
Is it surprising? Not so much, we previously spoke about a simple class of

solution of the boundary layer equations (3) which corresponds to the flow
258 P.-Y. Lagrée
along a wedge of half angle βπ/2 (see figure 6 left), the so called Falkner
Skan problem (41). A naive direct resolution gives a very stiff problem, and
in practice, for a given β it is impossible to find the f (0) which solves the
system 3 in the direct way when there is separation. We can nevertheless
obtain flow separation for some values ofβ. To obtain it, we have to solve in
∞
an inverse way, we impose the thickness 0 (1−f )dη, and we find the value
of β associated. Hence, a simple way to feel that the boundary layer must be
solved in inverse way is really the Falkner Skan case. It is representative by
many aspects of the boundary layer behavior: for a given external velocity
one has a given β and one computes the corresponding profile. But, we see
on figure 6 that if the external velocity is with a β to much negative, there
is no solution. Only for an ad hoc external velocity we have solution(s).

η
f )d

∞
(1 −
π 0

β

2

Β
Figure 6. Left the Falkner Skan problem: self similar flow on a wedge. Right,
the numerical solution, value of f (0) as a function of β. Not every external
velocity is compatible with the boundary layer, for example in the Falkner Skan
case, too small β (less than -0.199) are not relevant (small dashing). A larger
value of the outer velocity gradient (large dashing) gives solutions.
On figure 7 we present an example of inverse boundary layer computation

using the Keller Box method (21). The displacement δ1 is given, the velocity
is deduced. For enough large values of the increase of the boundary layer
displacement thickness, we have separation with reverse flow. The outer
velocity decreases and reincreases.
4.2 Some Explanations of the Upstream Influence Problem

Now, we have some responses with the inverse solution for Boundary
Layer. But we have not finished. The upstream influence paradox is still to
explain. It puzzled people.
• Some people think that there is always a subsonic part in the boundary

Figure 7. Examples of inverse boundary layer computation. Separation is not an

2
issue when displacement is prescribed. Here given δ̃1 = 1.73x̄1/2 + αe−25(x̄−1.5) ,
we compute the associated external velocity, and the skin friction. For α = 1.43
(left) is the incipient separation, for smaller increase of δ̃1 , there is no separation
just a decrease of velocity. For larger values (right α = 2.8) we have separation
with reverse flow. The outer velocity decreases and reincreases.
layer, so that the retroaction can travel back in this subsonic layer. In
fact it is not the good mechanism as the upstream influence would be of
same length than this subsonic layer is thick. But on the experiments, the
longitudinal scale is far larger than the boundary layer thickness.
• Garvine (16) proposed a simplified boundary layer model linearising around
u = 1 the supersonic boundary layer (neglecting thermal effects):
ỹ
2
∂x̄ ũ = −∂x̄ p̄ + ∂ỹ ũ, ṽ = − ∂x̄ ũdỹ
0
and writing the Ackeret formula (linking the pressure and the blowing ve-
locity) as:
1
p̄ = √ √ ṽ(δ̃)
Re M 2 − 1
he obtains after claiming δ̃ = cst (yes he did!) that the pressure gradient is
δ̃
− √Re√1M 2 −1 0 ũx̄x̄ dỹ so that a model equation of the interaction is:
δ
1
∂x̄ ũ = √ √ ux̄x̄ dỹ + ∂ỹ2 ũ.
Re M 2 − 1 0
He pointed out the come back of ellipticity due to this ux̄x̄ term. He then
obtained a set of eigen solutions with Laplace transform, in fact the expo-
nentially growing one of those solutions can be obtained in looking to eK x̄
260 P.-Y. Lagrée
solutions, so that solution behaves as:

√ √
Re M 2 −1
x̄
e δ̃ .
So the coupling of the two equations produces self induced explosive solu-
tions.
• Numerically those explosive solutions were obtained by Werle Dwoyer,
and Hankey (54) (among others). On figure 8 we have a clear example of
what happens when solving in a marching way the coupled system. Starting
from a given initial location they solved the coupled boundary layer system
with the so called tangent wedge law (valid for stronger shocks than the
linearised Ackeret formula). They showed that changing a bit one parameter
may cause different solutions. Those are called ”branching solutions”.
Figure 8. Branching solutions (54): changing a bit one parameter may cause
different solutions while solving the equations with a marching scheme.
• One may consider the most simple argument, see Le Balleur (26). He
considers the strong coupling of the boundary layer in Von Kármán form
(neglecting again thermal effects) with the Ackeret formula (linking the
perturbation of pressure at M > 1 due to the variations of the effective wall
(represented by δ1 ) as:
d δ̃1 δ̃1 2 dūe f2 H 1 dδ̃1

( ) + (1 + ) = , p̄ = √ √ , (6)
dx̄ H ūe H dx̄ δ̃1 ūe 2
Re M − 1 dx̄
so that, supposing that ūe is nearly one and ∂x̄ ūe = −∂x̄ p̄
d δ̃1 δ̃1 2 1 d2 δ̃1 f2 H

( )= (1 + ) √ √ + ,
dx̄ H ūe H Re M 2 − 1 dx̄2 δ̃1 ūe
this equation is ”not so far” from from the basic flow with subscript 0 and
ūe ∼ 1, so after linearisation.
d δ̃10 1 d2 δ̃1
δ̃1 = (H0 + 2) √ √ + ....
dx̄ ūe Re M 2 − 1 dx̄2
where we forget the contribution of the skin friction. So again, we obtain
exponential solutions (called supercritical by Crocco and Lees in 52) for the
disturbance of the displacement thickness δ1 :
√ √
Re M 2 −1
δ̃10 (H0 +2)
x̄
e .
It is nearly the same result than Garvine and than the one obtained numer-
ically by Werle et al.
• In fact, Lighthill in 53 (30) proposed a pre-theory of triple deck explaining
most of the mechanism (see in Stewartson 64 book as well) based on steady
perturbations of the Orr Sommerfeld equation.
• The real definitive theory is the Triple Deck (see Ruban’s contribution).
In this framework, those kind of explosive solutions are called ”self induced
solution” (see Neiland (35), Messiter (34) and Stewartson (46)).

Figure 9. Left, the Triple Deck scales. Right, ”triple decker ship of the line”
from HMS victory brochure Porthmouth (”vaisseau de ligne à trois ponts”). In
german ”Dreierdeck-Theorie”, a french translation of Triple Deck Theory may be
”Triple Pont” instead of ”Triple Couche”.
To present is quickly, we have the basic non dimensional Blasius profile

UB√(ỹ) in the boundary layer, where ỹ is the transverse variable scaled by
1/ Re (referred as ε4 and anticipating that ε = Re−1/8 ). Now suppose
that at longitudinal scale say x3 (referred as ε3 ) there is a perturbation of
this basic profile. We will call ”Main Deck” the region considered which is
of relative scale x3 but which is of boundary layer scale in the transverse
direction. As this scale is small, the boundary layer has not evolved, and at
262 P.-Y. Lagrée
first order there is no transverse velocity. At the longitudinal scale x3 there

is a perturbation of this basic profile of magnitude ε, then retaining all the
terms in the incompressibility and in the total derivative equation and as
pressure is O(ε2 ) gives the solution in the Main Deck (former Boundary
Layer)
∂ p̃
ũ = UB (ỹ) + εA(x)UB (ỹ) ṽ = −ε2 A (x)UB (ỹ), and = 0.
∂ ỹ
Note that this function −A is reminiscent to the Landau analysis of section
3. With this description, the velocity is not zero but εA(x)UB (0) on the
wall, so we have to introduce a new layer to full fit the no slip condition.
We reobtain Prandtl equations but with a new transverse scale δ3 = ε5
associated to the longitudinal one x3 = ε3 , in which the longitudinal velocity
is scalled by ε. In this new layer, the Lower Deck, the final system is then:
∂u ∂v ∂u ∂u dp ∂2u
+ = 0, u +v =− + 2. (7)
∂x ∂y ∂x ∂y dx ∂y
With no slip condition at the wall (u = v = 0), the entrance velocity profile
u(x → −∞, y) = UB (0)y, and the matching condition with the Main Deck:
u(x, y → ∞) = (y + A)UB (0). Note, that the system is parabolic, there is
no output condition needed to solve it.
Going back in the Main Deck, the disturbed velocity at the top of the
Main Deck, for ỹ → ∞:
ũ = 1; ṽ = −ε2 A (x),
there is no more longitudinal perturbation of the velocity at order ε, but

there is a transverse velocity, a kind of ”blowing velocity” at the edge of the
Main Deck. Note that the pressure remains of the same order ε2 .
The final layer is the Upper Deck of longitudinal size x3 = ε3 and of
same thickness in which we have a blowing velocity at the wall of order
ε2 . The velocity at the top of the Main Deck is then the velocity at the
bottom of the upper deck: −A . Depending on the ideal fluid régime, one
may compute the pressure (of order ε2 ). For a incompressible flow one has
the Hilbert relation: dA
1 dx
p= dξ.
π x−ξ
For a compressible supersonic flow, one has to use the Ackeret formula:
1 dA
p = −√ .
2
M − 1 dx
So one has to couple this relation linking a pressure and a displacement to

the Lower Deck problem 7.
One may linearize the Triple Deck around the Basic Flow u = UB (0)y,
say that u1 and v1 are the perturbations, the lower Deck :
∂u1 ∂v1 ∂u1 dp1 ∂ 2 u1

+ = 0, UB (0)y + v1 UB (0) = − + .
∂x ∂y ∂x dx ∂y 2
With at the wall (u1 = v1 = 0), at the entrance u1 (x → −∞, y) = 0, and

at the infinity u1 (x, y → ∞) = UB (0)A1 . We test eKx solutions on the
linearized system, with K > 0.
u1 = eKx φ (y), v1 = −eKx φ(y), p1 = eKx P,
with φ(0) = φ (0) = 0 and say φ (∞) = UB (0) so that A1 = eKx ; as the
incompressibility is fulfilled, the momentum gives (see Stewartson (47) or
Sychev et al. (49) for details):
∂ 2 φ (y)
= UB (0)Kyφ (y), with φ (0) = KP,
∂y 2
and p1 = − √KA 1
M 2 −1
so that we deduce that the supersonic case allows then
an eigen solution

K = (−3Ai (0)( M 2 − 1)UB (0)2 )3/4 .
This exponential behavior at the longitudinal Triple Deck scale is the ra-
tional explanation of the observed self induced separation. This upstream
influence is then understood as a not well posed problem. In fact, even if
each part of the flow seems hyperbolic/ parabolic, due to the interaction
one recovers the output influence.
This is the case in the supersonic flows, in shallow water flows at small
Froude number, in mixed convection. But there exist flows with no up-
stream influence: for example in the symmetrical pipe flows.
As a conclusion of this section, we see that we need two two ingredients.

First we need to solve the Boundary Layer in an inverse way. And second,
the Boundary Layer is no more driven by the Ideal Fluid. The Boundary
Layer can retroact on the Ideal Fluid. The retroaction explains the observed
self induced interaction. So we now introduce the concept of Interactive
Boundary Layer (IBL) which uses those ingredients.
264 P.-Y. Lagrée
5 Interactive Boundary Layer

5.1 Examples of Users
So it became clear that the interaction with the ideal fluid is not weak but
strong. In the early 60 Gad and Curle employed Von Kármán -Pohlhausen
method to try to solve the shock waves-boundary layer interaction, ”without
much success” (as quoted by Lees and Reeves (29). Lees and Reeves in 64
(29) did computations with integral methods, with more success, but the
details are not so clear. Reyhner Flügge Lotz 68 (38) did finite differences
on the Boundary layer and succeed by iteration to compute the supersonic
wedge interaction.
Among people working for applications in the aerospace area, some
names and teams are to be associated to Interactive Boundary Layer IBL/
Inviscid Viscous Interaction IVI. Among them:
• Le Balleur, from 1977 ((26), (27)) understood the interaction and using
Von Kármán profiles did a lot of practical computations at ONERA, in
supersonic and transsonic régimes.
• Veldman ((53), (52)) as well has is own codes at the National Aerospace
Laboratory NLR in Amsterdam,
• Carter (4), (14), Jameson (20) at Stanford.
• Cebeci did a huge work (several books on the interactive boundary layer
for example (7) (3)) and applied IBL at Boeing.
• Drela (13) and (12) developed a integral boundary layer code which is
now free: XFOIL.
• Lock & Williams in a review (31), present the english RAE point of view.
• And last but not least Neiland and Sychev (49) at the TsAGI in USSR.
Of course, this is again a very very partial list.
5.2 Interactive Boundary Layer

So the way to bypass Goldstein singularity is to adopt the Interactive
Boundary Layer point of view. It means that we use the classical Prantdl
boundary layer equations :
∂ ũ ∂ṽ ∂ ũ ∂ ũ dūe ∂ 2 ũ
+ = 0, ũ + ṽ = ūe + 2,
∂ x̄ ∂ ỹ ∂ x̄ ∂ ỹ dx̄ ∂ ỹ
with no slip boundary conditions (ũ = ṽ = 0 on the body f¯(x̄)), a first

given velocity profile: Blasius. The matching ũ(x̄, ỹ → ∞) → ūe (x̄).
A result of this computation is the transverse velocity at infinity,we ob-
tained the ”blowing velocity”.
d(δ̃1 ūe )
v̄e = Re−1/2 .
dx̄
Hence, the outer flow is no more only given by the wall f¯(x̄) (so that the
blowing velocity is f¯ ūe ) but, the wall is ”thickened” by the boundary layer
thickness (or ”blowing velocity”, or ”transpiration boundary condition”),
so that for a subsonic flow:
¯
1 f (x̄)ūe + Re−1/2 d(δ̃dx̄
1 ūe )
ūe = 1 + dξ
π x−ξ
or in a supersonic flow
1 d d(δ̃1 ūe )
ūe = 1 − √ [ f¯(x̄)ūe + Re−1/2 ].
M 2 − 1 dx̄ dx̄
Instead of the usual weak coupling with the hierarchy (figure 10 left), the
boundary layer retroacts on the ideal fluid (figure 10 right). So even if ūe
appears in the definition of himself through d(δ̃dx
1 ūe )
, it is not an issue because
of the iterations involved in the solution. The boundary layer thickness δ̃1
acts as a fictive wall, it disturbs the ideal fluid, the pressure (pressure and
velocity ūe (x̄) are linked) develops the boundary layer itself. It is a strong
interaction. The two layers are coupled. It explains the term ”Interactive
Boundary Layer”, or ”Viscous Inviscid Interaction”.
Most of the separation problems are then solved...
5.3 Link between Interactive Boundary Layer and Triple Deck

At separation, the displacement boundary layer thickness becomes very
thick. It is then not counterintuitive to think that the ideal fluid will be
drastically changed by the viscous layer. That is the picture for ”Triple
Deck”, but the scales are changed. In fact it is easy to show the that the
IBL equations give at large Reynolds number the Triple Deck structure.
The −A is the disturbance of the displacement thickness.
The IBL formulation emphasizes on the displacement thickness,
∞
−1/2
δ1 = (Re ) (1 − u(x, ỹ))dỹ
0
we have to decompose it into two parts as we cross the Lower and the
Main Decks. Let us introduce Ỹ so that we cut the integral in two parts
266 P.-Y. Lagrée

Figure 10. Left the Classical sequence, image taken from Van Dyke’s book (51).
Right the Interactive Boundary Layer, we do not follow the classical asymptotic
sequence (from left): the ideal fluid at order O(1) drives the boundary layer
at order O(1), in turn the boundary layer disturbs at the ideal fluid at order
O(Re−1/2 ), then this perturbation creates a corrective boundary layer at this
O(Re−1/2 ) order, etc. But, right, we couple the boundary layer and the ideal
fluid at first order.
Ỹ ∞
( 0 + Ỹ ). The first integral is estimated near the wall, so the Lower Deck
description (ỹ = εy) is valid there, but a good idea is to write the velocity
u(x, y) = UB (0)(y + A) + uc where uc is a correction, so the first integral:
Ỹ Ỹ /ε Ỹ /ε
( (1 − ũ(x̄, ỹ))dỹ) = ε( (1 − ε(UB (0)(y + A)))dy − εuc dy)
0 0 0
the second integral is in the Main Deck

∞ ∞
(1 − u(x, ỹ))dỹ = (1 − UB (ỹ) − εA(x)UB (ỹ))dỹ.
Ỹ Ỹ
Re summing the two integrals and changing the order of the terms allows
to recognise :
∞ ∞ Ỹ /ε
−1/2
δ1 = (Re ){ (1−UB (ỹ))dỹ+ (−εA(x)UB (ỹ))dỹ−ε2 uc dy)}.
0 0 0
or ∞
δ1 = (Re−1/2 ){ (1 − UB (ỹ))dỹ − εA(x) − O(ε2 )}.
0
the −εA contribution
∞ of the Triple Deck is the perturbation of the displace-
ment thickness 0 (1 − UB (ỹ))dỹ. The IBL equations (based on δ1 ) even if
they seem to be ill posed as they mix different order of magnitude may be
justified by the Triple Deck analysis (based on −A).
Recently, starting from NS equations, Dechaume Mauss and Cousteix

(10) and Cousteix & Mauss (9) showed that we may obtain the IBL system
using an other technique than ”Matched Asymptotic Expansion”. They
rather used the so called ”Successive Complementary Expansions Method”
(MASC in french).
5.4 Reduced Navier Stokes Prandtl, RNSP equations

In this subsection we focus on internal flows (in axi symmetrical pipes or
between two plates in 2D). Interestingly enough, we may write the Prandt
equations across the pipe itself and obtain a system that we call ”Reduced
Navier Stokes Prandtl” system (RNSP). This system, just as the Navier
Stokes system, degenerates in IBL and in various Double Deck or Triple
Deck descriptions. For example, in a pipe we scale the transverse derivative
by the radius (R0 ) of the pipe itself. In this case the non-dimensional
variables are given by :
U0
x = x̃R0 Re, r = r̃R0 , u = U0 ũ, v = Re ṽ, p = p0 +ρ0 U02 p̃, and Re = U0 R0 /ν.
p0 denoting the entry pressure. With these new variables, the following
partial differential system is obtained from the Navier-Stokes equations as
Re → ∞, the RNSP system
∂ ∂ r̃ṽ ∂ ũ ∂ ũ ∂ p̃ ∂ ∂ ũ ∂ p̃
ũ + = 0, ũ + ṽ =− + (r̃ ), 0=− . (8)
∂ x̃ r̃∂ r̃ ∂ x̃ ∂r̃ ∂ x̃ r̃∂ r̃ ∂ r̃ ∂ r̃
The associated boundary conditions are:

• the condition of axial symmetry : ∂r̃ ũ = 0 and ṽ = 0 at r̃ = 0,
• no-slip condition at the wall : ũ = ṽ = 0 at r̃ = 1 − f (x̃). Of course,
in order to be consistent f˜ is of order one, but smaller than one, and
the longitudinal scale has to be compatible.
• the entry velocity profiles (ũ(0, r̃) and ṽ(0, r̃)) are given : flat profile
or Poiseuille flow, but other profile is also possible,
• there is no outflow boundary condition because the system is parabolic.
The equations are solved by marching in the stream wise direction,
even if there is flow separation.
Rubin & Himansu (39) and Tannehil et al. (50) kept a transversal pres-
sure variation linked with the transverse velocity with O(Re−2 ) terms, but
as noted by Fletcher (15), this system contains a mix of orders of magni-
tude, and is not coherent from an asymptotical point of view. In Lagrée &
Lorthois (24) or (33), this system (8) is used to obtain most of the degen-
eracies of the full NS equations in an axisymmetrical pipe :
268 P.-Y. Lagrée
• this system allows to compute the entry problem in an unconstricted pipe.

This set of equations has been already used for studying entry effects by
Cebeci & Cousteix (7) and in Schlichting (42).
• this system, in a case of a constriction (of relative shape εf¯) situated near
the pipe entry, where the velocity profile is flat gives again the Interacting
Boundary Layer equations. This RNSP system may be itself splited in
an invisicid core and a boundary layer. This inviscid core interacts with
boundary layers near the wall. In fact, after rescaling: r = 1 − εȳ, u = ū,
v = −ε−1 v̄, x = ε2 x̄ and p = p̄ and assuming a flat entry velocity profile,
the RNSP leads to the final IBL formulation as follows:
∂ ū ∂v̄ ∂ ū ∂ ū dūe ∂ 2 ū
+ = 0, (ū + v̄ ) = ūe + 2, (9)
∂ x̄ ∂ ȳ ∂ x̄ ∂ ȳ dx̄ ∂ ȳ
1
ūe = (10)
(1 − 2ε(δ̄1 + f¯))
∞
where δ̄1 = 0 (1 − ūūe )dȳ, where f¯ represents the shape of a constriction
of the pipe and with the following boundary conditions:
ū(x̄, 0) = 0, v̄(x̄, 0) = 0 and ū(x̄, ∞) = ūe .
This case leads as well to a special Triple Deck case with p = A identified
by Ruban & Timoshin (40).
• this system is used as well in a case of a constriction situated far from
the pipe entry, where the flow is fully developed. In this region, the Double
Deck theory, also known as Smith’s theory of viscous perturbation on a
Poiseuille flow in a pipe, is valid ((43)). In this case A = 0, and it is shown
in (24) that the system (8) contains the Double Deck.
• Finally, it is shown in (24) that if the constriction is short compared to
R0 Re, the velocity profile at the entry is not important. In that case, the
Interacting Boundary Layer theory proves to be valid again: acceleration
is so high that the profile flattens, recreating an inviscid core and a thin
boundary layer near the wall.
The same system may be written in 2D, see (23) and (8) and see Bar-
renechea & Chouly (2) for a finite difference resolution.
5.5 Coupling the Solvers

Boxes We now have to solve numerically the problem which consists in
a coupling between a Boundary Layer and the Ideal Fluid response. As
there are two problems coupled, it is natural to define kind of ”boxes”. A
first ”box” is the Euler Solver. Given a wall, it computes the pressure and
the slip velocity. This box may be a subsonic, supersonic... a linear or not
solver. It does not matter, the input is the wall geometry, the output is the
slip velocity.
The second ”box” is of course the boundary layer box, given an outer
velocity, it computes the displacement thickness. The equations may be
laminar or turbulent with any turbulent model. It may be full finite differ-
ences resolution or Von Kármán integral method. This box may be used in
standard direct way: for a given slip velocities, it computes a displacement
thickness. This box may be used in reverse, given a displacement thickness
it computes what outer velocity produces it.
Coupling Now, we couple the boxes and present the various possibilities.
In fact we will use δ1 and u in the following figures (we forget all the tildes
in this section). We may use dδ 1
dx instead of δ1 , and instead of u we may use
dp
−p (by Bernoulli linearised) or we may use dx . There is no real influence of
the choice of δ1 instead of his slope, nor in u, p or his gradient (as we deal
with small perturbations).
• Now, having those boxes, we have to branch them. First, the classical
boundary layer theory may be represented as a ideal fluid box followed by
a boundary layer box, figure 11.
yw

ue δ1
Figure 11. Classical Boundary layer, the geometry gives the velocity which gives
the boundary layer.
• But as mentioned previously, branching the output of the boundary layer

to the input of the ideal fluid will give the second order effects but will not
allow the separation, figure 12.
yw + δ1

ue δ1
Figure 12. ”Direct method”: the geometry gives the velocity which gives the
boundary layer, the rebranching will give the second order effects.
• The good way to solve the boundary layer, is to solve it in inverse, we can
imagine that we solve the ideal fluid in inverse as well. This is the ”inverse
270 P.-Y. Lagrée
method” figure 13. in fact it is not a good idea as it is difficult to rewrite

the Euler codes.
y w + δ1 ue

δ1
Figure 13. ”Inverse method”, the total geometry (boundary layer thickness and
effective geometry) give the velocity which gives a total geometry, and so on.
• The good way to solve the boundary layer, is to solve it in inverse, the
good way to solve the ideal fluid is in the direct way. So we have to relax the
input depending on the difference of the outputs. This is the semi-inverse
coupling by Le Balleur (figure 14).
δ1n unBL

n+1
(unBL − une ) → δ1
yw + δ1n

une
Figure 14. ”Semi Inverse method”, inverse boundary layer, direct ideal fluid.
The difference of the two output velocities is used to update the displacement
thickness, and so on.
• There are other possibilities, one is the ”quasisimultaneous method” (52).

It means that during the coupling values computed downstream are rein-
jected, which is more useful in the subsonic case.
Semi inverse coupling The point to be clarified is how to update the

new δ1n+1 from δ1n and the difference (unBL − une ), the simplest way is to
write:
δ n+1 = δ n + λ(unBL − une )
One has to notice that by the Bernoulli relation variation of velocity are
opposite of variation of pressure so that we can write as well:
δ n+1 = δ n − λ(pnBL − pne ).
The choice of λ is such as we obtain stability for the iterative method.


δ
8H
τ

[
Figure 15. Incompressible flow. Top the velocity field ũ, ṽ (Prandtl trans-
form), bottom the wall, here a bump, the displacement thickness δ̃1 (start-
ing from Blasius value 1.7 in x̄ = 1), the skin friction (starting from Blasius
value 0.3 in x̄ = 1) and the outer velocity starting from Ideal Fluid value
1 in x̄ = 1. A positive disturbance of the wall increases the velocity and
decreases the displacement. Separation occurs after the bump.
Le Balleur (see (27) and Wigton and Holt (55)) analysis consist to lin-
earize the equation. He defines two operators, one for each box, first B ∗
defined as δ n = B ∗ pnBL and for the ideal fluid, he defines in the same vein
a linear response δ n = Bpne . Then the update is as:
δ n+1 = δ n − λ(1/B ∗ − 1/B)δ n
To make it clear, we use Fourier analysis for all the frequencies between
π/L and π/Δx (the smallest linked to the domain size, and the highest
linked to the discretisation step). Furthermore, the B operator may be
obtained in subsonic flow we have B = −1/k. The analysis is then very
simple, defining a ”gain” G = δ n+1 /δ n :
1
G = 1 − λ( + k),
B∗
272 P.-Y. Lagrée

Δδ
SH
τ

[
Figure 16. Supersonic flow on a flat plate with a bump. Top, the velocity
field ũ, ṽ (Prandtl transform), bottom the wall, here a bump, the perturba-
tion of displacement thickness from Blasius Δδ̃1 (starting from 0 in x̄ = 1),
the skin friction (starting from Blasius value 0.3 in x̄ = 1) and the outer
pressure starting from Ideal Fluid value 0 in x̄ = 1. Note the pressure
plateau (here tiny) associated to separation.
we want |G| < 1 for π/L < k < π/Δx. Often ((27), (55)), it was considered
that B ∗ was real (which is not true), so we can find an optimal λ.
For a supersonic flow we have B = (ik/(M 2 − 1))−1 . It is easy to show
that in this case it is impossible to find an optimal λ. The coupling is always
unstable. The good coupling is in fact with the derivative of the pressure:
d n d n
δ n+1 = δ n − μ( p − p )
dx BL dx e
then again we have:
δ n+1 = δ n − μik(1/B ∗ − 1/B)δ n
which allows to define a ”gain” G = δ n+1 /δ n . We want |G| < 1 for all the
space frequencies π/L < k < π/Δx. We can find an optimal μ.

δ
8H
τ

[
Figure 17. Subcritical flow on a flat plate. Top, the velocity field ũ, ṽ
(Prandtl transform), bottom the wall, here a bump, the displacement thick-
ness δ̃1 (starting from Blasius value 1.7 in x̄ = 1), the skin friction (starting
from Blasius value 0.3 in x̄ = 1) and the outer velocity starting from Ideal
Fluid value 1 in x̄ = 1. A positive disturbance of the wall increases the
velocity and decreases the displacement. Separation takes place after the
bump. There is no upstream influence.
In the following examples, we use this semi-inverse coupling. In the

examples taken from literature (paragraph 6.3) there is a mix between com-
putations in direct way when the flow is attached, and in indirect way when
the flow is near separation.
6 Examples
6.1 Some Numerical Examples of IBL
Bump on a Flat Plate in an Incompressible (Subsonic) Flow. As

a first example (fig 15), we present the results for the IBL on a flat plate
2
with a bump defined by f¯(x̄) = αe−25(x̄−2) ; with α increasing by steps of
274 P.-Y. Lagrée

δ
8H
τ

[
Figure 18. Supercritical flow on a flat plate. Top, the velocity field ũ, ṽ
(Prandtl transform), bottom the wall, here a bump, the displacement thick-
ness δ̃1 (starting from Blasius value 1.7 in x̄ = 1), the skin friction (starting
from Blasius value 0.3 in x̄ = 1) and the outer velocity starting from Ideal
Fluid value 1 in x̄ = 1. A positive disturbance of the wall decreases the
velocity and decreases the displacement. Separation occurs far before the
bump, note the long upstream influence and the large increase of δ̃1 .
0.01 and Re = 10000. The velocity is:

¯
1 f (x̄) + Re−1/2 d(δ̃dx̄
1 ūe )
ūe = 1 + dξ
π x−ξ
Before the bump there is a small decrease of the ūe velocity. In a pure
Hilbert case, the response in ūe is perfectly symmetrical, but here, due to the
boundary layer, the velocity is no more symmetrical. Due to the acceleration
on the bump, the displacement thickness first decreases and increases again
after the bump. It increases more. So, there is a small overshoot of the
thickness associated with the boundary layer separation. This makes the
outer velocity non symmetrical. The skin friction increases before the crest,
and decreases after. This is consistent with the fact that, for instance, before
the crest, the velocity increases, and the boundary layer thickness decreases,

Δδ
SH
τ

[
Figure 19. Supersonic flow on a flat plate with a wedge. Top, the velocity
field ũ, ṽ (Prandtl transform), bottom the wall, here a wedge in x̄ = 3.5, the
perturbation of displacement thickness Δδ̃1 (starting from 0 in x̄ = 1), the
skin friction (starting from Blasius value 0.3 in x̄ = 1) and the outer pressure
starting from Ideal Fluid value 0 in x̄ = 1. Note the plateau pressure: the
pressure is nearly constant in the separated bulb before the wedge, and note
the separation occurs far upstream of the wedge.
so the slope of the velocity in the boundary layer increases (it is more or
less the ratio of ūe and δ̃1 ), the reverse happens after. We notice that the
maximum of the skin friction is before the crest, after the inflexion point of
the bump, the velocity increases less, but the boundary layer continues to
decrease because of the inertia of the fluid, so the maximum of skin friction
is between the inflexion point of the bump and the crest. There is eventually
a separated bulb with negative skin friction.
Bump on a Flat Plate in a Supersonic Flow. As a second example

(fig 16), we present the results for the IBL on a flat plate with a bump
2
defined by f¯(x̄) = αe−25(x̄−3.5) ; but in the compressible supersonic case,
276 P.-Y. Lagrée
so that the edge velocity is:
1 d d(δ̃1 ūe )
ūe = 1 − √ [ f¯(x̄) + Re−1/2 ].
2
M −1 dx̄ dx̄
The bump creates upstream influence and a separated bulb far upstream.
The skin friction reincreases and then redecreases to create a second sepa-
rated bulb.
Bump on a Flat Plate in Subcritical Flow. Nearly the same occurs

in the case of the subcritical flow (F < 1) or in the case of symmetrical pipe
flows. The edge velocity is:
1
ūe = 1 + [f¯(x̄) + δ̃1 Re−1/2 ]
1−F
It means that the velocity increases and decreases after the crest (see figure
17). The skin friction is extremal just before the crest, and there may be
flow separation on the lee side. The behaviour is nearly the same than in
the incompressible case but there is no influence of the bump before the
beginning of it. In the incompressible case there was some small effect due
to the Hilbert integral, but here there is no effect before the bump.
Bump on a Flat Plate in a Supercritical Flow. In the supercriti-

cal flow, the equation is the same for the edge velocity, but the story is
completely different as F > 1. We observe a strong upstream influence on
figure 18. The velocity decreases due to the bump, and the skin friction is
negative upstream of the bump, the extremum is on the lee side, after the
bump. There is a huge jump in δ̃1 , a kind of hydraulic jump.
Wedge on a Flat Plate in a Supersonic Flow. As final example (fig

19), we present the results for the IBL on a flat plate with a wedge defined by
f¯(x̄) = α(x̄−3.5)+ ; with α increasing by steps of 0.01 and Re = 100000. For
enough large α we observe the ”plateau” of pressure which is the signature
of the self induced interaction and upstream influence. This increase of
pressure before the wedge creates a region of reverse flow.
6.2 Example in Internal Flows: Axi- symmetrical Flows, sym-

metrical and non symmetrical 2D Flows
The system RNSP 8 which is like a Boundary Layer in a whole pipe
degenerates in IBL equations like Navier Stokes. Figure 20 from (24), shows

∂ ũ/∂ ỹ

x̃

Figure 20. Longitudinal evolution of the wall shear stress near the incipient
separation case, in an axis symmetrical flow with a stenosis (a constriction). Re-
duced Navier Stokes Prandtl model, integral IBL, full IBL resolution (in RNSP
variables, the bump is located in x = 0.02, and its width is 0.00125), and Triple
Deck resolution. All the curves are rescaled in Triple Deck scales.
that the result from system (8) compares to IBL with integral resolution
and with full equations and triple Deck (case p = A).
The set of RNSP equations in 2D may be solved and compared with
Interactive Boundary Layer equations. This has been done for example in
(23). On figure 21 is the 2D symmetrical flow between two plates with a
constriction. We compare Navier Stokes solved with Castem2000 (5), IBL
(Integral resolution) and RNSP (finite differences). Pressures are nearly the
same.
If now, we relax the hypothesis of symmetry (see (22)), we compare the

solution of the integral IBL and Navier Stokes between two plates, the upper
one being flat, and the lower one with a bump (figure 21). In this case two
boundary layers, the upper and the lower interact (using twice eq. 4) we
just write the difference of pressures between the top and the bottom ph −pb
which is proportional to:
((δ h )2 − (f + δ b )2 ) δ h − f − δ b
1 b 1
+ 1 b 1
.
1 − (fb + δ1b ) − (δ1h ) 2
We note that this coupling relation produces upstream influence, it means

that before the bump, the flow ”knows” what is coming. This creates so-
lutions in ekx with k > 0 (see figure 22). The pressure on the flat plate
278 P.-Y. Lagrée

p̃(x̃)
P̃m P̃t

P̃g

(x − xg )/Rg
Figure 21. 2D symmetrical flow between two plates with a constriction.

A comparison between computed non-dimensional pressure for the three
models (NS with Castem 2000 (5), IBL integral and RNSP finite differences,
in this last case the wall has been smoothed in x = ±1 to avoid an infinite
slope, this was not the case for NS and IBL), here α = 0.75, Re = 500. The
upper half geometry is plotted as well (the smoothed one is not plotted).
It is observed that the ratio: pressure at the glottis divided by maximum
pressure drop is nearly constant (Ke = P̃t /P̃m 0.82). Likewise, the ratio:
pressure at the output divided by the pressure drop betwen the output and
the glottis is nearly constant (Kt = P̃t /(P̃t − P̃g ) 0.86).

(h0/U0)∂(u∗)/∂y∗

p∗/(ρU02 )

x x
Figure 22. Left Right Comparison of integral IBL and NS (with Castem 2000)
pressures. The IBL approach well predicts the over pressure on the flat wall and
the positions of the minima of of the pressures after the throat. Skin friction,
comparison of integral IBL and NS. The integral IBL over predicts the maximum
of skin friction but well predicts the position of the point of separation. The
incipient separation before the bump is well predicted.

Τ

Figure 23. A bump in a Poiseuille flow at the lower will disturb the core flow,
the pressure changes across the core flow, perturbations are induced at the upper
wall. Linear perturbation of skin friction τ1 (left) and pressure p1 (right) over a
bump f1 (x) = cos(πx/2)2 , for −1 < x < 1 in the Triple Deck framework. The
A = 0 is in plain line and the A case is in dashed line.
(on the top) decreases before the bump, and pressure on the bumpy wall
increases before the bump).
Note that we recover a result that looks like Smith (43) (or Sobey (44))
result in pipe flow, the transversal perturbation of pressure in a perturbed
Poiseuille flow is ph − pb = A /30 where −A is a displacement of the stream
lines as δ1b − δ1h is. Of course the two configurations are very different. On
figure 23 we plot the perturbation of a Poiseuille skin friction in the linear
case for the A = 0 symmetrical case and the non symmetrical A case (see
Smith (43)). We see the that the case A = 0 presents no upstream influence
as already mentioned, but we clearly see that the case with A promotes
upstream response of the flow (before the first position of the bump, the
pressure has increased and the skin friction has decreased).
6.3 Some other Numerical Examples

We just reproduce here some examples from literature using this IBL
theory. We select among others comparisons of experiments, IBL and Ideal
Fluid over an airfoil. On the curves, the experimental and the computation
are displayed showing a very good concordance. We present Drela & Giles
(12) on figure 24, comparisons from Le Balleur computations of figure 25,
and comparisons from Lock & Williams (31) on figure 26. On figure 27,
Aftosmis et al. (1) successfully compare IBL strategy with a Navier Stokes
solver.
280 P.-Y. Lagrée
Figure 24. Example of comparison of IBL computation, Drela & Giles (12)
Figure 25. Example of comparison of IBL computation, Le Balleur.

