Hydrodynamic stability

Jan Pralits

Department of Chemical, Civil and Environmental Engineering

University of Genoa, Italy
[email protected]

July 8-10, 2013

Corso di dottorato in Scienze e Tecnologie per l’Ingegneria (STI):

Fluidodinamica e Processi dell’Ingegneria Ambientale (FPIA)

Jan Pralits (University of Genoa) Hydrodynamic stability July 8-10, 2013 1 / 77

 Outline

Outline stability analysis

Topic : Hydrodynamic stability

Hours : 10h
Content :
1 Introduction
2 Definitions
3 Modal analysis (2h)
4 Nonmodal analysis (2.5h)
5 Optimal perturbations (Constrained optimization) (2.5h)
6 Exercises : (3h)
Aim : Overview of main concepts; Provide you with tools and let you test them
Book : Schmid P. J. & Henningson D. S., Stability and Transition in Shear Flows, Springer

Jan Pralits (University of Genoa) Hydrodynamic stability July 8-10, 2013 2 / 77

 Examples

Poiseuille flow

The evolution of the linearized equations give us the dynamics of infinitesimal perturbations,
potentially leading to transition.

Q1: What is the behaviour for t → ∞ ?

A1: Modal analysis will give the answer.

Q2: How large can the amplification be for finite t ?

A2: Nonmodal analysis will give the answer.

Poiseuille: Re=10000 Poiseuille: Re=10000 Poiseuille: Re=10000

2
1 10 0.6
|v(y,0)| d/dt(ln E)
|v(y,T)| λ
0.5
0.5
0.4

d/dt(ln(E))
1
0.3
10
E

0
y

0.2

0.1
−0.5
0
0
−1 10 −0.1
0 0.2 0.4 0.6 0.8 1 0 20 40 60 80 100 0 20 40 60 80 100
|v(y,t)| t t

Perturbations Energy Growth rate

Jan Pralits (University of Genoa) Hydrodynamic stability July 8-10, 2013 3 / 77

 Examples

Aeroelasticity

Q1: What is the behaviour for finite and infinite t ?

L

A1: Answer from nonmodal and modal stability analysis.

U∞ Θ M KΘ
a.c.

Q2: Can we determine an optimal way to control c.g.

instabilities ? e
Kw w
δ
A2: Constrained optimization is a useful tool.
Optimal perturbations ↔ Nonmodal growth
2
10

1
10

||w||

0
10

−1
10
0 2 4 6 8 10 12 14 16 18 20
t
Movie 2

Jan Pralits (University of Genoa) Hydrodynamic stability July 8-10, 2013 4 / 77

 Introduction

Hydrodynamic stability

Hydrodynamic stability theory is concerned with the respons of laminar flow to a disturbance of
small or moderate amplitude.

The flow is generally defined as

Stable : If the flow returns to its original laminar state.

Unstable: If the disturbance grows and causes the laminar

flow to change into a different state.

Stability theory deals with the mathematical analysis of the evolution of disturbances superposed
to a laminar base flow.

In many cases one assumes the disturbances to be small so that further simplifications can be
justified. In particular, a linear equation governing the evolution of disturbances is desirable.

As the disturbance velocities grow above a few % of the base flow, nonlinear effects become
important and linear equations no longer accurately predict the disturbance evolution.

Although the linear equations have a limited region of validity they are important in detecting
physical growth mechanisms and identifying dominant disturbance types.

Jan Pralits (University of Genoa) Hydrodynamic stability July 8-10, 2013 5 / 77

 Introduction

Reynolds pipe flow experiment (1883)

Original 1883 appartus

Dye into center of pipe
Critical Re = 13.000
Lower today due to traffic

Jan Pralits (University of Genoa) Hydrodynamic stability July 8-10, 2013 6 / 77

 Introduction

History of shear flow stability and transition

Reynolds pipe flow experiment (1883)

Rayleigh’s inflection point criterion (1887)

Orr (1907) Sommerfeld (1908) viscous eq.

Heisenberg (1924) viscous channel solution

Tollmien (1931) Schlichting (1933) viscous Boundary

Layer solution

Schubauer & Skramstad (1947) experimental

TS-wave verification

Klebanoff, Tidström & Sargent (1962) 3D breakdown

Jan Pralits (University of Genoa) Hydrodynamic stability July 8-10, 2013 7 / 77

 Introduction

Routes to transition : highly dependent on Tu

 

Tu ∼ 0.1%

 
 
 
 

Tu ∼ 10%

Jan Pralits (University of Genoa) Hydrodynamic stability July 8-10, 2013 8 / 77

 Introduction

More examples of instabilities I

Jan Pralits (University of Genoa) Hydrodynamic stability July 8-10, 2013 9 / 77

 Introduction

More examples of instabilities II

Movie 2

Jan Pralits (University of Genoa) Hydrodynamic stability July 8-10, 2013 10 / 77

 Definitions

Disturbance equations I

∂ui ∂ui ∂p 1 2
= −uj − + ∇ ui
∂t ∂xj ∂xi Re
∂ui
= 0
∂xi
ui (xi , 0) = ui0 (xi )
 

ui (xi , t) = 0 on solid boundaries

∗ ∗
Re = U∞ δ /ν ∗
ui = Ui + ui0 decomposition
p = P + p0
Introduce decomposition, drop primes, subtract eq’s for {Ui , P}
∂ui ∂ui ∂Ui ∂p 1 2 ∂ui
= −Uj − uj − + ∇ ui − uj
∂t ∂xj ∂xj ∂xi Re ∂xj
∂ui
= 0
∂xi

Jan Pralits (University of Genoa) Hydrodynamic stability July 8-10, 2013 11 / 77

 Definitions

Disturbance equations II

∂ui ∂ui ∂p 1 2
= −uj − + ∇ ui
∂t ∂xj ∂xi Re
∂ui
= 0
∂xi
ui (xi , 0) = ui0 (xi )
 

ui (xi , t) = 0 on solid boundaries

∗ ∗
Re = U∞ δ /ν ∗
ui = Ui + ui0 decomposition
0
p = P +p
Introduce decomposition, drop primes, linearize
∂ui ∂ui ∂Ui ∂p 1 2 ∂ui
= −Uj − uj − + ∇ ui − uj
∂t ∂xj ∂xj ∂xi Re ∂xj
∂ui
= 0
∂xi

Jan Pralits (University of Genoa) Hydrodynamic stability July 8-10, 2013 12 / 77

 Definitions

Disturbance equations III

∂ui ∂ui ∂p 1 2
= −uj − + ∇ ui
∂t ∂xj ∂xi Re
∂ui
= 0
∂xi
ui (xi , 0) = ui0 (xi )
 

ui (xi , t) = 0 on solid boundaries

∗ ∗
Re = U∞ δ /ν ∗
ui = Ui + ui0 decomposition
p = P + p0
Linearised Navier-Stokes equations,
∂ui ∂ui ∂Ui ∂p 1 2
= −Uj − uj − + ∇ ui
∂t ∂xj ∂xj ∂xi Re
∂ui
= 0
∂xi

