Slides KTH Stability and Transition

Non-modal stability
Luca Brandt & Peter Schmid

Linn FLOW Centre, KTH Mechanics
Imperial College London, UK
Outline
Stability of uid systems
modal limit
short Gme dynamics, matrix exponenGal
RecepGvity
resolvent norm
Resonance limit
Adjoint modes
SensiGvity
Structural sensiGvity
Base-ow sensiGvity
an example of instability and transiGon that is going to disappear...
Numerical simulaGon of a jet combustor: mixing of fuel and air

(Stanford University/NASA Ames)
Cirrus clouds developing in a jet stream over Saudi Arabia and the Red
Sea. The picture was taken from the Space ShuVle (NASA)
An example of Kelvin-Helmoltz shear instability
TransiGon to turbulence
Transition is a complex physical
process
critically depends on the
disturbance environment
is parameter-dependent
is important for the design of
fluid systems
Reynolds pipe ow experiment (1883)
Transition is a compex physical
process
fluid systems
Reynolds pipe ow experiment (1883)
Transition is a compex physical
process
fluid systems
TransiGon on a swept wing

SimulaGons by Hosseini (2013)
Classical TransiGon
outflow
turbulent boundary layer
Low levels of freestream turbulence

(<1%) exponential
growth of TS and
secondary instability
secondary 3D
instability
turbulent spots
Branch II
exponential
growth of 2D
TS waves
Branch I
flat plate
high velocity
low velocity
contours of 2
-vortices
inflow
hairpin vortices
Bypass TransiGon
High levels of freestream turbulence
(>1%) exponential
growth of TS waves is
bypassed
outflow
(decaying)
freestream turbulence
turbulent spots
(laminar) streaks
high velocity
low velocity
contours of 2
inflo
w
flat plate
Oscillators vs. noise ampliers

Open ows: global instability and transient growth
RotaGng cylinder
Hydrodynamic oscillators:

Global instability
Intrinsic frequency
Local absolute instability (WKB)

Boundary layer under
free-stream turbulence
Noise amplifiers:

Globally stable
Broad-band frequency spectrum
Local convecGve instability (WKB)

Globally transient growth of perturba4ons!
Two concepts of stability

Linear stability: we are interested in the minimum
critical parameter above which a specific initial
condition of infinitesimal amplitude grows
exponentially
Energy stability: we are interested in the maximum

critical parameter below which a general initial
condition of finite amplitude decays monotonically
Stability analysis
Hydrodynamic stability
SoluGons of Navier-Stokes: CoueVe, Poiseuille, boundary layer, Jet

Can we observe them in the lab?

CoueVe ow, Re=400
Boundary layer
Turbulent chaoGc moGons
Stability
Stable
Neutral
Unstable
CondiGonally
(nonlinearly) stable
We will talk about linear stability mostly
Overview
Stability analysis
Baseow and disturbances
Linearised equaGons
Normal modes, waves

Thermal instability
Benards problem, natural convecGon

AnalyGcal soluGon

Instability of shear ows (jet, wake, channel, boudnary layers)
Parallel ow assumpGon
Viscous and inviscid instability
Non-modal stability and transient growth
SensiGvity analysis

Laminar-turbulent transiGon
Stability analysis
1. Linear stability: steady base ow, soluGon of Navier-Stokes ekv.
U = (U, V, W ) ,
2. Disturbances u(xi , t),
T = T (xi )
T (xi , t)
3. Derive ekv for the disturabnce evoluGon u = F (u; U)
4. Assume small amplitudes, linear ekv.
u
= A(U)u
t
e i t
5. Assume u(xi , t) = u
= A(U)
u
egienvalue problem i u

Linear stability: normal modes
Homogeneous base ow in x och z U = (U (y), 0, 0) ,
(y)eikx+imz
PerturbaGons u(xi , t) = u
T = T (y)
i t
Wave-vector (k, m)
+i
k
m
i
>0
Unstable
Complex frequency
Wavenumber in x-direcGon
Vavenumber in z-direcGon
=0
Neutral
<0
Stable
Thermal instability: Benards problem