Figure 26. Example of comparison of IBL computation, Lock & Williams (31)
2#0

Figure 27. Example of comparison of IBL computation, pressures from (1)

(coupled-IBL approach) compared with results from the pure inviscid solver and
published data using the NSU3D RANS solver (see (1) for details).
282 P.-Y. Lagrée
7 Conclusion
Figure 28. Complexity and models from Jameson (20)
So we now know that the Boundary Layer equations are more than
useful. They can handle flow separation and compute reverse flow bubbles.
The strong coupling between the Ideal Fluid and the Boundary Layer allows
this. The explanation of this lies in the Triple Deck theory which couples a
thin wall layer (the Lower Deck) with the Ideal Fluid (Upper Deck) trough
a displacement of the stream lines in the Boundary Layer (the Main Deck).
Further more, the IBL allows as well some upstream influence. It means
that in some special régimes such as supersonic, hypersonic, supercritical
flows, in non-symmetrical 2D pipe flows, disturbances in the Boundary
Layer influence the flow pattern far upstream the position of the distur-
bance. In axisymmetrical pipe flows, in symmetrical 2D pipe flows, in sub-
critical flows, there is no upstream influence at all. In the subsonic case,
there is only a small influence due to the elliptic character of the flow (seen
with the Hilbert integral which is global).
The methodology of IBL, or Inviscid Viscous Interactions, may be sum-
marized in the figure 28 extracted from Jameson (20). Even if this paper
was written in 1983, it seems that most of the flying aircraft have been
defined since by Viscous Inviscid interactions. The Airbus A380 is one of
the first aircraft designed with ”full Navier Stokes” (in fact certainly crude
RANS models). In the late 90’, before the end of the century, a large effort
has been done on Navier Stokes solvers. Lot of people are working on this
equation. Tremendous progress have been done, and with Navier Stokes,
the complexity of the geometry is a problem with lot of solutions. So Navier
Stokes solvers are very promising, and give a lot of practical results.
To a certain extent, IBL-IVI methods are less versatile and require spe-
cific methods, they need a kind of savoir faire (Aftosmis et al. (1) point
some difficulties of the IBL). But when used, they are very good. For ex-
ample Le Balleur has codes which may compute even large stall on wings,
giving results very close to experiments. NS solver have difficulties to re-
produce those results. XFOIL free code from Drela allows now everybody
to do quick computations of flows over airfoils.
The review of Piomelli & Balaras (36), shows that up to now only very
simple models are used for boundary layer near the wall. They suggest a
coupling of a Large Eddy Simulation Navier Stokes with a boundary layer
code near the wall.
Bibliography
[1] Michael J. Aftosmis Marsha J. Berger, Juan J. Alonso, (2006) ”Applica-
tions of a Cartesian Mesh Boundary-Layer Approach for Complex Con-
figurations”, 44th AIAA Aerospace Sciences Meeting, Reno NV, January
9-12, 2006 AIAA 2006-0652
[2] G.R. Barrenechea, F. Chouly (2009) ”A finite element method for the
resolution of the Reduced Navier-Stokes/Prandtl equations” ZAMM -
Volume 89 Issue 1, Pages 54 - 68
[3] P. Bradshaw, T. Cebecci & J.H. Whitelaw (1981): ”Engineering calcu-
lation methods for turbulent flow” Academic Press.
[4] J.E. Carter (1979) ”A new boundary layer inviscid iteration technique
for separated flow”. AIAA-79-1450, Jul 1979.
[5] Castem 2000 http://www-cast3m.cea.fr/cast3m/index.jsp
[6] D. Catherall and K.W. Mangler (1966) ”The integration of the two-
dimensional laminar boundary-layer equations past the point of vanish-
ing skin friction. ”J. Fluid Mech. 26 (1966) pp.163182.
[7] Cebeci T. & Cousteix J. (2005): “Modeling and computation of bound-
ary layer flows“, Springer Verlag 502p.
[8] F. Chouly, A. Van Hirtum, X. Pelorson, Y. Payan, and P.-Y. Lagre
(2008): ”Numerical and experimental study of expiratory flow in the
case of major upper airway obstructions with fluid-structure interaction”
Journal of Fluids and Structures 24 (2008) pp. 250 -269
[9] Cousteix J., & Mauss J. (2007): ”Asymptotic analysis and boundary
layers”, Scientific Computations Springer, 432 p.
[10] A. Dechaume, J. Cousteix, J. Mauss, An interactive boundary layer
model compared to the triple deck theory, Eur. J. Mech. B Fluids 24
(4)(2005) 439 447
284 P.-Y. Lagrée
[11] Darrozès J.S. et François C. ”Mécanique des fluides incompressibles”.

Berlin : Springer Verlag, 1982. 461 p. Lecture notes in Physics 163. ISBN
3-540-11578-1
[12] Drela, M. and Giles, M. B. (1987). Viscous-inviscid analysis of transonic
and low reynolds number flows. AIAA J., vol 25:13471355.
[13] M. Drela. ”XFOIL: An analysis and design system for low Reynolds
number airfoils.” In T. J. Mueller, editor, Low Reynolds Number Aero-
dynamics. Springer-Verlag, Jun 1989. Lecture Notes in Engineering, No.
54.
[14] D.E. Edwards and J.E. Carter (1985) ”A quasi-simultaneous finite dif-
ference approach for strongly interacting flow.” In Third Symposium on
Numerical and Physical Aspects of Aerodynamic Flows, Long Beach,
California, Jan 1985.
[15] Fletcher C.A.J. (1991): “Computational techniques for fluid dynamics,
vol II“, Springer Verlag.
[16] R W Garvine (1968) ” Upstream influence in Viscous Interaction Prob-
lems” Phys of Fluids V 11 N 7,pp. 413 1423
[17] Gersten, K. and Hervig, H. (1992): “Strömungsmechanik : Grundla-
gen der Impuls-Wärme-und Stoffübertragung aus asymptotischer Sicht“,
Vieweg, Wiesbaden.
[18] S. Goldstein (1948) ”On laminar boundary-layer flow near a position
of separation”. Quarterly J. Mech. and Appl. Math., 1 pp. 4369
[19] Kundu P.K. Cohen I.M. (2008) ”Fluid Mechanics Fourth Edition”,
Academic Press.
[20] A. Jameson, The Evolution of Computational Methods in Aerodynam-
ics. ASME J. Appl . Mech. 50 (1983) 1052-1070.
[21] Keller H.B. (1978): ”Numerical methods in Boundary Layer theory”,
Ann. Rev. Fluid Mech, vol 10, pp. 417-433.
[22] P.-Y. Lagrée, A. Van Hirtum & X. Pelorson (2007): ”Asymmetrical
effects in a 2D stenosis”. European Journal of Mechanics - B/Fluids,
Volume 26, Issue 1, January-February 2007, Pages 83-92
[23] Lagrée P.-Y., Berger E., Deverge M., Vilain C. & Hirschberg A. (2005):
”Characterization of the pressure drop in a 2D symmetrical pipe: some
asymptotical, numerical and experimental comparisons”, ZAMM, 5, No.
2, pp. 141-146.
[24] P.-Y. Lagrée & Sylvie Lorthois (2005),”The RNS/Prandtl equations
and their link with other asymptotic descriptions: application to the
wall shear stress scaling in a constricted pipe”, International Journal of
Engineering Sciences, Vol 43/3-4 pp 352-378
[25] Landau L. & Lifshitz E. (1997) ”Fluid Mechanics” Butterworth-

Heinemann, 539 pages
[26] J.C. Le Balleur (1977), Couplage visqueux non-visqueux: Analyse
du problme incluant décollements et ondes de choc. La Recherche
Aérospatiale. no 6pp. 349-358.
[27] J.C. Le Balleur (1978), Couplage visqueux non-visqueux: Méthode
numérique et applications aux écoulements bidimensionnels
transsoniques et supersoniques. La recherche Aérospatiale. 1978-2,
Eng Trans ESA TT-496 (1978), pp. 6576.
[28] J.C. Le Balleur (1982),”Viscid-Inviscid coupling calculations for txo
or three dimensional flows” Von Kármán Institute for Fluid Dynamics
LEcture serires 1982-04.
[29] Lees L. & Reeves B.L. ”Supersonic Separated and reattaching Laminar
Flows”, AIAA Journal, Vol 2 11, 1964 pp 1907-1920.
[30] M. J. Lighthill (1953) ”On Boundary Layers and Upstream Influence.
II. Supersonic Flows without Separation” Proceedings of the Royal So-
ciety of London. Series A, Mathematical and Physical Sciences, Vol. 217,
No. 1131 (May 21, 1953), pp. 478-507
[31] R. Lock and B. Williams. (1987) ”Viscous-inviscid interactions in ex-
ternal aerodynamics. ”Prog. Aerospace Science, 24:51171, 1987.
[32] Lorthois S. & Lagrée P.-Y. (2000): ”Écoulement dans un convergent
axisymétrique: calcul de la contrainte de cisaillement pariétal maximale/
Flow in a axisymmetric convergent: evaluation of maximum wall shear
stress”, C. R. Acad. Sci. Paris, t328, Série II b, p33-40.
[33] Lorthois S., Lagrée P.-Y., Marc-Vergnes J.-P., & Cassot. F. (2000):
”Maximal wall shear stress in arterial stenoses: Application to the in-
ternal carotid arteries”, Journal of Biomechanical Egineering, Volume
122, Issue 6, pp. 661-666.
[34] A.F. Messiter (1970): Boundarylayer flow near the trailing edge of a
flat plate. SIAM J. Appl. Math., 18 :241257,
[35] Neiland V. Ya (1969): ”Propagation of perturbation upstream with
interaction between a hypersonic flow and a boundary layer”, Mekh.
Zhid. Gaz., Vol. 4, pp. 53-57.
[36] Piomelli U, Balaras E. (2002) ”Wall-layer models for large-eddy simu-
lations.” Annual Rev. Fluid Mech. 34:34974
[37] Prandtl L. (1928) ”Motion of fluids with little viscosity/ Vier Abhand-
lungen zur Hydrdynamik und Aerodynamik Göttingen 1927”, NACA
report 452.
[38] Reyhner, T.A. and & Flügge-Lotz, I., (1968). The interaction of a
shock wave with a laminar boundary layer. Int. J. Non-linear Mech. 3,
pp. 173199.
286 P.-Y. Lagrée
[39] Rubin S.G. & Himansu A. (1989): “Convergence properties of high

Reynolds number separated flow calculations“, Int J. Num Meth Fluids
vol 9 1395-1411.
[40] Ruban A.I. & Timoshin S.N. (1986): “Propagation of perturbations in
the boundary layer on the walls of a flat channel“, MZG 2, pp. 74-79.
[41] H. Schlichting (1987): ”Boundary Layer Theory”, 7th edition Mac
Graw Hill.
[42] H. Schlichting K. Gersten (2000): ”Boundary Layer Theory”, 8th edi-
tion Springer.
[43] Smith F. T. (1976): ”Flow through constricted or dilated pipes and
channels”, part 1 and 2, Q. J. Mech. Appl. Math. vol 29, pp 343- 364 &
365- 376.
[44] I.J. Sobey (2000): ”Introduction to interactive boundary layer theory”,
Oxford applied and engineering mathematics, 256 p.
[45] Stewartson K. (1964) ”the Theory of Laminar Boundary Layers in
Compressible Fluids”, Oxford University Press, London (1964).
[46] Stewartson K. & Williams P.G. (1969): ”Self - induced separa-
tion”,Proc. Roy. Soc. London, A 312, pp 181-206.
[47] Stewartson K. (1974): ”Multistructured boundary layer on flat plates
and related bodies” Advances Appl. Mech. 14 p145-239.
[48] Van Dyke M. (1962): ’Higher approximations in boundary layer the-
ory”, JFM 14, pp 161-177.
[49] Sychev V. V. , Ruban A. I. , Sychev V. V. & Korolev G. L. (1998):
”Asymptotic theory of separated flows”, Cambridge University Press.
[50] Tannehil J. C., Anderson D. A. & Plettcher R. H. (1997): ‘Computa-
tional fluid mechanics and heat transfer“, Taylor & Francis.
[51] Van Dyke M. (1975): ”Perturbation Methods in Fluid Mechanics”
Parabolic Press.
[52] A.E.P. Veldman (2001): ”Matched asymptotic expansions and the nu-
merical treatment of viscous-inviscid interaction”, Journal of Engineer-
ing Mathematics, Volume 39, Issue 1, March, pp. 189 - 206
[53] A.E.P. Veldman. The calculation of incompressible boundary layers
with strong viscous-inviscid interaction. In Conference on Computation
of Viscous-Inviscid In- teractions, 1980. AGARD-CP-291.
[54] Werle, M. J., ; Dwoyer, D. L., ; Hankey W. L., (1973) ”Initial conditions
for the hypersonic-shock/boundary-layer interaction problem.” AIAA
Journal 1973 0001-1452 vol.11 no.4 (525-530) doi: 10.2514/3.6780
[55] L.B. Wigton and M. Holt. (1981) ”Viscous-inviscid interaction in tran-
sonic flow”. AIAA-81- 1003, 1981.
Interaction mechanisms in the mixed
convection flow past a horizontal plate
*
Abstract Two different interaction mechanisms arise in the asymp-
totic analysis of mixed convection flow past a horizontal plate in the
limit of large Reynolds and Grashof numbers. A global interaction
mechanism between the wake flow and the potential flow and a local
triple deck interaction mechanism at the trailing edge. Both inter-
action mechanisms will be analyzed in the framework of matched
asymptotic expansions.
1 Introduction
Two different interaction mechanisms will be investigated by considering
the flow past a horizontal heated plate which is aligned under a small angle
of attack φ to the oncoming parallel flow with velocity Ũ∞ in a distin-
guished limit of large Reynolds Re and large Grashof number Gr. The
global structure of the flow field is shown in figure 1. According to the
method of matched asymptotic expansions in the limit of large Reynolds
numbers the flow field can be decomposed into the outer inviscid potential
flow, the boundary-layer flow along the plate, the wake behind the plate
and several sub-layers near the trailing edge of the plate.
However, the asymptotic approximation in the different layers cannot be
determined in a hierarchical order. The following two different interaction
mechanism can be identified:
Global interaction: The wake and the potential flow have to be de-
termined simultaneously since the temperature perturbation in the wake
causes a pressure difference across the wake which influences the global flow
field. On the other hand the potential flow determines the inclination of the
wake and thus the velocity and temperature distribution in the wake.
Local interaction: Near the trailing edge of the plate the boundary-
layer interacts with the potential flow according to the well known triple-
deck mechanism, see Stewartson (1969). Here the influence of buoyancy
© CISM, Udine 2010
288 H. Steinrück
forces on the interaction mechanism will be analyzed. In the context of a

sudden change of the temperature perturbation local interaction problems
of mixed convection flow have been studied by Lagree (1999).
g
y
U∞ = 1, θ∞ = 0 K
−1 Re−1/2 x
θp = 1 0
Figure 1. Mixed convection flow past a horizontal plate.
The starting point of the analysis are the Navier Stokes equations for
an incompressible fluid using Boussinesq’s approximation to take buoyancy
forces into account and the energy equation. We introduce a Cartesian co-
ordinate system such that its x-axis is horizontal and its origin is at the
trailing edge of the plate. In the following we will use dimensionless vari-
ables. All lengths (if not stated otherwise) are scaled with the plate length
L̃. Velocities are scaled with the velocity Ũ∞ of the unperturbed parallel
flow. The dimensionless velocity components u, v in x and y direction and
the dimensionless temperature perturbation θ = (T̃ − T̃∞ )/(T̃p − T̃∞ ) satisfy
the equations of motion, the energy equation and the continuity condition
1
uux + vuy = −px + (uxx + uyy ), (1a)
Re
1 Gr
uvx + vvy = −py + (vxx + vyy ) + θ, (1b)
Re Re2
1
uθx + vθy = (θxx + θyy ), (1c)
Re P r
ux + uy = 0, (1d)
subject to the asymptotic boundary conditions
u = 1, v = φ, θ = 0 as x2 + y 2 → ∞ (2)
and the no-slip boundary conditions at the plate
u(x, 0) = v(x, 0) = 0, θ(x, 0) = 1, −1 < x < 0. (3)

Interaction Mechanisms in Mixed Convection Flow… 289
The dimensionless parameters Reynolds number Re = Ũ∞ L̃/ν̃, Grashof

number Gr = g̃β̃ΔT̃ L̃3 /ν̃ 2 and Prandtl number P r = ν̃/ã are defined
as usual. Here β̃, ν̃ and ã denote the isobaric expansion coefficient, the
kinematic viscosity and the thermal diffusivity, respectively. Additionally
to the above mentioned dimensionless parameters the angle of attack φ
enters the problem. Although the Reynolds number is assumed to be large,
we restrict the analysis to laminar flow conditions. Data for the physical
quantities satisfying this assumption can be found in Savić and Steinrück
(2007).
In order to define a meaningful interaction parameter we estimate the
order of magnitude of the physical quantities involved in the global in-
teraction mechanism. The temperature perturbations are limited √ to the
boundary-layer and the wake, which are both of the thickness L̃/ Re. The
temperature perturbation in the wake causes a vertical hydro-static pres-
sure gradient in the wake which is of the order√ρ̃g̃β̃ΔT̃ . Thus there is
a pressure difference of order Δp̃h = ρ̃g̃ β̃ΔT̃ L̃/ Re across the wake, cf.
Schneider (2005). This pressure difference induces a velocity perturbation
of the outer potential flow field of the order Δp̃h /ρ̃Ũ∞ . This in turn causes
2
a small inclination of the wake of order K = Δp̃h /ρ̃Ũ∞ = Gr Re−5/2 , (see
table 1).
From the viewpoint of the outer (potential) flow the wake is located
around the streamline starting from the trailing edge of the plate. Due to
the (small) inclination of the wake there is a non-vanishing component of the
hydro-static pressure gradient in the main flow direction of the wake flow.
It is of the order ρ̃g̃ β̃ΔT̃ K. Referring this pressure gradient in main flow
2
direction to the reference value for a pressure gradient ρ̃Ũ∞ /L̃ we obtain
the interaction parameter
g̃β̃ΔT̃ L̃
κ2 = K = Gr2 Re−9/2 . (4)
ũ2∞
Thus the velocity profile and the temperature profile in the wake depend on
the potential flow, namely on the inclination of the streamline emanating
from the trailing edge. As a consequence the wake flow problem and the
potential flow problem form an interaction problem and thus both problems
have to be solved simultaneously. This problem has been first formulated
by Savić and Steinrück (2005) and solved numerically. However, solutions
of the potential flow problem exist only if the pressure perturbation across
the wake decay downstream. This is the case if the oncoming flow has a
positive angle of attack φ which is of the order of the
√ induced inclination of
the wake. Thus an inclination parameter λ = φK Re is introduced. The
global interaction problem is formulated in section 2 and and numerical
290 H. Steinrück
thickness of the wake L̃ Re−1/2

hydro-static pressure gradient in the ρ̃g̃β̃ΔT̃
wake
hydro-static pressuure jump across Δp̃h = ρ̃g̃β̃ΔT̃ L̃Re−1/2
the wake
velocity perturbation induced by Δp̃h
Δp̃h ρ̃Ũ∞
inclination of the wake centerline Δp̃h
induced by Δp̃h K= = Gr Re−5/2
2
ρ̃Ũ∞
hydro-static pressure gradient in di- ρ̃g̃β̃ΔT̃ K
rection of the wake centerline
g̃ β̃ΔT̃ L̃
interaction parameter κ2 = K = Gr2 Re−9/2
2
Ũ∞
Table 1. Magnitudes of physical quantities involved in the interaction mech-
anism
solutions are discussed. Keeping the inclination parameter λ constant it

turns out that solutions exist only for interaction parameters κ below a
critical value κc . At κ = κc a singularity in the wake is observed which will
be discussed.
Near the trailing edge of the plate the boundary layer interacts (locally)
with the potential flow and sub-layers according to triple deck theory, cf.
Stewartson (1969); Messiter (1970). It will be analyzed in section 3. As we
will see the inclination parameter λ will play no role in the trailing edge
analysis. The local interaction problem will be analyzed and a numerical
solution reveals that the interaction pressure is discontinuous at the trailing
edge. Thus new sub-layers are introduced to resolve the discontinuity.
2 The global interaction problem

2.1 The potential flow
Using the notation of complex valued functions of a complex variable
z = x + iy the potential flow can be written as

z
u − iv = 1 − iφ + K(u1 − i v1 ). (5)
z+1
The first two terms on the right side of equation (5) describe the potential
flow past a horizontal plate of a free stream with an angle φ to the horizontal
axis. The third term on the right side of (5) takes the buoyancy effects into
account. Along the plate the vertical velocity component v1 has to vanish.
From the view point of the potential flow the scaled pressure has a jump
discontinuity of size γw across the wake. Using the linearized Bernoulli
equation we have
−u1 (x, 0+) + u1 (x, 0−) = γw (x), x > 0, (6)
where γw is the dimensionless pressure jump across the wake, see (11).
If γw (x) is given, following Savić and Steinrück (2005), we obtain for the
dimensionless inclination of the wake

x
yw (x) = φ + Kv1 (x, 0), (7)
x+1
with

∞
1 x ξ + 1 γw (ξ) dξ
v1 (x, 0) = . (8)
2π 0 ξ x+1 x−ξ
2.2 The wake flow

√
In the wake we introduce the vertical coordinate Y = (y − yw (x)) Re
referred to the centerline y = yw (x) of the wake. Thus the equations for
the flow u ∼ U (x, Y ), v ∼ Re−1/2 V (x, Y ), pressure p ∼ K P (x, Y ) and
temperature field θ = Θ(x, Y ) are
Ux + VY = 0, (9a)
U Ux + V UY = Yw Θ + UY Y , UY (x, 0) = 0, U (x, ∞) = 1, (9b)

PY = Θ, (9c)
1
U Θx + V ΘY = ΘY Y , ΘY (x, 0) = 0, Θ(x, ∞) = 0. (9d)
Pr

√ √
Note that −px = −K(Px −PY yw Re). Using (9c) and with Yw = K Reyw
denoting the appropriately scaled centerline of the wake the momentum
equation in x-direction (9b) is obtained. If the wake is inclined the hydro-
static pressure gradient has a non-vanishing component in the main flow
direction. Thus the fluid in the wake is accelerated in case of positive
inclination Yw > 0 and decelerated in case of negative inclination Yw < 0.
At the trailing edge, x = 0, the velocity and temperature profiles are
given by the Blasius velocity profile UB (Y ) and the corresponding temper-
ature profile ΘB (Y ) for the case of forced convection,
U (0, Y ) = UB (Y ), Θ(0, Y ) = ΘB (Y ). (10)

292 H. Steinrück
The scaled hydro-static pressure difference across the wake is given by
∞
γw (x) = 2 Θ(x, Y ) dY. (11)
0
From the viewpoint of the potential flow γw can be interpreted as a vortex

distribution along the wake centerline, see Schneider (2005).
The inclination condition (7) can be rewritten as

x
Yw (x) = λ + κ2 v1 (x, 0), (12)
x+1
√
with λ = φK Re. From equation (12) it becomes evident when the wake
inclination influences the flow and temperature field in the wake. If λ is
of order one and κ2 1, the velocity and temperature profile in the wake
is affected by the wake inclination. The hydrostatic pressure difference
across the wake is also influenced and thus the outer potential flow u1 + iV1 .
However, in that case there is no interaction.
Wake-potential flow interaction is only present if κ is of the order one. In
the following we assume λ = 1 fixed and vary the interaction parameter κ.
2.3 Numerical solution

A necessary condition for the existence of the integral in (8), such that
∞ is that γw (x) decays to zero for x → ∞. Since the total enthalpy
v1 exists,
flux 0 uθ dy in the wake is constant, γw can only vanish, if the velocity u
in the wake tends to infinity. This is the case when λ > 0. Then in the far
field a similarity solution of the form
1 Y
U ∼ λ2/5 x1/5 F (η), Θ∼ D(η), η = λ1/5 (13)
λ1/5 x3/5 x2/5
exists, see Savić and Steinrück (2005), where F and D are the solutions of
the similarity equations
3 1 1 3
F + F F − F F + D = 0, D + F D = 0, (14)
5 5 Pr 5
with the boundary conditions
∞ ∞

F (0) = F (0) = F (∞) = 0, F D dη = UB ΘB dY. (15)
0 0
Thus in the far field the velocity and temperature profiles of the wake
flow tend to the velocity and temperature profiles of a two-dimensional
laminar plume. Since the flow and temperature profile of the wake flow
is symmetric with respect to the centerline, it is sufficient to integrate the
enthalpy flux only over one half of the wake.
For the numerical solution of the coupled wake (9), (10), (11), and po-
tential flow equations (8) and the interaction or inclination condition (12)
an iterative method is proposed.

(0) x
i) First a suitable wake centerline Yw = λ x+1 is chosen.
ii) The wake equations are integrated for a velocity U (i) and temperature
(i−1)
field Θ(i) by a marching technique for a prescribed inclination Yw
of the wake.
(i) ∞
iii) Then the pressure jump γw = 2 0 θ(i) dy across the wake is deter-
mined
(i)
iv) Evaluating (8) a new centerline Yw of the wake is determined and
steps ii) -iv) are repeated until convergence is obtained.
We note that for κ = 0 no iterations are necessary. In the following we
keep the inclination parameter λ = 1 and the Prandtl number P r = 0.71
fixed. The interaction parameter κ will be increased starting from zero.
In figure 2 the velocity at the centerline of the wake and the pressure
jump across the wake are shown. For κ = 0 the shape of the wake is given
by the well known 2-d potential flow solution of the flow past an inclined
plate Schneider (1978). The centerline velocity increases from u = 0 at the
trailing edge due to viscosity. Then buoyancy leads to further acceleration
and a velocity overshoot forms. Accordingly the vortex distribution γw (x)
(or the pressure jump across the wake) decreases.
Evaluating the integral (8) shows that the induced vertical velocity com-
ponent v1 is negative. Thus for κ sufficiently large the wake turns down-
wards about a plate length behind the trailing edge. After attaining a min-
imum the wake turns upwards again. Accordingly the graph of centerline
velocity first becomes flat. Increasing κ further a minimum forms. When
κ attains a critical value κ = κc this minimum becomes zero. Since this
solution is singular at the zero of the centerline velocity a further increase
of κ is not possible. The physical mechanism which causes the singularity is
the following: In the parts of the wake with downward inclination the wake
flow is decelerated. The deceleration of the wake causes the wake to broaden
there. The increase of the wake thickness causes finally an increase of the
hydro-static pressure jump across the wake. In the limiting case κ = κc ,
the wake thickness becomes infinite in wake coordinates and thus γw also
tends to infinity.
In order to compute solutions with κ close to the critical value κc a
different strategy has to be employed. First a value Umin for the minimum
294 H. Steinrück
of the centerline velocity is prescribed. We chose a suitable vertical velocity

(n)
perturbation v1 and determine κ such that the minimum of centerline
velocity has the prescribed value. This has to be done iteratively. Then a
(n+1)
new vorticity distribution is computed and a new vertical velocity v1 is
determined. The process is repeated until convergence is obtained.
2.4 Analysis of the wake singularity

In order to study the singularity it is convenient to transform the wake
equations to the von Mises coordinates and use W = u2 as dependent
variable. We define

u(x, Y ) = W (s, ψ(x, Y )), Θ(x, Y ) = Θ̃(s, ψ(x, Y )) (16)
with Y
s = x − x0 , ψ(x, Y ) = u(x, Y ) dY , (17)
0
where x0 is the location of the singularity for κ = κc and ψ is the stream

function of the wake flow. Thus we obtain
√ √
Ws = 2Yw Θ̃ + W Wψψ , Θ̃s = W Θ̃ψ (18)
ψ
with the boundary conditions Wψ (s, 0) = Θ̃ψ (s, 0) = 0 and W (s, ∞) = 1,

Θ̃(s, ∞) = 0. The pressure jump across the wake is given by
∞
Θ̃(s, ψ)
γw = 2 dψ. (19)
0 W (s, ψ)
At s = 0 for |κ − κc | 1 we have W (0, ψ) ∼ ε + W0 (ψ) with W0 (0) =

W0 (0) = 0. The parameter ε represents the value of the minimum of W . In
the limit κ = κc it vanishes.
We expand W asymptotically for |s| 1, ε 1:
W (s, ψ) ∼ ε + W0 (ψ) + Ŷw (s)W1 (ψ) + sW2 (ψ) + · · · ,

(20)
Θ̃(s, ψ) ∼ Θ̃0 (ψ) + sΘ̃1 (ψ) + · · ·
with Ŷw (s) = Yw (x) − Yw (x0 ). The local behavior of the centerline Ŷw is
not known a priorily and thus a corresponding term in the expansion of W
is added. Inserting into the differential equation (18) we obtain

W1 = 2Θ̃0 , W2 = W0 W0 , Θ̃1 = ( W0 Θ̃0 ) . (21)

κ=0

U (x, 0)
κ = 0.5

κ = 0.9141
κ = 0.9173

κ = 0.9174
Umin = 0.13
Umin = 0

x
a) Centerline velocity in the wake

Umin = 0

Umin = 0.13

κ = 0.9173 κ = 0.9174
γw
κ = 0.9141
κ = 0.5
κ=0

x
b) Pressure jump across the wake
Figure 2. Numerical solution of the wake problem for λ = 1, P r = 0.71 .
However, W1 , W2 and Θ̃1 do not satisfy the boundary condition at ψ = 0.

Thus a sub-layer has to be introduced. It turns out that the sub-layer does
not influence the leading order equations and thus it will not be discussed
here. Note that W2 (0) = 0.
296 H. Steinrück
Using the local asymptotic expansion we can determine the local be-
havior of γw . We choose some value ψ ∗ > 0 and approximate W ∼
ε + W0 ψ/2 + Ŷw (s)W1 (0) and Θ̃ ∼ Θ̃0 (0) for ψ < ψ ∗ and obtain
∞ ψ∗
Θ̃ Θ̃0 (0)
γw (s) = 2 √ dψ ∼ 2 dψ
0 W 0 ε + W0 ψ 2 /2 + Ŷw (s)W1 (0) (22)
√
2Θ̃0 (0)
∼ − ln |ε + Ŷw (s)W1 |.
W0 (0)
It can be shown that the singular part of γ(s) is independent of the choice
of ψ∗ .
Considering that u1 − iv1 is a potential flow, using the complex valued
function theory and u1 = −γw /2 we conclude
Θ̃0 (0)
u1 − i v1 = ln F (z, ε), (23)
2W0 (0)
where F (z; ε) is a complex valued function of z with
Θ̃0 (0)
|F (s)| = |ε + Ŷw (s) W1 |, − arg F (s) = Ŷw for s real, (24)
2W0 (0)
and Ŷw (0) = 0, Ŷw (0) = 0. This constitutes a problem for finding F (z, ε)
and Ŷw (s, ε) simultaneously.
We can express the solution F (z, ε) = εF̃ (z/ε), Ŷw (s, ε) = εỸw (s/ε) of
(24) for arbitrary values of ε by the solution F̃ and Ỹw of (24) for ε = 1. In
the limiting case ε = 0 we can guess the solution F (z, 0) = −iπ √Θ̃0 (0)
z.
2 2W0 (0)
Thus in that case the centerline of the wake has a corner of size
Θ̃0 π
[Yw ] = . (25)
2W0 (0)
As a consequence the centerline velocity behaves in the limiting case Umin =

0 locally like U ∼ |s| , see figure 2a and the hydro-static pressure difference
γw has a logarithmic singularity, see figure 2b. For ε > 0 the corner is
smoothed.
3 Local Interaction at the Trailing-Edge

For the analysis of the flow field near the trailing-edge the velocities, pressure
and temperature are decomposed into a symmetric and anti-symmetric part
u(x, y) + u(x, −y) u(x, y) − u(x, −y)

ū(x, y) = , Δu = . (26)
2 2Re−1/4 κ
All other dependent variables, with the exception of the vertical velocity
component v, are decomposed accordingly. We decompose the vertical ve-
locity v as
v(x, y) − v(x, −y) v(x, y) + v(x, −y)

v̄(x, y) = , Δv = . (27)
2 2Re−1/4 κ
We rewrite the basic equations for mixed convection flow in terms of the
symmetric and anti-symmetric part using the boundary-layer coordinates
x, y = Re−1/2 Y , u = U (x, Y ), v = Re−1/2 V (x, Y )

∂ Ū ∂ Ū κ2 ∂ΔU ∂ΔU
Ū + V̄ + ΔU + ΔV =
∂x ∂Y Re1/2 ∂x ∂Y
(28a)
∂ P̄ ∂ 2 Ū 1 ∂ 2 Ū
− + + ,
∂x ∂Y 2 Re ∂x2

1 ∂ V̄ ∂ V̄ κ2 ∂ΔV ∂ΔV
Ū + V̄ + ΔU + ΔV =
Re ∂x ∂Y Re3/2 ∂x ∂Y
(28b)
∂ P̄ κ2 1 ∂ 2 V̄ 1 ∂ 2 V̄
− + Δθ + + .
∂Y Re1/2 Re ∂Y 2
Re2 ∂x2
Thus in the equations for the symmetric part the reduced buoyancy
parameter κ appears in the terms of order Re−1/2 . However, these terms
do not influence the equations for leading order terms of the triple deck
analysis. For the anti-symmetric parts we obtain
∂ΔU ∂ Ū ∂ΔU ∂ Ū
Ū + ΔU + V̄ + ΔV =
∂x ∂x ∂Y ∂Y
(29a)
∂ΔP ∂ 2 ΔU 1 ∂ 2 ΔU
− + + ,
∂x ∂Y 2 Re ∂x2
298 H. Steinrück

1 ∂ΔV ∂ V̄ ∂ΔV ∂ V̄
Ū + ΔU + V̄ + ΔV =
Re ∂x ∂x ∂Y ∂Y
(29b)
∂ΔP 1 ∂ 2 V̄ 1 ∂ 2 ΔV
− + θ̄ + + .
∂Y Re ∂Y 2 Re2 ∂x2
If the symmetric parts of the flow and pressure field are known the
equations for the anti-symmetric part are linear and independent of κ.
In the limit of large Reynolds numbers Re the flow structure near the
trailing edge can be described by a triple deck problem, cf. Stewartson
(1969); Messiter (1970). The scaling of the different layers is sketched in
Figure 13 in Ruban (2010). We define the independent variables in the
different layers (lower, main and upper deck) as
x∗ = Re3/8 x, y∗ = Re3/8 y, Y∗ = Re1/2 y, Y∗ = Re5/8 y. (30)
Here we summarize the leading order terms of the asymptotic expansion

of the symmetric part of the velocity and pressure field in the triple deck
region: ⎧ −1/4
⎨ 1 + Re ū∗,2 (x∗ , y∗ ) + ...,
ū(x, y) = U (Y ) + Re−1/8 Ā(x∗ )UB (Y ) + ..., (31a)
⎩ B−1/8
Re Ū∗,1 (x∗ , Y∗ ) + ...,
⎧ −1/4
⎨ Re v̄∗,2 (x∗ , y∗ ) + ...,
v̄(x, y) = −Re−1/4 Ā (x∗ )UB (Y ) + ..., (31b)
⎩
Re−3/8 V̄∗,1 (x∗ , Y∗ ) + ...,
⎧ −1/4
⎨ Re p̄∗,2 (x∗ , y∗ ) + ...,
p̄(x, y) ∼ Re−1/4 P̄2 (x∗ ) + ..., (31c)
⎩ −1/4
Re P̄∗,2 (x∗ ) + ... ,
with p̄∗,2 (x∗ , 0) = P̄2 (x∗ ) = P̄∗,2 (x∗ ) and UB (Y ) is the Blasius velocity
profile in the self-similar boundary-layer at the trailing edge. For further
use we define the constant a0 = UB (0) = 0.332.
According to Stewartson (1969) the function Ā can be interpreted as
the correction of the negative displacement thickness. We recall that the
asymptotic behavior of the negative displacement thickness Ā is given by
1/3
Ā(x∗ ) ∼ A0 (x∗ ) as x∗ → ∞, (32)
with the constant A0 given in table 2. In analogy to the velocity profile of the
symmetric part in the main deck the temperature profile of the symmetric
part is given as
θ ∼ ΘB (Y ) + Re−1/8 Ā(x∗ )ΘB (Y ). (33)
In the following we will discuss the interaction problem for the anti-
symmetric part of the solution.
3.1 Main deck

We start with the main deck then using the upper deck we derive the in-
teraction law and finally derive the lower deck problem. The anti-symmetric
part of the pressure in the main deck can be expanded in the form
Δp = ΔP0 + Re−1/8 ΔP1 + ... . (34)
In contrast to (classical) triple deck problems the pressure is not constant
across the main deck. The pressure involved in the interaction mechanism
is of order Re−1/8 , i.e. ΔP1 . The y-momentum equation reduces to
∂ΔP0 ∂ΔP1
= Θ̄0 , = Θ̄1 , (35)
∂Y ∂Y
with Θ̄ = ΘB and Θ̄1 = ĀΘB . Using (33) we obtain
Y
ΔP0 (x∗ , Y ) = ΘB (Ỹ ) dỸ + Δp∗,0 (0, 0), (36a)
∞
ΔP1 (x∗ , Y ) = Ā(x∗ )ΘB (Y ) + ΔP1 (x∗ , 0). (36b)
The expansions for the velocity components Δu and Δv follow the same
lines as Stewartson (1969) and the solution of the equations of the leading
order terms can be expressed in terms of an yet undetermined function
ΔA of x∗ , which can be interpreted as the scaled difference of the negative
displacement thicknesses on the upper and lower side of the plate. The
leading order terms of the anti-symmetric part of the velocity components
are given as
c1
Δu = ln Re U (Y ) + ΔA(x∗ )UB (Y ) + C1 (Y ) + ..., (37a)
UB (0) B
Δv = −Re−1/8 ΔA (x∗ )UB (Y ) + .... . (37b)