Jan Pralits (University of Genoa) Hydrodynamic stability July 8-10, 2013 13 / 77

 Definitions

Stability definitions I
Z
1
E (t) = ui (t)ui (t) dΩ
2 Ω

E (t)
Stable : lim →0
t→∞ E (0)

Conditionally stable : ∃ δ > 0 : E (0) < δ ⇒ stable

Globally stable : Conditionally stable with δ → ∞

dE
Monotonically stable : Globally stable and ≤0 ∀t > 0
dt

Jan Pralits (University of Genoa) Hydrodynamic stability July 8-10, 2013 14 / 77

 Definitions

Critical Reynolds numbers

ReE : Re < ReE flow monotonically stable

ReG : Re < ReG flow globally stable

ReL : Re < ReL flow linearly stable (δ → 0)

                                                                                                     

Initial energy E vs the Reynolds number Re

Jan Pralits (University of Genoa) Hydrodynamic stability July 8-10, 2013 15 / 77

 Definitions

Critical Reynolds numbers

Flow ReE ReG Retr ReL

Hagen-Poiseuille 81.5 − 2000 ∞
Plane Poiseulle 49.6 − 1000 5772
Plane Couette 20.7 125 360 ∞

Critcial Reynolds numbers for a number of wall-bounded shear flows compiled from the literature.

Jan Pralits (University of Genoa) Hydrodynamic stability July 8-10, 2013 16 / 77

 Definitions

Reynolds-Orr equation

Scalar multiplication of linearised Navier-Stokes equations with ui

∂ui ∂Ui 1 ∂ui ∂ui

ui = −ui uj −
∂t ∂xj Re ∂xj ∂xj
 
∂ 1 1 1 ∂ui
+ − ui ui Uj − ui ui uj − ui pδij + ui
∂xj 2 2 Re ∂xj

integrate in space (Ω), vanishing perturbation at the boundaries ⇒

Z Z
dE ∂Ui 1 ∂ui ∂ui
= −ui uj dΩ − dΩ
dt Ω ∂xj Re Ω ∂xj ∂xj

Nonlinear terms have dropped out

RHS : exchange of energy with the base flow and energy dissipation due to viscosity

Theorem : Linear mechanisms required for energy growth

1 dE
Proof : independent of disturbance amplitude
E dt

Jan Pralits (University of Genoa) Hydrodynamic stability July 8-10, 2013 17 / 77

 Linear Inviscid Analysis

Inviscid Analysis

Jan Pralits (University of Genoa) Hydrodynamic stability July 8-10, 2013 18 / 77

 Linear Inviscid Analysis

Parallel shear flows : Ui = U(y )δ1i I

∂u ∂u ∂p
+U + vU 0 = −
∂t ∂x ∂x
∂v ∂v ∂p
+U + = −
∂t ∂x ∂y
∂w ∂w ∂p
+U + = −
∂t ∂x ∂z
∂u ∂v ∂w
+ + = 0
∂x ∂y ∂z

Initial conditions :
{u, v , w }(x, y , z, t = 0) = {u0 , v0 , w0 }(x, y , z)

Boundary conditions :
v(x, y = y1 , z, t) · n = 0 solid boundary 1
v(x, y = y2 , z, t) · n = 0 solid boundary 2

Jan Pralits (University of Genoa) Hydrodynamic stability July 8-10, 2013 19 / 77

 Linear Inviscid Analysis

Parallel shear flows : Ui = U(y )δ1i II

We can reduce the original 4 eq’s & 4 unknowns to a system of 2 eq’s and 2 unknowns
This is in two steps
1 Take the divergence of the momentum equations. This yields

∂v
∇2 p = −2U 0 .
∂x
2 The new pressure equation is introduced in the momentum equation for v . This yields
  
∂ ∂ ∂
+U ∇2 − U 00 v = 0.
∂t ∂x ∂x

The three-dimensional flow is then analyzed introducing the normal vorticity

∂u ∂w
η= − ,
∂z ∂x
where η satisfies  
∂ ∂ ∂v
+U η = −U 0 .
∂t ∂x ∂z
with the boundary conditions

v =η=0 at a solid wall and in the far field (or second solid wall)

Jan Pralits (University of Genoa) Hydrodynamic stability July 8-10, 2013 20 / 77

 Linear Inviscid Analysis

The Rayleigh equation I

Assume wave-like solutions:

v (x, y , z, t) = ṽ (y ) exp i(αx + βz − ωt)

Introduce the ansatz in the v equation.

We limit ourselves to study the v-equation. This yields  

(−iω + iαU)(D 2 − k 2 )ṽ − iαU 00 ṽ = 0

substitute ω = αc ⇒

U 00
 
D2 − k2 − ṽ = 0
U −c

Here, k 2 = α2 + beta2 and D i = ∂ i /dy i , and the boundary conditions are

ṽ (y = y1 ) = ṽ (y = y2 ) = 0 solid boundaries

Jan Pralits (University of Genoa) Hydrodynamic stability July 8-10, 2013 21 / 77

 Linear Inviscid Analysis

The Rayleigh equation II

The Rayleigh equation poses an eigenvalue problem of second order with c as the complex
eigenvalue. The coefficients of the operator are real. Any complex eigenvalue will therefore
appear as complex conjugate pairs. So, if c is an eigenvalue, so is c ∗ .

It has a regular singular point at U(yc ) = c, a condition where the order of the equation is
reduced (critical layer).

Analytical solution for the eigenfunctions exists (Tollmien, 1928)

Instability must depend on U(y ) (only parameter). U can be any base flow
We look for base flows where the perturbations become unstable

By definition perturbations in time behave as ∼ exp(−iαcr t)exp(αci t)

Take α > 0. If αci > 0 the corresponding mode grows exponentially in time

Jan Pralits (University of Genoa) Hydrodynamic stability July 8-10, 2013 22 / 77

 Linear Inviscid Analysis

Rayleigh’s inflection point criterion (1887) I

Here we consider a parallel shear flow in a domain y ∈ (−1, 1) and prove a necessary condition
for instability.

THEOREM : If there exist perturbations with ci > 0, then U 00 (y ) must vanish for some
ys ∈ [−1, 1]

PROOF :
The proof is given by multiplying the Rayleigh equation by ṽ ∗ and integrating y from −1 to 1.
This yields
1 U 00
Z  
− ṽ ∗ D 2 ṽ − k 2 ṽ − ṽ dy =
−1 U −c
1 1 U 00
Z Z
|D ṽ |2 + k 2 |ṽ |2 dy + |ṽ |2 dy


= 0
−1 −1 U −c

The first integral is positive definite. The equation equals zero if the second integrand of the
second equation changes sign.