Success story of linear stability
z
T0
x or y
T0
Steady soluGon with linear temperature prole and

zero velocity
Warmer lighter parGcle below

Cold heavier parGcles above
Termiskt instabilitet: Benards problem

Triumf av linjr stabilitetsteori
T0
x or y
T0
Destabilising force: Buoyancy

Restoring force: viscous forces and thermal diusion
Termiskt instabilitet: Benards problem

Triumf av linjr stabilitetsteori
T0
x or y
T0
Destabilising force: Buoyancy

Restoring force: viscous forces and thermal diusion
Rayleighs number
Ra =
d4
Thermal expansion coecient

Diusivity
Viscosity
Temperature gradient
Rayleigh-Benard instability
z
T0
x or y
T0
Rayleigh number: ratio between buoyancy forces (temperature
gradient) and viscous forces the governing parameter
Linear stability theory: above a critical Rayleigh number of

1708 the conductive state becomes unstable to infinitesimal
perturbations
Energy stability theory: below a critical Rayleigh number of
1708 finite-amplitude perturbations superimposed on the
conductive state decay monotonically in energy
Rayleigh number: ratio between buoyancy forces (temperature
gradient) and viscous forces the governing parameter
Linear stability theory: above a critical Rayleigh number of

1708 the conductive state becomes unstable to infinitesimal
perturbations
Energy stability theory: below a critical Rayleigh number of
1708 finite-amplitude perturbations superimposed on the
conductive state decay monotonically in energy
Experiments show the onset of convec>ve instabili>es at a

cri>cal Rayleigh number of about 1710 !!!
Plane Poiseuille ow
Reynolds number: ratio between inertial forces and viscous forces the
governing parameter
Plane Poiseuille ow
governing parameter
Linear stability theory: above a critical Reynolds number of

5772 the parabolic profile becomes unstable to infinitesimal
perturbations
Energy stability theory: below a critical Reynolds number of
49.6 finite-amplitude perturbations superimposed on the
parabolic profile decay monotonically in energy
Plane Poiseuille ow
governing parameter
Linear stability theory: above a critical Reynolds number of

5772 the parabolic profile becomes unstable to infinitesimal
perturbations
Energy stability theory: below a critical Reynolds number of
49.6 finite-amplitude perturbations superimposed on the
parabolic profile decay monotonically in energy
Experiments show turbulent patches at a cri>cal Reynolds

number of about 1000 !!!
Two opposite behaviors

Linear stability theory, energy stability theory and experiments
are in excellent agreement for Rayleigh-Bnard problem
Linear stability theory, energy stability theory and experiments
show significant discrepancies for plane Poiseuille flow
Questions
Can we explain the success and failure of stability theory for
the two examples?
Is there a beVer way to invesGgate the stability of plane
Poiseuille ow (and many other wall-bounded shear ows)?
Energy equation
Linear growth mechanisms
SubcriGcal transiGon
The nonlinear terms of the Navier-Stokes equaGons conserve
energy
During transiGon to turbulence we observe a substanGal
increase in kineGc perturbaGon energy, even for Reynolds
numbers below the criGcal one.
energy
The increase in energy for subcritical Reynolds numbers is
related to a linear process, without relying on an
exponential instability;
energy
The increase in energy for subcritical Reynolds numbers is
related to a linear process, without relying on an
exponential instability;
linear instability without an unstable eigenvalue!
Non-modal approach
Linear stability problem

Start from linearised Navier-Stokes about base ow U
wriVen as iniGal value problem
d
q = Lq
dt
With soluGon in the form of matrix exponenGal

q = exp(tL)q0 ;
q(t = 0) = q0
Norm of matrix exponenGal

q = exp(tL)q0
Input output analysis:

maximum possible amplicaGon at Gme t over
all iniGal condiGons
||q||2
|| exp(tL)q0 ||2
G(t) = max
= max
q0 ||q0 ||2
q0
||q0 ||2
Norm of matrix exponenGal

q = exp(tL)q0
Input output analysis:

maximum possible amplicaGon at Gme t over
all iniGal condiGons
||q||2
2
G(t) = max
=
||
exp(tL)||
q0 ||q0 ||2
Definition of the norm of a matrix
Matrix norm G = max Gw

w =1
Euclidean scalar product
2
2
= wjH wj
||w||2 = hw, wi
Matrix as transformaGon with associated amplicaGon