The term of the magnitude ln Re and the term C1 (Y ) are both independent
of x∗ and arise from matching the main deck solution with the solution of
the boundary-layer equations, for the anti-symmetric part of the flow field,
cf. (58). The constant c1 will be determined later in (53). In table 2 all
constants, their numerical value, and their definitions are summarized.
300 H. Steinrück
3.2 The upper deck

Since the flow in the upper deck is a potential flow with the velocity field
Δu − iΔv = Δu0 (0, 0) + Re−1/8 (Δu∗,1 (x∗ , y∗ ) − iΔv∗,1 (x∗ , y∗ )) + ..., (38)
where Δu0 (0, 0) = κ1 u1 (0, 0+), see (5), and Δu∗,1 (x∗ , 0) = −Δp∗,1 (x∗ , 0),
Δv∗,1 (x∗ , 0) = −ΔA (x∗ ) holds, the pressure Δp∗,1 (x∗ , 0) and the negative
displacement thickness ΔA (x∗ ) can be interpreted as the real and imaginary
part of a complex analytical function ΔΦ1 evaluated on the real axis. We
have
ΔΦ1 (x∗ , 0) = −Δp∗,1 (x∗ , 0) + iΔA (x∗ ) =

(39)

− ΔP1 (x∗ , 0) − Ā(x∗ ) + iΔA (x∗ ).
Considering ΔP1 (x∗ , 0) = 0 for x∗ > 0 and using the asymptotic behav-
ior of Ā for x∗ → ∞ we conclude that
ΔΦ1 (z) ∼ (a + ib)z 1/3 for z→∞ (40)

√ ΔA → 0 for
holds. The constants a and b are determined by using that
x(3) → −∞. They turn out to be a = A0 and b = − 3A0 . Thus the
asymptotic behavior of ΔP1 and ΔA is given by
ΔP1 (x∗ , 0) ∼ −2A0 |x∗ |1/3 for x∗ → −∞, (41)
√
ΔA (x∗ ) ∼ − 3A0 |x∗ |1/3 , for x∗ → ∞. (42)
Note that if ΔP∗,1 − Ā(x∗ ) tends to different constants for x∗ → ∞ and

x∗ → −∞ the real part of the complex function ΔΦ1 − (a + ib)z 1/3 would
tend to different constant values for x∗ → −∞ and x∗ → ∞. Thus the next
order term in the expansion of Φ1 for z → ∞ would be of the form i ln z
contradicting the requirement that ΔA vanishes for x∗ → −∞. Thus a
possible constant in the expansion of ΔP1 for x∗ → −∞ must be the same
constant as in the expansion of ΔP1 − Ā for x∗ → ∞. But using the fact
that ΔP1 (x∗ ) = 0 for x∗ > 0 and equation (4.4a) in Stewartson (1969) we
conclude that this constant has to vanish thus showing that ΔΦ1 (z) − (a +
ib)z 1/3 → 0 as z → ∞.
Finally the interaction law can be written in the form

√ 1/3
ΔA (x∗ ) + 3A0 h(x∗ )x∗ =
0
1 ΔP1 (ξ, 0) + 2A0 |ξ|1/3
− dξ (43)
π −∞ x∗ − ξ
∞
1 Ā(ξ) − A0 h(ξ)|ξ|1/3
+ dξ,
π −∞ x∗ − ξ
where h(x) denotes the Heaviside function with h(x) = 1 for x > 0 and
h(x) = 0 for x < 0.
We have written the interaction law in a form such that the singular parts
are separated and the integrand in the Hilbert integral decays sufficiently
fast to zero for x∗ → ±∞.
The asymptotic behavior of the vertical velocity component v for x∗ →
∞ and Y → ±∞ is given by
v = ±v̄ + κRe−1/4 Δv ∼

∼ − κRe−3/8 ΔA UB (Y ) ± Re−2/8 Ā UB (Y ) = (44)
√ 1/3 1 −2/3
= κRe−3/8 3A0 (x∗ ) UB (Y ) ± Re−2/8 A0 x∗ UB (Y ).
3
Rewriting (44) in the outer variables we have
√ 1
v(x, y) ∼ κRe−2/8 3A0 x1/3 ± Re−4/8 A0 x−2/3 . (45)
3
Applying the matching principle we conclude for the asymptotic behaviour
for of v1 given in (8)
√
v1 (x, 0) ∼ 3A0 x1/3 for x → 0 + . (46)
3.3 The lower deck

The equations for the velocity profile in the lower deck are given by the
momentum equation in the x-direction,
∂ΔU∗,0 ∂ Ū∗,1 ∂ΔU0,∗ ∂ Ū∗,1

Ū∗,1 + ΔU0,∗ + V̄∗,1 + ΔV0,∗ =
∂x∗ ∂x∗ ∂Y∗ ∂Y∗
(47)
∂ΔP∗,1 ∂ 2 ΔU∗,0
− + ,
∂x∗ ∂Y∗2
302 H. Steinrück
the continuity equation, and the momentum equation in the y-direction
which reduces to
∂ΔP∗,1
0=− + Θ̄B (0). (48)
∂Y
The boundary conditions are
ΔU∗,0 (x∗ , 0) = ΔV∗,0 (x∗ , 0) = 0, x∗ < 0 at the plate,

(49)
ΔU∗,0 (x∗ , 0) = ΔP∗,1 (x∗ , 0) = 0, x∗ > 0 in the wake.
We remark that changes of the temperature profile in the lower deck are
too small to influence the leading order terms of the hydrostatic pressure
distribution and thus a discussion of the energy equation is not necessary.
Integrating (48) we obtain the pressure difference in the lower deck
ΔP∗,1 (x∗ , Y+ ) = θB (0)Y∗ + ΔP∗,1 (x∗ , 0), (50)
which matches with the pressure difference in the main deck (36). Thus
ΔP1 (x∗ , 0) = ΔP∗,1 (x∗ , 0) and in the interaction law (43) ΔP1 can be re-
placed by ΔP∗,1 .
It remains to specify the asymptotic behavior of the velocity profile for
x∗ → −∞ and Y∗ → ∞.
Considering the asymptotic behavior of the pressure ΔP∗,1 and of Ū∗,1 ∼
UB (0)Y∗ for x∗ → −∞ we conclude that the asymptotic behavior of the flow
field ΔU∗,0 , ΔV∗,0 in the lower deck is self-similar. Using a scaled stream
function E defined by
Y∗
ΔU∗,0 ∼ E (η), with η= , (51)
|x∗ |1/3
we obtain the similarity equation for E

3E − UB (0) η2 E − ηE + E = 2A0 , (52)
with the boundary conditions E(0) = E (0) = 0.

The corresponding homogeneous equation has three linearly independent
solutions e1 (η) ∼ η ln η for η → ∞, e2 (η) = η and e3 (η). The third solution
e3 increases at least exponentially for η → ∞. In order to match the velocity
profile with the solution of the main deck problem e3 has to be eliminated.
Thus we have
2A0 2A0
E(η) = − + c1 e1 (η) + c2 e2 (η) ∼ − + c1 η ln η + c2 η, η → ∞.
UB (0) UB (0)
(53)
Since there are two boundary conditions at η = 0 the constants c1 and c2 are
uniquely defined. Their values can be found in table 2. The corresponding
velocity profile is shown in figure 4 labeled with x∗ = −∞. The asymptotic
behavior of the velocity profile for x∗ → −∞, Y∗ → ∞ is given by
c1
Δu0 (x∗ , Y∗ ) ∼ E (η) ∼ c1 ln Y∗ − ln |x∗ | + c1 + c2 . (54)
3
To supplement the lower deck equation (47) with the correct asymptotic
boundary condition for Y∗ → ∞ we need a condition which is satisfied by the
derivatives with respect to Y∗ of all linear combinations of the two admissible
fundamental solutions 1 and ln Y∗ of the linear ordinary differential equation
(52). Such a condition is given by
∂ 2 ΔU∗,0 ∂ΔU∗,0
Y∗ 2
+ → 0, for Y∗ → ∞. (55)
∂Y∗ ∂Y∗
The negative displacement thickness ΔA is given by

∂ΔU∗,0
ΔA(x∗ ) = lim ΔU∗,0 − Y∗ ln Y∗ . (56)
Y∗ →∞ ∂Y∗
The Y∗ independent part of the asymptotic behavior of U∗,0 can be inter-
preted as the asymptotic behavior of the negative displacement thickness
ΔA. Thus we have
c1
ΔA(x∗ ) ∼ (c1 + c2 ) − ln |x∗ | as x∗ → −∞. (57)
3
Expanding the boundary-layer of the anti-symmetric part of the flow
field near the trailing-edge 0 < −x∗ 1 the flow field has a viscous sub-
layer and an inviscid main part. The solution of the viscous sub-layer is
again given by the similarity solution E. Matching the sub-layer with the
main part introduces a logarithmic term in the inviscid main part. We have:

E |x|Y1/3 for Y ∼ |x|1/3 ,
ΔU = (58)
C1 (Y ) − 3Uc 1(0) ln |x∗ |UB (Y ) + ... for Y |x|1/3 ,
B
with the asymptotic behavior of C1 given by C1 Y ∼ c1 ln Y for Y → 0.

Matching the inviscid main part of the boundary-layer (58) yields the
anticipated expansion of the main deck (37a).
In figure 3a the negative displacement thickness ΔA and the interaction
pressure ΔP∗,1 (x∗ , 0) of the anti-symmetric part of the flow field are shown.
The asymptotic behavior for x∗ → −∞ of ΔA and ΔP∗,1 is shown
in figure 3b on a logarithmic and double logarithmic scale, respectively.
Velocity profiles are shown in figure 4.
304 H. Steinrück

a20 ΔA
a20 ΔA

ï
a0 ΔP∗,1
ï
3/4
3/4
a0 ΔP∗,1
ï
ï
ï ï
5/4
x∗ a 0
a) Behaviour of ΔA and ΔP∗,1 (x∗ , 0) near the trailing edge

ï
1/3
a0 ΔP∗,1

x ∗|
a20 ΔA 2 A 0|
− ï
3/4

a20 ΔA
ï

(− c1
3 ln |
ï 3/4 x
a0 ΔP∗,1 ∗| + ï
c1 +
ï c2 ) 2
a0
ï
ï ï ï ï ï ï ï ï
5/4
x∗ a0
b) Asymptotic behavior of ΔA, and ΔP∗,1 (x∗ , 0) for x∗ → −∞

Figure 3. Negative displacement thickness ΔA and interaction pressure
ΔP∗,1 (x∗ , 0)
.

X = 0+
X = 10−5 X = 10−4X = 10−3
X = 0.163 X = 0.057

X = 0−

X = −0.163

a0 ΔU∗
X = −3.24
ï
ï
X = −∞
ï
ï

η̃ = Y∗ / x2∗ + 1
−5/4
Figure 4. Velocity profiles ΔU∗,1 at different locations x∗ = a0 X
We remark that −ΔA can be considered as the inclination of the near

wake. Thus at the trailing edge the wake first bends downwards and then
√ 1/3
turns upwards. For x∗ → ∞ we have −ΔA ∼ 3A0 x∗ which matches

with yw = v1 (x, 0) for x → 0, see eq. (2.11) in Savić and Steinrück (2007).
3.4 The local behavior of the lower deck velocity field near the
trailing-edge
It turns out that the interaction pressure ΔP∗,1 has a jump disconti-
nuity at the trailing edge x∗ = 0. As a consequence the derivative of the
displacement thickness ΔA has a logarithmic singularity at x∗ = 0. To dis-
cuss the behavior of the velocity profile ΔU∗ we integrate the momentum
equation (47) across the jump discontinuity at x∗ . We use the fact that the
symmetric part of the flow field Ū∗,0 , V̄∗,0 is continuous. Let
[ΔU∗ ](X∗ ) = ΔU∗,0 (0+, Y∗ ) − ΔU∗,0 (0−, Y∗ ), (59a)
[ΔP∗ ] = ΔP∗,1 (0+) − ΔP∗,1 (0−), (59b)

V̂∗ (Y∗ ) = lim ΔV∗,0 (x∗ , Y∗ ) dx∗ (59c)
→0 −
denote the jump in the Δu-component, the jump in the interaction pressure
ΔP∗,1 and the integral of the ΔV∗ component across the jump discontinuity,
306 H. Steinrück
respectively. Integrating the momentum equation (47) across the disconti-

nuity at x∗ = 0 we obtain the ordinary differential equation for V̂∗
Ū∗s V̂∗ − V̂∗ Ū∗s

= [ΔP∗ ], (60)
with the general solution

Y∗
Ūs (Y∗
V̂∗ (Y∗ ) = [ΔP∗ ] dζ + B Ū∗s (Y∗ ) (61)
∞ (Ū∗s (ζ))2
where B is a constant and Ū∗s (Y∗ ) = Ū∗,1 (0, Y∗ ) . The jump in ΔU∗,0 is
therefore given by
[ΔU∗ ](Y∗ ) = −V̂∗ =

(62)
Y∗
1 dζ
−[ΔP∗ ] − [ΔP∗ ] + B Ū∗s (Y∗ )
Ū∗s (Y∗ ) ∞ (Ū∗s (ζ))2
Since the displacement thickness ΔA is continuous we conclude that [ΔU∗ ] →

0 as Y∗ → ∞ and thus B = 0. Considering the behavior of [ΔU ] for Y∗ → 0
we obtain
[ΔU ] ∼ C u,log ln y (5) + Cu,0 (63)
with
a2
C u,log = [ΔP∗ ] , (64)
a21

0
a2 Ū∗s
Cu,0 = [ΔP∗ ] ln a1 − a1 ln U∗s dζ , (65)
a21 ∞
)3
(Ū∗s

with a1 = Ū∗s (0), and a2 = Ū∗s (0) = ddx
P̄∗2
∗
(0−). Their numerical values can
be found in Sychev et al. (1998) or in Chow and Melnik (1976) and are listed
together with the numerical values of C u,log and Cu,0 in table 2. In order
to satisfy the boundary condition ΔU∗,0 a sub-layer has to be introduced.
3.5 Resolving the pressure discontinuity on main deck scales

In order to resolve the discontinuity of the interaction pressure additional
sub-layers will be introduced.
Due to the discontinuity of the difference pressure ΔP∗,1 in the lower deck
at the trailing-edge the pressure difference has a discontinuity in the main
deck as well. In the upper deck the pressure difference Δp∗,1 is singular
at (0, 0). Using the calculus of analytic functions of a complex variable
z∗ = x∗ + iy∗ we can infer the behavior of Δp∗,1 close to 0. The velocity

field in the upper deck is given locally by
[ΔP∗ ]
Δu∗,1 − iΔv∗,1 ∼ Ā(0) − i ln z∗ =
π
(66)
[ΔP∗ ] y∗
Ā(0) + arctan 2
− i ln x∗ + y∗ .2
π x∗
Thus the pressure and the derivative of the displacement thickness behave
locally like
[ΔP∗ ] y∗ [Δp]
Δp∗,1 ∼ −Ā(0) − arctan , ΔA (x∗ ) ∼ − ln |x∗ |. (67)
π x∗ π
In order to resolve the discontinuity in the main deck we introduce the
sub-layer with
X = Re1/2 x, Y = Re1/2 y. (68)
The velocity profile has to match with the (x∗ , Y )-region (37a),(37b). Thus
we use the following expansion of the anti-symmetric part
Δu ∼ ln Re U c1(0) UB (Y ) + ΔU0 (0, Y )+

B (69a)
Re−1/8 ln Re [ΔP ∗]
8π XUB (Y ) + Re
−1/8
ΔŨ1 (X, Y ) + . . . ,
[ΔP∗ ]
Δv ∼ −Re−1/8 ln Re UB (Y ) + Re−1/8 Ṽ1 (X, Y ) + · · · , (69b)
8π
Δp ∼ ΔP0 (0, Y ) + Re−1/8 ΔP̃1 (X, Y ) + · · · . (69c)
The term of order Re−1/8 ln Re in the vertical velocity component Δv arises
from matching with the main deck solution and the logarithmic behavior of
ΔA as x∗ → 0. As a consequence a term of the same magnitude must be
present in the expansion of the horizontal velocity component Δu. However,
in the expansion of the pressure Δp these “logarithmic” terms are missing.
The constant c1 has been introduced in (53).
We obtain the following equations for the leading order terms
∂ΔŨ1 ∂ΔP1
UB + ΔṼ1 UB = − , (70a)
∂X ∂X
∂ΔṼ1 ∂ΔP̃1
UB =− + Θ̄1 , (70b)
∂X ∂Y
308 H. Steinrück
ΔPh

ï Y
ï

X

Figure 5. Local behavior of the interaction pressure ΔP1 near the trailing-
edge. The solution ΔPh of the homogenous problem , cf. (73), is shown
∂ΔŨ1 ∂ΔṼ1
+ = 0. (70c)
∂X ∂Y
The flow in the (X, Y )-sub-layer is inviscid. But in contrast to the main
deck the y-momentum equation is not degenerate. Eliminating ΔŨ1 and
ΔṼ1 an elliptic equation for ΔP̃1 can be derived,

∂ 2 ΔP̃1 ∂ 2 ΔP̃1 ∂ Θ̄1 ∂ΔP̃1
UB + − + 2UB Θ̄1 − = 0. (71)
∂X 2 ∂Y 2 ∂Y ∂Y
The boundary and matching conditions can be expressed as

|ΔP∗ | Y
ΔP̃1 ∼ − arctan + Ā(0) (ΘB (Y ) − 1) , (72)
π X
√
for Y = 0 or R = X 2 + Y 2 → ∞. We note that (72) represents the solu-
tion of the Laplace equation. Equation (71) becomes the Laplace equation
if UB is zero, which is the case for Y → ∞. The matching condition (72) for
R → ∞ is obtained from matching ΔP̃1 with the upper deck solution (67)
and the main deck solution. The boundary condition at Y = 0, Y < 0 fol-
lows from (70a) and Ṽ1 (X, 0) = 0 for X < 0 which in turn is a consequence
that ΔA and thus ΔU0 is continuous at x∗ = 0.
For the numerical solution we decompose the solution of the linear elliptic
partial differential equation (71) into a particular solution and a solution of
the homogenous problem:
symbol numerical value reference

a0 0.3321 below (31c)
−1/3
A0 0.8920 a0 (32)
a1 1.343 a0 below (65)
7/4
a2 −0.301 a0 below (65)
c1 −1.644 a−1 0 (53)
c2 (0.724 − 1.644
3
ln a0 )a−1
0 (53)
−3/4
[ΔP∗ ] 5.27 a0 (59b)
−1/4
C u,log −0.168[ΔP∗ ] a0 (64)
−1/4
C u,0 (0.042 − 0.168 43 ln a0 )[ΔP∗ ]a0 (65)
Table 2. Constants and their definition
|ΔP∗ |
ΔP̃1 = − ΔPh (X, Y ) + Ā(0) (ΘB (Y ) − 1) , (73)
π
with ΔPh ∼ arctan Y /X for X 2 + Y 2 → ∞ and ΔP̃h (X, 0) = π for X < 0
and ΔPh (X, 0) = 0 for X > 0.
The local behavior near the singularity can be discussed by transforming
the equation (71) to polar coordinates R, ϕ. Expanding ΔPh ∼ ΔPh,0 (ϕ) +
O(R) for R 1 we obtain

sin ϕ ΔPh,0 − 2 cos ϕΔPh,0 = 0, ΔPh,0 (0) = 0, ΔPh,0 (π) = π, (74)
with the solution

1
ΔPh,0 = ϕ −sin 2ϕ. (75)
2
A numerical solution for ΔPh is shown in figure 5. The correct asymp-
totic behavior for R → 0 and R → ∞ could be verified.
4 Summary and Conclusions

The 2D laminar mixed convection flow past a horizontal plate has been
studied. In previous papers (Schneider (2005)) it has been shown that no
solution of the outer potential flow problem exists if the oncoming flow is
strictly horizontal in an unbounded domain. To circumvent this technical
problem two different strategies have been followed. In this paper the on-
coming flow was assumed to have a small angle of attack. In Schneider
(2005) the horizontal plate has been considered in a horizontal channel.
310 H. Steinrück
However in both approaches the wake and the potential flow have to be
considered simultaneously. It turns out that if the interaction parameter
approaches a critical value a singularity in the wake forms. The nature of
this singularity is essentially an inviscid phenomenon. The author believes
that also in the case when the plate is placed in a horizontal channel such
a singularity can occur, but it has not been investigated up to now.
To resolve the singularity in the flow field at the trailing edge a triple deck
problem has been formulated. This problem constitutes a local interaction
problem between the a sub-layer of the boundary-layer (lower deck) and the
potential flow (upper deck). Here the influence of the hydrostatic pressure
on the interaction mechanism has been studied. Surprisingly the interaction
pressure at the trailing edge turns out to have a jump discontinuity which
can be resolved be introducing additional sub-layers.
Bibliography
R. Chow and E. Melnik. Numerical solutions of triple-deck equations for
laminar trailing edge stall. Technical Report RE-526J, Grumman Re-
search Dept, 1976.
P.-Y. Lagree. Thermal mixed convection induced locally by a step change in
surface temperature in a Poiseuille flow in the framework of triple deck
theory. Int. J. Heat and Mass Transfer, 42:2509–2524, 1999.
A. F. Messiter. Boundary layer flow near the trailing edge of flat plate.
SIAM J. Appl. Math, 18:241–257, 1970.
A. Ruban. Asymptotic theory of separated flows. In H. Steinrück, editor,
Asymptotic Methods in Fluid Mechanics-Recent Advances, CISM Lecture
Notes 523, pages 311–408. Springer, 2010.
Lj. Savić and H. Steinrück. Mixed convection flow past a horizontal plate.
Theor. Appl. Mech., 32:1–19, 2005.
Lj. Savić and H. Steinrück. The trailing-edge problem for mixed-convection
flow past a horizontal plate. J. Fluid Mech., 588:309–330, 2007.
W. Schneider. Lift, thrust and heat transfer due to mixed convection flow
past a horizontal plate. J. Fluid Mech., 529:51–69, 2005.
W. Schneider. Mathematische Methoden der Strömungslehre. Vieweg, 1978.
K. Stewartson. On the flow near the trailing edge of a flat plate. Mathe-
matica, 16:106–121, 1969.
V. V. Sychev, A. I. Ruban, Vic. V. Sychev, and G. L. Korolev. Asymptotic
theory of separated flows. Cambridge University Press, 1998.
Asymptotic Theory of Separated Flows
Anatoly I. Ruban
Department of Mathematics, Imperial College London
1 Introduction
Separation is a fluid dynamic phenomenon that influences the behaviour of
a wide variety of liquid and gas flows. The difference between an attached
flow and its separated counterpart is demonstrated in Figure 1 where the
theoretical streamline pattern, given by the classical solution of the inviscid
flow theory1
a2
ϕ + iψ = V∞ z + , (1)
z
is compared with a real flow visualisation for a circular cylinder in a water
(a) Streamline pattern given by (1) (b) Visualisation of the cylin-

der flow by Taneda (1956);
Re = 26
Figure 1. Comparison of the theoretical attached and real separated flows
past a circular cylinder.
tank (Taneda, 1956). In the theoretically predicted form of the flow the
fluid particles follow closely the cylinder contour from the front stagnation
point all the way to the rear stagnation point. Contrary to that in the
experimentally observed flow the fluid particles brake away from the cylinder
1
See, for example, Lamb (1932).
© CISM, Udine 2010
312 A. Ruban
surface at a separation point and form a pair of eddies in the wake behind
the cylinder.
The irony of the situation is that the theoretical flow in Figure 1(a)
has been constructed based on the Euler equations. These are intended
for describing the motion of fluids with extremely small viscosity. While
the Euler equations admit solutions in the form of attached flows, such
flows can not be observed in practice except for some special cases. In
fact, experiments clearly indicate that the attached form of fluid motion
past rigid bodies is characteristic for relatively small values of the Reynolds
number. In particular, the flow past a circular cylinder assumes an attached
form only if the Reynolds number Re = V∞ a/ν, with a being the cylinder
radius, is smaller than a critical value of about six. The actual flow shown in
Figure 1(b) corresponds to Re = 26, and we see that the separation eddies
are already well developed. Further increase of the Reynolds number results
in an extention of the eddies, and then the flow looses its symmetry and
becomes unstable, but it never returns to an attached form (see Figure 2).
Figure 2. Visualisation of the cylinder flow by Werlé & Gallon (1972); the
Reynolds number Re = 2000.
Since many ”common” gases and liquids, such as air and water, have
extremely small viscosity, their flows are characterised by very large values of
the Reynolds number. This explains why most liquid and gas flows observed
in nature and encountered in engineering applications involve separation.
The difference between a separated flow and its theoretical unseparated
counterpart (constructed on the basis of inviscid flow analysis) concerns
not only the form of trajectories of fluid particles, but also the magnitudes
of aerodynamic forces acting on the body. For example, for bluff bodies
in an incompressible flow, it is known from experimental observations that
the drag force is never zero; furthermore, it does not approach zero as the
Asymptotic Theory of Separated Flows 313
Reynolds number becomes large. On the other hand, one of the most famous
results of the inviscid flow theory is d’Alembert’s paradox which states that
a rigid body does not experience any drag in incompressible steady flow.
This contradiction is associated with the assumption of an attached form of
the flow.
Separation imposes a considerable limitation on the operating character-
istics of aircraft wings, helicopter and turbine blades leading to a significant
degradation of their performance. It is well known that the separation is
normally accompanied by a loss of the lift force, sharp increase of the drag,
increase of the heat transfer at the reattachment point, development of flow
oscillations, etc.
It is hardly surprising that the problem of flow separation has attracted
considerable interest amongst researchers. The first theoretical model of a
separated flow was due to Helmholtz (1868) and Kirchhoff (1869) in the
framework of the classical theory of inviscid fluid flows. This model was
originally applied to the flow past a flat plate perpendicular to the free
stream when the separation is known to take place at the plate edges. The
Kirchhoff model may be, of course, applied to other body shapes. In partic-
ular, Levi-Civita (1907) used this model for the flow past a circular cylinder
(see Figure 3). A major conclusion that may be drawn from this theory is
that Euler equations allow for a family of solutions where the position of
separation point S remains a free parameter.
S
Figure 3. Separated flow past a circular cylinder (Levi-Civita, 1907).
The described non-uniqueness of the solution of the Euler equations

poses a theoretician a dilemma, how to find the location of the separation
point. Prandtl (1904) was the first to recognise that the resolution to this
problem lies with a specific behaviour of the boundary layer. According to
Prandtl’s theory at large values of the Reynolds number the main part of the
flow field may be treated as inviscid. However, for any Reynolds number,
no matter how large, there always exists a thin boundary layer developing
along the wall where the flow is viscous in nature. The behaviour of the
boundary layer depends on the pressure distribution along the wall. If the
314 A. Ruban
pressure gradient is favourable, i.e. the pressure decreases downstream,

then the boundary layer remains well attached to the wall. However with
adverse pressure gradient, when the pressure starts to rise in the direction of
the flow, the boundary layer tends to separate from the body surface. The
reason for separation may be explained as follows. Since the velocity in the
boundary layer drops towards the wall, the kinetic energy of fluid particles
inside the boundary layer appears to be less than that at the outer edge
of the boundary layer, in fact the closer a fluid particle is to the wall the
smaller appears to be its kinetic energy. This means that while the pressure
rise in the outer flow may be quite significant, the fluid particles inside the
boundary layer may not be able to get over it. Even a small increase of
pressure may cause the fluid particles near the wall to stop and then turn
back to form a reverse flow region characteristic of separated flows.
The separation point may be identified as a point of zero skin friction

∂u
τw = μ = 0. (2)
∂y y=0
Indeed with τw being positive upstream of this point, all the fluid particles
in the boundary layer move downstream along the wall (see Figure 4) and
the flow appears to be attached to the body surface. However, as soon as the
skin friction τw turns negative a layer of reversed flow (u < 0) emerges near
the wall, giving rise to a region of recirculation which, obviously, originates
from point S where condition (2) holds.
y
u
Figure 4. Boundary-layer separation.
It might seem surprising that the clear understanding of the physical

processes leading to the separation, could not be converted into a rational
mathematical theory for more than half a century. The fact is that the clas-
sical boundary-layer theory, which was intended by Prandtl for predicting
flow separation, was based on the hierarchical approach when the outer in-
viscid flow should be calculated first ignoring the existence of the boundary
layer, and only after that one can turn to the boundary layer analysis. In
the forties it became obvious that such a strategy leads to a mathematical

contradiction associated with the so called Goldstein’s singularity. The form
of the singularity was first described by Landau & Lifshitz (1944). They
demonstrated that the skin friction produced by the boundary layer on the
body surface decreases
√ as the separation is approached proportional to the
square root τw ∼ s of the distance s from the separation point. Simulta-
neously the velocity component normal√ to the body experience unbounded
growth being inversely proportional to s. This result was believed to ex-
plain why the separation changes the flow field so drastically. If the solution
of the boundary-layer equations could be continued smoothly through the
separation point then the reversed flow region would be confined within
the boundary layer, and its effect on the flow behaviour would be barely
noticeable.
Later Goldstein (1948) presented a more rigorous mathematical analysis
of the boundary-layer equations in a small vicinity of the zero skin friction
point. He not only confirmed the structure of the singularity predicted by
Landau & Lifshitz (1944), but also demonstrated (and this result appeared
to be of paramount importance for further development of the boundary-
layer separation theory) hat the singularity at the separation precludes the
solution to be continued beyond the point of zero skin friction, which sug-
gested that the entire approach to the separation problem had to be reex-
amined.
The Goldstein’s theoretical discovery came at the time when an impor-
tant development was taking place in experimental investigation of sepa-
rated flows. Most disputable was the effect of upstream influence through
the boundary layer in a supersonic flow prior to separation. It might be
observed in a number of physical situations, for example, when a shock
wave impinges the boundary layer on a rigid body surface. The effect was
extensively studied by many experimentalists during the forties and fifties2 ,
and it may be described using the following simple flow layout.
Let a flat plate be placed in a supersonic flow parallel to the free-stream
velocity V∞ as shown in Figure 5. Let further an oblique shock wave be
produced in the flow by an obstacle, say, a wedge, situated above the plate.
This shock impinges upon the boundary layer on the plate surface at point
A, and the focus of attention of the above mentioned studies was on the
behaviour of perturbations induced in the flow through the interaction be-
tween the shock wave and boundary layer.
2
A review of these efforts may be found, for example, in Chapman et al. (1956).
Early theoretical models to explain the phenomenon are reviewed in a resent paper
by Lighthill (2000).
316 A. Ruban
incident shock
V∞
A
Figure 5. Shock wave impinging upon the boundary layer.
If we attempt to describe the flow behaviour based on the classical

Prandtl’s hierarchical strategy then we have to start with the external invis-
cid flow. To the leading order we ignore the existence of the boundary layer
and recall that at a supersonic speed the governing Euler equations are hy-
perbolic; they do not allow for perturbations to propagate upstream. This
suggests that the inviscid flow should remain uniform everywhere in front
of the incident shock (see Figure 5). Now turning to the boundary layer
on the plate surface we note that with given (constant) pressure, Prandtl’s
equations governing the flow in the boundary layer are parabolic, and there-
fore the boundary layer also is incapable of conducting any perturbations
upstream of the cross-section AA .
Figure 6. Oblique shock wave interacting boundary layer. Visualisation

by Liepmann, Roshko and Dhawan (1952).
These theoretical arguments proved to fail completely in predicting the

real behaviour of the flow. The experiments invariably showed (see, for
example, Figure 6) that, unless the incident shock was very weak, the flow
separated from the plate surface some distance upstream of the incident
shock (see Figure 7). It was also established that the boundary layer was
perturbed even upstream of the separation point S, and the distance over
which the pressure perturbations were able to propagate upstream of point
S through the boundary layer proved to be significantly larger then the
boundary-layer thickness. An increase of the pressure in the boundary layer
prior to separation and, even more so, the separation of the boundary layer
cause the streamlines at the bottom of the inviscid flow region to deviate
from the wall giving rise to a secondary shock as shown in Figure 7. Together
with the primary shock they form a characteristic shock structure called the
λ-structure.
primary shock
secondary shock
Figure 7. Schematic representation of the separation provoked by imping-

ing shock, and formation of the shock λ-structure.
In order to find an explanation to this unexpected behaviour, the ex-

perimental data were carefully examined, and, in particular, it was noticed
that the boundary-layer separation process had a universal character being
solely determined by the state of the boundary layer immediately prior to
the separation. This conclusion was supported by an observation that in
a vicinity of the separation point S the flow remained unchanged when in-
stead of the impinging shock the separation was caused, for example, by the
forward facing step (see Figure 8) or other obstacle.
It was first suggested by Oswatitsch & Wieghardt (1948) that the ob-
served upstream influence through the boundary layer may be explained by
an interaction between the boundary layer and external inviscid part of the
flow. The impinging shock (Figure 7) or forward facing step (Figure 8) serve
to trigger the interaction, but once started the process proceeds very much
independently obeying its own rules. For this reason the interaction of the
318 A. Ruban
boundary layer with supersonic inviscid flow was termed by Chapman et al.
(1956) the free-interaction.
free-interaction
region
z }| {
Figure 8. Separation upstream of a forward-facing step.
The phenomenon of free-interaction may be described vaguely in the

following way. Let us suppose that for some reason the pressure at the outer
edge of the boundary layer starts to rise in the downstream direction. Since
the pressure perturbations can freely penetrate into the boundary layer, this
would lead to a deceleration of the fluid particles inside the boundary layer
and, as a consequence, to a displacement of the streamlines from the wall.
The response of the external supersonic flow to the displacement effect of the
boundary layer is such that it further increases the pressure, and the chain
of events repeats again. We shall see that this process, once initiated, is able
of maintaining a monotonic growth of the pressure leading ultimately to the
separation of the boundary layer. Asymptotic theory of this phenomenon,
which for obvious reasons is called the self-induced separation, was developed
independently by Neiland (1969) and Stewartson & Williams (1969). We
shall now turn to the mathematical description of the theory.
2 Self-Induced Separation
2.1 Formulation of the Problem
The theory that will be discussed here applies to boundary-layer sep-
aration in a wide variety of supersonic flow. However, for the purpose of
describing the theory it is convenient to choose a particular flow layout,
for example, the flow past a flat plate surface; see Figure 9. We shall as-
sume that the plate is aligned with oncoming flow which is supersonic. We
shall further assume that the boundary layer on the plate surface separates
due to some downstream disturbance. The position of the separation point
S depends on the nature of this perturbations, and in this study will be
treated as known.
In what follows we shall assume that the gas considered may be treated
as perfect. Its motion will be assumed steady and two-dimensional, in which
V∞
Figure 9. The flow layout.
case the governing Navier-Stokes equations may be written as

∂ û ∂ û ∂ p̂ ∂ 4 ∂ û 2 ∂v̂ ∂ ∂ û ∂v̂
ρ̂ û + v̂ =− + μ̂ − + μ̂ + ,
∂ x̂ ∂ ŷ ∂ x̂ ∂ x̂ 3 ∂ x̂ 3 ∂ ŷ ∂ ŷ ∂ ŷ ∂ x̂

∂v̂ ∂v̂ ∂ p̂ ∂ 4 ∂v̂ 2 ∂ û ∂ ∂ û ∂v̂
ρ̂ û + v̂ =− + μ̂ − + μ̂ + ,
∂ x̂ ∂ ŷ ∂ ŷ ∂ ŷ 3 ∂ ŷ 3 ∂ x̂ ∂ x̂ ∂ ŷ ∂ x̂

∂ ĥ ∂ ĥ ∂ p̂ ∂ p̂ 1 ∂ ∂ ĥ ∂ ∂ ĥ
ρ̂ û + v̂ = û + v̂ + μ̂ + μ̂ +
∂ x̂ ∂ ŷ ∂ x̂ ∂ ŷ P r ∂ x̂ ∂ x̂ ∂ ŷ ∂ ŷ
2
4 ∂ û 2 ∂v̂ ∂ û 4 ∂v̂ 2 ∂ û ∂v̂ ∂ û ∂v̂
+ μ̂ − + μ̂ − + μ̂ + ,
3 ∂ x̂ 3 ∂ ŷ ∂ x̂ 3 ∂ ŷ 3 ∂ x̂ ∂ ŷ ∂ ŷ ∂ x̂
∂ ρ̂û ∂ ρ̂v̂
+ = 0,
∂ x̂ ∂ ŷ
γ p̂
ĥ =
γ − 1 ρ̂
We denote the distance from the leading edge to the separation point S
by L; the velocity, density, viscosity and pressure in the unperturbed flow
upstream of the plate are denoted by V∞ , ρ∞ , μ∞ and p∞ respectively. To
study the flow we shall use Cartesian coordinate system x̂Oŷ with x̂ mea-
sured parallel to the plate from its leading edge, and ŷ in the perpendicular
direction. The velocity components in these coordinates are denoted by û
and v̂. Other quantities used in the Navier-Stokes equations are the gas
density ρ̂, pressure p̂, enthalpy ĥ and viscosity μ̂. As before, the “hat” de-
notes dimensional variables. The non-dimensional variables are introduced
as
û = V∞ u, v̂ = V∞ v, ρ̂ = ρ∞ ρ,
2 2
p̂ = p∞ + ρ∞ V∞ p, ĥ = V∞ h, μ̂ = μ∞ μ,
x̂ = Lx, ŷ = Ly.
320 A. Ruban
This allows to convert the Navier-Stokes equations into the non-dimensional

form

∂u ∂u ∂p 1 ∂ 4 ∂u 2 ∂v
ρ u +v =− + μ − +
∂x ∂y ∂x Re ∂x 3 ∂x 3 ∂y
(3a)
∂ ∂u ∂v
+ μ + ,
∂y ∂y ∂x

∂v ∂v ∂p 1 ∂ 4 ∂v 2 ∂u
ρ u +v =− + μ − +
∂x ∂y ∂y Re ∂y 3 ∂y 3 ∂x
(3b)
∂ ∂u ∂v
+ μ + ,
∂x ∂y ∂x

∂h ∂h ∂p ∂p
ρ u +v =u +v +
∂x ∂y ∂x ∂y

1 1 ∂ ∂h ∂ ∂h
+ μ + μ + (3c)
Re P r ∂x ∂x ∂y ∂y
2
4 ∂u 2 ∂v ∂u 4 ∂v 2 ∂u ∂v ∂u ∂v
+μ − +μ − +μ + ,
3 ∂x 3 ∂y ∂x 3 ∂y 3 ∂x ∂y ∂y ∂x
∂ρu ∂ρv
+ = 0, (3d)
∂x ∂y
1 1 γ p
h= 2
+ . (3e)
(γ − 1)M∞ ρ γ − 1 ρ
Here M∞ is the free-stream Mach number defined as

V∞ p∞
M∞ = , a∞ = γ .
a∞ ρ∞
We shall assume that M∞ is a finite quantity greater than one.