Jan Pralits (University of Genoa) Hydrodynamic stability July 8-10, 2013 23 / 77

 Linear Inviscid Analysis

Rayleigh’s inflection point criterion (1887) II

This is analyzed by multiplying and dividing the second integral with U − c ∗ . This yields
1 1 U 00 (U − c ∗ )
Z Z
|D ṽ |2 + k 2 |ṽ |2 dy + |ṽ |2 dy


= 0
−1 −1 (U − c)(U − c ∗ )

The real part is

1 U 00 (U − cr ) 2 1
Z Z
|D ṽ |2 + k 2 |ṽ |2 dy ,


|ṽ | dy = −
−1 |U − c|2 −1

the imaginary part states : U 00 must change sign to render the integral equal to zero if c 6= 0.
1 U 00 ci
Z
|ṽ |2 dy = 0.
−1 |U − c|2

Jan Pralits (University of Genoa) Hydrodynamic stability July 8-10, 2013 24 / 77

 Linear Inviscid Analysis

Fjortofts criterion (1950) I

Here we consider the same flow as in the Rayleigh’s criterion.

THEOREM : Given a monotonic mean velocity profile U(y ), a necessary condition for instability
is that U 00 (U − Us ) < 0 for some y ∈ [−1, 1], with Us = U(ys ) as the mean velocity at the
inflection point, i.e. U 00 (ys ) = 0

PROOF : Consider again the real part

1 U 00 (U − cr ) 2 1
Z Z
|D ṽ |2 + k 2 |ṽ |2 dy ,


|ṽ | dy = −
−1 |U − c|2 −1

We add to the left side the following integral which is identically 0 (Rayleigh’s i.p. criteria)
1 U 00
Z
(cr − Us ) |ṽ |2 dy = 0.
−1 |U − c|2

We then get
1 U 00 (U − Us ) 2 1
Z Z
|D ṽ |2 + k 2 |ṽ |2 dy ,


|ṽ | dy = −
−1 |U − c|2 −1

For the integral on the LHS to be negative the value of U 00 (U − Us ) must be negative somewhere
in the flow.

Jan Pralits (University of Genoa) Hydrodynamic stability July 8-10, 2013 25 / 77

 Linear Inviscid Analysis

Fjortofts criterion (1950) II

Here are two examples of parallel monotonic shear flow.

   

Both profiles lead to unstable solutions according to Rayleigh’s criterion; however the inflection
point has to be a maximum of the spanwise vorticity (not a minimum).

LEFT : unstable according to Fjortoft RIGHT : stable according to Fjortoft

Jan Pralits (University of Genoa) Hydrodynamic stability July 8-10, 2013 26 / 77

 Linear Inviscid Analysis

Solutions to piecewise linear velocity profiles I

Before computers were available to researchers in the field   of hydrodynamic stability theory, a
 
common technique to solve inviscid stability problems was   to approximate continuous mean
velocity profiles by piecewise linear profiles. It allows to     find analytical expression for the
dispersion relation c(α, β) and the eigenfunctions.  
 
 
       
y  

General considerations:  I

 
U 00 = 0 which simplifies the Rayleigh equation
(except at the connecting points) II   U(y)  
 

Matching conditions must be imposed where U is

continuous but U 00 is discontinuous
III

Jan Pralits (University of Genoa) Hydrodynamic stability July 8-10, 2013 27 / 77

 Linear Inviscid Analysis

Solutions to piecewise linear velocity profiles II

Matching condition
We can rewrite the Rayleigh equation as

D[(U − c)D ṽ − U 0 ṽ ] = (U − c)k 2 ṽ

and integrating over the discontinuity in U and/or U 0 located at yD we get

Z yD +
y +
[(U − c)D ṽ − U 0 ṽ ]yD − = k 2 (U − c)ṽ dy
D yD −

As  → 0 the RHS → 0 which gives the first matching condition

J(U − c)D ṽ − U 0 ṽ K = 0, Condition 1

which is equivalent to matching the pressure across the discontinuity which in

Fourier-transformed form reads
iα 0
p̃ = (U ṽ − (U − c)D ṽ ).
k2

Jan Pralits (University of Genoa) Hydrodynamic stability July 8-10, 2013 28 / 77

 Linear Inviscid Analysis

Solutions to piecewise linear velocity profiles III

A second condition is derived by dividing the pressure p̃ by iα(U − c)/k 2 . This yields

k 2 p̃ U 0 ṽ
 
D ṽ ṽ
− = − = D
iα(U − c)2 U −c (U − c)2 U −c

Integrating across the discontinuity in the velocity profile gives

y
D + k2 yD +
 Z
ṽ p̃
=− dy
U −c yD − iα yD − (U − c)2

Again, as  → 0 we obtain the second matching condition

s {
ṽ
= 0, Condition 2
U −c

which, for continuous U, corresponds to matching ṽ .

Jan Pralits (University of Genoa) Hydrodynamic stability July 8-10, 2013 29 / 77

 Linear Inviscid Analysis

Solutions to piecewise linear velocity profiles IV

Summary :
To solve the Rayleigh equation for a piecewise linear velocity profile we need to solve

(D 2 − k 2 )ṽ = 0

in each subdomain and impose boundary and matching conditions

J(U − c)D ṽ − U 0 ṽ K = 0,
s {
ṽ
= 0,
U −c

to determine the coefficients of the fundamental solution and finally the dispersion relation c(k).

Jan Pralits (University of Genoa) Hydrodynamic stability July 8-10, 2013 30 / 77

 Linear Inviscid Analysis
 
 
 
Solutions to piecewise linear velocity profiles
  V
 
 
Exercise : piecewise linear mixing layer  
 
       
y  

Velocity profile
  I
1  for y >1
U(y ) = y for −1 ≤ y ≤ 1  
−1 for y < −1

II   U(y)  
 
Boundary conditions

ṽ → 0 as y → ±∞
III
A general solution can be written
 
ṽI = A exp(−ky ) for y >1
ṽII = B exp(−ky ) + C exp(ky ) for −1 ≤ y ≤ 1
ṽIII = D exp(ky ) for y < −1

Derive
c = c(k)
Make a plot of c(k) for k ∈ [0, 2] and discuss the results.

Jan Pralits (University of Genoa) Hydrodynamic stability July 8-10, 2013 31 / 77

 Linear Inviscid Analysis

Solutions to piecewise linear velocity profiles VI

Results : Piecewise mixing layer

s 2  
1 1
c=± 1− − exp(−4k)
2k 4k 2

For 0 ≤ k ≤ 0.6392 the expression under the square root is negative resulting in purely
imaginary eigenvalues
For k > 0.6392 the eigenvalues are real, and all disturbances are neutral
As the wave number goes to zero, the wavelength associated with the disturbances is much
larger than the length scale associated with U(y ). The limit of small k is equivalent to the
limit of zero thickness of region II .
1 0.8
c+ c+
c− c−
0.8
0.6

0.6
0.4
0.4

0.2
0.2
cr
ci

0 0

−0.2
−0.2

−0.4
−0.4
−0.6

−0.6
−0.8

−1 −0.8
0 0.2 0.4 0.6 0.8 1 1.2 1.4 1.6 1.8 2 0 0.2 0.4 0.6 0.8 1 1.2 1.4 1.6 1.8 2
k k