H
Gw 2
w G Gw
=
w 2
wH w
1/2
Eigenvalues of GH G or singular values of G

:
GH Gvi =
Largest amplicaGon for the largest singular value:

1
= max
w
2 =1
Gw
Gv1 2
=
,
v1 2
Input and output basis vi , u i : G = U V H
1,
i vi
v1
Gv1 = u1
Eigenvalues vs Propagator Norm

q = exp(tL)q0
Matrix exponenGal dicult to compute

System eigenvalues used
L = SS
Eigenvalue decomposition
|| exp(tL)||2 = || exp(tSS
S: Column eigenvector

: Diagonal eigenvalues
)||2 = ||S exp(t)S
1 2
||
Traditional stability analysis:

Behavior deduced by system eigenvalues
Eigenvalues vs Propagator Norm

Upper and lower bounds of G(t)
Lower bound
e2t
max
|| exp(tL)||2
The energy cannot decay at a faster rate than that given by

the least stable eigenvalue max
Upper bound
|| exp(tL)||2 = || exp(tSS
||S||2 ||S
1 2 2t
|| e
)||2
max
Bounds of the matrix exponenGal

e2t
max
|| exp(tL)||2 ||S||2 ||S
Condition number:
1 2 2t
|| e
(S) = ||S||2 ||S
max
1 2
||
Two distinct cases:

(S) = 1
upper and lower bound coincide:

the energy amplica4on is governed by the least stable
eigenvalue
Bounds of the matrix exponenGal

e2t
max
|| exp(tL)||2 ||S||2 ||S
Condition number:
1 2 2t
|| e
(S) = ||S||2 ||S
max
1 2
||
Two distinct cases:

(S)
upper and lower bound can dier signicantly:

the energy amplica4on is governed by the least stable
eigenvalue only at large 4mes
Non-normality
2
(S) = ||S|| ||S
(S) = ||S||2 ||S
1 2
|| = 1
1 2
||
Normal stability problem:

orthogonal eigenvectors
Eigenvalues capture the dynamics
Non-Normal stability problem:

non-orthogonal eigenvectors
Eigenvalues capture the asympto4c
dynamics, not the transient behavior
Non-modal transient growth

The non-normality of the system can give rise to transient
energy amplicaGon
4me
Although we observe exponential decay for large times, the non-orthogonal

superposition of eigenvectors can lead to short-time growth of energy.
Short Gme dynamics

Taylor expansion of matrix exponenGal at t=0
E(t) = ||S||2 = hq, qi =

hexp(tL)q0 , exp(tL)q0 i
h(I + tL)q0 , (I + tL)q0 i
hq0 , q0 i + thq0 , (L + LH )q0 i

LH
Adjoint matrix defined by the norm used (Energy)
Short Gme dynamics

IniGal energy growth rate
1 dE
hq0 , (L + LH )q0 i
| 0+ =
E dt
hq0 , q0 i
1 dE
| 0+ =
E dt
H
(L
+
L
)
max
(L + LH ) Hermitian matrix: numerical abscissa of L
The numerical range

GeneralizaGon of the numerical abscissa
d
d
d
2
||q|| = h q, qi + hq, qi =
dt
dt
dt
hLq, qi + hq, Lqi =
2<{hLq, qi}
Definition of numerical range:

Rayleigh quotient of L
F(L) =
hLq, qi
z|z =
hq, qi
The numerical range
hLq, qi
z|z =
hq, qi
ProperGes of the numerical range

F(L) =

1. The numerical range is convex
2. The numerical range contains the spectrum of L
3. If L is normal, the numerical range is the convex hull of the
spectrum
convex
Non-normal system
Normal system
The numerical range