The asymptotic analysis of the Navier-Stokes equations will be con-
ducted assuming that the Reynolds number
ρ∞ U∞ L
Re = → ∞.
μ∞
2.2 The flow upstream of the interaction region

Before analysing the free-interaction region which occupies a small vicin-
ity of the separation point S we need to consider the boundary layer on the
plate surface upstream of the interaction, as this boundary layer plays a
role of the background on which the interaction develops. The flow out-
side the boundary layer is almost uniform with small perturbations of order
O(Re−1/2 ) caused by the presence of the boundary layer.
Asymptotic analysis of the boundary layer is based on the limit proce-
dure
x = O(1), Y = Re1/2 y = O(1), Re → ∞,
and the solution of the Navier-Stokes equations may may be sought in the
form of the asymptotic expansions
u(x, y; Re) = U0 (x, Y ) + · · · , v(x, y; Re) = Re−1/2 V0 (x, Y ) + · · · ,
ρ(x, y; Re) = ρ0 (x, Y ) + · · · , p(x, y; Re) = Re−1/2 P1 (x, Y ) + · · · , (4)
h(x, y; Re) = h0 (x, Y ) + · · · , μ(x, y; Re) = μ0 (x, Y ) + · · · .
Substitution of (4) into the Navier-Stokes (3) equations leads to the classical
boundary-layer equations

∂U0 ∂U0 ∂ ∂U0
ρ0 U0 + ρ0 V0 = μ0 , (5a)
∂x ∂Y ∂Y ∂Y
2
∂h0 ∂h0 1 ∂ ∂h0 ∂U0
ρ0 U0 + ρ0 V0 = μ0 + μ0 , (5b)
∂x ∂Y P r ∂Y ∂Y ∂Y
∂(ρ0 U0 ) ∂(ρ0 V0 )
+ = 0, (5c)
∂x ∂Y
1 1
h0 = 2
. (5d)
(γ − 1)M∞ ρ0
They should be solved with the free-stream conditions at the leading edge
of the flat plate
1
U0 = 1, h0 = 2
at x = 0, Y ∈ [0, ∞) (6)
(γ − 1)M∞
as well as at the outer edge of the boundary layer

1
U0 = 1, h0 = at Y = ∞, x ∈ [0, ∞). (7)
(γ − 1)M∞
2
On the plate surface the no-slip conditions hold

U0 = V0 = 0 at Y = 0, x ∈ [0, 1]. (8)
It should be supplemented with a thermal condition. For example, we can
assume that the wall temperature is given as a function of x, i.e.
h0 = F (x) at Y = 0, x ∈ [0, 1] (9a)
322 A. Ruban
or the wall is thermally isolated
∂h0
= 0 at Y = 0, x ∈ [0, 1]. (9b)
∂Y
Boundary-value problem (5)–(9) admits a self-similar solution for ther-
mally isolated wall (9b) as well as in the case when the wall temperature
is known to be constant, i.e. function F (x) in (9a) does not depend on x.
However, we do not need to restrict ourselves to these particular flow condi-
tions. If they are not satisfied then the boundary-value problem (5)–(9) may
be solved numerically. To proceed further we only need to know that the
solution remains smooth when the trailing edge of the plate is approached.
Therefore we shall assume that the sought functions U0 , h0 , ρ0 and μ0 may
be represented in the form of Taylor expansions
⎫
U0 (x, Y ) = U00 (Y ) + (−s)U01 (Y ) + · · · ,⎪
⎪
⎪
h (x, Y ) = h (Y ) + (−s)h (Y ) + · · · , ⎬
0 00 01
as s = x − 1 → 0− . (10)
ρ0 (x, Y ) = ρ00 (Y ) + (−s)ρ01 (Y ) + · · · , ⎪
⎪
⎪
⎭
μ0 (x, Y ) = μ00 (Y ) + (−s)μ01 (Y ) + · · ·
The leading order terms in (10) exhibit the following behaviour near the
plate surface
⎫
U00 (Y ) = λY + · · · ,⎪ ⎪
⎪
h00 (Y ) = hw + · · · , ⎬
as Y → 0, (11)
ρ00 (Y ) = ρw + · · · , ⎪
⎪
⎪
⎭
μ00 (Y ) = μw + · · ·
where λ, hw , ρw , μw are positive constants representing the dimensionless
skin friction, enthalpy, density and viscosity on the wall surface.
2.3 Inspection analysis of the interaction process

Let us suppose that for some reason a small pressure rise Δp 1 forms
at the outer edge of the boundary layer as shown in Figure 10. Let us further
suppose that it acts over a short distance with Δx 1. It is apparent that
this pressure rise will result in a deceleration of fluid particles inside the
boundary layer. To estimate the corresponding velocity variation Δu one
needs to compare the first convective term on the left hand side of the
longitudinal momentum equation (3a) with the pressure gradient
∂u ∂p
ρu ∼ . (12)
∂x ∂x
Indeed, the convective term serves to describe acceleration/deceleration of

fluid particles, and since this process is caused by pressure variations, the
balance expressed by (12) should hold. Here symbol ∼ is used to show that
the quantities in (12) are same order of magnitude.
Δp > 0
Figure 10. Viscous-inviscid interaction region.
Taking into account that the perturbations are small, and therefore the
velocity u and density ρ may be represented by their initial profiles given
by the leading order terms in (10), we have
∂u ∂p
ρ00 U00 ∼ .
∂x ∂x
Approximating the derivatives by finite differences, we have
Δu Δp
ρ00 U00 ∼ ,
Δx Δx
and it follows that
Δp
Δu ∼ . (13)
ρ00 U00
Since everywhere in the boundary layer, except near the wall, both ρ00 and
U00 are order one quantities, we can finally write
Δu ∼ Δp. (14)
Applying the same arguments to the energy equation (3c), we find
Δh ∼ Δp,
and then it follows from the state equation (3e) that
Δρ ∼ Δp. (15)
Let us now we consider a small filament in the boundary layer confined
between two neighbouring streamlines with δ being the initial distance be-
tween them, as shown in Figure 11. Using the mass conservation law, we
can write
ρ00 U00 δi = (ρ00 + Δρ)(U00 + Δu)(δi + Δδi ).
324 A. Ruban
Neglecting squares of perturbations in this equation, we can deduce that

Δδi Δρ Δu
∼ + .
δi ρ00 U00
Since both ρ00 and U00 are order one quantities and their variations are
given by (14) and (15), we can conclude that the thickness of the filament
increases by the value
Δδi ∼ δi Δp.
The integral effect of the thickening of all the filaments in boundary layer
is
Δδ = Δδi ∼ δi Δp ∼ Δp δi ∼ Re−1/2 Δp. (16)
i i i
δi + Δδi
Re−1/2
δi
Δx
Figure 11. Thickening of a stream filament in the main part of the bound-
ary layer.
The above analysis, obviously, is invalid near the bottom of the boundary
layer. Indeed, according to (11) the initial velocity U00 tends to zero as
Y → 0, and equation (13) predicts unbounded growth of the perturbation
velocity. Of course, before it happens non-linear effects take over. In a thin
sublayer near the wall, where Δu ∼ U00 , equation (13) may be written as

Δu ∼ U00 ∼ Δp. (17)
Combining (17) with the formula for U00 in (11), we can deduce that

Y ∼ Δp,
which means that the thickness of the near-wall sublayer

y = Re−1/2 Y ∼ Re−1/2 Δp. (18)
To estimate the displacement effect of the sublayer we again use the mass
conservation law. Treating the sublayer as one filament (see Figure 12) and
taking into account that along this filament Δu ∼ U00 , we have to conclude
that the variation of the filament thickness is of the same order as its initial
value given by (18), i.e.

Δδ ∼ Re−1/2 Δp. (19)
Re−1/2
y y + Δδ
Δx
Figure 12. Thickening of the sublayer.
Comparing (19) with (16), we see that for any Δp 1 the contribution
of the sublayer into the displacement effect of the boundary layer is signifi-
cantly larger than that of the main part of the boundary layer. Hence, the
slope angle θ of the streamlines at the outer edge of the boundary layer
should be estimated based on the displacement effect of the sublayer (19).
We have √
Δδ Re−1/2 Δp
θ∼ ∼ ,
Δx Δx
Using further the Ackeret formula, we find that the pressure perturbations
in the interaction region
√
Re−1/2 Δp
Δp ∼ θ ∼ . (20)
Δx
It remains to recall that the sublayer is adjacent to the wall and therefore
should be viscous,
∂u 1 ∂ ∂u
ρu ∼ μ . (21)
∂x Re ∂y ∂y
Indeed, if the flow were inviscid, then the Bernoulli equation could be used
along each streamline. Near the wall the velocity small, and therefore, we
can write this equation in the incompressible form
u2 p U2
+ = 00 . (22)
2 ρ 2
Here it is taken into account that p denotes the perturbation of the pressure
with respect to its value in the oncoming flow; everywhere upstream of the
interaction region p = 0.
326 A. Ruban
Writing equation (22) in the form
u2 U2 p
= 00 − , (23)
2 2 ρ
we can see that for any increase of pressure, no matter how small, we can
always find a streamline close enough to the wall for which the right hand
side of (23) appears to be negative. Since this is impossible, we have to
conclude that the flow in the sublayer should be viscous.
Taking into account that the density ρ and viscosity μ are order one
quantities everywhere in the boundary layer, and representing equation (21)
in the finite differences form, we have
Δu 1 Δu
U00 ∼ .
Δx Re (Δy)2
Since in the sublayer Δy ∼ y, this equation may be written as
U00 1 1
∼ . (24)
Δx Re y 2
Let us now summarise the results of the above analysis. Taking into
account that in the sublayer the velocity u is same order as initial velocity
U00 at the bottom of the boundary layer approaching the interaction region,
we will write equation (17) as

u ∼ Δp. (25)
Equation (18) for the thickness of the viscous sublayer is written as

y ∼ Re−1/2 Δp. (26)
From equation (20), representing the pressure induced in the interaction
region, it follows that
Re−1/2
Δp ∼ . (27)
Δx
In order to close this set of equations we need to add equation (24) which
we will now write as
u 1
∼ . (28)
Δx Re y 2
As a result we have four equations (25) – (28) for four unknowns, the velocity
u in the sublayer, characteristic thickness of the sublayer y, induced pressure
Δp and longitudinal extent Δx of the interaction region. These equations
may be treated as algebraic equations. They are easily solved to give
u ∼ Re−1/8 , y ∼ Re−5/8 , Δp ∼ Re−1/4 , Δx ∼ Re−3/8 . (29)
2.4 Triple-Deck Model

The inspection analysis presented in the previous section serves two pur-
poses. Firstly, it allows to obtain estimates for fluid dynamic quantities in
different parts of the flow, which we will use when constructing the corre-
sponding asymptotic expansion of the sought functions. Secondly, it reveals
physical nature of the processes involved. In this particular problem it shows
that under certain conditions the boundary layer may come into interaction
with external inviscid flow; this interaction being termed viscous-inviscid
interaction. The region of the interaction occupies an O(Re−3/8 ) vicinity of
the separation point, and has a three-tiered structure shown in Figure 13.
It is composed of the viscous sublayer (shown as region 1 ), the main part
of the boundary layer (region 2 ) and an inviscid potential flow (region 3 )
situated outside the boundary layer.
Re−3/8
3
Re −5/8
Re−1/2
1
Re−3/8
Figure 13. Three-tiered structure of the interaction region.
The characteristic thickness of the viscous sublayer is estimated as be-

ing an O(Re−5/8 ) quantity, so that it occupies an O(Re−1/8 ) portion of the
boundary layer and is comprised of the stream filaments immediately adja-
cent to the wall. The flow velocity in this region is O(Re−1/8 ) relative to the
free-stream velocity, and due to the slow motion of gas here the flow exhibits
high sensitivity to pressure variations. Even a small variation of pressure
along the wall may cause significant deceleration/acceleration of fluid parti-
cles there. As a result the flow filaments change their thickness leading to a
deformation of streamlines. This process is termed the displacement effect
of the boundary layer.
The main part of the boundary layer, the middle tier of the interac-
328 A. Ruban
tive structure, represents a continuation of the conventional boundary layer

developing along the plate. Its thickness is estimated as O(Re−1/2 ) and
the velocity is an order one quantity. The flow in this tier is significantly
less sensitive to the pressure variations. It does not produce any noticeable
contribution to the displacement effect of the boundary layer, which means
that all the streamlines in the middle tier are parallel to each other and
carry the deformation produced by the displacement effect of the viscous
sublayer.
Finally, the upper tier is situated in the potential flow region outside
the boundary layer. It serves to “convert” the perturbations in the form of
the streamlines into perturbations of pressure. These are then transmitted
through the main part of the boundary layer back to the sublayer, enhancing
the process of fluid deceleration. This process is self-sustained, and it drives
the boundary layer towards the separation.
We shall start our analysis with the viscous sublayer, region 1.
Viscous sublayer Estimates (29) suggest that asymptotic analysis of the

Navier-Stokes equations in the viscous sublayer (region 1 ) should be based
on the limit procedure, where
x−1 y
x∗ = = O(1), Y∗ = = O(1), Re → ∞. (30)
Re−3/8 Re−5/8
The solution of the Navier-Stokes equations in this region will be sought in
the form of asymptotic expansions
u(x, y; Re) = Re−1/8 U ∗ (x∗ , Y∗ ) + · · · ,

v(x, y; Re) = Re−3/8 V ∗ (x∗ , Y∗ ) + · · · ,
p(x, y; Re) = Re−1/4 P ∗ (x∗ , Y∗ ) + · · · ,
(31)
h(x, y; Re) = h∗ (x∗ , Y∗ ) + · · · ,
ρ(x, y; Re) = ρ∗ (x∗ , Y∗ ) + · · · ,
μ(x, y; Re) = μ∗ (x∗ , Y∗ ) + · · · .
The form of the asymptotic expansions for u and p directly follows from
estimates (29). In order to find an estimate for v we used the continuity
equation (3d), viz.
∂ρu ∂ρv
∼ .
∂x ∂y
Being written in finite-difference from, it gives
ρu ρv u
∼ =⇒ v∼y ∼ Re−3/8 .
Δx y Δx
As far as the enthalpy h, density ρ and viscosity μ are concerned, in normal

circumstances (when there is no extreme heating or cooling of the wall) they
remain order one functions throughout the boundary layer.
Substitution of (31) into the Navier-Stokes equations (3) leads to

∂U ∗ ∂U ∗ ∂P ∗ ∂ ∂U ∗
ρ∗ U ∗ + ρ∗ V ∗ =− + μ∗ , (32a)
∂x∗ ∂Y∗ ∂x∗ ∂Y∗ ∂Y∗
∂P ∗
= 0, (32b)
∂Y∗

∂h∗ ∂h∗ 1 ∂ ∂h∗
ρ∗ U ∗ + ρ∗ V ∗ =− μ∗ , (32c)
∂x∗ ∂Y∗ P r ∂Y∗ ∂Y∗
∂ρ∗ U ∗ ∂ρ∗ V ∗
+ = 0, (32d)
∂x∗ ∂Y∗
1 1
h∗ = 2 ρ∗
. (32e)
(γ − 1)M∞
Since the flow in the viscous sublayer is slow, it should behave as in-
compressible. To confirm this proposition, we need to consider the energy
equation (32c). It is a parabolic equation, and requires an initial condition
(x∗ → −∞), condition on the wall (Y∗ = 0) and condition at the outer edge
of the sublayer (Y∗ → ∞). We start with the initial condition. It may be
formulated by matching with the solution in the boundary layer upstream
of the interaction region. This solution may be termed the outer solution
for our purposes. According to (4), the outer asymptotic expansion for the
enthalpy has the form
h(x, y; Re) = h0 (x, Y ) + · · · (33)
In order to perform the matching, we have to re-expand (33) in terms of

the inner variables (30). We start with the longitudinal coordinate
x = 1 + Re−3/8 x∗ .
Since x − 1 is small, we can use the Taylor expansion (10) for the enthalpy.
Being substituted into (33), it gives
h(x, y; Re) = h00 (Y ) + O(Re−3/8 ). (34)
Now we recall that region 1 is significantly thinner as compared to the

boundary layer. Indeed, in the boundary layer
y = Re−1/2 Y,
330 A. Ruban
while in region 1
y = Re−5/8 Y∗ .
Comparing these formulae, we see that
Y = Re−1/8 Y∗ ,
which means that Y is small, and we can use the asymptotic formula (11)
for h00 in (33). We have
h(x, y; Re) = hw + · · · .
This is the inner re-expansion of the outer solution. If we compare it with

the asymptotic expansion (31) for the enthalpy in region 1, we can conclude
that the sought matching condition may be written as
h∗ → hw as x∗ → −∞. (35)
On the plate surface we can either prescribe the wall temperature
h∗ = hw at Y∗ = 0, (36a)
or assume the wall thermally isolated
∂h∗
=0 at Y∗ = 0. (36b)
∂Y∗
Concerning condition (36) it should be noted that even when the wall tem-
perature is not constant, its variation over a short distance occupied by the
interaction region is small in normal circumstances, and therefore, should
be neglected in the leading order approximation.
It remains to consider the outer edge of the viscous sublayer. At large
values of Y∗ the energy equation (32c) turns into the inviscid the form
∂h∗ ∂h∗
U∗ +V∗ = 0, (37)
∂x∗ ∂Y∗
which shows that the enthalpy h∗ does not chance along streamlines. All
the streamlines, except in the separation region, originate from an upstream
location, where condition (35) holds. Integrating (37) with (35), we arrive
at a conclusion that the sought boundary condition may be written as
h∗ → hw as Y∗ → ∞. (38)
The solution of the energy equation (32c), satisfying boundary conditions

(35), (36) and (38), is simply
h∗ = hw . (39)
Thus, we have demonstrated that the enthalpy h∗ is constant everywhere
inside the viscous sublayer (region 1 ). Using now the state equation (32e),
we can see that the density ρ∗ is also constant, which means that the flow
in region 1 may be treated as incompressible. We can further recall that
the viscosity μ∗ is a function of temperature only, and since h∗ is constant,
μ∗ should also be constant. To be consistent with (11), we write
ρ∗ = ρw , μ∗ = μw . (40)
Now we turn to the momentum (32a) and continuity (32d) equations.
Using (39), (40), we can write these equations as
∂U ∗ ∂U ∗ dP ∗ ∂ 2U ∗
ρw U ∗ + ρw V ∗ =− + μw , (41a)
∂x∗ ∂Y∗ dx∗ ∂Y∗2
∂U ∗ ∂V ∗
+ = 0. (41b)
∂x∗ ∂Y∗
According to (32b), the pressure does not change across the viscous sublayer.
Function P ∗ is a function of x∗ only, which is why the pressure gradient on
the right hand side of (41a) is expressed by an ordinary derivative dP ∗ /dx∗ .
Equations (41) should be solved with the no-slip condition on the plate
surface
U ∗ = V ∗ = 0 at Y∗ = 0. (42)
We also need to formulate an initial condition for U ∗ at x∗ = −∞. This is
done by matching with the solution in the boundary layer upstream of the
interaction region. According to (4), the longitudinal velocity component u
is represented in the boundary layer in the form of an asymptotic expansion
u(x, y; Re) = U0 (x, Y ) + · · · . (43)
To perform the matching, we need to re-expand (43) in terms of the inner
variables (30). We have
u(x, y; Re) = U0 (1 + Re−3/8 x∗ , Re−1/8 Y∗ ) + · · · .
Taking into account that the first argument of function U0 is close to one,
and the second is small, we can use the corresponding asymptotic formulae
(10), (11), which yields
u(x, y; Re) = Re−1/8 λY∗ + · · · . (44)
332 A. Ruban
It remains to compare (44) with the asymptotic expansion (31) for u in

region 1, and we will find that the sought initial condition is written as
U ∗ = λY∗ at x∗ = −∞. (45)
Now we shall study asymptotic behaviour of the solution of equations

(41) at the outer edge of the viscous sublayer (Y∗ → ∞). For this purpose
it is convenient to introduce the stream function Ψ∗ (x∗ , Y∗ ). Its existence
follows from the continuity equation (41b), and we can write
∂Ψ∗ ∂Ψ∗
U∗ = , V∗ =− . (46)
∂Y∗ ∂x∗
Let us try to use for the stream function the following asymptotic expansion
Ψ∗ (x∗ , Y∗ ) = A0 (x∗ )Y∗α + · · · as Y∗ → ∞. (47)
Here parameter α and function A0 (x∗ ) are expected to be found by using

the momentum equation (41a).
Substitution of (47) into (46) yields
U ∗ = αA0 (x∗ )Y∗α−1 + · · · , V ∗ = −A0 (x∗ )Y∗α + · · · . (48)
Therefore, the convective terms on the left hand side of equation (41a) and
the viscous term on the right hand side are written as
∂U ∗ dA0 2α−2
ρw U ∗ = ρw αA0 Y + ··· ,
∂x∗ dx∗ ∗
∂U ∗ dA0 2α−2
ρw V ∗ = −ρw α(α − 1)A0 Y + ··· ,
∂Y∗ dx∗ ∗
∂2U ∗
μw = α(α − 1)(α − 2)A0 Y∗α−3 + · · · .
∂Y∗2
We see that if we assume, subject to subsequent confirmation, that α > 1,
then the convective terms will dominate not only over the viscous term, but
also over the pressure gradient, which remains finite as Y∗ → ∞. We have
dA0
O(Y∗2α−2 ) : A0 = 0.
dx∗
The initial condition for this equation may be obtained by substituting
the first of formulae (48) into (45). We find

λ/α if α = 2,
A0 (−∞) =
0 if α
= 2.
Hence, a non-trivial solution exists only if α = 2, in which case A0 = 12 λ,
and (47) turns into
1
Ψ∗ (x∗ , Y∗ ) = λY 2 + · · · as Y∗ → ∞.
2 ∗
Now we shall try to find the next order term in this expansion:
1
Ψ∗ (x∗ , Y∗ ) = λY 2 + A1 (x∗ )Y∗α + · · · as Y∗ → ∞. (49)
2 ∗
In order to ensure that the second term in (49) is small as compared with
the first one, we have to assume that α < 2. Substitution of (49) into (46)
yields
U ∗ = λY∗ + αA1 (x∗ )Y∗α−1 + · · · , V ∗ = −A1 (x∗ )Y∗α + · · · .
Therefore, the convective terms on the left hand side of equation (41a) and
the viscous term on the right hand side are written as
∂U ∗ dA1 α
ρw U ∗ = ρw λα Y + ··· ,
∂x∗ dx∗ ∗
∂U ∗ dA1 α
ρw V ∗ = −ρw λ Y + ··· ,
∂Y∗ dx∗ ∗
∂ 2U ∗
μw = α(α − 1)(α − 2)A1 Y∗α−3 + · · · .
∂Y∗2
Hence, the convective terms remain dominant provided that α > 0, in which
case the momentum equation (41a) reduces at
dA1
O(Y∗α ) : (α − 1) = 0.
dx∗
Since the initial condition (45) does not contain any terms except the
one which matches with the leading order term in (49), we have to conclude
that
A1 (−∞) = 0.
We see that a non-trivial solution for A1 exists only if α = 1. Function
A1 (x∗ ) remains arbitrary in the framework of the asymptotic analysis of
equations (41). We, of course, expect that this function will be found as a
part of the solution of the problem as a whole.
Redenoting A1 (x∗ ) as A(x∗ ) renders (49) in the form
1
Ψ∗ = λY 2 + A(x∗ )Y∗ + · · · as Y∗ → ∞. (50)
2 ∗
334 A. Ruban
The physical content of function A(x∗ ) may be clarified by calculating the

the streamline slope angle
v
ϑ = arctan .
u
In the viscous sublayer the velocity components u and v are given by the
asymptotic expansions (31):
u = Re−1/8 U ∗ (x∗ , Y∗ ) + · · · , v = Re−3/8 V ∗ (x∗ , Y∗ ) + · · · . (51)
Substitution of (51) into (46) shows that at the outer edge of the viscous
sublayer
dA
U ∗ (x∗ , Y∗ ) = λY∗ + A(x∗ ) + · · · , V ∗ (x∗ , Y∗ ) = − Y∗ + · · · , (52)
dx∗
and we can see that the streamline slope angle ϑ, which is zero on the wall,
reaches at the outer edge of the viscous sublayer the following value
∗
1 dA
v −1/4 V −1/4
ϑ = arctan = Re + · · · = Re − + · · · . (53)
u U ∗ Y∗ →∞ λ dx∗
In view of this formula, function A(x∗ ) is called the displacement func-

tion.
Main part of the boundary layer Region 2, the middle tier of the
triple-deck structure (see Figure 13), is a continuation of the conventional
boundary layer developing on the plate surface before the interaction. Its
thickness is estimated as y = O(Re−1/2 ). The longitudinal extent of region
2 coincides with that of the entire interaction region and is estimated as
|x − 1| = O(Re−3/8 ). Consequently, the asymptotic analysis of the Navier-
Stokes equations (3) in the main part of the boundary layer should be based
on the limit procedure
x∗ = Re3/8 (x − 1) = O(1), Y = Re1/2 y = O(1), Re → ∞. (54)
The form of the asymptotic expansions of the fluid dynamic functions in

region 2 may be predicted by analysing the solution in the overlap region
that lies between regions 1 and 2. In particular, we found that at the outer
edge of region 1 the velocity components are given by (52). Substituting
(52) into (51), we find that in the overlap region
dA
u = Re−1/8 λY∗ + Re−1/8 A(x∗ ) + · · · , v = Re−3/8 − Y∗ + · · · .
dx∗
(55)
We now express (55) in term of variables (54) of region 2, recalling that

Y = Re−1/8 Y∗ . We find that at the bottom of the middle tier (region 2 )
dA
u = λY + Re−1/8 A(x∗ ) + · · · , v = Re−1/4 − Y + ··· . (56)
dx∗
This suggests that the solution in region 2 should be written in the form of
asymptotic expansions
u(x, y; Re) = U00 (Y ) + Re−1/8 U1 (x∗ , Y ) + · · ·

(57)
v(x, y, Re) = Re−1/4 V1 (x∗ , Y ) + · · · .
The leading order term U00 (Y ) in the expansion for u(x, y; Re) coincides
with the velocity profile (10) in the boundary layer immediately before the
interaction region. According to (11)
U00 = λY + · · · as Y → 0. (58)
It further follows from (57) that the perturbation terms U1 (x∗ , Y ) and
V1 (x∗ , Y ) in (57) satisfy the following boundary conditions at the bottom
of region 2,
⎫
U1 = A(x∗ ) + · · · , ⎬
dA as Y → 0. (59)
V1 = − Y + · · · .⎭
dx∗
By analogy with the longitudinal velocity component u in (57), we shall

seek the enthalpy h, density ρ and viscosity μ in region 2 in the form of
asymptotic expansions
h(x, y; Re) = h00 (Y ) + Re−1/8 h1 (x∗ , Y ) + · · · ,

ρ(x, y; Re) = ρ00 (Y ) + Re−1/8 ρ1 (x∗ , Y ) + · · · , (60)
−1/8
μ(x, y; Re) = μ00 (Y ) + Re μ1 (x∗ , Y ) + · · · .
Finally, taking into account that in view of (32b) the pressure perturbations
in region 2 should be same order as in region 1, we write
p(x, y; Re) = Re−1/4 P1 (x∗ , Y ) + · · · . (61)
Substitution of (57), (60) and (61) into the Navier-Stokes equations (3)
336 A. Ruban
results in
∂ U1
U00 (Y ) + V1 U00 (Y ) = 0, (62a)
∂x∗
∂ P1
= 0, (62b)
∂Y
∂ h1
U00 (Y ) + V1 h00 (Y ) = 0, (62c)
∂x∗
∂ U1 ∂ ρ1 ∂ V1
ρ00 (Y ) + U00 (Y ) + ρ00 (Y ) + V1 ρ00 (Y ) = 0, (62d)
∂x∗ ∂x∗ ∂Y
1 1 1 ρ1
h00 = , h1 = − . (62e)
2
(γ − 1)M∞ ρ00 (γ − 1)M∞ ρ200
2
These equations are easily solved using the following elimination process.
We first substitute (62e) into the energy equation (62c). This leads to
∂ ρ1
U00 (Y ) + V1 ρ00 (Y ) = 0,
∂x∗
showing that the continuity equation (62d) may be written as
∂ U1 ∂ V1
+ = 0. (63)
∂x∗ ∂Y
Now, using (63), we can eliminate ∂ U1 /∂x∗ from the longitudinal mo-
mentum equation (62a). This results in
∂ V1
U00 (Y ) − V1 U00 (Y ) = 0. (64)
∂Y
2
Dividing both terms in (64) by U00 , we have
1 ∂ V1 U
− V1 00
2 = 0,
U00 (Y ) ∂Y U00
or equivalently

∂ V1
= 0.
∂Y U00
We see that the ratio V1 /U00 is a function of x∗ only, say G(x∗ ), i.e.
V1
= G(x∗ ).
U00
This function may be found by making use of the behaviour of V1 and U00
at the bottom of region 2, as given by (58), (59). We find that
1 dA
G(x∗ ) = − .
λ dx∗
Hence, everywhere across region 2
V1 1 dA
=− .
U00 λ dx∗
Let us now return to the asymptotic expansions (55) of the velocity
components, and calculate the streamline slope angle in region 2 :

v V1 1 dA
ϑ = arctan = Re−1/4 + · · · = Re−1/4 − + ··· . (65)
u U00 λ dx∗
We see that ϑ does not depend on Y , i.e. stays unchanged across the
main part of the boundary layer. This confirms an important result of
the inspection analysis that the displacement effect of the main part of the
boundary layer is relatively small, and may be neglected to the leading
order. The streamline slope angle (53), produced by the viscous sublayer,
is simply transported by the main part of the boundary layer towards the
bottom of the the upper tier of the triple-deck structure, shown as region 3
in Figure 13
Interaction law The flow in region 3 is governed by the linearised equa-

tions of inviscid gas motion. The solution of these equation results in the
following formula for the pressure at the ”bottom” of region 3
2 ϑ
p̂ = p∞ + ρ∞ V∞ (66)
2 −1
M∞
known as the Ackeret formula (see, for example, Siebert (1948)). Since
the pressure does not change across regions 2 and 1, we can apply formula
(66) directly to the pressure in the viscous sublayer. Converting (66) into
non-dimensional form and using (65), we have
1 dA
P∗ = − . (67)
2
λ M∞ − 1 dx∗
This equation establishes a relationship between the the displacement effect
of the boundary layer and the pressure induced in the inviscid from. This
is why it is referred to as the interaction law.
338 A. Ruban
Canonical form of the interaction problem We can now collect to-

gether all the equations governing the interaction process. We fond that
in order to describe the flow in the viscous sublayer we have to solve the
momentum and continuity equations (41)
∂U ∗ ∂U ∗ dP ∗ ∂ 2U ∗
ρw U ∗ + ρw V ∗ =− + μw , (68)
∂x∗ ∂Y∗ dx∗ ∂Y∗2
∂U ∗ ∂V ∗
+ = 0, (69)
∂x∗ ∂Y∗
subject to the no-slip conditions (42)
U∗ = V ∗ = 0 at Y∗ = 0, (70)
and the initial condition (45)
U ∗ = λY∗ at x∗ = −∞. (71)
Unlike in classical Prandtl’s formulation, the pressure gradient in (68) is not

known in advance. Instead the interaction law (67) should be used
1 dA
P∗ = − . (72)
λ M∞2 − 1 dx∗
It relates the induces pressure P ∗ to the displacement function A(x∗ ). This

function may be calculated with the help of the first of equations (52)
U ∗ (x∗ , Y∗ ) = λY∗ + A(x∗ ) + · · · . (73)
Affine transformations
−1/4 −1/2 1/4 −1/2 1/4 −1/2

μw ρw μw ρw μw ρw
x∗ = X̄, Y∗ = Ȳ , U∗ = Ū ,
λ5/4 β 3/4 λ3/4 β 1/4 λ−1/4 β 1/4
3/4 −1/2 1/4 −1/2 1/2 −1/2
μw ρw μw ρw μw ρw
V∗ = V̄ , A= Ā, P∗ = P̄ ,
λ−3/4 β −1/4 λ−1/4 β 1/4 λ−1/2 β 1/2

where β = 2 − 1, allow to represent the interaction problem (68)–(73)
M∞
in the following canonical form.
⎫
∂ Ū ∂ Ū dP̄ ∂ 2 Ū ⎪
⎪
Ū + V̄ =− + , ⎪
⎪
∂ X̄ ∂ Ȳ dX̄ ∂ Ȳ 2 ⎪
⎪
⎪
⎪
∂ Ū ∂ V̄ ⎪
⎪
+ = 0, ⎪
⎪
∂ X̄ ∂ Ȳ ⎪
⎪
⎪
⎪
⎬
dĀ (74)
P̄ = − , ⎪
⎪
dX̄ ⎪
⎪
⎪
⎪
Ū = V̄ = 0 at Ȳ = 0, ⎪
⎪
⎪
⎪
⎪
Ū = Ȳ + · · · as X̄ → −∞,⎪
⎪
⎪
⎪
⎪
⎭
Ū = Ȳ + Ā(X̄) + · · · as Ȳ → ∞.
Numerical results Solution of the interaction problem (74) requires spe-

cial numerical techniques. These will not be discussed here; an interested
reader is referred to Chapter 7 of the monograph by Sychev et al. (1998).
The calculation results are displayed in Figure 14, where the pressure distri-
bution along the interaction region is shown together with the skin friction.
The latter is defined as
∂ Ū
τ= .
∂ Ȳ Ȳ =0
1
1.5
1 0.5
P̄
τ
0.5
0
0 −0.2
−12 −7 −2 3 8 13 −12 −7 −2 3 8 13
X̄ X̄
(a) Pressure distribution (b) Skin friction
Figure 14. Results of the numerical solution of the interaction problem
(74).
We see that far upstream, where the interaction region matches with
the unperturbed boundary layer, the skin friction τ = 1 and the pressure
perturbation function P̄ = 0. As the pressure starts to rise in the inter-
action region, it cause the flow in the viscous sublayer to decelerate. This
340 A. Ruban
is revealed by the observed decrease of the skin friction. The skin friction
decreases crossing zero at the separation point. It should be noted that the
interaction problem (74) is invariant with respect to arbitrary shift in the
direction parallel to the body surface,
X̄ −→ X̄ + C,
with C being an arbitrary constant. Keeping this in mind, we have chosen

the separation point in Figure 14 to be at X̄ = 0.
While the pressure continues to grow monotonically, the skin friction
reaches a minimum and then starts to rise slowly. While it remains negative,
the fluid near the wall is moving in the direction opposite to the rest of the
flow. Further downstream |τ | → 0, which means that the fluid slows down,
and the pressure develops a ‘plateau’ which is clearly seen in Figure 14. At
the separation point the pressure was found to be
P̄separation = 1.046,
and in the plateau region
P̄plateau = 1.723 (75)
Exercise 1. Using (75) and the Ackeret formula (66), find the shape of
the boundary between the separation region and the main flow above it.
3 Marginal Separation Theory

The Marginal Separation theory was developed independently by Ruban
(1981, 1982) and Stewartson et al. (1982) who originally applied it to the
boundary-layer separation at the leading edge of a thin aerofoil. Later it
became clear that the theory describes a variety of subsonic and supersonic
flows with a small separation “bubble” forming on a smooth part of the
body surface. In this presentation we shall follow the original papers by
Ruban (1981, 1982) and Stewartson et al. (1982), and use as an example the
leading edge separation. It was first observed Jones (1934) in wind tunnel
and real flight experiments. Since then many researchers were involved in
experimental study of the flow round the leading edge of an aerofoil. A
comprehensive account of these efforts may be found, for example, in the
review by Tani (1964).
M
S
(a)
S
(b)
S
(c)
Figure 15. The flow past a thin aerofoil: (a) with a short separation bubble;
(b) with an extended separation region that forms after the bubble bursting;
(c) with a long separation bubble.
Experiments show that for thick aerofoils (with thickness to cord ratio
larger than 15%) the boundary layer first separates near the trailing edge.
For thin aerofoils (with thickness to cord ratio larger than 12%) this is the
leading edge separation that causes the stall of the aerodynamic character-
istics. For the stall to take place the angle of attack α should exceed a
critical value αc . When α is small, the flow over the aerofoil remains fully
attached, and the pressure has its maximum at the front stagnation point
S; see Fig. 15(a). As one moves from this point around the aerofoil nose,
the pressure first drops dramatically reaching a minimum at some point M
on the upper side of the aerofoil, and then starts to recover, so that down-
stream of point M the boundary layer finds itself under an adverse pressure
gradient. Its magnitude increases with increase of the angle of attack, and
342 A. Ruban
the boundary layer separates at α = αs .