Jan Pralits (University of Genoa) Hydrodynamic stability July 8-10, 2013 32 / 77

 Linear Viscous Analysis

Viscous Analysis

Only linear or parabolic velocity profiles satisfy the steady viscous equations (Couette,
Poiseuille)

Inviscid criteria state that Poiseuille flow is stable

Common sense would suggest that viscosity acts as a damping

However, viscous Poiseuille flow undergoes transition: viscosity destabilizes the flow

Jan Pralits (University of Genoa) Hydrodynamic stability July 8-10, 2013 33 / 77

 Linear Viscous Analysis

Parallel shear flows : Ui = U(y )δ1i I

∂u ∂u ∂p 1 2
+U + vU 0 = − + ∇ u
∂t ∂x ∂x Re
∂v ∂v ∂p 1 2
+U + = − + ∇ v
∂t ∂x ∂y Re
∂w ∂w ∂p 1 2
+U + = − + ∇ w
∂t ∂x ∂z Re
∂u ∂v ∂w
+ + = 0
∂x ∂y ∂z

Initial conditions :
{u, v , w }(x, y , z, t = 0) = {u0 , v0 , w0 }(x, y , z)

Boundary conditions : depend on flow case

{u, v , w }(x, y = y1 , z, t) = 0 solid boundaries
Semi-infinite domain :
{u, v , w }(x, y → ∞, z, t) → 0 free stream
Closed domain :
{u, v , w }(x, y = y2 , z, t) = 0 solid boundary 2

Jan Pralits (University of Genoa) Hydrodynamic stability July 8-10, 2013 34 / 77

 Linear Viscous Analysis

Parallel shear flows : Ui = U(y )δ1i II

We can reduce the original 4 eq’s & 4 unknowns to a system of 2 eq’s and 2 unknowns
This is in two steps
1 Take the divergence of the momentum equations. This yields

∂v
∇2 p = −2U 0 .
∂x
2 The new pressure equation is introduced in the momentum equation for v . This yields
  
∂ ∂ ∂ 1 4
+U ∇2 − U 00 − ∇ v = 0.
∂t ∂x ∂x Re

The three-dimensional flow is then analyzed introducing the normal vorticity

∂u ∂w
η= − ,
∂z ∂x
where η satisfies  
∂ ∂ 1 2 ∂v
+U − ∇ η = −U 0 .
∂t ∂x Re ∂z
with the boundary conditions

v = v0 = η = 0 at a solid wall and in the far field (or second solid wall)

Jan Pralits (University of Genoa) Hydrodynamic stability July 8-10, 2013 35 / 77

 Linear Viscous Analysis

Orr-Sommerfeld and Squire equations

Assume wave-like solutions:

v (x, y , z, t) = ṽ (y ) exp i(αx + βz − ωt)

Introduce the ansatz in the equations for {v , η}. This yields

 
1
(−iω + iαU)(D 2 − k 2 ) − iαU 00 − (D 2 − k 2 )2 ṽ = 0
Re
 
1
(−iω + iαU) − (D 2 − k 2 ) η = −iβU 0 ṽ
Re

Here, k 2 = α2 + β 2 and D i = ∂ i /dy i .

Orr-Sommerfeld modes : {ṽn , η̃np , ωn }N

n=1

Squire modes : {ṽ = 0, η̃m , ωm }M

m=1

Jan Pralits (University of Genoa) Hydrodynamic stability July 8-10, 2013 36 / 77

 Linear Viscous Analysis

Squire modes I

THEOREM : Squire modes are always damped, i.e. ci < 0 ∀α, β, Re

Rewriting the homogeneous Squire equation we get

1
(U − c)η̃ = (D 2 − k 2 )η̃
iαRe
Multiplying by η̃ ∗ and integrating
Z 1 Z 1 1
Z 1
c |η̃|2 dy = U|η̃|2 dy − η̃ ∗ (D 2 − k 2 )η̃dy
−1 −1 iαRe −1

Taking the imaginary part and integrating by parts yields

Z 1 1
|η̃|2 dy = − k 2 |η̃|2 + |D η̃|2 < 0


ci
−1 αRe

Jan Pralits (University of Genoa) Hydrodynamic stability July 8-10, 2013 37 / 77

 Linear Viscous Analysis

Squire’s transformation and theorem I

Let’s consider 3D and 2D Orr-Sommerfeld equation with ω = αc

1
(U − c)(D 2 − k 2 )ṽ − U 00 ṽ − (D 2 − k 2 )2 ṽ = 0
iαRe
1
(U − c)(D 2 − α22D )ṽ − U 00 ṽ − (D 2 − α22D )2 ṽ = 0
iα2D Re2D

p
α2D = k= α2 + β 2
α2D Re2D = αRe
⇒
α
Re2D = Re < Re
k

Jan Pralits (University of Genoa) Hydrodynamic stability July 8-10, 2013 38 / 77

 Linear Viscous Analysis

Squire’s transformation and theorem II

Each 3D Orr-Sommerfeld mode corresponds to a 2D Orr-Sommerfeld mode at a lower Re, i.e.

α
Re2D = Re < Re
k

We can therefore define a critical Reynolds number for parallel shear flows as

Rec ≡ min ReL (α, β) = min ReL (α, 0)

α,β α

since the growth rate increases with the Reynolds number.

Jan Pralits (University of Genoa) Hydrodynamic stability July 8-10, 2013 39 / 77

 Linear Viscous Analysis

Discretization of the equations in y

The Orr-Sommerfeld equations

 
1
(−iω + iαU)(D 2 − k 2 ) − iαU 00 − (D 2 − k 2 )2 ṽ = 0
Re
 
1
(−iω + iαU) − (D 2 − k 2 ) η = −iβU 0 ṽ
Re

including boundary conditions ṽ = D ṽ = η = 0 y = ±1, can, after suitable discretization

(Chebyshev polynomials, finite-differences), be written on the following compact form

ωq̃ = Aq̃ with q̃ = (ṽ , η̃)

where A is a matrix ∈ C2N×2N . This is an eigenvalue problem from which a solution is obtained
for the eigenvalue ωn and eigenvector q̃n . Note that N is the number of discrete points in the
wall-normal direction.