It can substantially larger than the convex hull of the spectrum:
positive energy growth even if stable eigenvalues!
d
||q||2 = 2<{hLq, qi}
dt
Normal system
Non-normal system
Non-normal stability problems

The numerical abscissa (numerical range) governs the
short-time behavior. The sign determines the initial
energy growth or decay
The least stable eigenvalue governs the long-time

behavior. The sign of the real part of max determines
the initial energy growth or decay
Rayleigh-Benard convecGon
T0
Normal stability problem
x or y
T0
The numerical range is the convex hull of the spectrum

The numerical range and the spectrum cross into the unstable
half-plane at the same Rayleigh number
Energy growth and instability occur at the Rayleigh number
The spectrum governs the flow behavior at all times
Ralin = Raen = 1708
Plane Poiseuille ow
Non-Normal stability problem
The numerical range is larger than the convex hull of the
spectrum
The numerical range crosses into the unstable half-plane
before the spectrum
Initial Energy growth occur before asymptotic instability
The spectrum governs the flow behavior only at long times
Reen = 49.6 Relin = 5772
Non-modal analysis
1 dE
| 0+ =
E dt
H
(L
+
L
)
max
Short-time: numerical abscissa
|| exp(tL)||2
Any time: matrix exponential
G(t ! 1) = lim || exp(tL)||2 = et

t!1
Long time: eigenvalues
max
Results for Poiseuille ow

G(t) envelope over many individual growth curves
For each point, a specific initial condition reaches its
maximum energy amplification at this point in time
Solid line: envelope G(t)

Dashed lines: evoluGon of
selected iniGal condiGons
Re = 1000, = 1

G(t) envelope over many individual growth curves
Potential for strong amplification of spanwise periodic
disturbances
Re = 2500, = 1,
=1
Re = 2500, = 0,
=2
Bypass transiGon
SupercriGcal Poiseuille ow, Re = 10000
Over short time horizon,
streamwise independent streaks
are favored
Streak
transient
growth
Modal growth
of Tollmien-Schlichting waves
Over long time horizon, waves

with weak or no streamwise
dependence are favored
Classical TransiGon
outflow
Low levels of freestream turbulence

(<1%) exponential
growth of TS and
secondary instability
secondary 3D
instability
turbulent spots
Branch II
exponential
growth of 2D
TS waves
Branch I
flat plate
high velocity
low velocity
contours of 2
-vortices
inflow
hairpin vortices
Bypass TransiGon
High levels of freestream turbulence
(>1%) exponential
growth of TS waves is
bypassed
outflow
(decaying)
freestream turbulence
turbulent spots
(laminar) streaks
high velocity
low velocity
contours of 2
inflo
w
flat plate
OpGmal iniGal condiGon

Initial condition that results in the maximum energy
amplification at a given time
q(t ) = exp(t L)q0
||q(t )||2 = ||q0 ||2 = 1
exp(t L) q0 = || exp(t L)|| q(t )
propagator input
amplicaGon output

exp(t L) q0 = || exp(t L)|| q(t )
propagator input
amplicaGon output
Singular value decomposition of a matrix
A = U V H
AV = U
1
v1
u1
= ||A||2

exp(t L) q0 = || exp(t L)|| q(t )
propagator input
amplicaGon output
svd(exp(t L)) = U V H
G(t ) = || exp(t L)||
exp(t L)
v1
u1

exp(t L) q0 = || exp(t L)|| q(t )
propagator input
amplicaGon output
svd(exp(t L)) = U V H
G(t ) = || exp(t L)||
exp(t L)
v1
Optimal initial condition

left principal
singular vector
u1
Optimal final condition

right principal
singular vector
OpGmal disturbance: Poiseuille ow

Re = 5000; = 0,
Optimal initial condition:
Counter-rotating streamwise vortices
Cross-stream
perturbaGon
velocity
Optimal final condition:

streamwise streaks
Streamwise
perturbaGon
velocity
=2
Non-Modal Growth
Lift-up mechanism in shear layers
(Ellingsen & Palm 1975, Landahl 1980)
Streamwise
vortices
Particle displaced in the