When it happens, one can observe the appearance of a closed region of
recirculating flow on the upper surface of the aerofoil (Fig. 15a). This region
is referred to as a short bubble. Its length does not exceed 1% of the aerofoil
chord, and therefore, has an extremely weak influence on the flow field and
the values of the aerodynamic forces produced by the aerofoil. However,
the short separation bubble only exists within an interval α ∈ (αs , αc ), and
when the angle of attack increases beyond αc , the bubble suddenly bursts.
As a result, a new flow regime is formed with an extended separation region
which covers the entire upper surface of the aerofoil (Fig. 15b) or, at least,
a significant part of it (Fig. 15c). In either case the flow transformation
is accompanied by an abrupt decrease in the lift produced by the aerofoil
and a significant increase in the drag. This phenomenon is known as the
leading edge stall. If encountered in a real flight, the consequences are most
likely to be catastrophic, which explains why the problem has been receiving
undiminishing attention of aeronautic engineers.
We shall now proceed to the theoretical analysis of the described phe-
nomenon.
3.1 Statement of the Problem. Inviscid-Flow Region

Let an aerofoil, placed in a two-dimensional flow of incompressible fluid,
be of a relative thickness ε, so that its equation may be written as

εY+ (x ) for the upper surface,

y = (76)
εY− (x ) for the lower surface.
Here we use Cartesian coordinate system O x y with the origin O placed

at the leading edge aerofoil and the x -axis directed tangentially to the
middle line of the aerofoil at its leading edge (shown as the dashed line in
Figure 16). In what follows we shall use dimensionless variables; distances
along the x and y axes are referred to the aerofoil chord c, the velocity
components to the free-stream velocity V∞ , and the increment in pressure
2
(relative to its value at infinity) to ρV∞ . In this study the fluid density ρ is
assumed constant all over the flow field.
We shall further assume that the nose of the aerofoil is parabolic, in
which case the scaling parameter ε in (76) may be chosen such that
√
Y± (x ) = ± 2x + · · · as x → 0.
With this choice the radius of the curvature of the leading edge appears to
be r = c ε2 .
y
1 x
V∞
O
α
Figure 16. Incompressible flow past a thin aerofoil.
In what follows the flow past the aerofoil will be investigated using the
asymptotic analysis of the Navier-Stokes equations with a limit procedure,
where the Reynolds number, based on the radius of the leading edge of the
profile, tends to infinity, and the thickness of the aerofoil tends to zero, i.e.
V∞ r
Re = → ∞, ε → 0.
ν
Here ν is the kinematic viscosity of the fluid.

Following the routine of the method of matched asymptotic expansions,
we have to consider first the main inviscid-flow region whose dimensions are
comparable with the aerofoil chord: x = O(1), y = O(1). If the angle of
attack α is an order ε quantity, that is it can be expressed as
α = εα∗ ,
with α∗ = O(1), then the flow in this region can be described by the classical
thin aerofoil theory. The asymptotic expansions for the velocity components
(u, c) and pressure p are written as.
u = 1 + ε u1 (x, y) + · · · , v = ε v1 (x, y) + · · · ,
p = ε p1 (x, y) + · · · .
(77)
Substituting (77) into the Navier-Stokes equations, and assuming that ε → 0
and Re → ∞, one can find that the pressure p1 satisfies the Laplace equation
∂ 2 p1 ∂ 2 p1
2
+ = 0.
∂x ∂y 2
The solution of this equation, satisfying the impermeability condition of the

aerofoil surface and Kutta condition at the trailing edge of the aerofoil, is
344 A. Ruban
written as
1
1 Y+ (ζ) − Y− (ζ)
p1 (x, 0±) = dζ±
2π ζ −x
0
(78)
1

1−x 1 ζ Y+ (ζ) + Y− (ζ)
± − α∗ + dζ .
x 2π 1−ζ ζ −x
0
Here “plus” corresponds to the upper surface of the aerofoil and “minus”
to the lower.
The asymptotic expansion for the pressure in (77) shows that the pres-
sure perturbations in the main inviscid-flow region are too small to the
cause the boundary-layer separation. However, the solution in this region
develops a singularity near the leading edge. Indeed, it follows from (78)
that
k
p1 (x, 0±) = ± √ ,
2x
where
1
√ 1 G(x ) 1 dY+ dY−
k = 2 α∗ + dx , G=− + . (79)
π x (1 − x ) 2 dx dx
0
Consequently, the inviscid flow should be reexamined a region, where
x = ε2 X , y = ε2 Y ,
with X and Y being order one quantities. In this new region (in what
follows we shall call it region
√ 1) the aerofoil contour is represented by the
infinite parabola Y = ± 2X ; see Figure 17. The tangential component of
the velocity vector on the surface of the parabola is given by
Y+k
Ue = √ . (80)
Y 2 + 1
Here Y is the distance from the point on the surface of the parabola where
Ue is calculated to the X -axis; parameter k measures the degree of non-
symmetry of the flow. It is related to the angle of attack and aerofoil shape
through equation (79). It is easily seen from (80) that at the stagnation
point Y = −k. Differentiating (80) and setting the derivative to zero, one
can find that the maximum of the velocity is achieved at the point where
Y = 1/k.
Y
y
x
M 1
k
X
O
k
O
Figure 17. The flow near the leading edge of a thin aerofoil.
3.2 Boundary Layer

When analysing the fluid motion in the boundary layer, that forms on
the aerofoil surface near its leading edge, it is convenient to use the “body-
fitted” coordinates. These are curvilinear orthogonal coordinate (x, y) with
x measured along the the aerofoil surface from the front stagnation point
O, as shown in Figure 17, and y in the direction normal to the aerofoil
contour. We shall denote the velocity components in these coordinates as
Vτ and Vn , respectively, and the stream function as ψ. All the variables are
assumed dimensionless. We take the radius r = ε2 c of the leading edge of
the aerofoil as the unit of length; the velocity components are referred to
V∞ , the stream function to rV∞ and the pressure increment with respect
2
to its value in the free-stream, p∞ is referred to ρV∞ . The Navier-Stokes
equations written in these coordinates have the form

Vτ ∂Vτ ∂Vτ κVτ Vn 1 ∂p 1 1 ∂ 1 ∂Vτ
+ Vn + =− + +
H1 ∂x ∂y H1 H1 ∂x Re H1 ∂x H1 ∂x
(81a)
∂ 2 Vτ ∂ Vτ κ ∂Vn 1 ∂ κVn
+ +κ + 2 + ,
∂y 2 ∂y H1 H1 ∂x H1 ∂x H1

Vτ ∂Vn ∂Vn κVτ2 ∂p 1 1 ∂ 1 ∂Vn
+ Vn − =− + +
H1 ∂x ∂y H1 ∂y Re H1 ∂x H1 ∂x
(81b)
∂ 2 Vn ∂ Vn κ ∂Vτ 1 ∂ κVτ
+ + κ − − ,
∂y 2 ∂y H1 H12 ∂x H1 ∂x H1
1 ∂Vτ ∂Vn κVn
+ + = 0. (81c)
H1 ∂x ∂y H1
346 A. Ruban
Here κ is the local curvature of the body contour, and H1 is the Lamé
coefficient,
H1 = 1 + κ(x)y.
The velocity components are related to the stream function through the
equations
∂ψ ∂ψ
= −H1 Vn , = Vτ . (82)
∂x ∂y
In the boundary layer the asymptotic expansion of the stream function
has the form
ψ = Re−1/2 Ψ(x, Y ) + · · · , with y = Re−1/2 Y. (83)
Substitution of (83) into (82) and then into the Navier-Stokes equations
(81), results in the classical boundary-layer equation
∂Ψ ∂ 2 Ψ ∂Ψ ∂ 2 Ψ dpe ∂ 3Ψ
− 2
=− + . (84a)
∂Y ∂x∂Y ∂x ∂Y dx ∂Y 3
It should be solved with the initial condition at the stagnation point O,
Ψ=0 at x = 0, (84b)
no-slip conditions on the aerofoil surface,
∂Ψ
Ψ= = 0 at Y = 0, (84c)
∂Y
and the matching condition with the solution in the inviscid flow region,
∂Ψ
= Ue (x) at Y = ∞. (84d)
∂Y
The pressure gradient dpe /dx on the right hand side of equation (84a)
does not depend on Y , and may be calculated with the help of the Bernoulli
equation,
pe + 12 Ue2 = 12 . (85)
Differentiating (85), we have
dpe dUe
= −Ue .
dx dx
Remind that the velocity at the outer edge of the boundary layer, Ue (x),
is given by equation (80). It is zero at the front stagnation point O, but
rapidly grows as the fluid flows around√ the leading edge of the aerofoil. It
then reaches a maximum value of 1 + k 2 at point M on the upper side
of the aerofoil, for which Y = 1/k; see Figure 17. Downstream from this
point Ue (x) shows monotonic decay, and tends to unity as x → ∞.
The results of numerical solution of problem (84) are shown in Figures 18
in the form of the skin friction distribution along the aerofoil surface. The
skin friction is calculated as

∂ 2 Ψ
τ= . (86)
∂Y 2 Y =0
It appears that there exists a critical value of parameter k = k0 = 1.1575.
If k > k0 , then the solution terminates at a finite position where the skin
friction turns zero and the Goldstein (1948) singularity develops in the so-
lution. The larger the parameter k, the earlier on the aerofoil surface the
singularity is encountered.

k = 1.3

k = 1.15
τ × 100
τ

k = k0

x x
a) Skin friction distribution on b) Local behaviour of the solu-
the aerofoil nose for k = 1.15 tion near point x = x0 ; the
and k = 1.3 graphs of
Figure 18. Results of the numerical solution of problem (84).
If, on the other hand, k < k0 , then the solution exists for all values
of x. Interestingly enough, in this case the skin friction develops a mini-
mum, which tends to zero as k → k0 − 0. We denote the coordinate of the
point where τ first becomes zero by x0 . The calculations show that for the
boundary layer on the parabola surface x0 = 8.265.
The solution ahead of the point of zero skin friction According to

the classical viewpoint dating back to the original work of Prandtl (1904),
348 A. Ruban
the position of zero skin friction on the body surface gives the position of flow
separation. It is therefore of considerable interest to study the behaviour of
the solution of the problem (84) in the vicinity of this point.
We shall denote the point of zero skin friction as xs . For k > k0 it
lies upstream of x0 , and tends to x0 as k → k0 + 0. We first consider the
region lying before the line x = xs ; our task will be to find an asymptotic
expansion of the stream function Ψ as x → xs −0. At any point (x, Y ) in the
region upstream of x = xs , the longitudinal velocity component u = ∂Ψ/∂Y
is positive. Consequently, the boundary-layer equation (84a) possesses the
standard properties of equations of parabolic type. Its solution Ψ(x, Y ) near
the line x = xs depends upon the velocity distribution Ue (x) at the outer
edge of the boundary layer throughout the whole range of values of x from
the stagnation point O to the point x = xs of zero skin friction. On the
other hand, the asymptotic procedure utilised below is restricted to analysis
of a small vicinity of the line x = xs only, and therefore, does not take into
account all the boundary conditions affecting the solution for Ψ(x, Y ). This
means that the sought asymptotic expansion of Ψ(x, Y ) must be based on
the eigensolutions of the local problem. The coefficients multiplying the
eigenfunctions remain arbitrary from the viewpoint of the local analysis.
At the same time, they can be determined uniquely if the solution of the
problem (84) is constructed in the entire region x ∈ [0, xs ).
We start by noticing that the pressure gradient dpe /dx is a smooth
function. Near point x = xs it may be represented in the form of Taylor
expansion
dpe
= λ0 + λ1 s + · · · as s → ∞. (87)
dx
Here s = x − xs is the distance from the point x = xs . The leading order
term, λ0 in (87) coincides with the pressure gradient at point x = xs . For
all k ≤ k0 , when the zero skin point exists, λ0 > 0.
Setting Y = 0 in (84a) and using the no-slip conditions (84c), it is easy
to find that at any point on the aerofoil surface
∂ 3Ψ dpe
= . (88)
∂Y 3 dx
At the point of zero skin friction, in addition to the no-slip conditions (84a)
we also know that
∂ 2 Ψ
τ (xs ) = = 0. (89)
∂Y 2 Y =0
Substituting (87) into (88), and integrating the resulting equation with
(84a) and (89), we find that the leading order term of the asymptotic rep-
resentation of Ψ(x, Y ) near the point x = xs may be written as
Ψ = 16 λ0 Y 3 + · · · . (90)
We can now determine the thickness of the viscous flow region that forms
inside the boundary layer upstream of the point x = xs . In this region the
viscous term on the right hand side of equation (84a) should be comparable
with either term on the left hand side of (84a). Using, for example, the first
of these, we can write
∂Ψ ∂ 2 Ψ ∂ 3Ψ
∼ . (91)
∂Y ∂x∂Y ∂Y 3
It follows from (90) that the coefficient ∂Ψ/∂Y in the convective term on
the left hand side of (91) may be estimated as an order O(Y 2 ) quantity,
which allows to express (91) in a more simple form,
∂2Ψ ∂ 3Ψ
Y2 ∼ . (92)
∂x∂Y ∂Y 3
Approximating the derivatives in (92) by finite differences, we arrive at a
simple algebraic equation
Ψ Ψ
Y2 ∼ 3
(x − xs )Y Y
which, being solved for Y , shows that the thickness of the viscous region
decreases, as the point x = xs is approached, according to the law
! "
Y = O (−s)1/4 . (93)
Guided by (90) and (93), we seek the asymptotic expansion of Ψ(x, Y )

in the viscous region in the form
Ψ(x, Y ) = (−s)3/4 16 λ0 η 3 +(−s)α f1 (η)+(−s)2α−3/4 f2 (η)+· · ·

as s → −0.
(94)
The first term in (94) is obtained by simply expressing (90) in terms of a
new independent variable
Y
η= , (95)
(−s)1/4
which, according to (93) remains an order one quantity in the viscous region.
The second term represents an eigenfunction with an unknown eigenvalue
α. The third term is a forced term, arising in the expansion (94) due to the
nonlinearity of the boundary-layer equation (84a).
350 A. Ruban
Substituting (94) together with (95) into (84a), and setting s → −0, we
find that function f1 (η) satisfies the following ordinary differential equation

1 1 1 2
f1 − λ0 η 3 f1 + λ0 α + η f1 − λ0 αηf1 = 0. (96)
8 2 4
As the region considered adjoins the aerofoil surface, we pose the no-slip
conditions
f1 (0) = f1 (0) = 0, (97)
which are deduced by substituting (94), (95) into (84c).
Equation (96) is linear and homogeneous. Three complementary solu-
tions to this equation, f11 (η), f12 (η) and f13 (η), can be chosen such that

f11 (0) = 1, f11 (0) = 0, f11 (0) = 0,

f12 (0) = 0, f12 (0) = 1, f12 (0) = 0, (98)

f13 (0) = 0, f13 (0) = 0, f13 (0) = 1.
The first two solutions do not satisfy the boundary conditions (97) and
must be rejected. As for the third solution, it can easily be verified by
direct substitution into (96) to be simply f13 = 12 η 2 . Hence, a non-trivial
solution of equation (96) with boundary conditions (97) exists for all α, and
may be written in the form
f1 (η) = 12 a0 η 2 , (99)
where a0 is an arbitrary constant.

In order to determine the eigenvalue α, it is necessary to consider the
third term of the expansion (94). It satisfies the equation

1 3 1 2 3 1
f2 − λ0 η f2 + λ0 α− η f2 − λ0 2α− ηf2 = (1 − 2α)a20 η 2 , (100)
8 4 4 4
which also has to be solved with the no-slip conditions
f2 (0) = f2 (0) = 0. (101)
Equation (100) is not homogeneous; in addition to three complementary

solutions, f21 (η), f22 (η) and f22 (η), of the homogeneous part of the equa-
tion, it requires a particular integral, f2p (η). We can choose, for example,
a20
f2p (η) = η,
2λ0
and then the general solution to (100) is written as

a20
f2 (η) = C1 f21 (η) + C2 f22 (η) + C3 f23 (η) + η. (102)
2λ0
Here C1 , C2 and C3 are arbitrary constants. Choosing again the comple-
mentary solutions according to rule

f21 (0) = 1, f21 (0) = 0, f21 (0) = 0,

f22 (0) = 0, f22 (0) = 1, f22 (0) = 0,

f23 (0) = 0, f23 (0) = 0, f23 (0) = 1,
and applying the no-slip conditions (101), we find

a20
C1 = 0, C2 = − .
2λ0
Taking further into account that f23 (η) = 12 η 2 , we can express (102) in the
form
a2
f2 (η) = 0 (η − f22 ) + 12 b0 η 2 , (103)
2λ0
with constant b0 = C3 remaining arbitrary.
Our task now will be to determine the function f22 (η). It satisfies the
homogenous version of equation (100), namely,

1 3 1 2 3
f22 = λ0 η f22 − λ0 α − η f22 + λ0 2α − ηf22 , (104)
8 4 4
which should be solved with the initial conditions

f22 (0) = 0, f22 (0) = 1, f22 (0) = 0. (105)
If one wants to find f22 (η) in the form of the power series, then in view of
(105) the first term should be written as
f22 (η) = η + · · · . (106)
Using (106) on the right hand side of (104), we find

1 2
f22 = λ0 α − η + ··· .
2
Integration of this equation with conditions (105) yields the second term in
(106),
λ0
f22 (η) = η + (2α − 1)η 5 + · · · .
5!
352 A. Ruban
This procedure can be repeated, leading to a conclusion that the power

series for f22 (η) has the form
∞

f22 (η) = cn η 4n+1 , (107)
n=0
where c0 = 1.
A recurrent equation for the coefficients cn of the series is obtained by
substituting cn η 4n+1 into the right hand side of the equation (104) and
cn+1 η 4n+5 into its left hand side. We find that
! "
λ0 n − (2α − 1) n − 14
cn+1 = cn .
32 n + 34 n + 54 (n + 1)
Using the method of mathematical induction, it may be shown that

n
λ0 (1 − 2α)n
cn = − .
32 (5/4)n (4n − 1) n!
Here (a)n denotes a quantity defined as
(a)n = a(a + 1)(a + 2) . . . (a + n − 1), (a)0 = 1. (108)
It may be expressed in terms of the Euler Gamma function3 ,
Γ(a + n)
(a)n = .
Γ(a)
Therefore,
n
Γ(5/4) λ0 Γ(n + 1 − 2α)
cn = − . (109)
Γ(1 − 2α) 32 Γ(n + 5/4) (4n − 1) n!
It remains to substitute (109) back into (107), and we will have

∞ n
Γ(5/4) λ0 Γ(n + 1 − 2α)
f22 (η) = − η 4n+1 . (110)
Γ(1 − 2α) n=0 32 Γ(n + 5/4) (4n − 1) n!
Now our task will be to determine the asymptotic behaviour of f22 (η)
as η → ∞. For this purpose we shall express f22 (η) through the Kummer’s
3
Here we use a well known property of the Gamma function, zΓ(z) = Γ(z + 1).
function M (a, b, z), whose properties are well known4 . Remind that the
Kummer’s function is a solution of the confluent hypergeometric equation
d2 w dw
z 2
+ (b − z) − aw = 0,
dz dz
which remains regular at point z = 0. In fact, it may be represented by the
Taylor series
a (a)2 2 (a)n n
M (a, b, z) = 1 + z + z + ···+ z + ··· =
b (b)2 2! (b)n n!
∞ (111)
Γ(b) Γ(a + n) n
= z
Γ(a) n=0 Γ(b + n) n!
convergent at any finite point z in the complex plane.

It may be easily verified that the series (111) for the Kummer function
M (a, b, z) may be converted into the series (110) for function f22 (η) with
the help of the following integral transformation,
η
2 5 λ0
f22 (η) = η − η ξ −2 M 1 − 2α, , ξ 4 − 1 dξ. (112)
4 32
0
In general case the Kummer’s function grows exponentially
Γ(b) z a−b
M (a, b, z) = e z + ··· (113)
Γ(a)
as z tends to infinity along a ray that lies in the right half (z > 0) of the
complex plane z. Through (112) this makes function f22 (η) also to grow
exponentially as η → ∞. However, if
a = −m, m = 0, 1, 2 . . . , (114)
then Γ(a) = 0, and formula (113) cannot be used. In this case (a)m+1
and all the subsequent members of the sequence (108) vanish, reducing the
Taylor expansion (111) to a polynomial of degree m.
It is important to notice that regardless of the choice of parameter a, the
solution (94), (99), (103), (112) for the viscous region does not satisfy the
boundary condition (84d) at the outer edge of the boundary layer. There-
fore, in addition to viscous region (which is shown as region 2a in Figure 19),
4
See, for example, Abramowitz & Stegun (1965).
354 A. Ruban
Y
1
2c Re−1/2
2b
2d
2 2a x
xs
Figure 19. Asymptotic regions’ layout near the point of zero skin friction.
it is necessary to consider the main part of the boundary layer (region 2b),
where the asymptotic analysis of the boundary-layer equations (84) is based
on the limit
Y = O(1), s = x − xs → 0 − .
Since the boundary layer, we are dealing with, is exposed to a finite
pressure gradient, the fluid velocity in region 2b should remain finite. It
cannot be matched with the exponentially growing solution in region 2a.
Hence, an acceptable solution for region 2a may only be obtained under the
condition (114). Setting a = 1 − 2α in (114), leads to a conclusion that the
sought eigenvalues are
m+1
α= , m = 0, 1, 2 . . . . (115)
2
Goldstein’s singularity Let us now return to the expansion (94). As any

other asymptotic expansion, it should obey the rule that each subsequent
term in (94) should be smaller than the previous one. This requirement is
satisfied if α > 3/4. Consequently, the first eigenvalue is
α = 1.
The coefficient a0 in the eigenfunction (99) depends on the distribution of

the velocity Ue (x) at the outer edge of the boundary layer on the entire
interval x ∈ [0, xs ]. We shall see that through special adjustment of Ue (x)
the coefficient a0 may be made zero. However, in general case a0
= 0, and
the solution (94), (99), (103), (112) in region 2a assumes the form
Ψ(x, Y ) = (−s)3/4 61 λ0 η3 + (−s)f1 (η) + (−s)5/4 f2 (η) + · · · ,

a20 5 (116)
f1 (η) = 12 a0 η 2 , f2 (η) = 12 b0 η 2 − η .
240
In addition to the terms shown, the expansion (116) contains the sum of
an infinite number of successive eigenfunctions, and also includes additional
terms produced by the higher order terms in the expansion of the pressure
gradient (87). All of these, however, are small as compared to (−s)5/4 f2 (η)
and can be disregarded.
Let us now turn to region 2b, where Y = O(1); see Figure 19. The form
of the asymptotic expansion of the stream function Ψ(x, Y ) in this region
may be determined with the help of the following standard procedure based
on the principle of matched asymptotic expansions. We perform the change
of variables η = Y /(−s)1/4 in (116) and, assuming Y = O(1), collect terms
of the same order as s → 0−. We find

1 a2 1 ! "
Ψ(x, Y ) = λ0 Y 3 − 0 Y 5 + · · · + (−s)1/2 a0 Y 2 + · · · + O (−s)3/4 .
6 240 2
This suggests that the solution in region 2b should be sought in the form
Ψ(x, Y ) = Ψ00 (Y ) + (−s)1/2 Ψ01 (Y ) + · · · as s → 0−, (117)
where functions Ψ00 (Y ) and Ψ01 (Y ) are such that

⎫
1 a2
Ψ00 (Y ) = λ0 Y 3 − 0 Y 5 + · · · ,⎪
⎬
6 240 as Y → 0. (118)
1 2 ⎪
⎭
Ψ01 (Y ) = a0 Y + · · ·
2
With (117) the four terms in the boundary-layer equation (84a) are
calculated as
∂Ψ ∂ 2 Ψ dpe
= − 21 (−s)−1/2 Ψ00 Ψ01 + · · · , = λ0 + · · · ,
∂Y ∂x∂Y dx
∂Ψ ∂ 2 Ψ ∂ 3Ψ
= − 21 (−s)−1/2 Ψ01 Ψ00 + · · · , = Ψ
00 + · · · .
∂x ∂Y 2 ∂Y 3
Clearly, the pressure gradient and the viscous term lose their significance as
s → 0−, and equation (84a) reduces to
Ψ00 Ψ01 − Ψ01 Ψ00 = 0.
Rearranging this equation as

Ψ01
= 0,
Ψ00
356 A. Ruban
and integrating it with the initial conditions (118), we find

a0
Ψ01 = Ψ (Y ). (119)
λ0 00
Of course, the velocity profile Ψ00 (Y ) at the point of zero skin friction,
x = xs , may be only found by calculating the boundary-layer equations
(84) from the stagnation point x = 0 to the point x = xs .
Substitution of (119) back into (117) renders the solution in region 2b
in the form
a0
Ψ(x, Y ) = Ψ00 (Y ) + (−s)1/2 Ψ00 (Y ) + · · · as s → 0 − . (120)
λ0
If we now substitute (84) into (83) and make use of (120), then we will
find that the tangential and normal velocity components in region 2b are
a0
Vτ = Ψ00 (Y ) + · · · , Vn = Re−1/2 (−s)−1/2 Ψ (Y ) + · · · .
2λ0 00
We see that Vn develops a singularity on approach to point x = xs . Singular
behaviour is exhibited by other fluid functions as well. Let us, for example,
calculate the skin friction (86). For this purpose we have to use the solution
(116) in region 2a. We find
! "
τ = (−s)1/2 a0 + O (−s)3/4 as s → 0 − .
Since the skin friction τ is always positive upstream of the point x = xs

(see Figure 18), we have to conclude that a0 > 0.
The square root form of the singularity at the position of zero skin fric-
tion was first predicted by Landau & Lifshitz (1944). Four years later Gold-
stein (1948) confirmed their conclusions based on a more rigorous analysis
of the boundary-layer equations. Goldstein also demonstrated that if the
solution develops this form of singularity, then it cannot be continued down-
stream of the point of zero skin friction.
Indeed, let us assume that the continuation is possible. Then, keeping
in mind that the solution of the boundary-layer equation (84a) is sought
in the class of continuous functions, we have to assume that the leading
order term of the asymptotic expansion of the stream function Ψ(x, Y ) in
region 2c (see Figure 19) is written as
Ψ(x, Y ) = Ψ00 (Y ) + · · · as s → 0+, Y = O(1). (121)
The velocity profile in the boundary-layer cross-section x = xs is given

by the derivative Ψ00 (Y ). According to (118), near the aerofoil surface it
behaves as Ψ00 (Y ) = 12 Y 2 . Therefore, the thickness of the viscous region 2d

(see Figure 19) may be estimated using again equation (92). We find that
Y ∼ s1/4 , which means that the asymptotic analysis of the boundary-layer
equation (84a) is based in region 2d on the limit
Y
s = x − xs → 0+, ξ= = O(1).
s1/4
We shall seek the stream function in this region in the form of the asymptotic
expansion
Ψ(x, Y ) = s3/4 61 λ0 ξ 3 + sμ fˆ1 (ξ) + s2μ−3/4 fˆ2 (ξ) + · · · , (122)
with the exponent μ being unknown in advance. The functions fˆ1 (ξ) and
fˆ2 (ξ) are determined in the same way as the functions f1 (η) and f2 (η) in
the solution (94) for region 2a. We have
fˆ1 = 12 â0 ξ 2 , (123)

ξ
â2 5 λ0
fˆ2 = 0 ξ 2 ζ −2 M 1 − 2μ, , − ζ 4 − 1 dζ + 12 b̂0 ξ 2 , (124)
2λ0 4 32
0
where â0 and b̂0 are arbitrary constants.

It remains to carry out the matching of the solutions in regions 2c and
2d. In order to perform this procedure it is necessary to determine the
asymptotic behaviour of the function fˆ2 (ξ) at the outer edge of region 2d.
It is known the asymptotic behaviour of the Kummer’s function M (a, b, z)
depends on a choice of the ray along which z tends to infinity. If the ray
lies in the left half-plane (z < 0), then
1
Γ(b)
M (a, b, z) = (−z)−a 1 + O .
Γ(b − a) z
Using this formula in (124), we find that

2μ−1 8μ−3
â2 Γ(5/4) λ0 ξ
fˆ2 = 0 + ··· as ξ → ∞. (125)
2λ0 Γ(1/4 + 2μ) 32 8μ − 5
Finally, we substitute (123) and (125) into (122), and recall that ξ = Y /s1/4 .
As a result we find that at the outer edge of region 2d
2μ−1 8μ−3
â2 Γ(5/4) λ0 Y
Ψ = 16 λ0 Y 3 + 0 + ··· . (126)
2λ0 Γ(1/4 + 2μ) 32 8μ − 5
358 A. Ruban
On the other hand, the asymptotic representation of the stream function

Ψ(x, Y ) at the ‘bottom’ of region 2c has the form
a20 5
Ψ = 16 λ0 Y 3 − Y + ··· . (127)
240
It is obtained by using the first of formulae (118) in the solution (121) in
region 2c.
According to the principle of matched asymptotic expansions, expres-
sions (126) and (127) should coincide with one another. This is only possible
if μ = 1 and5
â20 = −a20 . (128)
While the constant a0 is found in the process of solving the boundary-layer
equation (84a) from the stagnation point x = 0 to the point of zero skin
friction x = xs , in order to find constant â0 one has to use equation (128).
As this equation does not allow for a real solution, we can conclude that
the solution of the boundary-layer equation (84a) only exists upstream of
the line x = xs and cannot be extended beyond this line.
A Weaker Singularity Remind that the behaviour of the boundary layer

at the leading edge of the aerofoil depends on the parameter k, which is de-
termined by the angle of attack and the profile shape by means of equation
(79). If k exceeds the critical value k0 = 1.1575, then the solution termi-
nates at the point of zero skin friction, where the Goldstein singularity is
encountered. If, on the other hand, k is smaller than k0 , then the skin fric-
tion stays positive and the solution exists for all x > 0. In this latter case
the skin friction τ (x) develops a minimum when k becomes close enough to
k0 . The value of the minimum decreases with the parameter k approaching
k0 , as shown in Figure 18. Finally, when k reaches its critical value k0 , the
minimal skin friction becomes zero at point x0 = 8.265.
Let us denote the solution of the boundary-layer problem (84) at k = k0
by Ψ0 (x, Y ), and study its behaviour near the point x = x0 separately. We
start by noting that the procedure used above for constructing the solu-
tion in the vicinity of the point of zero skin friction remains applicable for
this particular case. However, one needs to bear in mind that the function
Ψ0 (x, Y ) may thought of as the limit of the solution of the boundary-layer
equation (84a) as k → k0 − 0. Since for all k < k0 the solution continues
through the line x = x0 , the same should be true for the limit solution
5
In order to simplify the coefficient in the second term in (126) we used a well known
property of the Gamma function, Γ(1 + z) = zΓ(z).
Ψ0 (x, Y ). In view of equation (128), the latter is possible only if the coeffi-
cient a0 in the first eigenfunction in the expansion (116) is zero.
This conjecture is confirmed by the numerical solution of boundary-layer
equation (84a). We have seen that for all k > k0 , Goldstein’s singularity
develops in the solution at the point of zero skin friction x = xs . However, as
the parameter k decreases, the point of zero friction xs moves downstream,
approaching the point x0 . Simultaneously, the singularity becomes weaker,
i.e. the coefficient a0 decreases. At k = k0 it becomes equal to zero, and
the first eigenfunction in (116) disappears.
The second eigenvalue
3
α= (129)
2
corresponds to m = 2 in (115). With (129) the solution in region 2a (see
Figure 19) turns into
Ψ0 (x, Y ) =(−s)3/4 61 λ0 η 3 + (−s)3/2 f1 (η)+

+ (−s)7/4 F1 (η) + (−s)9/4 f2 (η) + · · · ,
2λ0 λ1 7 (130)
f1 = 12 a0 η 2 , F1 = − 16 λ1 η 3 + η ,
7!
a2 λ0 a20 9
f2 = 12 b0 η 2 − 0 η5 + η .
5! 8!
To maintain uniformity in notations we still denote the coefficient in the
eigenfunction by a0 as before; however, it should be kept in mind that now
this is the coefficient multiplying the second eigenfunction. The appearance
of an additional term (−s)7/4 F1 (η) in the expansion (130) for Ψ0 (x, Y )
is related to the linear term λ1 s in the Taylor expansion of the pressure
gradient (87).
The solution in region 2b (see Figure 19) is expressed by the asymptotic
expansion
Ψ0 (x, Y ) = Ψ00 (Y ) + (−s)Ψ01 (Y ) + · · · as s → 0 − . (131)
Substituting (131) into the boundary-layer equation (84a), we find
−Ψ00 Ψ01 + Ψ01 Ψ00 = −λ0 + Ψ

00 ,
or, equivalently,

Ψ01 λ0 − Ψ00
= 2 .
Ψ00 Ψ00
360 A. Ruban
Integration of this equation yields
Y
Ψ00 (Y ) − λ0
Ψ01 (Y ) = Ψ00 (Y ) C− ! "2 dY
. (132)
Ψ00 (Y )
0
Matching the solution (131), (132) in region 2b with the solution (130)
in region 2a shows, firstly, that the constant C in (132) is
a0
C= , (133)
λ0
and secondly, that
2λ0 λ1 7 λ0 a20 9
Ψ00 (Y ) = 16 λ0 Y 3 + Y + Y + ··· as Y → 0. (134)
7! 8!
This completes the construction of the solution before the point of zero
skin friction, x = x0 . It is interesting to notice that substitution of (133) into
(132) and then into (131) yields an asymptotic representation of Ψ0 (x, Y ),
Y
Ψ00 (Y ) − λ0 a0
Ψ0 (x, Y ) = Ψ00(Y ) + sΨ00 (Y ) ! "2 dY
− + · · · , (135)
Ψ00 (Y ) λ0
0
which proves to be valid both in region 2b and region 2a; see Exercise 2.
Let us now show that this solution can be continued downstream of point
x = x0 . We start with region 2c (see Figure 19), where Ψ0 (x, Y ) is sought
in the form
# 01 (Y ) + · · ·
Ψ0 (x, Y ) = Ψ00 (Y ) + sΨ as s → 0 + . (136)
Function Ψ# 01 (Y ) is found through substitution of (136) into the boundary-

layer equation (84a), and integrating the resulting equation for Ψ# 01 (Y ). We
find
Y
# # Ψ00 (Y ) − λ0
Ψ01 (Y ) = Ψ00 (Y ) C + ! "2 dY . (137)
Ψ00 (Y )
0
In region 2d the solution is represented by the asymptotic expansion
Ψ0 (x, Y ) = s3/4 16 λ0 ξ 3 + s3/2 fˆ1 (ξ) + s7/4 F#1 (ξ) + s9/4 fˆ2 (ξ) + · · ·
(138a)
as s → 0+,
with ξ = Y /s1/4 . Substituting (138a) into the boundary-layer equation

(84a) and solving the resulting equations for fˆ1 (ξ), F#1 (ξ) and fˆ2 (ξ) with
the no-slip conditions on the aerofoil surface (84c), we find
2λ0 λ1 7
fˆ1 = 12 â0 ξ 2 , F#1 = 16 λ1 ξ 3 + ξ ,
7!
(138b)
â2 λ0 â20 9
fˆ2 = 12 b̂0 ξ 2 + 0 ξ 5 + ξ .
5! 8!
It remains to match the expansions (136) and (138). We recall that it is
this procedure that was found impossible to perform when an attempt was
made to extend the solution with Goldstein’s singularity downstream from
the point of zero friction. The matching resulted in equation (128), which
could not be solved for â0 in terms of real numbers. Now instead of (128)
we obtain
â20 = a20 . (139)
# in (137) is given by
The matching also shows that constant C
# = â0 .
C (140)
λ0
Substitution of (140) back into (137) and then into (136) results in the
formula
Y
Ψ00 (Y ) − λ0 â0
Ψ0 (x, Y ) = Ψ00 (Y ) + sΨ00 (Y ) !