Jan Pralits (University of Genoa) Hydrodynamic stability July 8-10, 2013 40 / 77

 Linear Viscous Analysis

Solutions of Eigenvalue analysis I

Plane Poiseuille flow

Neutral curve & spectrum (Re = 10.000, α = 1, β = 0)

A (cr → 0), P (cr → 1), S (cr = 2/3), Mack (1976)

Jan Pralits (University of Genoa) Hydrodynamic stability July 8-10, 2013 41 / 77

 Linear Viscous Analysis

Solutions of Eigenvalue analysis II

A, P, S- Eigenfunctions for PPF

Re = 5000, α = 1, β = 1

Jan Pralits (University of Genoa) Hydrodynamic stability July 8-10, 2013 42 / 77

 Linear Viscous Analysis

Solutions of Eigenvalue analysis III

Blasius boundary layer

Neutral curve & spectrum (Re = 500, α = 0.2, β = 0)

 
 

Jan Pralits (University of Genoa) Hydrodynamic stability July 8-10, 2013 43 / 77

 Linear Viscous Analysis

Critical Reynolds numbers

Flow αcrit Recrit crcrit

Plane Poiseulle 1.02 5772 0.264
Blasius boundary layer flow 0.303 519.4 0.397

Plane Poiseuille Flow & Blasius boundary layer

   

Jan Pralits (University of Genoa) Hydrodynamic stability July 8-10, 2013 44 / 77

 Linear Viscous Analysis

Continuous spectrum

As y → ∞ the OSE reduces to Example : Blasius boundary layer

2 2 2 2 2
(D − k ) ṽ = iαRe[(U∞ − c)(D − k )]ṽ

If we assume that

ṽ (y ) = v̂ exp(λn y )

then the solution is analytical with eigenvalues

q
λ1,2 = ± iαRe(U∞ − c) + k 2 , λ3,4 = ±k

Assuming that iαRe(U∞ − c) + k 2 is real and negative

 
which means that ṽ and D ṽ are bounded, λ1,2 = ±iC

⇒ αReci + k 2 < 0, αRe(U∞ − cr ) = 0

From which we can derive analytically c(k, Re)

k2
c = U∞ − i(1 + ξ 2 )
αRe

Jan Pralits (University of Genoa) Hydrodynamic stability July 8-10, 2013 45 / 77

 Summary modal analysis

Summary

We are considering the stability of the linearized system

Stability in the limit in which E (0) → 0
Reynolds-Orr equation: linear mechanism required for energy growth
Modal analysis: we consider v ∼ exp(iαx + iβz − iωt)
Eigenvalue problem
Rayleigh & Fjortoft: Inflection point criteria for instability
Piecewise linear profiles: approximate analytical solutions exist
Squire’s theorem: 2D perturbations are more unstable
Finite domain (ex. channel): all discrete modes
Semi-infinite domain (ex. boundary layer): discrete and continuous modes
ReL sometimes far from Retr . Modal analysis cannot tell the whole story

Jan Pralits (University of Genoa) Hydrodynamic stability July 8-10, 2013 46 / 77

 Nonmodal stability analysis

Nonmodal stability analysis

Jan Pralits (University of Genoa) Hydrodynamic stability July 8-10, 2013 47 / 77

 Nonmodal stability analysis

Critical Reynolds numbers

Flow ReE ReG Retr ReL

Hagen-Poiseuille 81.5 − 2000 ∞
Plane Poiseulle 49.6 − 1000 5772
Plane Couette 20.7 125 360 ∞

Critcial Reynolds numbers for a number of wall-bounded shear flows compiled from the literature.

Jan Pralits (University of Genoa) Hydrodynamic stability July 8-10, 2013 48 / 77

 Nonmodal stability analysis

Time scale of viscous linear instability

Maximum growth rate in plane Poiseuille flow occurs at Re ≈ 46950.

It takes ≈ 90 time units, corresponding to a propagation about ≈ 54 times

the channel half width, for the wave to double its amplitude.

Viscous instability acts on a slow time scale

Are we missing some faster dynamics ?

Jan Pralits (University of Genoa) Hydrodynamic stability July 8-10, 2013 49 / 77

 Nonmodal stability analysis

Comments on classical stability theory

Looking at eigenvalues of the linear stability operator gives us

information about the asymptotic behavior of the solution, as t → ∞

No information is provided about the short-time dynamics if t remains

finite

What if the linear solution experiences transient amplifications before

eventually going to zero ?

Is linearization still valid in this case ?

Jan Pralits (University of Genoa) Hydrodynamic stability July 8-10, 2013 50 / 77

 Nonmodal stability analysis

Linearization & diagonalization I

To analyze the failure of linear stability theory for the case of plane Poiseuille flow, we need to
scrutinize the steps involved in the analysis. Linear stability theory is a two-step procedure,
consisting of a linearization and a diagonalization step.

Linearization
The linearization step decomposes the flow field into a (steady) base flow and a small amplitude
perturbation of order O()
Q(x, t) = Q(x) + q(x, t) + O(2 ).
Substituting into the Navier-Stokes equations and extracting the terms of order O() yields the
linearized Navier-Stokes equations governing the evolution of small disturbances

∂q
= Lq.
∂t
Note that L = L(Q).

Jan Pralits (University of Genoa) Hydrodynamic stability July 8-10, 2013 51 / 77

 Nonmodal stability analysis

Linearization & diagonalization II

Diagonalization The formal solution of the linearized Navier-Stokes equations can be written

q = exp(tL)q0

where q0 is the initial condition. The operator exponential propagates the initial condition
forward in time.

Note:
∞
1 1 X 1
exp(tL) = I + tL + t 2 L2 + ... = (tL)n .
1! 2! n=0
n!

We simplify the linear operator L by transforming it into diagonal form, thus decoupling the
degrees of freedom. This allows the analysis of individual modes. If LS = SΛ then we have

L = SΛS −1

where Λ represents a diagonal operator of eigenvalues, and S consists of the eigenfunctions.

Jan Pralits (University of Genoa) Hydrodynamic stability July 8-10, 2013 52 / 77

 Nonmodal stability analysis

Linearization & diagonalization III

So far exp(tL) has been diagonalized as L = SΛS −1 .

Most conclusions about the behavior of exp(tL) are drawn from Λ with little regard given to the
similarity transformation based on S that diagonalized the linear operator L.

Questions

1 When is the above two-step procedure appropriate

and accurate?

2 When can we deduce the behavior of exp(tL) entirely

from Λ?

Evaluating the bounds on exp(tL) can help us.

Jan Pralits (University of Genoa) Hydrodynamic stability July 8-10, 2013 53 / 77

 Nonmodal stability analysis

Bounds on the operator exponential

Let’s determine a lower and upper bound of the operator exponential norm.

e tλmax ≤ k exp(tL)k = kS exp(tΛ)S −1 k ≤ kSkkS −1 ke tλmax

Note
Lower bound: It cannot decay faster than the least stable mode λmax

The term kSkkS −1 k = κ(S) is called the condition number and κ(S) ≥ 1.

Classification
If κ(S) = 1 then the upper and lower bound coincide. The temporal behavior is governed by
the exponential behavior for all times.

If κ(S) > 1 then only the asymptotic behavior is given by the least stable mode.

Short explanations:
If L = SΛS −1 and L2 = (SΛS −1 )(SΛS −1 ) = SΛ2 S −1 then Ln = SΛn S −1
So I + tL + 1/(2!)t 2 L2 + ... = S(I + tΛ + 1/(2!)t 2 Λ2 + ...)S −1 = S exp(tΛ)S −1

Jan Pralits (University of Genoa) Hydrodynamic stability July 8-10, 2013 54 / 77

 Nonmodal stability analysis

Definition of non-normality

Linear operators with κ(S) = 1 are called Normal and have orthogonal eigenvectors

Linear operators with κ(S) > 1 are called Non-normal and have non-orthogonal eigenvectors

Alternatively
An operator is non-normal if LL? 6= L? L

Linear operators which satisfy L = L? are called self-adjoint.