wall-normal direction
retain their horizontal
momentum
Streamwise vortices
induce streamwise
streaks
Streamwise
streaks
In boundary layers: wall-normal shear is O(Re)

streak growth O(Re)
Outline
modal limit
RecepGvity
resolvent norm
Resonance limit
Adjoint modes
SensiGvity
Base-ow sensiGvity
RecepGvity
Interested in the response of a uid system to external excitaGons (free-
stream turbulence, roughness, acousGc waves)
d
q = Lq + f
dt
External forces
General solution for zero initial condition: convolution integral
qp =
Zt
0
exp((
t)L)f ( )d
OpGmal response
Linear problem: consider the case of harmonic forcing f = fei!t
Regime solution given by the resolvent
qp = (i!
L)
Written as an input-output
problem: optimal response
||
qp ||
||(i! L)
R(!) = max
= max
f ||f ||
f
||f||
f||
= ||(i!
L)
||
Bounds of resolvent norm

Diagonalize the system matrix L
L = SS
1
||(i!
dist{i!, }

L)
|| = ||S(i!
1
|| (S)
dist{i!, }

L = SS
1
||(i!
dist{i!, }
(S) = 1

L)
|| = ||S(i!
1
|| (S)
dist{i!, }
Normal system: upper and lower bound coincide:

the classical resonance condi4ons holds
minimum distance from the spectrum

L = SS
1
||(i!
dist{i!, }
(S)

L)
|| = ||S(i!
1
|| (S)
dist{i!, }
Non-Normal system: upper and lower bound dier

we can have a pseudo-resonance
Strong amplica4on also far from system eigenfrequency
OpGmal forcing
Singular value decomposiGon svd of the resolvent norm
(i!
L)
f = ||(i!
Transfer funcGon Input
AmplicaGon
||
qp || = ||f|| = 1
(i!
L)
L)
|| qp
Output
R(! ) = ||(i!
v1
u1
Optimal harmonic forcing Optimal harmonic response

left principal
right principal
singular vector
singular vector
L)
||

Largest possible amplification versus forcing frequency
R(!)
= 0,
=2
= 1,
=1

Resolvent norm in the complex plane
R(!)
Re = 2500, = 0,
=2
Re = 2500, = 1,
=1
Component-wise energy transfer

Dene transfer funcGon with input/output matrices
qp = (i!
L)
f
H(!) = C(i!
L)
Take the worst case amplicaGon over all frequencies
||H||1 =
max
1<!<1
max (H)
Component-wise input-output analysis

fx
u
v
w
Jovanovic & Bamieh, JFM, 2004
fy
fz
Adjoint system
hp, qi = q H p
Inner product
Adjoint system
hp, Lqi = q H Lp = (LH q)H p = hLH p, qi
Eigenmodes/values
Bi-orthogonality
i qi = Lqi ;
+
j qj
+
hqm
, qn i =
= LH qj+
mn
Adjoint system
Eigenmodes/values
i qi = Lqi ;
Bi-orthogonality
h(LH
+
hqm
, (L
m )qn i
+
= hqm
, (L
+
)q
m m,
+
j qj
L+
= LH qj+
qn i = 0
n )qn i
=0
+
)hq
m
m , qn i = 0
+
hqm
, qn i =
mn
Expansion using eigenmodes

hqm , qn i =
Bi-orthogonality
Expand solution
q(t) =
mn
i q i e
mt
Given initial condition
q0
compute coefficients
using bi-orthogonality
n
hqn+ , q0 i =
X
i
m hqn+ , qm i = n
Use of adjoint eld for recepGvity

Linearized Navier-Stokes
(u, p)
(f + , m+ )
Differentiable fields
Sum and multiply
0=
@u
+ L(U, Re)u + rp = 0
@t
ru=0
@u
+ L(U, Re)u + rp f + + (r u)m+
@t
Integrate by parts over

time and space

Integrate by parts over time and space
Z t Z
@u
+ L(U, Re)u + rp f + + (r u)m+
@t
0
D
Z tZ +
Z t
Z
+
@f
@u
f
u
+ L+ (U, Re)f + + rm+ + p(r f + ) +
+
rJ
@t
@t
0
D
0
D
where
1 2
L(U, Re)u = U ru + u rU
r u
Re
1 2 +
L+ (U, Re)f + = U rf + rU f + +
r f
Re
J = U (u f + ) +
1
(rf + u
Re
ru f + ) + m+ u + pf +