"2 dY + + · · · , (141)
Ψ00 (Y ) λ0
0
which, similar to (135), may be used not only in the main part of the
boundary layer (region 2c) but also in the viscous sublayer (region 2d).
Equation (138) shows that the solution of the boundary-layer problem
(84), which is unique before the point of zero friction, can be continued
downstream of this point in two ways. The first one is given by
â0 = −a0 .
In this case the formulae (135) and (141) may be written together
Y
Ψ00 (Y ) − λ0 a0
Ψ0 (x, Y ) = Ψ00 (Y ) + sΨ00 (Y ) !

"2 dY − + O(s2 ), (142)
Ψ00 (Y ) λ0
0
showing that the solution is smooth in the vicinity of the point of zero skin
friction, x = x0 . It follows from (142) that the skin friction changes sign at
362 A. Ruban
this point, namely,

∂ 2 Ψ0
τ= = −a0 s + O(s2 ), (143)
∂Y 2 Y =0
showing that a region of reverse flow forms downstream of x = x0 . Solutions

of this kind are typical for the triple-deck theory. In particular, in the first
part of this presentation we were dealing with the self-induced separation of
the boundary layer in supersonic flow. Figure 14 shows a smooth behaviour
of the flow near the separation point.
However, there exists a second branch of the solution with
â0 = a0 .
In this case, combining (135) and (141) together, we have
Y
a0 Ψ00 (Y ) − λ0
Ψ0 = Ψ00 (Y ) + Ψ00 (Y ) |s| + s !
2
"2 dY + O(s ). (144)
λ0 Ψ00 (Y )
0
This solution has a singularity

at x = x0 . In particular, let us calculate the
angle θ = arctan Vn /Vτ made by the streamlines with the aerofoil contour.
It follows from (82), (83) and (144) that
θ = Re−1/2 Θ, (145a)
where
Y
∂Ψ0 /∂x a0 λ0 − Ψ
00 (Y )

Θ=− = − sign(s) + ! "2 dY + O(s). (145b)
∂Ψ0 /∂Y λ0 Ψ00 (Y )
0
We can see that Θ has a discontinuity at x = x0 .

It further follows from (144) that the skin friction
τ = a0 |s| + O(s2 ) as s → 0. (146)
A characteristic minimum in the distribution of the skin friction (146) shows

that it is the singular solution (144) that represents the limiting solution of
the boundary-layer equations (84) as k → k0 − 0.
The formation of the singularity in the boundary layer Let us now

see what happens if the parameter k does not coincide with its critical value
k0 but the difference Δk = k − k0 is small. In this case the velocity at the

outer edge of the boundary layer (80) may be represented by the Taylor
expansion ! "
Ue (x, k) = Ue,0 (x) + ΔkUe,1 (x) + O (Δk)2 , (147)
where
∂Ue (x, k)
Ue,0 (x) = Ue (x, k0 ), Ue,1 (x) = .
∂k k=k0
Using the Bernoulli equation

1 2
2 Ue + pe = 12 ,
one can then conclude that the pressure in the boundary layer, pe (x), is also
representable by the Taylor expansion
pe (x) = p0 (x) + Δkp1 (x) + · · · . (148)
Here ! "
1 2
p0 = 2 1 − Ue,0 , p1 = −Ue,0 Ue,1 .
Being guided by (147) and (148), we seek a solution of the boundary-
layer problem (84) in the form
Ψ = Ψ0 (x, Y ) + ΔkΨ1 (x, Y ) + · · · as Δk → 0. (149)
The leading order term Ψ0 (x, Y ) has been studied in the preceding section.
We shall now consider the function Ψ1 (x, Y ). Substitution of (149) together
with (147) and (148) into (84) yields
∂Ψ0 ∂ 2 Ψ1 ∂ 2 Ψ0 ∂Ψ1 ∂Ψ0 ∂ 2 Ψ1 ∂ 2 Ψ0 ∂Ψ1

+ − 2
− =
∂Y ∂x∂Y ∂x∂Y ∂Y ∂x ∂Y ∂Y 2 ∂x (150a)
dp1 ∂ 3 Ψ1
− + ,
dx ∂Y 3
Ψ1 = 0 at x = 0, (150b)
∂Ψ1
Ψ1 = =0 at Y = 0, (150c)
∂Y
∂Ψ1
= Ue,1 (x) at Y = ∞. (150d)
∂Y
The solution of (150) near point x = x0 may be constructed in the same
way as it was done for the leading order problem. We start with region
364 A. Ruban
2a (see Figure 19), where the asymptotic expansion of function Ψ1 (x, Y ) is
sought in the form
Ψ1 (x, Y ) = (−s)β g1 (η) + (−s)β+3/4 g2 (η) + · · · as s → 0−, (151)
with
Y
η= (152)
(−s)1/4
assumed an order one quantity. The leading order term in (151) represents
an eigenfunction of the local solution; constant β is the eigenvalue to be
determined.
Substitution of (151) together with (130) into (150a) results in the fol-
lowing equation for g1 (η):

1 1 1 2
g1 − λ0 η3 g1 + λ0 β + η g1 − λ0 βηg1 = 0.
8 2 4
For any β, its solution satisfying the no-slip conditions
g1 (0) = g1 (0) = 0,
is written as
g1 = 12 a1 η 2 , (153)
where a1 is an arbitrary constant. In order to find the eigenvalue β one
needs to consider the second term in (151). Function g2 (η) satisfies the
equation

1 1 3 β 1
g2 − λ0 η 3 g2 + λ0 β+1 η2 g2 −λ0 β+ ηg2 = − + a0 a1 η 2 . (154)
8 2 4 2 4
It should be solved with the no-slip conditions on the aerofoil surface
g2 (0) = g2 (0) = 0.
The solution to (154) may be constructed in the same way as it was done
with equation (100). We find that
a0 a1
g2 (η) = (η − g22 ) + 12 b1 η 2 , (155)
λ0
where function g22 (η) allows for the power series representation
∞
n
λ0 (−β − 1/2)n
g22 (η) = − η4n+1 , (156)
n=0
32 (5/4)n (4n − 1) n!
and may be expressed through the Kummer function,

η
2 5 λ0
g22 (η) = η − η ξ −2 M − β − 1/2, , ξ 4 − 1 dξ. (157)
4 32
0
It follows from (157) that g22 (η) grows exponentially as η → ∞ for all
values of β except
β = m − 12 , m = 0, 1, 2 . . . . (158)
Equation (158) gives the sought sequence of the eigenvalues; the first of
these being
β = − 12 . (159)
With (159) all the coefficients in (156), except the first one (n = 0), are
zeros, and we have
g22 = η. (160)
Using (160) in (155), and substituting (155) together with (159) and (153)
into (151), we arrive at a conclusion that the solution of equation (150a) in
region 2a has the form
Ψ1 (x, Y ) = (−s)−1/2 21 a1 η 2 + (−s)1/4 12 b1 η 2 + · · · . (161)
In order to predict the form of the solution in region 2b (see Figure 19),
we rearrange (161) with the help of (152). We have
Ψ1 (x, Y ) = (−s)−1 21 a1 Y 2 + (−s)−1/4 12 b1 Y 2 + · · · ,
which suggests that in region 2b, where Y = O(1), the asymptotic expansion
of Ψ1 (x, Y ) should be written in the form
! "
Ψ1 (x, Y ) = (−s)−1 Ψ11 (Y ) + O (−s)−1/4 as s → 0−, (162)
where function Ψ11 (Y ) is such that
Ψ11 (Y ) = 12 a1 Y 2 + · · · as Y → 0. (163)
Substitution of (163) and (131) into (150a) results in
Ψ00 Ψ11 − Ψ00 Ψ11 = 0.
The solution of this equation, satisfying the boundary condition (163), has
the form
a1
Ψ11 = Ψ (Y ). (164)
λ0 00
366 A. Ruban
It remains to substitute (164) back into (162), and we can conclude that in
region 2b
a1 ! "
Ψ1 (x, Y ) = (−s)−1 Ψ00 (Y ) + O (−s)−3/4 as s → 0−, (165)
λ0
Similar to (135), equation (165) proves to be valid not only in region 2b
but also in region 2a; see Exercise 2. This statement is easily verified by
substituting (134) into (165) and comparing the resulting expression with
(161).
Let us now return to the expansion (149). As any other asymptotic
expansion it is expected to have a proper ordering of the terms, namely,
each subsequent term should be much smaller than the previous one. The
assumption that the expansion (149) satisfies this requirement was used
when deriving equation (150a), and became the basis of the entire proce-
dure employed for analysing the behaviour of the function Ψ1 (x, Y ). For
each point (x, Y ) situated upstream of the point x = x0 this assumption
is indeed satisfied thanks to the smallness of Δk. However, as x → x0 − 0
the function Ψ1 (x, Y ) increases without bound, leading to violation of the
supposed relationship between the terms in (149). Consequently, we have
to examine the vicinity of the point x = x0 separately.
In order to find the size of a new region that has to be introduced near
x = x0 , we notice that the x-dependence of both Ψ0 (x, Y ) and Ψ1 (x, Y ) is
determined by the eigenfunctions in the solutions (130) and (161) for region
2a. Comparing the contribution of the eigenfunctions in the asymptotic
expansion (149),
(−s)3/2 12 a0 η2 ∼ Δk (−s)−1/2 a1 η 2 .
we find that the longitudinal extent of the new region is estimated as
√
|x − x0 | ∼ ,
where = |Δk|.
The asymptotic structure of this region is shown in Figure 20. It is
composed of√two
layers. The first one (region 3) is a continuation of region 2a
into the O vicinity of the point x = x0 . Since η = Y /(−s)1/4 is an
order one quantity in region 2a, the thickness of region 3 may be estimated
as
Y ∼ |x − x0 |1/4 ∼ 1/8 .
We see that in region 3 the asymptotic analysis of the boundary-layer equa-
tions (84) is based on the limit
x − x0 Y
x∗ = = O(1), Y∗ = = O(1), → 0. (166)
1/2 1/8
1
2c
4
2b 2d
3
2 2a
√

√
Figure 20. The O vicinity of the point x = x0 .
In order to find the form of the asymptotic expansion of the stream

function Ψ(x, Y ) in this region, we shall re-expand the solution in region 2a
in terms of the variables (166). We start by substituting (130) and (161)
back into (149),
2λ0 λ1 7
Ψ = (−s)3/4 16 λ0 η 3 + (−s)3/2 12 a0 η 2 + (−s)7/4 − 16 λ1 η 3 + η +
7!
a2 λ0 a20 9
+ (−s)9/4 12 b0 η 2 − 0 η5 + η +
5! 8!
+ Δk (−s)−1/2 21 a1 η 2 + (−s)1/4 21 b1 η 2 + · · · . (167)
We then note that

Y Y∗
(−s) = x0 − x = 1/2 (−x∗ ), η= = , (168)
(−s)1/4 (−x∗ )1/4
and express Δk in the form
Δk = sign(Δk). (169)
These, being substituted into (167), lead to

1
Ψ = 3/8 λ0 Y∗3 +
6

6/8 1 1
+ (−x∗ ) a0 Y∗2 + (−x∗ )−1 sign(Δk) a1 Y∗2 +
2 2

1 2λ0 λ1 7
+ 7/8 λ1 x∗ Y∗3 + Y∗ +
6 7!
2
a λ0 a20 9 1
+ 9/8 0 x∗ Y∗5 + Y + (−x∗ )7/4 b0 Y∗2 + (170)
5! 8! ∗ 2

1
+ (−x∗ )−1/4 sign(Δk) b1 Y∗2 + · · · . (171)
2
368 A. Ruban
This suggests that in region 3 the solution of the boundary layer equations
(84) should be sought in the form
1
Ψ =3/8 λ0 Y∗3 + 6/8 Ψ∗1 (x∗ , Y∗ )+
6 (172)
7/8 1 3 2λ0 λ1 7
+ λ1 x∗ Y∗ + Y∗ + 9/8 Ψ∗2 (x∗ , Y∗ ) + · · · .
6 7!
It also follows from (171) that functions Ψ∗1 (x∗ , Y∗ ) and Ψ∗2 (x∗ , Y∗ ) satisfy
the following matching conditions with the solution in region 2a,
1$ %
Ψ∗1 = a0 (−x∗ ) + sign(Δk)a1 (−x∗ )−1 Y∗2 + · · · as x∗ → −∞
2
(173a)
and
a20 λ0 a20 9
Ψ∗2 = x∗ Y∗5 + Y +
5! 8! ∗
1$ %
+ b0 (−x∗ )7/4 + sign(Δk)b1 (−x∗ )−1/4 Y∗2 + · · · as x∗ → −∞.
2
(173b)
We are dealing here with the classical formulation of the boundary-layer

equations with the pressure gradient assumed to be known. Differentiating
(148) with respect to x, we have
dpe dp0 dp1
= + Δk + ··· .
dx dx dx
In a small vicinity of point x = x0 the first term is given by the Taylor
expansion (87), while the second term remain finite. Therefore in region 3
the pressure gradient
dpe
= λ0 + 1/2 λ1 x∗ + O(). (174)
dx
Substitution of (172) and (174) into (84a) gives in the leading order approx-
imation
∂ 2 Ψ∗1 ∂Ψ∗1 ∂ 3 Ψ∗1
1
λ Y2
2 0 ∗ ∂x ∂Y
− λ0 Y∗ = . (175)
∗ ∗ ∂x∗ ∂Y∗3
The solution to this equation should satisfy the no-slip conditions on the
aerofoil surface,
∂Ψ∗1
Ψ∗1 = = 0 at Y∗ = 0. (176)
∂Y∗
Since equation (175) is linear, and its coefficients do not depend on x∗ ,

the solution to this equation may be constructed using the method of Fourier
Transforms. We define the Fourier Transform of function Ψ∗1 (x∗ , Y∗ ) as
∞
Ψ̆1 (k, Y∗ ) = Ψ∗1 (x∗ , Y∗ )e−ikx∗ dx∗ , (177)
−∞
with the inverse transformation being

∞
1
Ψ∗1 (x∗ , Y∗ ) = Ψ̆1 (k, Y∗ )eikx∗ dk.
2π
−∞
Here the Fourier variable k is assumed real.

Applying the Fourier transformation to the equation (175) renders it in
the form
1 2 dΨ̆1 d3 Ψ̆1
2
ikλ0 Y∗ − ikλ0 Y∗ Ψ̆ 1 = , (178)
dY∗ dY∗3
with the no-slip conditions (176) turning into
dΨ̆1
Ψ̆1 = = 0 at Y∗ = 0. (179)
dY∗
Since the ordinary differential equation (178) is linear and homogeneous, its
general solution may be written as
Ψ̆1 = C1 ψ11 + C2 ψ12 + C3 ψ13 , (180)
where ψ11 , ψ12 and ψ13 are the three complementary solutions of (178).
They can be chosen such that

ψ11 (0) = 1, ψ11 (0) = 0, ψ11 (0) = 0,

ψ12 (0) = 0, ψ12 (0) = 1, ψ12 (0) = 0,

ψ13 (0) = 0, ψ13 (0) = 0, ψ13 (0) = 1.
The first two solutions do not satisfy the conditions (179), and therefore, we
have to set C1 = C2 = 0 in (180). As far as the third solution is concerned,
it is written as
ψ13 = 12 Y∗2 ,
which is easily verified by direct substitution into (178). This reduces (180)
to
Ψ̆1 = 12 C3 Y∗2 . (181)
370 A. Ruban
Notice that factor C3 is an arbitrary function of parameter k. Applying

the inverse transform to (181), we can conclude that
Ψ∗1 (x∗ , Y∗ ) = 12 A∗ (x∗ )Y∗2 , (182)
where A∗ (x∗ ) is an arbitrary function of x∗ , except it is known from (173a)
that
A(x∗ ) = a0 (−x∗ ) + sign(Δk)a1 (−x∗ )−1 + · · · as x∗ → −∞. (183)
In order to find A∗ (x∗ ) we need to consider the next term in (172).
Function Ψ∗2 satisfies the equation
1 2 ∂ 2 Ψ∗2 ∂Ψ∗2 ∂ 3 Ψ∗2 dA∗ 2

2 λ0 Y∗ − λ0 Y∗ = − 12 A∗ Y , (184a)
∂x∗ ∂Y∗ ∂x∗ ∂Y∗3 dx∗ ∗
and the no-slip conditions
∂Ψ∗2
Ψ∗2 = =0 at Y∗ = 0. (184b)
∂Y∗
When applying the Fourier Transforms method to (184) one has to keep
in mind that the Fourier integral (177) converges only if the function it is
applied to tends to zero as |x∗ | → ∞. It follows from (173b) that function Ψ∗2
does not satisfy this requirement. Therefore, we introduce a new unknown
function Ψ2 such that
λ0 a20 9 a20
Ψ∗2 = Ψ2 + Y + x∗ Y∗5 + 12 B∗ (x∗ )Y∗2 + G∗ (x∗ )Y∗ , (185)
8! ∗ 5!
where
A2∗ − a20 x2∗ − 2a0 a1 sign(Δk)
G∗ (x∗ ) = . (186)
2λ0
It follows from (183) that G∗ (x∗ ) → 0 as x∗ → −∞. Notice that the
solution of the boundary-value problem (184) is not unique. It can always
be supplemented with 12 B∗ (x∗ )Y∗2 , where B∗ (x∗ ) is an arbitrary function.
It is inserted into (185) to ensure that Ψ∗2 tends to zero as x∗ → −∞.
Comparing (185) with (173b), we see that
B∗ (x∗ ) = b0 (−x∗ )7/4 + sign(Δk)b1 (−x∗ )−1/4 + · · · as x∗ → −∞.
The transformation (185) renders the boundary-value problem in the
form
2
1 2 ∂ Ψ2 ∂Ψ2 ∂ 3 Ψ2
2
λ0 Y∗ − λ0 Y∗ = , (187a)
∂x∗ ∂Y∗ ∂x∗ ∂Y∗3
∂Ψ2
Ψ2 = 0, = −G∗ (x∗ ) at Y∗ = 0. (187b)
∂Y∗
Applying the Fourier transformation to the equation (187a) and boundary
conditions (187b) yields
1 2 dΨ̆2 d3 Ψ̆2
2 ikλ0 Y∗ − ikλ0 Y∗ Ψ̆2 = , (188a)
dY∗ dY∗3
dΨ̆2
Ψ̆2 = 0, = −Ğ(k) at Y∗ = 0. (188b)
dY∗
Here Ψ̆2 (k, Y∗ ) and Ğ(k) are Fourier Transforms of functions Ψ2 and G∗ ,
respectively.
The general solution of equation (188a) is written as
Ψ̆2 = C1 ψ21 + C2 ψ22 + C3 ψ23 , (189)
where ψ21 , ψ22 and ψ23 are the three complementary solutions of the equa-
tion (188a). We shall choose them using the initial conditions

ψ21 (0) = 1, ψ21 (0) = 0, ψ21 (0) = 0,

ψ22 (0) = 0, ψ22 (0) = 1, ψ22 (0) = 0, (190)

ψ23 (0) = 0, ψ23 (0) = 0, ψ23 (0) = 1.
Then it follows from (188b) that
C1 = 0, C2 = −Ğ(k). (191)
Factor C3 remains an arbitrary function of k, and the third complementary

solution ψ23 is easily seen to be
ψ23 = 12 Y∗2 . (192)
Substituting (191) and (192) back into (189), we have
Ψ̆2 = −Ğ(k)ψ22 + 12 C3 (k)Y∗2 . (193)
Our task now will be to determine the function ψ22 . The power series
presentation of this function may be constructed in the same way as it was
done with function f22 (η) representing the second complementary solution
of equation (100). We start with the conditions (190). They show that the
first term in the series is
ψ22 = Y∗ + · · · . (194)
Using (194) on the left hand side of (188a) yields
d3 ψ22
= − 21 ikλ0 Y∗2 .
dY∗3
372 A. Ruban
Integration of this equation gives the second term,

ikλ0 5
ψ22 = Y∗ − Y + ··· .
120 ∗
This procedure can be repeated as many time as one wishes, leading to a
conclusion that the power series of ψ22 should be written in the form
∞

ψ22 (Y∗ ) = cn Y∗4n+1 . (195)
n=0
Alternatively we can write

∞

ψ22 (Y∗ ) = Y∗ + cn+1 Y∗4n+5 . (196)
n=0
Using (195) on the left hand side of the equation (188a) and (196) on the
right hand side, we find that the coefficients of the series (195) satisfy the
following recurrent equation
ikλ0 n − 14
cn+1 = cn , (197a)
32 n + 34 n + 54 (n + 1)
which has to be solved starting with
c0 = 1. (197b)
By direct substitution one can easily verify that the solution of (197) is
n
ikλ0 Γ(5/4)
cn = − . (198)
32 Γ(n + 5/4)(4n − 1)n!
Substituting (198) back into (195), we have
∞ n
5 ikλ0 Y∗4n+1
ψ22 (Y∗ ) = −Γ 4 . (199)
n=0
32 Γ(n + 5/4)(4n − 1)n!
We shall show now that the function ψ22 (Y∗ ) may be expressed through
the Bessel function. It is known (see, for example Abramowitz & Stegun,
1965) that the Bessel function of the first kind and order ν may be repre-
sented by the power series
∞
2n+ν
(−1)n z
Jν (z) = , (200)
n=0
Γ(n + ν + 1) n! 2
which converges in the entire complex plane z except z = ∞. It is also

known that in the limit, |z| → ∞,

2 1
Jν (z) = cos z − 12 νπ − 14 π + e|(z)| O ,
πz z (201)
− π < arg z < π.
In order to reproduce the Gamma function, Γ(n + 5/4), in (199), we
choose ν = 1/4, and then, writing the first term in the series (200) sepa-
rately, we will have
1/4 ∞ 2n+1/4
1 z (−1)n z
J1/4 (z) = + . (202)
Γ(5/4) 2 n=1
Γ(n + 5/4) n! 2
It follows from (202) that

z 3/2 ∞ 2n−1/2
2 J1/4 (z) 1 (−1)n z
− dz = 4 .
z (z/2)1/4 Γ(5/4) n=1
Γ(n + 5/4)(4n − 1)n! 2
0
Therefore
∞
2n+1/2
(−1)n z
=
n=0
Γ(n + 5/4)(4n − 1)n! 2
z 3/2 1/2 (203)
z 2 J1/4 (z) 1 1 z
= 1/4
− dz − .
8 z (z/2) Γ(5/4) Γ(5/4) 2
0
We substitute (201) into (203), and restrict our attention to the leading
order exponentially growing term. Using the integration by parts with
9/4
2
u= , dv = cos z − 38 π dz,
z
13/4
9 2
du = − dz, v = sin z − 38 π ,
8 z
we find that
∞
2n+1/2
(−1)n z
=
n=0
Γ(n + 5/4)(4n − 1)n! 2
5/4
1 2
= √ sin z − 38 π + · · · as |z| → ∞.
4 π z
374 A. Ruban
It remains to set z = 12 iΩ1/2 Y∗2 , where Ω = 12 ikλ0 , and we will see that
function ψ22 , given by (199), grows exponentially as Y∗ → ∞, namely,
5/4
Γ(5/4) 4 1 1/2 2
ψ22 (Y∗ ) = − √ iπ/4 1/4 sin 2 iΩ Y∗ − 38 π + · · · .
2 πe Ω iΩ1/2 Y∗2
(204)
This means that the exponential growth of the solution (193) of the equation
(187a) can only be suppressed by setting
Ğ(k) = 0. (205)
Applying the inverse Fourier transformation to (205), we have
G∗ (x∗ ) = 0.
It remains to recall that the function G∗ (x∗) is defined by equation (186).

We see that
A2∗ − a20 x2∗ − 2a0 a1 sign(Δk) = 0. (206)
The solution of quadratic equation (206), satisfying the condition (183), is
written as
a1
A∗ = a0 x2∗ + 2 sign(Δk). (207)
a0
This completes the flow analysis in region 3. To conclude, we substitute
(182) into (172), and neglecting the higher order terms, we have the solution
in region 3 in the form
Ψ = 3/8 16 λ0 Y∗3 + 6/8 12 A∗ (x∗ )Y∗2 + · · · , (208)
with A∗ (x∗ ) being given by (207).

Let us now turn our attention to region 4 (see Figure 20), where the
asymptotic analysis of the boundary-layer equation (84a) is based on the
limit
x − x0
x∗ = 1/2 = O(1), Y = O(1), → 0. (209)

We start by re-expanding the solution in region 2b (see Figure 20) in terms
of variables (209), namely, we substitute (135), (165) and (169) into (149).
Taking into account that s = x − x0 = 1/2 x∗ , we will have
Ψ = Ψ00 (Y )+
Y −1

1/2 Ψ00 − λ0 a0 (−x∗ ) + a1 sign(Δk)(−x∗ )
+ Ψ00 (Y ) x∗ 2 dY + +· · · .
Ψ00 λ0
0
This suggests, firstly, that the asymptotic expansion of the stream function
Ψ(x, Y ) in region 4 should be sought in the form
Ψ(x, Y ) = Ψ00 (Y ) + 1/2 Ψ1 (x∗ , Y ) + · · · . (210)
Secondly, we see that the matching condition with the solution in region 2b
reads
Y
Ψ00 − λ0 A∗ (x∗ )
Ψ1 = Ψ00 (Y ) x∗
2 dY + +··· as x∗ → −∞. (211)
Ψ00 λ0
0
Here we use the fact that function A∗ (x∗ ) is represented at large negative
x∗ by (183).
We will also need the matching condition with the solution (208) in
region 3. It can be formulated with the help of a usual routine. We start
by expressing the solution in region 3 in terms of the variables of region 4.
Since the scaled tangential coordinate x∗ is common for regions 3 and 4, we
just need to recall that the relationship between the normal coordinates is
given by (166), i.e.
Y
Y∗ = 1/8 . (212)

Substitution of (212) into (208) yields
Ψ = 16 λ0 Y 3 + 1/2 A∗ (x∗ )Y 2 + · · · . (213)
Now we turn our attention to the asymptotic expansion (210) of the stream
function in region 4. At the “bottom” of region 4 the leading order term in
(210) may be simplified with the help of (134), which turns (210) into
Ψ(x, Y ) = 16 λ0 Y 3 + 1/2 Ψ1 (x∗ , Y ) + · · · . (214)
It remains to compare (213) with (214), and we can conclude that the sought
matching condition is written as
Ψ1 (x∗ , Y ) = 12 A∗ (x∗ )Y 2 + · · · as Y → 0. (215)
The equation for function Ψ1 (x∗ , Y ) is obtained by substituting (210)

together with (174) into the boundary-layer equation (84a). We find that
∂ 2 Ψ1 ∂ Ψ1
Ψ00 − Ψ00 = −λ0 + Ψ
00 , (216)
∂x∗ ∂Y ∂x∗
376 A. Ruban
or, equivalently,

∂2 Ψ1 Ψ − λ0
= 00 2 . (217)
∂x∗ ∂Y Ψ00 Ψ00
It follows from (215) and (134) that

Ψ1 A∗ (x∗ )
= .
Ψ00 Y =0 λ0
Therefore, integrating (217) with respect to Y , we have
Y
∂ Ψ1 Ψ
00 − λ0 A∗ (x∗ )
= 2 dY + . (218)
∂x∗ Ψ00 Ψ00 λ0
0
It remains to integrate (218) with respect to x∗ . When performing the

integration, condition (211) has to be used. We find that
Y
Ψ00 − λ0 A∗ (x∗ )
Ψ1 (x∗ , Y∗ ) = Ψ00 (Y ) x∗
2 dY + , (219)
Ψ00 λ0
0
where the function A∗ (x∗ ) is given by (207).

Substituting (219) back into (210), we have the solution in region 4 in
the form
Y
1/2 Ψ00 − λ0 A∗ (x∗ )
Ψ(x, Y ) = Ψ00 (Y )+ Ψ00 (Y ) x∗ 2 dY + +· · · . (220)
Ψ00 λ0
0
Through making use of (134) one can verify that in region 3, where Y =
1/2 Y∗ , the expansion (220) reduces to (208). This means that we can use
(220) not only in region 4 but also in region 3. Substituting (207) into (220),
and returning to the original variable x = x0 + 1/2 x∗ , we have
Ψ(x, Y ) = Ψ00 +
Y
a0 a1 Ψ00 − λ0
(221)
+ Ψ00 (x − x0 ) + 2 Δk + (x − x0 )
2
2 dY + · · · .
λ0 a0 Ψ00
0
Equation (221) represents an asymptotic solution of the boundary-layer

equations (84). It holds in a small vicinity of point x = x0 provided that
Δk is small. The domain of existence of the solution is determined by the
sign of the argument of the square root in (221). If a1 Δk > 0, then the
solution exists for all x. The skin friction

∂ 2 Ψ a1
τ= = a0 (x − x0 )2 + 2 Δk (222)
∂Y 2 Y =0 a0
remains positive everywhere and at the point x = x0 it reaches the mini-
mum, whose value decreases as Δk → 0 according to the rule

τmin = 2a0 a1 Δk. (223)
If, however, a1 Δk < 0, then the solution exists only up to the point of zero
friction,
|a1 Δk|
xs = x0 − 2 ;
a0
downstream of this point the argument of the square root becomes nega-
tive. On approach to the point x = xs the solution develops Goldstein’s
singularity. Indeed, it follows from (222) that
1/4 √
τ = 8a30 |a1 Δk| xs − x + · · · as x → xs − 0.
Notice that the singularity becomes progressively weaker as Δk → 0.
According to the numerical calculations (see Figure 18) the first of the
situations described is realized for Δk < 0 and the second for Δk < 0. This
means that the constant a1 is negative.
Remind that in the case when Δk = 0, the boundary-layer equations (84)
admit two solutions. One of them, given by (142), passes smoothly through
the point x = x0 , while the second solution, given by (144), develops a
singularity at this point. Comparing (142) and (144) with (221), one can see
that this is the singular solution (144) that represents the limiting solution
of the boundary-layer equations as k → k0 − 0.
In order to determine the constants a0 and a1 we need to compare the
analytical solution with the results of the numerical calculations of the
boundary-layer equations. Firstly, the skin friction distribution in Figure 18
calculated for k = k0 should exhibit near x = x0 the behaviour predicted
by formula (146). Using this fact, we found that the the flow at the leading
edge of the aerofoil, a0 = 0.0085. Then constant a1 was found using formula
(223). It appeared that a1 = −1.24.
3.3 Viscous-Inviscid Interaction

Up to this point the analysis of the flow has been carried out in the frame-
work of the classical boundary-layer theory. We started with the inviscid
378 A. Ruban
region where the solution of the Euler equations was constructed using the
impermeability conditions on the aerofoil surface, i.e. the existence of the
boundary layer was completely ignored. As a result we found the veloc-
ity distribution (80) on the aerofoil surface, while the pressure distribution
was calculated using the Bernoulli equation. Then, as a second step, the
boundary layer was analysed. For this purpose the classical Prandtl’s equa-
tions (84) were used, with the pressure, pe , and the velocity at the outer
edge of the boundary layer, Ue , assumed uninfluenced by the presence of
the boundary layer. This assumption is based on the observation that the
boundary layer is only capable of causing an O(Re−1/2 ) displacement of the
streamlines from the aerofoil surface. Consequently, as long as the solution
in the boundary layer remains regular, its influence on the pressure field
remains weak.
The situation changes as the parameter k approaches its critical value k0 ,
and the singularity forms in the boundary layer. Remind that when k = k0 ,
the streamlines in the boundary layer develop a corner. The deflection angle
θ may be calculated using (145). We have

2a0
θ = Re−1/2 Θ − Θ = −Re−1/2 . (224)
x=x0 +0 x=x0 −0 λ0
Then it follows from the small perturbation inviscid flow theory (see Exer-
cise 3) that the pressure induced by the boundary layer
2

−1/2 2a0 U0
p = Re ln |x − x0 | + O(1) as x → x0 . (225)
πλ0
Here U0 = Ue (x0 ), and λ0 is the leading order term in the pressure gradient
(87); for an aerofoil with parabolic nose U0 = 1.286 and λ0 = 0.024.
We see that the induced pressure gradient exhibits an unbounded growth
as the singularity is approached, namely,
dp 2a0 U02 1

= Re−1/2 + ··· . (226)
dx πλ0 x − x0
Consequently, one can expect that there exists a small vicinity of the point
x = x0 , where the pressure perturbations, induced by the displacement
effect of the boundary layer, become large enough to start influencing the
flow inside the boundary layer in the leading order approximation.
In order to determine the size of this
√ region let us return to the procedure
used to analyse the flow in the O( ) vicinity of the singular point (see
Figure 20), and recall that in regions 3 and 4 the dependence of the fluid
functions on the longitudinal coordinate x∗ is expressed through function
A∗ (x∗ ). The latter is determined by solving equation (184a). As soon as

the induced pressure gradient becomes large enough “to find its way” into
this equation, the process of viscous-inviscid interaction come into play.
The induced pressure gradient (226) can be compared with any term
in equation (184a). Let us consider, for example, the first term on the
right hand side of (184a). It is produced by the viscous term ∂ 3 Ψ/∂Y 3
in equation (84a). In region 3 the stream function is represented by the
asymptotic expansion (172). Being differentiated three times with respect
to Y , it gives

∂3Ψ 3 ∗
3/8 ∂ Ψ1 1/2 λ0 λ1 3 ∂ 3 Ψ∗2
3
= λ 0 + 3
+ λ1 x∗ + Y∗ + 3/4 + ··· .
∂Y ∂Y∗ 12 ∂Y∗3
The fourth term in the above expansion is a part of equation (184a). When
comparing it with the induced pressure gradient (226),
Re−1/2
∼ 3/4 , (227)
x − x0
we have to take into account that in the region considered,

√
|x − x0 | ∼ . (228)
Solving (227) and (228) for and |x − x0 |, we find
= |Δk| ∼ Re−2/5 , (229)

−1/5
|x − x0 | ∼ Re . (230)
Therefore, in what follows we shall assume that
k = k0 + Re−2/5 k1 , (231)
where constant k1 remains an order one quantity as Re → ∞. Our task

will be to study the interaction region which, apparently, has a three-tiered
structure. In addition to regions 3 and 4, that lie inside the boundary
layer (see Figure 20), we also need to introduce region 5 (see Figure 21).
The latter is a part of the potential flow region, and serves to convert the
perturbations in the streamline shape into the perturbations of pressure.
Before proceeding further it should be mentioned that the induced pres-
sure gradient is only important in the interaction region. Outside this region
the classical boundary-layer theory holds, which means that the solutions
for regions 2a and 2b, constructed in the previous section, are still valid.
380 A. Ruban
1 5
2c
4
2b 2d
3
2 2a
−1/5
Re
Figure 21. The interaction region.
Upper layer (region 5) If we return to region 1 (see Figure 17), where

X and Y are order one quantities, and ignore the existence of the boundary
layer, then the solution of the Euler equations in this region would be regular
at point x = x0 . Keeping in mind that Vn is zero along the aerofoil contour,
we can write the Taylor expansions for the velocity components and pressure
in a small vicinity of the point x = x0 in the form
& ' &
Vτ = U0 + a1 (x − x0 ) + a2 y + a3 (x − x0 )2 +
' (232a)
+ a4 (x − x0 )y + a5 y 2 + · · · ,
& '
Vn = b2 y + b4 (x − x0 )y + b5 y 2 + · · · , (232b)
& ' & 2
p = Pe0 + c1 (x − x0 ) + c2 y + c3 (x − x0 ) +
' (232c)
+ c4 (x − x0 )y + c5 y 2 + · · · .
Remind that U0 denotes the value of the tangential inviscid flow velocity at
point x = x0 on the aerofoil surface; Pe0 is the corresponding value of the
pressure. These are related to one another through the Bernoulli equation,
Pe0 = 12 (1 − U02 ).
Setting Y = 0 in (232c), yields the pressure at the outer edge of the
boundary layer in the form

pe (x) = p = Pe0 + c1 (x − x0 ) + c3 (x − x0 )2 + · · · . (233)
y=0
Comparing (233) with (87), we can see that
c1 = λ0 , c3 = 12 λ1 . (234)
Let us now consider the perturbations produced in region 1 due to the

presence of the boundary layer. We have already analysed the perturba-
tions of pressure. Now we need to extend the analysis to other fluid dy-
namic functions as well. We know that at the critical value of the angle
of attack (k = k0 ) the streamlines in the boundary layer develop a corner

with the deflection angle given by (224). Using (224) in (315), we find that
perturbations induced by the displacement effect of the boundary layer
2a0 U0
Vτ = −Re−1/2 ln (x − x0 )2 + y 2 + · · · , (235a)
πλ0

2a0 U0 y
Vn = −Re−1/2 π − arctan + ··· , (235b)
πλ0 x − x0
2a0 U02
p = Re−1/2 ln (x − x0 )2 + y 2 + · · · . (235c)
πλ0
Let us now turn our attention to region 5; see Figure 21. The longitudinal
extent of this region is given by (230). We shall see that the flow in region 5 is
described by the Laplace equation (241). The principle of least degeneration
requires the lateral size of region 5 to be comparable with its longitudinal
size,
y ∼ |x − x0 | ∼ Re−1/5 .
This means that the asymptotic analysis of the Navier-Stokes equations (81)
in region 5 has to be based on the limit procedure
x − x0 y
x∗ = = O(1), y∗ = = O(1), Re → ∞. (236)
Re−1/5 Re−1/5
In order to predict the form of the asymptotic expansions for the velocity
components and pressure in region 5, we shall express (232) and (235) in
terms of the new variables (236) and combine these together. Being applied
to the tangential velocity component Vτ this procedure yields
& ' & '
Vτ = U0 +Re−1/5 a1 x∗ + a2 y∗ + Re−2/5 a3 x2∗ + a4 x∗ y∗ + a5 y∗2 +
2a0 U0 2a0 U0 2
+ Re−1/2 ln Re − Re−1/2 ln x∗ + y∗2 + · · · .
5πλ0 πλ
Similarly, for the normal velocity component Vn and pressure p we find