Summary common measures

κ(S) = kSkkS −1 k

kLL? − L? Lk

Short explanations:
Definition of adjoint: hv , Lui = hL∗ v , ui for two arbitrary fields u and v , and h, i denotes a chosen inner
product.
If L is a real valued matrix then L∗ = LT

Jan Pralits (University of Genoa) Hydrodynamic stability July 8-10, 2013 55 / 77

 Nonmodal stability analysis

Non-orthogonal superposition
Let us assume that we expand an initial condition q
of unit length in a non-orthogonal (two-dimensional)
basis as shown in the Figure.
Φ1 and Φ2 are two solutions which decay in time. In
terms of eigenvalues they are both stable.

The non-orthogonal superposition of

exponentially decaying solutions can give rise to
short-term transient growth.

Eigenvalues alone only describe the asymptotic

fate of the disturbance, but fail to capture
 
transient effects.

The source of the transient amplification of the

initial condition lies in the nonorthogonality of
the eigenfunction basis.

Jan Pralits (University of Genoa) Hydrodynamic stability July 8-10, 2013 56 / 77

 Nonmodal stability analysis

Norm of the operator exponential: Definition of Gain

The correct way to analyze the behavior of exp(tL) is to compute its potential to amplify a given
disturbance over time.

We will measure the size of the disturbance by an appropriate norm (see below) and define as the
maximum amplification the ratio of disturbance size to its initial size optimized over all possible
initial conditions.

We have
kqk k exp(tL)q0 k
max = max = k exp(tL)k ≡ G (t)
∀q0 kq0 k ∀q0 kq0 k
The quantity G (t) represents the maximum possible amplification of unit-norm initial conditions
over a time period t and is denote the gain.

Short explanations:
Using the inequality k exp(tL)q0 k ≤ k exp(tL)kkq0 k
then
k exp(tL)q0 k k exp(tL)kkq0 k
≤ = k exp(tL)k
kq0 k kq0 k

Jan Pralits (University of Genoa) Hydrodynamic stability July 8-10, 2013 57 / 77

 Nonmodal stability analysis

General properties of the norm

The choice of inner product will quantitatively influence the maximum amplification potential of
the underlying operator. Therefore, the norm and inner product have to be chosen carefully.

Ex. in shear flows the disturbance kinetic energy is normally chosen.

Basic requirements
kqk ≥ 0
and
kqk = 0 if and only if q = 0.

Note that the norm has to include all components of q. Otherwise, infinite transient growth is
possible, by choosing a disturbance with infinite amplitudes in components that are not
accounted for in the norm.

Jan Pralits (University of Genoa) Hydrodynamic stability July 8-10, 2013 58 / 77

 Nonmodal stability analysis

Algorithm : gain (simple)

1 Compute the first N eigenvalues (λ) and eigenvectors (q) of the flow, where L = SΛS −1

S = [q1 , q2 , ... qN ] and Λ = diag (λ1 , λ2 , ..., λN )

2 Invert S
3 Form the matrix
exp(tλ1 )
 

S 
 ..  −1
S
.
exp(tλN )
4 Compute the norm of the above matrix

G (t) = kS exp(tΛ)S −1 k

5 Advance in time and go back to step (3)

Jan Pralits (University of Genoa) Hydrodynamic stability July 8-10, 2013 59 / 77

 Nonmodal stability analysis

The energy norm

From now on we consider disturbances which behave as

p
q(y , t) exp i(αx + βz) and k= α2 + β 2

We choose a formulation of the linearized Navier-Stokes equations in terms of the normal velocity
v and the normal vorticity η = ∂u/∂z − ∂w /∂x. The linear operator is similar to the classical
Orr-Sommerfeld operator.

Our state vector is q = (v , η)T and the kinetic energy in these variables is
Z
1
|Dv |2 + k 2 |v |2 + |η|2 dΩ


E (t) =
2k 2 Ω
Z  H 
−D2 + k 2 0
 
1 v v
= kqk2E =
2k 2 Ω η 0 1 η
Z
1
= q H M q dΩ
2k 2 Ω
Here, D denotes differentiation, k wave-number modulus and M a positive definite weight matrix.

Jan Pralits (University of Genoa) Hydrodynamic stability July 8-10, 2013 60 / 77

 Nonmodal stability analysis

Reduction to a 2-norm

The problem is simplified by transforming the energy norm to a standard L2 -norm. Using
Cholesky decomposition we can write M = F H F . We then get
Z Z
1 1
kqk2E = q H H
F Fq dΩ = (Fq)H Fq dΩ
2k 2 Ω 2k 2 Ω

Recalling the definition of G (t) we have

kqk2E kFqk22 kF exp(tL)q0 k22

G (t) = max = max = max
∀q0 kq k2 ∀q0 2
kFq0 k2 ∀q0 kFq0 k22
0 E
kF exp(tL)F −1 F q0 k22
= max = kF exp(tL)F −1 k22
∀q0 kFq0 k22

Note that the energy weight is accounted for by the matrices F −1 and F .

Jan Pralits (University of Genoa) Hydrodynamic stability July 8-10, 2013 61 / 77

 Nonmodal stability analysis

Projections onto eigenvectors

Computationally, it is not practical to compute exp(tL). A better solution is to decompose q into
a large, but finite, number of eigenvectors of L. This can be written
N
X
q(y , t) = κn (t)q̄n (y )
n=1

for the first N eigenfunctions of L. In the following we need to consider the expansion coefficients
κ and the matrix exponential exp(tΛ). The latter is much easier to compute.
The energy norm is now written
N N
Z Z ! !
1 1 X
∗
X
kqk2E = q H
M q dΩ = κn (t)q̄ H
n M κm (t)q̄ m dΩ
2k 2 Ω 2k 2 Ω n=1 m=1
N
1 X ∗
= κ (t)Mmn κm (t),
2k n,m=1 n
2

where Z
Mmn = q̄nH M q̄m dΩ.
Ω

Finally, with Mmn = F H F , we get

G (t) = kF exp(tΛ)F −1 k22

Jan Pralits (University of Genoa) Hydrodynamic stability July 8-10, 2013 62 / 77

 Nonmodal stability analysis

Algorithm : gain (energy norm)

1 Compute the first N eigenvalues and eigenvectors of the flow (L)
q̄j , λj for j = 1, ..., N
2 Compute the entries of the matrix Mmn
Z
Mmn = q̄nH M q̄m dΩ
Ω

3 Decompose Mmn into F H F

Mmn = UΣU H (SVD)
F = UΣ1/2
4 Invert F
5 Form the matrix
exp(tλ1 )
 

F
 ..  −1
F
.
exp(tλN )
6 Compute the L2 -norm of the above matrix

G (t) = kF exp(tΛ)F −1 k22

7 Advance in time and go back to step (5)
Jan Pralits (University of Genoa) Hydrodynamic stability July 8-10, 2013 63 / 77
 Nonmodal stability analysis

A note on the result

The quantity G (t) gives the maximum amplification

optimized over all possible initial conditions. In
general, each point on the curve G (t) is arrived at by
a different initial condition, and G (t) represents the
envelope of individual growth curves, see figure.