Integrate by parts over time and space
Z t Z
@u
+ L(U, Re)u + rp f + + (r u)m+
@t
0
D
Z tZ +
Z t
Z
+
@f
@u
f
u
+ L+ (U, Re)f + + rm+ + p(r f + ) +
+
rJ
@t
@t
0
D
0
D
Definition of adjoint problem
Assume volume forcing, mass source and integrate in time
@u
+ L(U, Re)u + rp = f
@t
u(t) f + (t)
u(0) f + (0) =
Z tZ
0
ru=Q
f f + + Qm+ +
J n

u(t) f + (t)
u(0) f + (0) =
Z tZ
0
f f + + Qm+ +
Assume initial condition for adjoint system
u(t) u(t) = u(0) f + (0) +
Z tZ
0
J n
f + (t) = u(t)
f f + + Qm+ +
J n

u(t) f + (t)
u(0) f + (0) =
Z tZ
0
f f + + Qm+ +
u(t) u(t) = u(0) f + (0) +
Z tZ
0
J n
f + (t) = u(t)
f f + + Qm+ +
J n
Adjoint velocity gives sensitivity to initial condition and forcing
u2 (t)
= f + (0)
u(0)
u2 (t)
= f+
f

u(t) f + (t)
u(0) f + (0) =
Z tZ
0
f f + + Qm+ +
u(t) u(t) = u(0) f + (0) +
Z tZ
0
u2 (t)
= m+
Q
J n
f + (t) = u(t)
f f + + Qm+ +
Adjoint pressure gives sensitivity to mass source
J n

u(t) f + (t)
u(0) f + (0) =
Z tZ
0
f f + + Qm+ +
u(t) u(t) = u(0) f + (0) +
Z tZ
0
f + (t) = u(t)
f f + + Qm+ +
Gradient of adjoint field gives sensitivity to boundary

conditions
u2 (t)
1
uwall
Re
rf + + m+ n
J n
J n
Outline
modal limit
RecepGvity
resolvent norm
Resonance limit
Adjoint modes
SensiGvity
Base-ow sensiGvity
SensiGvity to internal changes
A(p)q = Bq
Governing parameter:
Reynolds number, base
ow, wavenumber
Perturbation expansion
(A + A)(q + q) = ( +
)B(q + q)
A(p)q = Bq
ow, wavenumber
(A + A)(q + q) = ( +
(A
B)q + (A
B) q + ( A
)B(q + q)
B)q + ( A
B) q = 0
Higher order
A(p)q = Bq
ow, wavenumber
(A + A)(q + q) = ( +
(A
B) q + ( A
)B(q + q)
B)q 0
(A
B)q 0
B) q + ( A
Use adjoint, left eigenvector
q + (A
q + (A
B) = 0
(A+
B) q + q + ( A
B + )q + = 0
B)q 0
A(p)q = Bq
Perturbation expansion,
linearize and use adjoint
q + Aq
= +
q Bq
rp
q + rp A q
=
q + Bq
Gradient: constraint
opGmizaGon (C. Cossu)
SensiGvity to a scalar parameter

Complex Ginzburg-Landau
ut = ( @x + @xx + (x))u
= U + icu
Eigenvalue sensitivity
rp
q + rp A q
=
q + Bq
A
u= u
r A =
@x
r = u
+ r A u
=
r = <{r }
Sensitivity of growth rate
A+ u
+ =
=
u
+ @x u
r ! = ={r }
Sensitivity of frequency
+ i!
SensiGvity to base ow modicaGons