& '
Vn = Re−1/5 b2 y∗ + Re−2/5 b4 x∗ y∗ + b5 y∗2 −

2a0 U0 y∗
− Re−1/2 π − arctan + ··· ,
πλ0 x∗
& ' & '
p = Pe0 +Re−1/5 c1 x∗ + c2 y∗ + Re−2/5 c3 x2∗ + c4 x∗ y∗ + c5 y∗2 −
2a0 U02 2a0 U02 2
− Re−1/2 ln Re + Re−1/2 ln x∗ + y∗2 + · · · .
5πλ0 πλ
382 A. Ruban
This suggests that the solution in region 5 should be sought in the form
& '
Vτ =U0 + Re−1/5 a1 x∗ + a2 y∗ +
& '
+ Re−2/5 a3 x2∗ + a4 x∗ y∗ + a5 y∗2 + (237a)
2a 0 U0
+ Re−1/2 ln Re + Re−1/2 u∗1 (x∗ , y∗ ) + · · · ,
5πλ0
& '
Vn =Re−1/5 b2 y∗ + Re−2/5 b4 x∗ y∗ + b5 y∗2 +
(237b)
+ Re−1/2 v1∗ (x∗ , y∗ ) + · · · ,
& '
p =Pe0 + Re−1/5 c1 x∗ + c2 y∗ +
& '
+ Re−2/5 c3 x2∗ + c4 x∗ y∗ + c5 y∗2 − (237c)
2
2a 0 U 0
− Re−1/2 ln Re + Re−1/2 p∗1 (x∗ , y∗ ) + · · · ,
5πλ0
where functions u∗1 , v1∗ and p∗1 satisfy the following matching conditions with
the solution in region 1,
⎫
∗ 2a0 U0 2 ⎪
⎪
u1 = − ln x∗ + y∗ + · · · ,
2
⎪
⎪
πλ0 ⎪
⎪
⎪
⎬
∗ 2a0 U0 y∗
v1 = − π − arctan + ··· , as x2∗ + y∗2 → ∞. (238)
πλ0 x∗ ⎪
⎪
⎪
⎪
2a0 U02 2 ⎪
⎪
p∗1 = ln x∗ + y∗2 + · · · ⎪
⎭
πλ0
The equations for these functions are deduced by substituting (237) into
the Navier-Stokes equations (81), and working with the O(Re−3/10 ) terms.
We find
∂u∗1 ∂p∗ ∂v1∗ ∂p∗ ∂u∗1 ∂v ∗
U0 = − 1, U0 = − 1, + 1 = 0. (239)
∂x∗ ∂x∗ ∂x∗ ∂y∗ ∂x∗ ∂y∗
The set of equations (239) is easily reduced to a single equation for
function p∗1 . We start by eliminating u∗1 . This is done by solving the last of
equations (239) and substituting the result into the first equation. We have
∂v1∗ ∂p∗1 ∂v1∗ ∂p∗
U0 = , U0 = − 1. (240)
∂y∗ ∂x∗ ∂x∗ ∂y∗
Now we eliminate v1∗ by cross-differentiating equations (240). We find that
the pressure p∗1 satisfies the Laplace equation
∂ 2 p∗1 ∂ 2 p∗1
+ = 0. (241)
∂x2∗ ∂y∗2
When solving this equation we will use the method of Fourier Trans-
forms. The later is applicable to functions that tend to zero as |x∗ | → ∞.
It follows from (238) that p∗1 does not belong to this category. Therefore,
we shall differentiate (241) with respect to x∗

∂2 ∂p∗1 ∂2 ∂p∗1
+ = 0, (242)
∂x2∗ ∂x∗ ∂y∗2 ∂x∗
and treat ∂p∗1 /∂x∗ as the sought function. The “far-field” boundary condi-
tion for this function is written as
∂p∗1 2a0 U02 x∗

= → 0 as x2∗ + y∗2 → ∞. (243)
∂x∗ πλ0 x2∗ + y∗2
The equation (242) also requires a boundary condition at y∗ = 0. It will

be formulated in the course of the flow analysis in regions 3 and 4; see
Figure 21.
Viscous syblayer (region 3) When analysing the interactive flow regimes

with the parameter k in the range (229), we can write the independent vari-
ables (166) in region 3 as
x − x0 y
x∗ = , Y∗ = . (244)
Re−1/5 Re−11/20
Remind that (x, y) are curvilinear orthogonal coordinates introduced as

shown in Figure (17).
In order to predict the form of the asymptotic expansion of the stream
function in region 3 we shall again recast the solution in region 2a in terms
of variables (244). Recall that when the flow was analysed in the frame of
classical boundary-layer theory (see Section ?), this procedure led to the
equation (171). Now instead of (168) and (169) we have to use
Y∗
(−s) = Re−1/5 (−x∗ ), η= , Δk = k − k0 = Re−2/5 k1 ,
(−x∗ )1/4
384 A. Ruban
which, being substituted into (167), yield
1
ψ = Re−1/2 Ψ = Re−13/20 λ0 Y∗3 +
6

−16/20 1 1
+ Re (−x∗ ) a0 Y∗2 + (−x∗ )−1 k1 a1 Y∗2 +
2 2

1 2λ λ
0 1 7
+ Re−17/20 λ1 x∗ Y∗3 + Y∗ +
6 7!
2
a λ0 a20 9 1
+ Re−19/20 0 x∗ Y∗5 + Y + (−x∗ )7/4 b0 Y∗2 +
5! 8! ∗ 2

1
(−x∗ )−1/4 k1 b1 Y∗2 + · · · .
2
This suggests that the asymptotic expansion of the stream function ψ in

region 3 should be sought in the form
1
ψ = Re−13/20 λ0 Y∗3 + Re−16/20 Ψ∗1 (x∗ , Y∗ )+
6

−17/20 1 3 2λ0 λ1 7
+ Re λ1 x∗ Y∗ + Y∗ + Re−19/20 Ψ∗2 (x∗ , Y∗ ) + · · · ,
6 7!
(245)
where functions Ψ∗1 (x∗ , Y∗ ) and Ψ∗2 (x∗ , Y∗ ) are such that
1$ %
Ψ∗1 = a0 (−x∗ ) + k1 a1 (−x∗ )−1 Y∗2 + · · · as x∗ → −∞ (246a)
2
and
a20 λ0 a20 9
Ψ∗2 = x∗ Y∗5 + Y +
5! 8! ∗
1 $ %
+ b0 (−x∗ )7/4 + k1 b1 (−x∗ )−1/4 Y∗2 + · · · as x∗ → −∞. (246b)
2
We shall represent the pressure in region 3 by the asymptotic expansion
p = Pe0 + Re−1/5 λ0 x∗ + Re−2/5 21 λ1 x2∗ −

2a0 U02
− Re−1/2 ln Re + Re−1/2 P ∗ (x∗ , Y∗ ) + · · · ,
5πλ0
(247)
the form of which is obtained by setting y∗ = 0 in (237c) and using (234).
We are now ready to formulate the equations of fluid motion in region 3.
We start by substituting (245), (244) into (82). This leads to the following
expressions of the velocity components in region 3,
1 ∂Ψ∗1
Vτ =Re−2/20 λ0 Y∗2 + Re−5/20 +
2 ∂Y∗
2λ0 λ1 6 ∂Ψ∗2
(248a)
+ Re−6/20 12 λ1 x∗ Y∗2 + Y∗ + Re−8/20 + ··· ,
6! ∂Y∗
∂Ψ∗1 ∂Ψ∗2
Vn = −Re−12/20 − Re−13/20 61 λ1 Y∗3 − Re−15/20 + ··· . (248b)
∂x∗ ∂x∗
We then substitute (248) together with (247) into the Navier-Stokes equa-
tions (81). We find from the x-momentum equation (81a) that the equation
(175) for Ψ∗1 retains its form
2 ∗
1 2 ∂ Ψ1 ∂Ψ∗1 ∂ 3 Ψ∗1
2 λ0 Y∗ ∂x ∂Y − λ0 Y∗ = .
∗ ∗ ∂x∗ ∂Y∗3
Its solution , satisfying the no-slip conditions
∂Ψ∗1
Ψ∗1 = = 0 at Y∗ = 0,
∂Y∗
is written as
Ψ∗1 = 12 A∗ (x∗ )Y∗2 . (249)
At this stage, the function A∗ (x∗ ) remains arbitrary; we can only claim that
in view of (246a),
A∗ (x∗ ) = a0 (−x∗ ) + k1 a1 (−x∗ )−1 + · · · as x∗ → −∞. (250)
The equation (184a) for function Ψ∗2 now acquires an additional term,
the induced pressure gradient,
∂ 2 Ψ∗2 ∂Ψ∗2 ∂ 3 Ψ∗2 dA∗ 2 ∂P ∗

1
λ Y2
2 0 ∗ ∂x ∂Y
− λ 0 Y∗ = − 12 A∗ Y − . (251)
∗ ∗ ∂x∗ ∂Y∗ 3 dx∗ ∗ ∂x∗
Substitution of (247), (248) into the y-momentum equation (81b) shows

that P ∗ does not change across region 3, namely,
∂P ∗
= 0. (252)
∂Y∗
386 A. Ruban
Main part of the boundary layer (region 4) When the parameter k
belongs to the range (229), the asymptotic expansion (210) of the stream
function in region 4 is written as
ψ = Re−1/2 Ψ00 (Y ) + Re−7/10 Ψ1 (x∗ , Y ) + · · · . (253)
Here
x − x0 y
x∗ = , Y = .
Re−1/5 Re−1/2
By analogy with (247) we shall seek the asymptotic expansion for the pres-
sure in region 4 in the form
p = Pe0 + Re−1/5 λ0 x∗ + Re−2/5 21 λ1 x2∗ −

2a0 U02
− Re−1/2 ln Re + Re−1/2 P (x∗ , Y ) + · · · .
5πλ0
(254)
Substitution of (253) into (82), yields
∂ Ψ1 ∂ Ψ1
Vτ = Ψ00 (Y ) + Re−1/5 + ··· , Vn = −Re−1/2 + · · · . (255)
∂Y ∂x∗
We then substitute (255) and (254) into the x-momentum equation (81a).
We find that the equation (216) retains its form6 . Since the boundary
conditions (211) and (215) also remain unchanged, we can use for Ψ1 the
solution given by (219),
Y
Ψ00 − λ0 A∗ (x∗ )
Ψ1 (x∗ , Y∗ ) = Ψ00 (Y ) x∗
2 dY + . (256)
Ψ00 λ0
0
It follows from (255) and (256) that the angle made by the streamlines
with the aerofoil contour is calculated as
Y
Vn λ0 − Ψ
00 1 dA∗
ϑ= = Re−1/2
2 dY − + ··· .
Vτ Ψ00 λ0 dx∗
0
It obviously changes with Y , but the curvature of the streamlines
∂ϑ 1 d2 A∗
= −Re−3/10 + ··· (257)
∂x λ0 dx2∗
6
The induced pressure gradient ∂P ∗ /∂x∗ is simply too weak to affect the flow in region 4.
stays constant across region 4.
To complete the flow analysis in region 4 we substitute (255) and (254)
into the y-momentum equation (81b). Restricting our attention to the lead-
ing order terms, we find
∂P ! "2
= κ(x0 ) Ψ00 (Y ) .
∂Y
Here κ(x0 ) is the curvature of the aerofoil contour at point x = x0 . We see
that while the pressure P changes with Y , the pressure gradient ∂ P /∂x∗
remains constant across region 4:

∂ ∂P
= 0.
∂Y ∂x∗
The Interaction Problem In order to describe the flow behaviour in

the interaction region, one needs to solve equation (251) in region 3,
1 2 ∂ 2 Ψ∗2 ∂Ψ∗2 ∂ 3 Ψ∗2 1 dA∗ 2 ∂P ∗

2 λ0 Y∗ − λ 0 Y∗ = − A ∗ Y − , (258)
∂x∗ ∂Y∗ ∂x∗ ∂Y∗3 2
dx∗ ∗ ∂x∗
subject to the no-slip conditions on the aerofoil surface
∂Ψ∗2
Ψ∗2 = = 0 at Y∗ = 0, (259)
∂Y∗
and the matching condition (246b) with the solution in region 2a,
a20 λ0 a20 9
Ψ∗2 = x∗ Y∗5 + Y +
5! 8! ∗ (260)
1 $ %
+ b0 (−x∗ )7/4 + k1 b1 (−x∗ )−1/4 Y∗2 + · · · as x∗ → −∞.
2
As the pressure gradient ∂P ∗ /∂x∗ in (258) is not known, we also need
to solve equation (242) in region 5,

∂ 2 ∂p∗1 ∂ 2 ∂p∗1
+ = 0. (261)
∂x2∗ ∂x∗ ∂y∗2 ∂x∗
The solution has to satisfy the attenuation condition (243) in the “far-field”,
∂p∗1
→0 as x2∗ + y∗2 → ∞, (262)
∂x∗
and a matching condition with the solution in region 4, which will now be
deduced.
388 A. Ruban
We know that the curvature of the streamlines in region 4 is given by

(257),
∂ϑ 1 d2 A∗
= −Re−3/10 + ··· (263)
∂x λ0 dx2∗
It follows from (237a), (237b) that at the “bottom” of region 5 the angle
made by the streamlines with the aerofoil surface is calculated as

Vn v∗ (x∗ , 0)
ϑ= = Re−1/2 1 + ··· . (264)
Vτ y∗ =0 U0
Differentiation of (264) with respect to x results in

∂ϑ 1 ∂v1∗
= Re−3/10 + ··· . (265)
∂x U0 ∂x∗ y∗ =0
Comparing (265) with (263), we can conclude that the sought matching
condition is written as

∂v1∗ U 0 d2 A ∗
=− .
∂x∗ y∗ =0 λ0 dx2∗
It remains to reformulate this condition for the pressure gradient. For this
purpose the second equation in (240) is used. Differentiating the equation
with respect to x∗ and setting y∗ = 0 we find
∗
∂ ∂p1 U 2 d3 A ∗
= 0 at y∗ = 0. (266)
∂y∗ ∂x∗ λ0 dx3∗
This completes the formulation of the viscous-inviscid interaction prob-

lem. In order to describe the flow in the interaction region, we need to solve
simultaneously equation (258) subject to the boundary conditions (259),
(260) and equation (261) subject to the boundary conditions (262), (266).
Before solving the interaction problem, we shall express it in canonic
form. In the viscous sublayer (region 3) we introduce a new unknown func-
tion Ψ2 such that
λ0 a20 9 a20
Ψ∗2 = Ψ2 + Y + x∗ Y∗5 + 12 B∗ (x∗ )Y∗2 + G∗ (x∗ )Y∗ ,
8! ∗ 5!
with
A2∗ − a20 x2∗ − 2k1 a0 a1
G∗ (x∗ ) = ,
2λ0
and perform the affine transformations

11/10 9/5 3/5 4/5 7/5 6/5
a0 U0 a0 U0 ∂P ∗ a U
Ψ2 = 17/10
Ψ̄, A∗ = 1/5
A, = 0 4/50 R(X),
λ0 λ0 ∂x∗ λ0
6/5 8/5 4/5 1/5
a0 U0 U0 U0
G∗ = 7/5
G(X), x∗ = 2/5 1/5
X, Y∗ = 1/10 3/10
Ȳ .
λ0 a0 λ0 a0 λ0
As a result the equation (258) and boundary conditions (259), (260) assume
the form
1 2 ∂ 2 Ψ̄ ∂ Ψ̄ ∂ 3 Ψ̄
2 Ȳ − Ȳ = − R(X), (267a)
∂X∂ Ȳ ∂X ∂ Ȳ 3
∂ Ψ̄
Ψ̄ = 0, = −G(X) at Ȳ = 0, (267b)
∂ Ȳ
Ψ̄ → 0 as X → −∞. (267c)
Here
G(X) = 1
2 A2 − X 2 + 2a , (268)
with parameter a defined as
2/5
(−a1 )λ0
a = k1 1/5 8/5
. (269)
a 0 U0
In the new variables the condition (250) is written as
A(X) = (−X) − a(−X)−1 + · · · as X → −∞, (270)
which means that G(X) tends to zero as X → −∞.

In the upper tier (region 5) the affine transformations are written as
7/5 6/5 4/5 4/5
∂p∗1 a U U0 U0
= 0 4/50 r(X, ȳ), x∗ = 2/5 1/5
X, y∗ = 2/5 1/5
ȳ.
∂x∗ λ0 a0 λ0 a0 λ0
They render the equation (261) and boundary conditions (262), (266) in the
form
∂2r ∂ 2r
2
+ 2 = 0, (271a)
∂X ∂ ȳ
r → 0 as X 2 + ȳ 2 → ∞, (271b)

∂r dQ
= , (271c)
∂ ȳ ȳ=0 dX
390 A. Ruban
where Q(X) denotes the second derivative of function A(X),

Q(X) = A (X).
Finally, taking into account that the pressure gradient does not change
across the middle tier (region 4), we can write

r = R(X). (272)
ȳ=0
Notice that being written in the new variables, the equations (267), (271)
describing the flow in the interaction region, involve a single controlling
parameter (269), which measures the deviation of the angle of attack from
its critical value; as a1 < 0, the parameter a increases with k1 , and hence,
with the angle of attack.
We shall now try to solve the equations of viscous-inviscid interaction
with the help of Fourier Transforms. We start with the inviscid flow in
region 5. It is easily seen that the Fourier transformation renders (271) in
the form
d2 r̆
− k 2 r̆ = 0, (273a)
dȳ 2
r̆ = 0 at ȳ = ∞, (273b)
dr̆
= ik Q̆ at ȳ = 0. (273c)
dȳ
Here r̆(k, ȳ) stands for the Fourier Transform of the function r(X, ȳ) defined
as
∞
r̆(k, ȳ) = r(X, ȳ)e−ikX dX,
−∞
and Q̆(k) denotes the Fourier Transform of the function Q(X).

The general solution of the equation (273a) has the form
r̆ = C1 ekȳ + C2 e−kȳ . (274)
In order to satisfy the boundary condition (273b) one has to set C1 = 0 for
all k > 0, and C2 = 0 for all k < 0. Therefore, we shall write (274) in the
form
r̆ = Ce−|k|ȳ . (275)
Substitution of (275) into (273c) yields
ik
C =− Q̆. (276)
|k|
It remains to substitute (276) back into (275), and we will have the solution
of the problem (273) in the form
ik
r̆ = − Q̆e−|k|ȳ . (277)
|k|
Let us now consider the flow in the viscous sublayer (region 3). The
Fourier transformation converts the problem (267) for this region into
1 2 dΨ̆ d3 Ψ̆
2 ik Ȳ − ik Ȳ Ψ̆ = − R̆, (278a)
dȲ dȲ 3
dΨ̆
Ψ̆ = 0, = −Ğ(k) at Ȳ = 0. (278b)
dȲ
Here Ψ̆(k, Ȳ ), Ğ(k) and R̆(k) are the Fourier Transforms of functions Ψ̄,
G(X) and R(X), respectively.
Unlike (188a) the equation (278a) is not homogeneous. Therefore, its
general solution is written as
Ψ̆ = C1 ψ21 + C2 ψ22 + C3 ψ23 + ψ2p , (279)
where ψ2p is a particular solution of the equation (278a) and ψ21 , ψ22 and
ψ23 are the complementary solutions of the homogeneous part of (278a).
These may be chosen such that

ψ21 (0) = 1, ψ21 (0) = 0, ψ21 (0) = 0,

ψ22 (0) = 0, ψ22 (0) = 1, ψ22 (0) = 0,

ψ23 (0) = 0, ψ23 (0) = 0, ψ23 (0) = 1,

ψ2p (0) = 0, ψ2p (0) = 0, ψ2p (0) = 0. (280)
It then follows from the boundary conditions (278b) that C1 = 0, and

C2 = −Ğ(k). Factor C3 remains arbitrary, and the third complementary
solution is easily seen to be ψ23 = 12 Ȳ 2 . Taking these into account, we can
write (279) as
Ψ̆ = −Ğ(k)ψ22 + 12 C3 (k)Ȳ 2 + ψ2p . (281)
The behaviour of the function ψ22 (Ȳ ) has been already studied in detail.
In particular, we found that its asymptotic behaviour at the outer edge of
region 3 is given by (204), namely,
5/4
Γ(5/4) 4 1 1/2 2
ψ22 (Ȳ ) = − √ i π 1/4 sin 2 iΩ Ȳ − 38 π + · · ·
2 πe 4 Ω iΩ1/2 Ȳ 2 (282)
as Ȳ → ∞,
392 A. Ruban
with Ω = 12 ik.
Now our task will be to study the particular solution ψ2p (Ȳ ). Function
ψ2p (Ȳ ) satisfies the equation (278a). If we set Ȳ = 0 in this equation, we
will find that
d3 ψ2p
= R̆. (283)
dȲ 3
Integration of (283) with the boundary conditions (280) yields
ψ2p (Ȳ ) = 16 R̆Ȳ 3 . (284)
We then use (284) to calculate the left hand side in (278a), and perform the
integration again. We find
2
ψ2p (Ȳ ) = 16 R̆Ȳ 3 +
ik R̆Ȳ 7 .
7!
The above procedure may be repeated as many times as one wishes, leading
to a conclusion that the power series of function ψ2p should be sought in
the form
∞

ψ2p (Ȳ ) = 16 R̆Ȳ 3 + cn+1 Ȳ 7+4n . (285)
n=0
Alternatively, we can write (285) as
∞

ψ2p (Ȳ ) = cn Ȳ 3+4n . (286)
n=0
Using (286) on the left hand side of the equation (278a), and (285) on the
right hand side, results in the following recurrent equation for the coefficients
cn ,
ik n + 14
cn+1 = cn ,
32 n + 4 n + 54 n + 32
7
which has to be solved starting with c0 = 16 R̆. The solution is written as

n
R̆ ik Γ(3/2)Γ(7/4)
cn = . (287)
6 32 Γ(n + 3/2)Γ(n + 7/4)(4n + 1)
It remains to substitute (287) back into (286), and we will have7
√ ∞ n
π ik Ȳ 4n+3
ψ2p (Ȳ ) = R̆Γ 34 . (288)
16 n=0
32 Γ(n + 3/2)Γ(n + 7/4)(4n + 1)
7
Here it is taken into account that
√
Γ(3/2) = 12 π and Γ(7/4) = 3
4
Γ(3/4).
We shall now show that the function ψ2p (Ȳ ) may be expressed in terms
of the Struve function Hν (z). It is known that the Struve function may be
represented by the power series
∞ 2n+ν+1
(−1)n z
Hν (z) = , (289)
n=0
Γ(n + 3/2)Γ(n + ν + 3/2) 2
which converges for all finite values of z. It is also known that

2
Hν (z) = sin z − 12 νπ − 14 π + · · · as |z| → ∞, (| arg z| < π).
πz
(290)
We choose the order of the Struve function to be ν = 14 , and then it
easily follows from (289) that
∞
2n+3/2 z 7/4
(−1)n z z 2
= H1/4 (z)dz.
n=0
Γ(n + 3/2)Γ(n + 7/4)(4n + 1) 2 8 z
0
(291)
Substituting (290) into the integral in (291) and performing the integration
by parts with
9/4
2
u= , dv = sin z − 38 π dz,
z
13/4
9 2
du = − dz, v = − cos z − 38 π ,
8 z
we find that
∞
2n+3/2
(−1)n z
=
n=0
Γ(n + 3/2)Γ(n + 7/4)(4n + 1) 2
5/4 (292)
1 2
=− √ cos z − 38 π + · · · as |z| → ∞.
4 π z
If we now set
Ω = 12 ik, z = 12 iΩ1/2 Ȳ 2 , (293)
then the equation (292) will turn into
∞ n
ik Ȳ 4n+3
=
n=0
32 Γ(n + 3/2)Γ(n + 7/4)(4n + 1)
5/4 (294)
2 4
1 1/2 2 3
= −√ i3π cos 2 iΩ Ȳ − 8 π + ··· .
πe 4 Ω3/4 iΩ1/2 Ȳ 2
394 A. Ruban
It remains to substitute (294) into (288), and we can conclude that

5/4
R̆ Γ(3/4) 4
ψ2p (Ȳ ) = − 3 cos 1
2
iΩ1/2 Ȳ 2 − 38 π + · · · (295)
8ei 4 π Ω3/4 iΩ1/2 Ȳ 2
as Ȳ → ∞ . To progress further we need to define more precisely the way

an analytic branch of the square root of 12 ik is introduced in (293). For our
purposes it is convenient to make a branch cut in the complex plane k along
positive imaginary semi-axis, as shown in Figure 22(a). Expressing k in the
form k = reiϑ , we find that for k lying on the real axis, equations (293) give
⎧
⎨ 1 k ei 34 π Ȳ 2 if k > 0,
z = 22
⎩ 1 (−k) ei 4 π Ȳ 2 if k < 0.
1
2 2
This means that as Ȳ tends to infinity the corresponding point in the com-
plex z-plane runs to infinity along one of the rays shown in Figure 22(b);
which one depends on the sign of k. On both rays {iz} < 0. Consequently,
k z
k
k>0 k<0
r
ϑ
a) Complex plane of the Fourier b) Complex plane z

variable k.
Figure 22. Graphical illustration of equations (293).
⎫
ei(z− 38 π) − e−i(z− 38 π) ⎪
1 3 ⎪
⎪
sin z − 38 π = = − e−i(z− 8 π ) + · · · ,⎬
2i 2i (296)
ei(z− 38 π) + e−i(z− 38 π) ⎪
⎪
3
cos z − 8 π =
1 3
= e−i(z− 8 π) + · · · ⎪
⎭
2 2
as Ȳ → ∞.
Substituting (296) into (295) and (282), and then into (281), we find
that at the outer edge of region 3 (see Figure 21),
5/4
1 Γ(5/4) Γ(3/4) 4 3
Ψ̆ = − i 3 π √ Ğ(k)+ 1/2
R̆(k) 1/2 2
e−i(z− 8 π) +· · · .
4e 4 Ω 1/4 π 4Ω iΩ Ȳ
In order to ensure that the solution in region 3 can be matched with the
solution in region 4, we have to suppress the exponential growth of the
function Ψ̆, which is done by setting
Γ(5/4) Γ(3/4)
√ Ğ(k) + R̆(k) = 0. (297)
π 4Ω1/2
The above equation relates (in the Fourier space) the pressure gradient
in region 3 with the function A(X). The second relationship between these
functions is given by the solution (277) in region 5. The solutions in regions 3
and 5 are linked to one another through the equation (272). Writing this
equation in terms of the Fourier Transforms, and using (277), we find that

ik
R̆(k) = r̆ = − Q̆(k). (298)
ȳ=0 |k|
The Fourier Transform of the pressure gradient, R̆(k), may be easily elimi-
nated from (298) and (297), leading to
√
π (ik)1/2
Ğ(k) = Λ Q̆(k). (299)
2 |k|
Here Λ is a constant given by
Γ(3/4)
Λ= √ .
2Γ(5/4)
We shall now try to express (299) in physical variables. Applying the
inverse Fourier transformation to the equation (299), renders it in the form8
√ ∞
1
2 2

π 1 (ik)1/2
2 A − X + 2a = Λ Q̆(k)eikX dk. (300)
2 2π |k|
−∞
Here Q̆(k) is the Fourier Transform of A :

∞
Q̆(k) = A (ξ)e−ikξ dξ. (301)
−∞
8 1
` ´
Recall that Ğ(k) is the Fourier Transform of the function G(X) = 2
A2 − X 2 + 2a .
396 A. Ruban
If we substitute (301) into the integral on the right hand side of (300) and
change the order of integration, we will find that
∞
2 2 Λ
A − X + 2a = √ A (ξ)I(X, ξ)dξ, (302)
2 π
−∞
where
∞
(ik)1/2 ik(X−ξ)
I(X, ξ) = e dk. (303)
|k|
−∞
It may be shown (see Exercise 4) that

0 if ξ < X,
I(X, ξ) = (304)
π
2 ξ−X , if ξ > X.
Substitution of (304) into (302) results in the following integro-differential

equation for function A(X),
∞
2 2 A (ξ)
A − X + 2a = Λ √ dξ. (305)
ξ−X
X
This equation does not allow for further simplification, and should be
solved numerically. When performing the calculations one needs to use
appropriate boundary conditions. The first of these is given by (270),
A(X) = (−X) − a(−X)−1 + · · · as X → −∞, (306)
and represents the condition of matching with the solution in the boundary
layer upstream of the interaction region. It should be noted, however, that
this condition alone does not make the solution of (305) unique. In addition
to solutions with short separation bubbles, we are interested in, it also allows
for the solutions with semi-infinite separation regions. Indeed, if, to make
it simple, we choose a = 0, then the equation (305) with the boundary
condition (306) will admit the solution A = −X. The latter matches with
the smooth branch (143) of the solution to the boundary-layer equation
outside the interaction region. In order to ensure the matching with the
singular branch (146), we shall require
A(X) = X + · · · as X → ∞. (307)
For more detailed analysis of the asymptotic behaviour of A(X) at large
values of |X| see Exercise 5.
The results of the calculations are presented in Figures (23)–(26). Fig-

ure (23) shows how function A(X) changes as the parameter a is gradually
increased. It should be noted that the function A(X) plays a dual role in
the Marginal Separation theory. Firstly, through equation (257), it defines
the shape of the streamlines in region 4, and therefore, similar to the corre-
sponding function in the conventional triple-deck theory, it may be termed
the displacement function. However, in the Marginal Separation theory
it also appears to be proportional to the shear stress on the aerofoil sur-
face. Indeed, according to (245), (249) and (244) the two-term asymptotic
expansion of the stream function in the region 3 is written as
ψ = Re−13/20 61 λ0 Y∗3 + Re−16/20 21 A∗ (x∗ )Y∗2 + · · · , y = Re−11/20 Y∗ .
Consequently, the dimensionless skin friction

3/5 4/5
1 ∂ 2 ψ −1/5 −1/5 a0 U0
τ=√ = Re A (x ) = Re A(X).
Re ∂y 2 y=0
∗ ∗ 1/5
λ0
1

2
A(X)
3

6
5

4

X
Figure 23. Solutions of the equation (305) on the upper branch of the
fundamental curve: 1) a = −0.5; 2) a = 0.0; 3) a = 0.5; 4) a = 1.0; 5)
a = as = 1.139; 6) a = ac = 1.330.
Graph 2 in Figure 23 is plotted for a = 0, which corresponds to the

critical value of the angle of attack, as estimated based on the classical
boundary-layer theory. When the viscous-inviscid interaction is ignored,
the Prandtl equations yield a singular solution with the skin friction given
by (146). We see that the interaction acts to smooth out the singularity. The
398 A. Ruban
minimal skin friction is lifted, and the skin friction appears to be positive
for all values of X ∈ (−∞, ∞). For graph 5 the parameter a has been
adjusted in such a way that the minimal skin friction returns back to zero
to capture the incipience of the separation. We found that this happens
at point X = 0.406 when the parameter a reaches the value as = 1.139.
Finally, graph 6 is plotted for the critical value of the parameter ac = 1.330.
It shows a region of negative A between X = −0.566 and X = 1.605,
occupied by the separation bubble. Interestingly enough, the solution does
not exist beyond a = ac .
This important result is illustrated by Figure 23, where the so called fun-
damental curve is displayed. This curve shows the entire set of admissible
solutions of the Marginal Separation theory. It is constructed in the follow-
ing way. Given a, the solution of the boundary value problem (305), (306),
(307) yields the distribution of the shear stress A(X) along the aerofoil sur-
face. Each such solution is represented by a point on the fundamental curve,
which is obtained by taking the value of A(X) at X = 0, and plotting A(0)
against the parameter a. The numbered circles on the fundamental curve
1 A(0)
1.5
2
1
3
0.5
4
−0.5 0 0.5 1 5 ac a
as
9
8
−0.5 6
−1 7
Figure 24. The fundamental curve.
represent the corresponding graphs of Figure 23. Notice that for a = as

(graph 5) the minimal skin friction is zero, but A(0) is still positive. This
explains why the corresponding point on the fundamental curve is situated
above the abscissa.
When the parameter a reaches its critical value ac the fundamental curve
turns back to form the second branch of the solutions. In figure 25 we
compare the solutions on the upper and lower branches for a = as . The first
of these represents the flow with the separation region just about to appear;

A(X)

5

7

X
Figure 25. Comparison of the solutions on the upper and lower branches
for a = as = 1.139.
in the second the separation region is already well developed. In fact, the
length of the separation region grows monotonically as an “observer” follows
the fundamental curve from the point 5 towards the critical point 6 and then
all the way along the lower branch. This trend is demonstrated by Figure 26,
where in addition to the solution at point 7 the solutions at points 8 and 9,
that lie on a small loop on the lower branch, are shown.
Despite the parameter a is still rather large on the loop, the solution
already shows an asymptotic behaviour characteristic of small a. When
a = 0 the equation (305) admits two solutions
A = −X and A = X. (308)
Neither of these satisfy both boundary conditions (306), (307), but they
are clearly visible in Figure 26 as two major fragments of the curves 8 and
9, being connected to one another through a sharp jump in a region that
becomes progressively shorter and moves to the right as a → 0−.
Summarising the results of the above analysis, we can conclude that
according to the Marginal Separation theory the flow near the leading edge
of a thin aerofoil exhibits a hysteresis behaviour. It should be noted that
hysteresis is routinely observed in experiments with aerofoils. These are
normally conducted in such a way that the angle of attack is at first increased
slowly enough to keep the flow quasi-steady, and then, after achieving the
aerofoil stall, it is gradually decreased. What one usually observes is that
the angle of attack at which the separation region forms on the upper surface
of the aerofoil does not coincide with the angle of attack at which the flow
400 A. Ruban

9
8
A(X)

X
Figure 26. Solutions on the lower branch of the fundamental curve: 7)
a = as = 1.139; 8) a = 0.600; 9) a = 0.680.
returns back to attached form. As a result the graph of the lift force versus
the angle of attack assumes the shape of a hysteresis curve. Within the
hysteresis loop, for each value of the angle of attack two flow states become
possible. The choice between them depends on the flow history of the
development of the flow.
Still experimental observations of the short separation bubbles show that
these are formed in a smooth manner without abrupt change of the flow
field, which is only possible if the solution remains on the upper branch of
the fundamental curve when the parameter a passes through as . Of course,
the flow cannot continue to change smoothly when the parameter a passes
through the critical value ac . The non-existence of the solution to (305),
(306), (307) for a > ac suggests that the flow has to undergo a sudden
change, known from experiment as the “bubble bursting”.
In order to calculate the values of the angle of attack at the incipience of
the short separation bubble and at the bubble bursting, we need to return
to the equation (79). Combining it with (231) and (269), we find that the
scaled angle of attack α∗ = α/ε is given by
1/5 8/5 1
1 a U0 1 G(x )
α∗ = √ k0 + Re−2/5 0 2/5
a − dx . (309)
2 π x (1 − x )
(−a1 )λ0
0
Remind that constants k0 , U0 , λ0 , a0 and a1 are found by solving the

Prandtl’s equations of the classical boundary-layer theory. For an aerofoil
with parabolic nose

k0 = 1.1575, U0 = 1.286, λ0 = 0.024, a0 = 0.0085, a1 = −1.24.
(310)
Using (310) in the equation (309), renders it in the form
1
−2/5 1 G(x )
α∗ = 0.8185 + 1.4610 · Re a− dx .
π x (1 − x )
0
If we assume, to make it even more simple, that the aerofoil is symmetric

(G ≡ 0), then we will have
α∗ = 0.8185 + 1.4610 · Re−2/5 a + · · · as Re → ∞. (311)
Setting a = as = 1.139 in (311) gives the angle of attack at moment of the
formation of the short separation bubble

αs = ε 0.8185 + 1.6641 · Re−2/5 + · · · . (312)
If, instead, we choose a = ac = 1.33, then we will find that the bubble
bursting has to be expected at the angle of attack

αc = ε 0.8185 + 1.9431 · Re−2/5 + · · · . (313)
When applying these results to real flows one needs to keep in mind
that the Marginal Separation theory relies on the assumption of laminar
flow. This assumption is well justified for attached flows near the leading
edge of a thin aerofoil. The reason is that the characteristic length scale
here is not the aerofoil cord, c, but a much smaller quantity, the radius of
the aerofoil nose, r = c ε2 . Under conditions typical of aerodynamic appli-
cations, the Reynolds number Re = V∞ r/ν, calculated based on r, is not
large enough for the attached boundary layer to become turbulent. In fact,
the flow is observed to remain laminar even after the formation of the short
separation bubble. However, before the bubble bursting, the flow in the
bubble changes its character and becomes partly turbulent. The separated
flows are known to be less stable and can undergo a rather rapid laminar-
turbulent transition. In the leading edge bubble the flow is observed to
separate laminar, becomes turbulent before the reattachement. The tran-
sition has a significant effect on the behaviour of the separation bubble.
For example, when the turbulence is enhanced by introducing an additional
acoustic noise in the wind tunnel test section, it always delays the bubble
bursting. This explains why the formula (312) proves to be fairly accurate
(see Hsiao & Pauley, 1994), while the formula (313) might underestimate
the critical angle of attack αc .
402 A. Ruban
Acknowledgment: The author would like to thank S. Braun for his help
in preparing the manuscript.
Exercise 2. Simplify (135) with the help of the asymptotic expansion

(134), and confirm that the resulting expression reproduces the solution
(130) in region 2a.
Exercise 3. Consider two-dimensional steady incompressible inviscid fluid

flow near a corner point O on a rigid body surface as shown in Figure 27.
Assuming the wall deflection angle θ small, represent the solution of the
y z
θ
O x
Figure 27. Flow past a corner point of a rigid body surface.
Euler equations in the form of asymptotic expansions

u = U0 + θu1 (x, y) + · · · , v = θv1 (x, y) + · · · , p = θp1 (x, y) + · · · .
Here u, v and p are the non-dimensional velocity components and pressure;
U0 denotes the value of the tangential velocity u that would be observed in
the flow if the angle θ was zero.
Prove that f (z) = p1 + iU0 v1 is an analytic function of the complex
variable z = x + iy, and try to find the solution for f (z) near the corner in
the form
p1 + iU0 v1 = (Cr + iCi ) ln z + (Dr + iDi ) + · · · as z → 0. (314)
Set z = reiϑ in (314), and using the impermeability condition on the body
surface upstream and downstream of the corner point O,
v
= tan θ ≈ θ at y = 0, x > 0,
u
v
=0 at y = 0, x < 0,
u
show that
U02
Cr = − , Ci = 0, Di = U02 .
π
Hence, deduce that in a small vicinity of the corner point

⎫
U02
p = −θ ln r + · · · , ⎪ ⎪
⎪
⎪
π ⎪
⎪
U0 ⎬
u=θ ln r + · · · , as r → 0. (315)
π ⎪
⎪
⎪
⎪
ϑ ⎪
v = θU0 1 − + · · ·⎪
⎭
π
Exercise 4. Split the integral (303) into two
0 ∞
(ik)1/2 ik(X−ξ) (ik)1/2 ik(X−ξ)
I(X, ξ) = e dk + e dk, (316)
−k k
−∞ 0
and assume, first, that ξ < X. In this case, when calculating the first
integral, I1 ,in (316) change the contour of integration from the negative
real semi-axis (contour C1 in Figure 28a) to the positive imaginary semi-
axis, C1 . Observe that, according to Jordan’s lemma, the integral along the
k N
CR1 CR2
C1 k C2
s
C1 ϑ C2
a) Contour change for the first b) Contour change for the second
integral in (316). integral in (316).
Figure 28. Calculation of the integrals in (316) for ξ < X.
quarter-circle, CR1 , tends to zero as its radius R tends to infinity. Introduce

a new integration variable s, such that k = is, with s being the distance
404 A. Ruban
from a point on C1 to the coordinate origin, and confirm that the adopted
rule of calculating (ik)1/2 gives on C1 ,
(ik)1/2 = −is1/2 .
Hence, show that

∞
e−(X−ξ)s
I1 = −i √ ds.
s
0
When calculating the second integral, I2 , in (316), change the contour

of integration as shown in Figure 28(b). Show that
∞
e−(X−ξ)s
I2 = i √ ds,
s
0
and conclude that
I(X, ξ) = I1 + I2 = 0 for ξ < X.
For the case when ξ > X, change the contour of integration for the two
integrals in (316) as indicated in Figure 29, and show that
√
π
I1 = I2 = √ .
ξ−X
C3 C4
C3 C4
CR3 CR4
Figure 29. Calculation of the integrals in (316) for ξ > X.