Optimal disturbances
The initial condition yielding the gain at a specific
time (tspec ) is called optimal disturbance. It can be
evaluated by performing a Singular Value
Decomposition as

F exp(tΛ)F −1 = UΣV H ,
 
or equivalently

F exp(tΛ)F −1 V = UΣ.
Figure : Amplification G (t) for
Poiseuille flow with Re = 1000, α = 1 We identify the left singular vector associated with
(solid line) and growth curves of the largest singular value (which is identical to the
selected initial conditions (dashed norm of the matrix exponential) as the desired initial
lines). condition that will result in maximum amplification at
time tspec . Note : U and V are unitary matrices.

Jan Pralits (University of Genoa) Hydrodynamic stability July 8-10, 2013 64 / 77

 Nonmodal stability analysis

Algorithm : optimal disturbance at t = tspec

1 Compute the first N eigenvalues and eigenvectors of the flow (L)
q̄j , λj for j = 1, ..., N
2 Compute the entries of the matrix Mmn
Z
Mmn = q̄nH M q̄m dΩ
Ω

3 Decompose Mmn into F H F

Mmn = UΣU H (SVD)
F = UΣ1/2
4 Invert F
5 Form the matrix
exp(tspec λ1 )
 

F
 ..  −1
F
.
exp(tspec λN )
6 Compute the singular value decomposition of the above matrix
F exp(tspec Λ)F −1 = UΣV H
7 Extract the first column of V as the optimal initial condition at t = tspec
Jan Pralits (University of Genoa) Hydrodynamic stability July 8-10, 2013 65 / 77
 Constrained optimization

Constrained Optimization

Jan Pralits (University of Genoa) Hydrodynamic stability July 8-10, 2013 66 / 77

 Constrained optimization

Motivation

Q: Why is constrained optimization useful in problems concerning stability analysis ?

A1: Gives a general framework to compute optimal perturbations. Alternative to the previously
shown nonmodal stability analysis and can be applied to nonlinear state equations.

A2: Gives a framework to compute optimal control of instabilities

U Θ M K
∞ Θ
a.c.
c.g.

e
Kw w
δ

Movie 2

Jan Pralits (University of Genoa) Hydrodynamic stability July 8-10, 2013 67 / 77

 Constrained optimization

Definition of the optimization problem

Given the state vector φ ∈ RN
and the control vector g ∈ RK
minimize the cost function J (φ
φ, g)
constrained by the state equation φ, g) = 0
F(φ

The goal is to reach a local minimum of J (φ

φ, g) acting on the control variables g.

The solution of the constrained problem is usually very different from the solution of the
unconstrained problem as seen from the example below.

Exercise Minimize the cost function J (φ, g ) = φ2 + 32g 2 constrained by F (φ, g ) = φg − 1 = 0.

J F J&F
3 3 3
200
130

200
130
100

100

−8 6
70

70
50

50

200

200
−4

130
100

130
100
20

20
10

10

70

70
50

50
20

20
10

10
2 2 4 2
−6

5
5

−2

0
1 −4

0
1 1
−2 0 0

5
5
200
130
100

100
130
200

0 0 0
5
θ

200

200
130
100

100
130
70
70

50
50

20
20

10
10

50
70
70
50

20
20

10
10
5
0 −2 0
−1 −1 2 −4 −1
−2
0

0
4 −6
−2 −2 −2

5
5

6
−4

−8

200
200
2

130

130
100

100
200
130

130
200
100

100

70
70
50

50
20
20
70
70
50

50
20

20

10

10
10

10

−3 −3 −3
−3 −2 −1 0 1 2 3 −3 −2 −1 0 1 2 3 −3 −2 −1 0 1 2 3
g g g

What is the value of φ and g in the constrained case ?

Jan Pralits (University of Genoa) Hydrodynamic stability July 8-10, 2013 68 / 77
 Constrained optimization

Lagrangian and optimality condition

J&F

Scope: descend as low as possible on the J level curves, remaining 3

200

200
130
100

130
100
70

70
50

50
20

20
10

10
on the path given by F = 0. If the level lines of J and the path are 2

0
1
continuous, then at the point where the minimum is reached, the 0

5
path is tangent to the level curve of the optimal J . 0

200

200
130
100

100
130
50
70
70
50

20
20
10
5

10
0
−1

0
This implies that at optimality the gradient of the cost function −2

200
200
130

130
100

100
and the gradient of F are parallel in the φ − g plane, i.e.

70
70
50

50
20
20
10

10
−3
−3 −2 −1 0 1 2 3
g
   
∂J ∂J ∂F ∂F
, =a ,
∂g ∂φ ∂g ∂φ
The above relation gives an Optimality System:

∂J ∂F
−a =0
∂g ∂g
∂J ∂F
−a =0
∂φ ∂φ
F =0

Lagrange remarked: if the new cost function L = J − a F is considered, then the above
conditions coincide with the optimality conditions for the unconstrained optimization of
L(φ, g , a) if the all the variables are considered as independent.
L is usually referred to as the Lagrangian and a is usually called Lagrange multiplier.

Jan Pralits (University of Genoa) Hydrodynamic stability July 8-10, 2013 69 / 77

 Constrained optimization

Lagrangian and optimality condition: a variational approach

We again consider minimizing J (φ, g ) constrained by F (φ, g ). The Lagrangian L is written
L(φ, g , a) = J (φ, g ) − a F (φ, g )
where φ, g and a are considered independent variables. We set the variation of L equal to zero
∂L ∂L ∂L
δL(φ, g , a) = δφ + δg + δa = 0
∂φ ∂g ∂a
By definition:
∂L L(φ + δφ, g , a) − L(φ, g , a)
δφ = lim = 0, ∀δφ
∂φ →0 
In practice:
 
∂L ∂J ∂F ∂L ∂J ∂F
δφ = −a δφ = 0 → = −a = 0, ∀δφ
∂φ ∂φ ∂φ ∂φ ∂φ ∂φ
Applied to all terms yields
∂L ∂J ∂F
= −a =0
∂g ∂g ∂g
∂L ∂J ∂F
= −a =0
∂φ ∂φ ∂φ
∂L
= F =0
∂a
This is exactly the system we obtained in the previous example !!!
Jan Pralits (University of Genoa) Hydrodynamic stability July 8-10, 2013 70 / 77
 Constrained optimization

Lagrangian and optimality condition

Application of the Lagrangian to the model problem:
We again consider the problem of minimizing the cost function J (φ, g ) = φ2 + 32g 2 constrained
by F (φ, g ) = φg − 1 = 0.