1 2
ut + U ru + urU =
r u
Re
Linearized Navier-Stokes
Eigenvalue sensitivity
q + Aq
= +
q Bq
A=
( U )r
u
u
r( U )
A
u= u
A+ u
+ =
=
rU = r
u+ u
(r
u) H u
+
Relate mean ow modicaGon to passive control: small control forcing
+ i!
Flow chart for sensiGvity/recepGvity

Base flow calculations
Newton, SFD
Linearized NS
Adjoint NS
Global modes
Adjoint modes
Same for optimal

IC and response
Same for optimal

IC and response
Sensitivity
Localized forcing
base flow modifications
Receptivity
Forcing, boundary
conditions, mass sources
Flow around a cylinder

1. base ow

1. base ow
2. global modes

1. base ow
2. global modes
3. adjoint modes and receptivity
u
+
v+
1. base ow
2. global modes
4. sensiGvity, wavemaker
uwavemaker
vwavemaker
1. base ow
2. global modes
SpaGal feedback
1. base ow
2. global modes
5. sensiGvity to base ow modicaGons
Oscillators vs. noise ampliers

Open ows: global instability and transient growth
RotaGng cylinder
Hydrodynamic oscillators:

Global instability
Intrinsic frequency
Local absolute instability (WKB)

Boundary layer under
free-stream turbulence
Noise amplifiers:

Globally stable
Broad-band frequency spectrum
Local convecGve instability (WKB)

Globally transient growth of perturba4ons!
Stability analysis
Oscillators Modal analysis

- Largest Eigenvalue gives the asymptoGc behavior

Noise amplifiers Non-modal analysis

IniGal condiGon that gives the maximum energy
growth at a xed nal Gme.
OpGmal forcing (pseudo-spectra)
Forcing funcGon the gives the maximum
energy of the regime response when the
forcing is applied with a xed frequency

Instability mechanisms:
Globally unstable ows
The ow elds behave like an oscillator
Huerre & Monkewitz, Annu. Rev. Fluid Mech., 1990
In weakly parallel ows the WKBJ approach idenGes a
specic spaGal posiGon in the absolutely unstable
region which acts as a wavemaker.
Chomaz, Annu. Rev. Fluid Mech., 2005

A concept similar to that of wavemaker can be
introduced by invesGgaGng where in space a
modicaGon in the structure of the problem produces
the largest drio of the eigenvalue:
determine the region where feedback from velocity to
force is most eecGve

Giannep & Luchini, Journal of Fluid Mech., 2007, Pralits et al, 2010
Lagrange idenGty and

adjoint equaGons
Using dierenGaGon by parts
Introduce the adjoint fields f+ and m+, and

the adjoint linearised Navier-Stokes L+
Structural sensiGvity w.r.t.

perturbaGons
Perturbed eigenvalue problem
Structural perturbaGon: local force related to local
velocity

Linear variaGon of eigenvalue-eigenfuncGon expressed as

Using
Lagrange idenGty for perturbaGon eld
The sensiGvity is a spaGal map given by the product above

Dierent norms can be displayed (Spectral norm=worst case here)
ProducGon of perturbaGon kineGc

energy
Classic approach in stability analysis
Work of Reynolds stresses against the base ow shear

Wavemaker: where in space a modicaGon in the structure of the
problem produces the largest drio of the eigenvalue:

determine the region where feedback from velocity to
force is most eecGve

Comparison with kineGc energy producGon

Shedding mode I: wavemaker vs. energy growth
Structural sensi4vity
Produc4on
Mode generated in region of absolute instability

Pralits et al, JFM, 2010
Comparison with kineGc energy producGon

Shedding mode I: wavemaker vs. energy growth
Structural sensi4vity
Produc4on
Same region identified


base ow variaGons
Consider variaGons of the eigenvalue due to structural
variaGons of the base ow.

Assume arbitrary variaGons of the base ow
C bilinear operator expressing variaGons of L w.r.t.

Eigenvalue drio

base ow variaGons
Consider variaGons of the eigenvalue due to structural
variaGons of the base ow.

Assume arbitrary variaGons of the base ow
C bilinear operator expressing variaGons of L w.r.t.