Exercise 5. Notice that the right hand side of the equation (305) tends
to zero as X → ∞. This means that the first two terms of the asymptotic
expansion of A(X) at large positive values of X may be written as
a
A(X) = X − + ··· as X → ∞. (317)
X
Differentiate (317) twice, and substitute the result into the integral on the
right hand side of (305). Perform the integration using the following chain
of substitutions: ξ = Xs, then s = t2 + 1 and, finally, t = tan θ. Conclude
that
A2 − X 2 + 2a = − 43 aΛπX −5/2 + · · · as X → ∞. (318)
Now, assume that X → −∞, and split the integral on the right hand
side of (305) as
∞ −Δ Δ ∞
A (ξ) A (ξ) A (ξ) A (ξ)
√ dξ = √ dξ + √ dξ + √ dξ, (319)
ξ−X ξ−X ξ−X ξ−X
X X −Δ Δ
where parameter Δ is such that
Δ (−X), but Δ (−X)1/3 . (320)
Consider, first, the middle integral. Notice that thanks to the first in-
equality in (320) it may be approximated as
Δ Δ )
A (ξ) 1 1 (
I2 ≈ √ dξ = √ A (ξ)dξ = √ A (Δ) − A (−Δ) .
−X −X −X
−Δ −Δ
Keeping in mind that Δ is large, calculate A (Δ) and A (−Δ) using (317)
and (306), respectively. Conclude that
2
I2 = √ + ··· . (321)
−X
When evaluating the first integral in (319), calculate A (ξ) with the help
of (306). Then use the fact that (−ξ) ≤ Δ everywhere in the integration
interval, and show that
−Δ
2a dξ 4a
|I1 | ≤ √ = 3 (−X) − Δ.
Δ3 ξ−X Δ
X
406 A. Ruban
Finally, consider the third integral,

∞
A (ξ)
I3 = √ dξ. (322)
ξ−X
Δ
Use the asymptotic expansion (317) to calculate A (ξ) in (322), and intro-
duce a new integration variable t through the substitution ξ = (−X)(t2 −1).
This leads to
∞
4a dt
I3 = − .
(−X)5/2 q (t − 1)3 (t + 1)3
Δ
1+ (−X)
Notice that t + 1 > 2 everywhere in the integration interval. Hence, deduce

that
∞
4a 1 dt a 1
|I3 | ≤ 3
≈ .
5/2
(−X) 8 q (t − 1) 4(−X) Δ2
1/2
Δ
1+ (−X)
Confirm that under the conditions (320), integrals I1 and I3 are much
smaller than I2 . Substitute (321) into (319), and conclude that
2Λ
A2 − X 2 + 2a = √ + ··· as X → −∞.
−X
Exercise 6. Show that the equation (305) may be inverted to take the
form
∞
1 A2 (ξ) − ξ 2 + 2a
A (X) = 1 − √ dξ.
πΛ ξ−X
X
For this purpose introduce a new function S(X) such that S = G (X), and
deduce from the equation (299) that
2 |k|
Q̆ = √ S̆. (323)
πΛ (ik)3/2
Apply the inverse Fourier transformation to (323), and using the technique
described in Exercise 4, show that
∞
1 S(ξ)
A (X) = − √ dξ. (324)
2πΛ ξ−X
X
Now, consider the integral

∞
G(ξ)
F (X) = √ dξ, (325)
ξ−X
X
Recall that the function G(X) was introduced through the equation (268),
and it follows from (318) that G(X) = O(X −5/2 ) as X → ∞. Keeping this
in mind, apply the integration by parts to (325), and show that
∞
F (X) = −2 ξ − XS(ξ)dξ. (326)
X
Then differentiate (326) with respect to X, and deduce that

∞
S(ξ)
F (X) = √ dξ. (327)
ξ−X
X
Finally, substitute (327) into (324) and integrate the resulting equation
using the fact that, according to (317), A = 1 at X = ∞.
Bibliography
Abramowitz, M. & Stegun, I. A. 1965 Handbook of Mathematical Functions,
3rd edn. Dover Publications.
Chapman, D. R., Kuehn, D. M. & Larson, H. K. 1956 Investigation of
separated flows in supersonic and subsonic streams with emphasis on
the effect of transition NACA Rep. 1356.
Goldstein, S. 1948 On laminar boundary-layer flow near a position of sep-
aration. Q. J. Mech. Appl. Math. 1 (1), 43–69.
Helmholtz, H. 1868 Über diskontinuirliche Flüssigkeitsbewegungen. Monats-
bericht Akad. Wiss. Berlin pp. 215–228.
Hsiao, C. T. & Pauley, L. L. 1994 Comparison of the triple-deck theory,
interactive boundary layer method, and Navier-Stokes computations for
marginal separation. Trans. ASME J. Fluids Eng. 116, 22–28.
Jones, B. M. 1934 Stalling. J. Roy. Aero. Soc. 38, 753–770.
Kirchhoff, G. 1869 Zur Theorie Freier Flüssigkeitsstrahlen. J. für die Reine
und Angew. Math. 70 (4), 289–298.
Lamb, H. 1932 Hydrodynamics, 6th edn. Cambridge University Press.
Landau, L. D. & Lifshitz, E. M. 1944 Mechanics of Continuous Media.
Gostekhizdat, Moscow.
408 A. Ruban
Levi-Civita, T. 1907 Scie e leggi di resistenza. Rendconte de Circolo Matem-

atico di Palermo XXIII, 1–37.
Lighthill, J. 2000 Upstream influence in boundary layers 45 years ago. Phil.
Trans. R. Soc. London, A 358 (1777), 3047–3061.
Neiland, V. Y. 1969 Theory of laminar boundary layer separation in super-
sonic flow. Izv. Akad. Nauk SSSR, Mech. Zhidk. Gaza (4), 53–57.
Oswatitsch, K. & Wieghardt, K. 1948 Theoretical analysis of stationary
potential flows and boundary layers at high speed NACA Tech. Memo.
1189.
Ruban, A. I. 1981 A singular solution of the boundary-layer equations which
can be extended continuously through the point of zero skin friction. Izv.
Akad. Nauk SSSR, Mech. Zhid. Gaza (6), 42–52.
Ruban, A. I. 1982 Asymptotic theory of short separation bubbles at the
leading edge of a thin airfoil. Izv. Akad. Nauk SSSR, Mech. Zhid. Gaza
(1), 42–52.
Siebert, H. M. 1948 High-Speed Aerodynamics. Prentice Hall, New York.
Stewartson, K., Smith, F. T. & Kaups, K. 1982 Marginal separation. Stud.
Appl. Math. 67 (1), 45–61.
Stewartson, K. & Williams, P. G. 1969 Self-induced separation. Proc. Roy.
Soc. London A 312, 181–206.
Sychev, V. V., Ruban, A. I., Sychev, Vic. V. & Korolev, G. L. 1998 Asymp-
totic Theory of Separated Flows. Cambridge University Press.
Taneda, S. 1956 Experimental investigation of the wakes behind cylinders
and plates at low reynolds numbers. J. Phys. Soc. Jpn. 11, 302–307.
Tani, I. 1964 Low-speed flows involving bubble separations. Progr. Aero-
naut. Sci. 5, 70–103.
Werlé, H. & Gallon, M. 1972 Controle d’ecoulements par jet transversal.
aéronaut. astronaut. Aéronaut. Astronaut. 34, 21–33.
Weakly 3D effects upstream a surface
mounted obstacle in transonic flows
A. Kluwick* and M. Kornfeld*

*
Institute of Fluid Mechanics and Heat Transfer, Vienna University of
Technology, Vienna, Austria
Abstract Steady transonic flows through channels so narrow that
the classical boundary layer approach fails are considered. The
resulting viscous inviscid interaction problem for weakly three di-
mensional laminar flows is formulated for perfect gases under the
requirement that the channel is sufficiently narrow so that the flow
outside the viscous wall layers becomes two-dimensional in the lead-
ing order approximation. The behavior of the flow upstream of a
surface mounted three-dimensional obstacle will be demonstrated.
1 Problem formulation
The viscous inviscid interactions of steady weakly three-dimensional tran-
sonic flows in narrow channels are considered which are triggered, for ex-
ample, by a shallow deformation of the channel walls. Using asymptotic
analysis for large Reynolds number Re = ũr L̃/ν̃ 1 Kluwick and Git-
tler, assuming two dimensional steady flows of a perfect gas, showed that
a consistent interaction theory can be formulated in which the flow inside
the inviscid core region is almost one-dimensional, A. Kluwick [2001]. The
former theory can be extended to the weakly three dimensional case if the
heights H̃ and h̃ of the channel and the surface mounted obstacle are of
orders 3 L̃ and 7 L̃ and if the length ΔX and width ΔZ of the obstacle are
of orders 3 L̃ and 2 L̃ with = Re−1/12 1. Here ũr , L̃ and ν̃ denote
the flow velocity in the core region just upstream of the local interaction
region, a characteristic length associated with the unperturbed boundary
layer adjacent to the channel wall and a reference value of the kinematic
viscosity. As in A. Kluwick [2001] the field quantities inside the inviscid
core region do not depend on the distance measured perpendicular to the
channel wall in leading order resulting, however, in a two-dimensional rather
than one-dimensional flow behavior.
The interaction region exhibits a triple deck structure, where as in the clas-
sical triple deck theory, e.g. see Stewartson [1974], the role of the main deck
© CISM, Udine 2010
410 A. Kluwick and M. Kornfeld
is to transfer the displacement effects excerted by the lower deck unchanged

to the upper deck and to transfer the resulting pressure disturbances again
unchanged back to the lower deck. Here, the fluid motion is governed by
a weakly three dimensional and incompressible form of the boundary layer
equations
∂U ∂V
+ = 0,
∂X ∂Y
∂U ∂U ∂P ∂2U
U +V =− + , (1)
∂X ∂Y ∂X ∂Y 2
∂W ∂W ∂P ∂ 2W
U +V =− + ,
∂X ∂Y ∂Z ∂Y 2
where (X, Y, Z), (U, V, W ) and P denote Cartesian coordinates parallel and
perpendicular to the channel wall and the lateral direction, the correspond-
ing velocity components and the pressure. All quantities are suitable scaled.
The boundary conditions include the no slip condition on the channel walls,
the requirement that the unperturbed velocity profile is recovered in the
limit X → −∞ and a matching condition between the lower and main deck
for large Y
Y = S(X, Z) : U = V = W = 0, X → −∞ : U = Y,

1 X ∂P (2)
Y → ∞ : U = Y + A(X, Z), W =− dζ = 0,
Y −∞ ∂Z
where A(X, Z) denotes the perturbation of the displacement thickness caused

by the interaction process. The flow in the upper deck is a quasi planar flow
weakly perturbed by the boundary layer displacement. As a result, pres-
sure disturbances resulting from the boundary layer displacement can be
calculated from elementary properties of the massflux velocity relationship
and the well known leading order approximation between the pressure and
velocity disturbances.
Taylor series expansions for |M∞ − 1| 1 up to second order and substi-
tution of the scaled quantities used in equation (1) and (2) then yields

2 1 X ∂wu1 ∂wu1 |K| ∂P
P +2sign(K)P −Λ A(X, Z) + dX̄ = 0, =− ,
f −∞ ∂Z ∂X 2Γ ∂Z
(3)
where Λ represents a transonic similarity parameter, Γ denotes the funda-
mental derivative of gasdynamics, e.g. see Kluwick [1993], wu1 the lateral
velocity component in the upper deck, f a scaling parameter and K is
Weakly 3D Effects in Transonic Flows 411
5
4.5
Y 3.5
upper deck
3
H ∼
3
Y
2.5
S(X,Z) −Z 6 2
7 1.5
2
1
main deck
0.5
lower deck
X 0
-1.5 -1 -0.5
X
0 0.5 1 1.5
L=1 3
Figure 2. Stream lines and separa-
Figure 1. Triple deck structure of
tion bubble through the centerline
interaction region; := Re−1/12 .
of the hump
proportional to the difference between the Mach number M∞ in the unper-

turbed core region and its critical value M = 1. For K → ±∞ one recovers
the cases of purely sub- and supersonic flow, respectively.
2 Results
For the numerical solution of the boundary layer equations (1) to (3) a
pseudo spectral method is used. Figure 3 shows the streamlines (dashed
lines) as well as the separation stream line (solid line) in a cross section
through the √ center line of the surface mounted hump (Z = 0) with S(X, Z) =
h · cos2 ( π2 X 2 + Z 2 ) and a height h = 2.25 for Λ = 2.5, K = 1 (subsonic
flow). The analytical separation angle ϑ (dashed dotted line), obtained by
Oswatitsch [1980], at the lee side of the hump matches the angle of the
separation streamline at the wall quite well.
For a constriction with h = 1 (to avoid the above mentioned effect of flow
separation) the pressure perturbation increases with decreasing absolute
values |K| under both, sub- and supersonic flow conditions, see Figure 4.
In the subsonic case, if |K| is small enough, i.e. the flow is close enough to
the point of transition P = −1, the flow exhibits a local supersonic region in
the upper deck. There is no corresponding local subsonic region in the core
region, i.e a pressure perturbation greater than P = 1, under supersonic
flow conditions for this surface perturbation.
Another remarkable effect is the phenomenon of upstream influence. For su-
personic flow conditions a first mathematical explanation of this effect was
given by Lighthill [1953]. For the two-dimensional subsonic case A. Kluwick
[to appear 2010] showed, that there is strictly no upstream influence. In
contrast to these results one finds an upstream influence in the three-
0.2
0.4
0.18 analytical K = −1 K = −2
separation angle 0.2
0.16 K = −10
0.14
0
0.12
separation streamline -0.2
(numerical result)
Y
P
0.1
-0.4
0.08
-0.6
K = −0.1
K = −0.001
0.06
ϑ -0.8
0.04
0.02 -1 K →0
0 -1.2
0.55 0.56 0.57 0.58 0.59 0.6 0.61 0.62 0.63 0.64 0.65 -10 -5 -1 0 1 5 10 15 20
X X
Figure 3. Stream lines near the sep- Figure 4. Pressure perturbation at

aration point and separation angle. Z = 0 for the supersonic case.
also under subsonic flow conditions. A deeper insight into the effect of
upstream influence is gained through studying the linearized problem in
spectral space. The pressure perturbation in spectral space is found to be
sign(K) Λ2 k 2
P ∗∗ = 1 2
S ∗∗ (4)
−k 2 − sign(K) |K| l + sign(K) Λ2 γ −4/3 (ik)7/3
with γ −4/3 = − 3Ai1 (0) , where Ai (0) denotes the first derivative of the
Airy function and k, l the spectral variables corresponding to the physical
variables X and Z, respectively. Investigation of the poles of the pressure
perturbation (4) shows that there is at least one pole causing an upstream
influence even in the 3D subsonic case. Consequently, the upstream behav-
γ4
ior for −X 1 and Z = 0 is found to be P ∼ − X1 e8 Λ3 X for the supersonic
case K < 0 and P ∼ X12 for the subsonic case K > 0. In the supersonic
case the pressure decays algebraic-exponentially with the same exponent as
in the two-dimensional case, which is not surprising since a weakly three-
dimensional problem is studied and solutions deviating only slightly from
strictly two-dimensional solutions are expected. In the subsonic case an al-
gebraic decay, which is typical for the decay of perturbations under subsonic
conditions, is found.
Both analytical results match the numerical results, found upstream of a
perturbation caused by a surface mounted hump, cf. Figure 6 and Figure
5, quite well. Following the idea of Lighthill to formulate a free interaction
theory under supersonic conditions we are looking for solutions of the form
1 8 γ 43 X 1 γ4
U ∼Y + e Λ cos(β(X)Z)Ũ (Y ), V ∼ − e8 Λ3 X cos(β(X)Z)Ṽ (Y ),
X X
1 8 γ 43 X 1 γ4
P ∼ − e Λ cos(β(X)Z), A ∼ + e8 Λ3 X cos(β(X)Z)Ũ (∞),
X X
(5)
Weakly 3D Effects in Transonic Flows 413
-2
10 -2
10
-3
10
10
-3 2D; K < 0
-4
10
-4
10 -5
10
log(P)
3D; K > 0
log(P)
-6
10
-5
10
3D; K < 0
-7
10
-6
10
-8
10
-9
10
-7
10
1 0
-40 -35 -30 -25 -20 -15 -10 -5 -1 0
-10 -10
log(X) X
Figure 5. Decay of the effect of the Figure 6. Decay of the effect of the
upstream influence for K > 0 upstream influence for K < 0
suggested by the previous results. Evaluating the governing equations with

the given ansatz one finally obtains

1 γ 4 |K|
β(X) ∼ √ 8 3 . (6)
3 Λ |X|
Consequently, the three-dimensional structure of the perturbations is get-
ting weaker as −X increases, which is confirmed by numerical results. In
the limiting case |K| → 0 a two-dimensional behavior with no upstream
influence is obtained.
With the above discussed results of the effect of upstream influence there
arises the question if there exists a weakly three-dimensional regularized
shock profile similar to the two-dimensional case, e.g. a regularized shock
with a curved shock front. Since the weakly three-dimensional problem
covers the two-dimensional shock solution each additional weakly three-
dimensional solution would cause a non-uniqueness. In this case the as-
pects of the stability of the solutions as well as the transition between these
solutions are of interest.
Acknowledgement: This work has been financed by the Austrian Sci-
ence Fund in the framework of the WK Differential Equations.
Bibliography
G. Meyer A. Kluwick. Shock regularisation in dense gases by viscous inviscid
preparation. J. Fluid Mech., 664:473–507, 2010.
Ph. Gittler A. Kluwick. Transonic laminar interacting boundary layers in
narrow channels. ZAMM, 81:473–474, 2001.
A. Kluwick. Transonic nozzle flow of dense gases. J. Fluid Mech., 247:
661–699, 1993.
M.J. Lighthill. On boundary layers and upstream influence. Proc. Roy. Soc.
A, 217, pages 344–356, 1953.
K. Oswatitsch. Contributions to the Development of Gasdynamics. Vieweg,

1980.
K. Stewartson. Multistructured boundary layers on flat plates and related
bodies. Adv. Appl. Mech., 14:145–239, 1974.
Self-Similar Blow-up Structures in Unsteady
Marginally Separated Flows
Mario Aigner∗ and Stefan Braun

Institute of Fluid Mechanics and Heat Transfer, Vienna University of Technology
Abstract In the present study we consider unsteady three-dimen-

sional marginally separated boundary layer flows in the limit of high
Reynolds numbers. Special emphasis is placed on solutions which
blow up within finite time. In analogy to strictly two-dimensional
flows the blow-up profile is self-similar and numerical results strong-
ly suggest its uniqueness. Furthermore, the leading order blow-up
structure is generic in the sense that it determines the terminal form
of both local and global three-dimensional flows.
1 Fundamental Equations
For a comprehensive description of the underlying theory based on matched
asymptotic expansions the reader is referred to Ruban (2010) of this mono-
graph and to Sychev et al. (1998) for the case of steady planar flows. Ex-
tensions of the theory to incorporate transient and three-dimensional effects
can be found in Smith (1982), Ruban (1983), Duck (1990) and Braun &
Kluwick (2004), respectively. Our main concern is the investigation of the
Cauchy problem associated with the evolution of the displacement function
or equivalently the local wall shear stress A = A(x, t) and A = A(x, z, t)
governed by the fundamental equations
A2 − x2 + Γ = λJ∗ [∂x2 A] − γI[∂t A] + g(x, t) for x ∈ R, t ≥ 0, (1)
for a strictly planar and
A2 −x2 +Γ = − 2π
λ
JK[ΔA]−γI[∂t A]+g(x, z, t) for (x, z) ∈ R2 , t ≥ 0, (2)
for a local three-dimensionally perturbed two-dimensional boundary layer

flow. Here Δ denotes the Laplace operator in spatial coordinates and x,
z, t, Γ, λ, γ denote non-dimensionalised, suitably scaled coordinates in
∗
This work was funded by the Austrian Science Fund (FWF project P21426-N22).
© CISM, Udine 2010
416 M. Aigner and S. Braun
stream- and spanwise direction, the time, a control parameter and positive
constants. The functions g in (1) and (2), which remain bounded ∀t ≥ 0,
account for flow control devices which provide physically meaningful initial
conditions for A at t = 0, see Braun & Kluwick (2004). Also, we have
formally introduced the following integral operators
∞
∗ 1
J [f ](x, ·) = f (ξ, ·)dξ,
(ξ − x)1/2
x
x x
1 1
J[f ](x, ·) = f (ξ, ·)dξ, I[f ](x, ·) = f (ξ, ·)dξ,
(x − ξ)1/2 (x − ξ)1/4

−∞ −∞
x−ξ
K[f ](x, z) = f (ξ, η)dξdη,
((x − ξ) + (z − η)2 )3/2
2
R2
(3)
sometimes referred to as fractional derivatives, Weyl operators or Riesz
potentials.
Remark 1.1. If we consider problem (1) in the stationary case (i.e. ∂t A ≡ 0
and g ≡ 0) the resulting equation coincides exactly with the fundamental
problem derived in Ruban (2010) of this monograph (cf. equation (305)
therein, where x = X, Γ = 2a and λ = Λ).
It is well known (see Smith (1982) and Duck (1990)) that solutions to equa-
tions (1) and (2) may blow up at a finite time t = ts > 0. In this short
treatise we focus on the terminal structure of these solutions as t → ts (see
section 2).
Remark 1.2. Matching to the classical boundary layer flow requires the
far-field behaviour for equations (1) and (2) to be
A(x, ·) ∼ |x| as x → ±∞ and A(x, z, ·) ∼ |x| as x → ±∞ (4)
whereas A(x, z, ·) stays at least bounded as z → ±∞.
Remark 1.3. Duck (1990) considered a fully three-dimensional incompress-
ible marginally separated boundary layer flow in a line of symmetry which
led to a different z-dependent left hand side than in (2). As in the usual ap-
proach use was made of the interaction law, which relates the pressure P to
the displacement function A via P = K[∂x A]. In this case the fundamental
equation reads (omitting positive coefficients)
x
2 2 2 2
A + γ(z, t) − μ (βx + z ) = J[∂x P + ∂z2 P dξ] − I[∂t A] (5)
−∞
Self Similar Blow-up Structures in Unsteady Flows 417
for (x, z) ∈ R2 , t ≥ 0 and with the far-field behaviour A(x, z, ·) ∼ μ(βx2 +

z 2 )1/2 as x2 + z 2 → ∞. Here μ is a lengthscale parameter and β takes into
account the aspect ratio of the elliptic domain. The function γ is analogous
to the control parameter Γ in equation (2) and remains bounded ∀t ≥ 0.
Remark 1.4. Substituting the interaction law into the fundamental equa-
tion (5) and integrating by parts shows the right-hand side operators in (2)
and (5) to be equivalent. We thus obtain that the only difference in the fun-
damental equations between a locally and globally three-dimensional flow
studied by Duck (1990) lies in the z-dependent left-hand side and the far-
field behaviour.
2 Finite Time Blow-up and its Self-Similarity

Under certain conditions there is strong numerical evidence – as demon-
strated in Smith (1982) and Duck (1990), for a more recent investigation
see also Scheichl et al. (2008) – that solutions to equations (1), (2) and (5)
only exist within a finite time interval 0 t < ts . As the blow-up time is
approached, the displacement function A becomes unbounded at a single
point xs and (xs , zs ) (in case of three-dimensional flow), respectively. An
order of magnitude estimate, performed in Smith (1982) and Duck (1990),
suggests the appropriate similarity variables as t → ts
τ = |ts − t|, x̄ = x − xs = τ 4/9 x̂, z̄ = z − zs = τ 4/9 ẑ,
(6)
A(x, z, t) = τ −2/3 Â(x̂, ẑ) + o(τ −2/3 ).
Substitution of (6) into (1) and taking the formal limit τ → 0 yields
Â2 = λJ∗ [Â ] − 23 γI[Â + 23 x̂Â ] x̂ ∈ R, (7)
for the planar flow case, where (·) denotes the derivative with respect to x̂
and Â(x̂) ∼ c± |x̂|−3/2 as x̂ → ±∞, c± = const. ∈ R, arising from matching
with the outer region where A remains an O(1) quantity. A detailed inves-
tigation of the far-field behaviour is presented in Scheichl et al. (2008).
Most interestingly, the scalings (6) lead to the same terminal structure for
both fundamental equations, (2) and (5), initially describing different (local
and global) three-dimensional marginally separated flows. Their common
blow-up profile Â(x̂, ẑ) then is governed by
Â2 = − 2π
λ
JK[ΔÂ] − 23 γI[Â + 23 (x̂∂x̂ + ẑ∂ẑ )Â] (x̂, ẑ) ∈ R2 . (8)
Remark 2.1. As indicated above, the singular behaviour of A as t → ts
is restricted to a single point (xs , zs ) and hence everywhere else A remains
bounded. Thus, by matching Â with A at |(x̄, z̄)| = O(1), and using polar
coordinates (x̂, ẑ) → (r̂, φ̂) the far-field condition for (8)
Â = Â(r̂, φ̂) ∼ c(φ̂) r̂−3/2 as r̂ → ∞ (9)
can be deduced, where c = c(φ̂) is assumed to be smooth in some sense.

This is in agreement with the far-field behaviour given in Duck (1990), such
that equations (2) and (5) evolve into the same blow-up structure.
Remark 2.2. Let Â = Â(x̂) be independent of ẑ, then equations (1) and
(2) are equivalent (with the same far-field condition). In other words, the
fundamental equation for a planar flow can be considered as a special case
of the locally three-dimensional problem. This also holds for the blow-up
profile (cf. equations (7) and (8)).
Remark 2.3. It should be mentioned, that because of (9) Â can not be

expected to be in L1 (R2 ) but in L2 (R2 ) and hence the Fourier-Transform of
Â does not exist in the usual sense.
Remark 2.4. A solution of (8) exhibits several symmetry and invariance

properties: the change of variables z → −(z − z0) (reflection and translation
invariance), t− ts → −(t− ts − t0 ) (translation and time-reversal invariance)
leaves the equation unchanged, where z0 and t0 can be chosen arbitrarily.
3 Numerical Solution
A first attempt to gain insight into the behaviour of equations (1) and (7)
can be found in Smith (1982), which has been extended by Scheichl et al.
(2008) to a more detailed survey. Therein a finite differencing scheme to
solve equations (1) and (7) is presented. Performed numerical experiments
give strong evidence for the existence and uniqueness of a non-trivial solu-
tion of (7).
To solve the singular, nonlinear, homogeneous, partial integro-differential
equation (8) numerically we follow the approach of Scheichl et al. (2008)
for a discretization of the singular integrals appearing in (3). A nontrivial so-
lution is sought in the form Â(x̂, ẑ) = Â(x̂∗ , ẑ ∗ )φ(x̂, ẑ), where A(x̂∗ , ẑ ∗ )
= 0
for a fixed (x̂∗ , ẑ ∗ ) (and since we want Â to be smooth in some sense, this
is true for a whole neighbourhood). Thus φ
≡ 0 with φ(x̂∗ , ẑ ∗ ) = 1.
Another key for a stable scheme is to find an appropriate, bijective map
ψ : C → R2 , where C ⊂ R2 is compact, such that one can impose bound-
ary conditions Â(ψ(u)) = 0 for u ∈ ∂C. The applied methods lead to an
inhomogeneous system (in the sense, that the null vector is not a possible
Self Similar Blow-up Structures in Unsteady Flows 419
solution) for the (discretized) unknown φ, which is solved using a Newton

algorithm. The results are depicted in Figures 1 and 2 (calculated on a
50×40 x̂-ẑ grid). Most interestingly, the behaviour of Â at the line of

4

3
2

1

x̂
Figure 1. The solution Â = Â(x̂, 0) of (8) (with dots) in comparison to

that of (7) Â = Â(x̂) (solid line), Scheichl et al. (2008).
symmetry ẑ = 0 is similar to the solution Â(x̂) obtained in Scheichl et al.

(2008) (due to the Fredholm operator K this can not be seen directly from
equation (8)). From Figure 2 one would claim the z-behaviour to be simi-
lar to a function g = g(ẑ) with algebraic decay and in further consequence
that the solution can be written as Â(x̂, ẑ) = Â(x̂, 0)g(ẑ). Again this is not
possible, since K and the ẑ-derivative in the argument of I do not allow the
equation to be separated into a pure x̂ and ẑ dependent part.
Remark 3.1. The convergence of the used Newton method depends quite
sensitively on the map ψ, due to which the x̂-ẑ grid on the whole R2 is non-
equidistant. As a consequence, increasing the grid points does not directly
imply a more accurate solution. One additionally has to use an accordingly
modified ψ.
Despite all the faced difficulties, the numerical solution still provides a good
first insight into equation (8) and gives further ideas for some analytical
investigation.

Â
4

3

2

1

ẑ
Figure 2. The solution Â = Â(x̂i , ẑ) of (8) for i = 1, . . . , 4 as indicated in

Fig. 1.
Bibliography
Braun, S., Kluwick, A. Unsteady three-dimensional marginal separation
caused by surface-mounted obstacles and/or local suction, J. Fluid Mech.
514, 121-152, 2004.
Duck, P.W. Unsteady three-dimensional marginal separation, including
breakdown, J. Fluid Mech. 220, 85-98, 1990.
Ruban, A.I. Stability of preseparation boundary layer on the leading edge
of a thin airfoil, Izv. Akad. Nauk SSSR: Mekh. Zhidk. Gaza 6, 55-63,
(Engl. transl. Fluid Dyn. 17, 835-843), 1983.
Ruban, A.I. Asymptotic Theory of Separated Flows in Steinrück, ed. Asymp-
totic Methods in Fluid Mechanics: Survey and Recent Advances, CISM
Courses and Lectures Vol. 523, Springer, 311–408, 2010.
Scheichl, S., Braun, S., Kluwick, A. On a similarity solution in the theory
of unsteady marginal separation, Acta Mech. 201, 153-170, 2008.
Smith, F.T. Concerning dynamic stall, Aeron. Quart. 33, 331-352, 1982.
Sychev, V.V., Ruban, A.I., Sychev, Vik.V., Korolev, G.L. Asymptotic The-
ory of Separated Flows, Cambridge University Press, Cambridge, 1998.

Asymptotic Methods in Fluid Mechanics by Steinruck PDF

Uploaded by

Copyright:

Available Formats

Asymptotic Methods in Fluid Mechanics by Steinruck PDF

Uploaded by

Document Information

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Asymptotic Methods in Fluid Mechanics by Steinruck PDF

Uploaded by

Copyright:

Available Formats

~ SpringerWienNewYork

Rational asymptotic methods developed in the ﬁfties and sixties of

Introduction to Matched Asymptotic Expansions

Inviscid potential flow around a thin profile. The flow potential is

Flow between two rotating discs: Ekman-layer. A common rea-

Model equation: turbulent pipe ﬂow. Here the asymptotic behav-

condition has to be satisﬁed. Thus again a local expansion turns out to be

2 Flow around a thin elliptical airfoil ε 1

Figure 1. Elliptical thin airfoil

As a ﬁrst example we consider the two-dimensional inviscid irrotational

and the incident ﬂow condition

2.1 Asymptotic expansion of the ﬂow potential (regular expan-

powers of ε is employed. In the limiting case ε = 0 the undisturbed parallel

 ïWHUPVRXWHUH[S ε = 0.3

2.2 Local expansion at leading/trailing edge

φ(x, y) = φ(−1, 0) + ε2 Φ(X, Y ). (16)

Thus the kinematic boundary condition in local coordinates reads

Φ(X, Y ) = Φ0 + εΦ1 + ε2 Φ2 + ... (19)

Figure 3. Airfoil in local coordinates

2.3 Matching procedure

z(ε) = −1 + ε2 ξ(ε) → −1, Z(ε) = ξ(ε) → ∞ as ε → 0, (23)

and insert it into the global and local expansion, respectively.

Thus both expansions agree in the overlap region if

Uloc,0 = Uloc,1 = 0, and C = 0 (27)

Matching principle: n-terms of the outer expansion rewritten in the

We demonstrate the Matching Principle at the surface velocity of a thin

(3) ε2 x2 1 (−1 + ε2 X)2

Expanding the above expression into two terms yields

Thus both expressions agree.

2.4 Composite approximation

3 Flow between rotating discs - Ekman layers

Figure 4. Flow between two coaxially rotating discs

ũ = r̃ ω̃U (Z), ṽ = r̃ Ω̃ + r̃ ω̃V (Z), w̃ = L̃ω̃W (Z), (33a)

Ro (2U V + W VZ ) = − 2U + Ek VZZ , (34b)

U (±1) = 0, V (±1) = ∓1, W (±1) = 0. (35)

Assuming a small diﬀerence in the speeds of rotation of the two discs

3.1 Small Ekman numbers

3.2 Core region

3.3 Boundary layers

U ∼ e−ξ sin ξ − e−η sin η, V ∼ e−ξ cos ξ − e−η cos η, (49a)

4 Model equation for fully developed turbulent

At the channel wall ỹ = 0 the no slip boundary condition ũ = 0 holds and at

d˜ dp̃ τ̃w ũ2 1 μ̃

subject to the boundary conditions

4.1 Defect Layer

For the term of order ε we obtain the equation

with l2 = l (0).

4.2 Viscous wall layer

with a stretching σ(ε) which will be determined appropriately. Inserting

with σ(ε)γ(ε) = ε. Integration of (67) yields

4.4 Turbulence asymptotics

defect layer. From dimensional analysis one obtains

˜ = Reτ . To get the behavior of Φ in the overlap

Michael J. Ward*‡ and Mary-Catherine Kropinski†‡

contexts. One primary goal of this chapter is to highlight how a common

2 Inﬁnite Logarithmic Expansions: Simple Pipe Flow

Here Ω ∈ R2 is the dimensionless pipe cross-section and Ωε is the cross-

Here AΩ is the cross-sectional area of the pipe-core geometry. For laminar

Here ν = O(1/ log ε) is a gauge function to be chosen, and we assume that

ïWHUPVRXWHUH[S ε = 0.3