The optimality system, using the Lagrangian as defined The solution is

previously, can be written (φ, g )1 = (2.38, 0.42)
(φ, g )2 = (−2.38, −0.42)
∂L ∂J ∂F
= −a = 64g − a φ = 0 3
J&F
∂g ∂g ∂g

200

130
100

100
130

200
70

70
50
50

20
20

10

10
∂L ∂J ∂F 2

5
= −a = 2φ − a g = 0
∂φ ∂φ ∂φ

0
1
0

∂L

5
0

θ
= F = φg − 1 = 0

200

200
130
100

100
130
5
70
50

50
70
20

20
10

10
0

∂a −1

0
This system of 3 unknowns and 3 equations can be solved −2

5
analytically.

200

130

100
130

200
100

70
50

70
50
20

20
10

10
−3
−3 −2 −1 0 1 2 3
g

Jan Pralits (University of Genoa) Hydrodynamic stability July 8-10, 2013 71 / 77

 Constrained optimization

Lagrangian and optimality condition in N dimensions I

So fare we have looked at the static case in 1 dimension. The above approach can easily be
generalized to a N-dimensional state vector φ and a K -dimensional control vector g. We
therefore have to consider a Lagrange multiplier vector a with the same dimension as the vector
of state equations F, i.e. N.

The corresponding Lagrangian can now be written

L(φ
φ, g, a) = J (φ
φ, g) − a · F(φ
φ, g)
where · denotes a scalar product. Optimality conditions are given on L considering φ , g and a as
independent variables and therefore enforcing that

∂L ∂L ∂L
= 0, (j = 1, . . . , N), = 0, (k = 1, . . . , K ), = 0, (i = 1, . . . , N)
∂φj ∂gk ∂ai

When enforced these conditions using the variational approach. The system reads:
 T
∂L ∂F ∂J
=0 → a= adjoint equations
φ
∂φ φ
∂φ φ
∂φ
 T
∂L ∂F ∂J
=0 → a= optimality condition
∂g ∂g ∂g
∂L
=0 → F=0 state equation
∂a

Jan Pralits (University of Genoa) Hydrodynamic stability July 8-10, 2013 72 / 77

 Constrained optimization

Lagrangian and optimality condition in N dimensions II

What is the adjoint equation ?
By definition the adjoint of a linear operator A is a linear operator A∗ which satisfies the
following identity:
hv , A ui = hA∗ v , ui.
The h, i denotes an inner product.

φ, g) and the linear operator can be

In our case the state equation, in general, is written as F(φ
written ∂F/∂φφ. If we define the inner product as ha, bi = aT b, then the adjoint identity can be
written
∂F T
 
∂F
ha, φi = h
δφ φi
a, δφ
φ
∂φ φ
∂φ
The adjoint operator does not really have any physical meaning but is very useful in different
fields of analysis.

Jan Pralits (University of Genoa) Hydrodynamic stability July 8-10, 2013 73 / 77

 Constrained optimization

Lagrangian and optimality condition in N dimensions III

An example why the adjoint is useful
Consider the following optimization problem where φ , c and g have dimension N.

φ, g) = cT φ ,
J (φ (1)
A φ = g, (2)

A simple optimization update (steepest descent) is given by:

∂J i
 
gi+1 = gi − ρ
∂g

A straightforward way to compute ∂J /∂g is Instead, solve an additional linear system

given by finite differences
AT a = c.
∂J φ, g + en ) − J (φ
J (φ φ, g)
· en = By simple linear algebra we find
∂g 

where n = 1, .., N,   1 and en is a Cartesian J = cT φ = (AT a)T φ = aT A φ = aT g

unit vector.
Now J depends explicitly on the vector g, and
This has computational cost N. This means
∂J
that you must solve (2) N times. = a.
∂g

Computational cost 1, independently of N.

Jan Pralits (University of Genoa) Hydrodynamic stability July 8-10, 2013 74 / 77

 Constrained optimization

Lagrangian and optimality condition: IVP (ODE systems) I

The initial value problem of an ODE system describes the dynamics of a system which evolves in
time. For simplicity let us consider the linear system

φ
dφ
φ, g) =
F(φ − Lφφ = 0, 0≤t≤T
dt
φ (0) = g

Let us optimize the initial condition g in order to maximize the ”energy” of φ at the final time T
to the input ”energy”. In a similar manner we can define a minimization problem where the cost
function is g·g
J =
φ (T ) · φ (T )
The Lagrangian of the unconstrained problem can now be written, by first introducing the
Lagrange multipliers a(t) and b, as
Z T 
φ
dφ

L(φ
φ, g, a, b) = J (φ
φ, g) − a· − Lφ
φ dt − b · [φ
φ(0) − g ]
0 dt

In general this problem definition considers optimal (transient) energy growth and the
corresponding optimal perturbation. This analysis coincides with the analysis of maximum
nonmodal growth for a prescribed final time T . With a converged solution we have the so called
gain as G (T ) = J −1 .

Jan Pralits (University of Genoa) Hydrodynamic stability July 8-10, 2013 75 / 77

 Constrained optimization

Lagrangian and optimality condition: IVP (ODE systems) II

The optimality system is derived using a variational approach. Further, integration by parts must
be used to ”move” the derivatives from φ to a.

This derivation will be shown on the white board...

The optimality system finally reads

∂L ∂L dφφ
= 0, =0 → − Lφ
φ = 0, φ (0) = g state equation
∂a ∂b dt
∂L da −2φ φ(T )g · g
=0 → − − LT a = 0, a(T ) = adjoint equations
φ
∂φ dt φ(T ) · φ (T ))2
(φ
∂L φ (T ) · φ (T )
=0 → g = −a(0) optimality condition
∂g 2

Note that the adjoint equation is integrated ”backwards” in time. How about the solution
procedure ?

Jan Pralits (University of Genoa) Hydrodynamic stability July 8-10, 2013 76 / 77

 Constrained optimization

Lagrangian and optimality condition: IVP (ODE systems) III

Solution procedure Example

err=10−10
 
Initial: i = 0, g = g 1, J0 = 1015 , L =
0.1
0
p
−0.15
Unstable case
do 10
4

1 i =i +1 10
3

φ
dφ

E(t)
2
10
2 − Lφφ = 0, 0 ≤ t ≤ T ,
dt 1
10
with φ (0) = g
0
10
g·g 0 10 20 30 40 50
3 Ji = 
t

φ (T ) · φ (T ) L =
−0.1 p
0 −0.15
Stable case
2
da 10
4 − − LT a = 0, 0 ≤ t ≤ T ,
dt 1
10
g·g
with a(T ) = −2φφ(T )
φ(T ) · φ (T ))2

E(t)
0
(φ 10

−1
10
φ (T ) · φ (T )
5 g = −a(0)
2 −2
10
0 10 20 30 40 50
t
while (J i − J i−1 )/J i > err
T = 50 and p = 0, 1, 2, 3, 4, 5

Jan Pralits (University of Genoa) Hydrodynamic stability July 8-10, 2013 77 / 77