Eigenvalue drio

base ow variaGons
Structural variaGons of the base ow:
linearized steady Navier-Stokes equaGons

Using Lagrange idenGty

with

Passive control
Wake control by means of small obstacles in the ow
Small control cylinder as localized structural perturbaGon:
First term of Lamb-Oseen expansion for drag at low Re
Structural perturbaGon is reacGng force, aligned with local

velocity and acGng at perturbaGon and base-ow level
DNS of passive control
Structural sensiGvity for Shedding

Mode I
Real part
Imaginary part
= 1.8
SensiGvity w.r.t. perturbaGons (spectral norm)

Mode I
Real part
Imaginary part
= 1.8
SensiGvity w.r.t. base ow (spectral norm)

Mode I
Real part
Imaginary part
= 1.8
Total SensiGvity (spectral norm)

Mode II
Real part
Imaginary part
= 4.85
SensiGvity w.r.t. perturbaGons (spectral norm)

Mode II
Real part
Imaginary part
= 4.85
SensiGvity w.r.t. base ow (spectral norm)

Mode II
Real part
Imaginary part
= 4.85
Total SensiGvity (spectral norm)
Passive control based on

sensiGvity analysis
SimulaGons of a secondary small cylinder
posiGoned as indicated by sensiGvity maps
1
0
velocity
0.5
0
-0.5
Time traces in the wake
-1
-2
-3
-1
-4
-1
-0.5
0.5
50
100
200
250
time
1.4
150
1.2
0.5
velocity
0.8
0.6
0.4
-0.5
0.2
-1
0
-1
-0.5
0.5
20
40
60
80
100
time
120
140
160
180
Instability mechanisms:
Noise ampliers
The aim of the present work is to extend the sensiGvity
analysis to ow behaving as noise ampliers
Huerre & Monkewitz, Annu. Rev. Fluid Mech., 1990
The target of the analysis is the largest singular value of
the system. We consider the resolvent norm in frequency
domain.
Schmid & Henningson, Stability and TransiGon of shear ows, 2001

A concept similar to that of wavemaker can be introduced
by invesGgaGng where in space a modicaGon of the base
ow produces the largest drio of the opGmal response:
wave-amplier

Brandt et al., Journal of Fluid Mech., 2011
Gradient of the resolvent norm

We assume
u = R( , U) f , R = S( , U)
, S=
M + L(U)
OpGmal forcing and response
R Rf , f
(u, u)
(R f , R f )
=
=
(f , f )
(f , f )
(f , f )
SensiGvity to base-ow modicaGons:

Overlap of opGmal forcing and response
=2
{(f ) u
Constrained opGmizaGon
K=
u , S(U) u
a , S (U) a
f}
u
f ,
Brandt et al, JFM, 2011
Boundary layer over at plate

Op4mal forcing: LiH-Up eect
Streamwise vorGces induce streamwise streaks:

component-wise non-normality, zero frequency and =1
Forcing acGve upstream (x<0) and above the at plate
Boundary layer over at plate

Op4mal forcing: TS-waves
ConvecGve non-normality, 2D forcing at frequency F=100.

Forcing acGve between branch I and II
SensiGvity to base-ow modicaGons

streamwise velocity component
Lio-up
TS-waves
Largest sensiGvity for TS-instability
Streak amplicaGon: very robust mechanism

Outline
modal limit
RecepGvity
resolvent norm
Resonance limit
Adjoint modes
SensiGvity
Base-ow sensiGvity
SOME MORE
Time-dependent ows
Generalize to Gme-periodic and Gme-dependent ows
Relate non-modal analysis to opGmizaGon problems
d
q = L(t)q
dt
With solution
q(t) = A(t)q0
Propagator
Initial
condition
Input-output analysis
Seek opGmal energy amplicaGon
hq, qi
G(t) = max
q0 hq0 , q0 i
hq, qi
G(t) = max
q0 hq0 , q0 i
hA(t)q0 , A(t)q0 i
= max
q0
hq0 , q0 i
hq, qi
G(t) = max
q0 hq0 , q0 i
hA(t)q0 , A(t)q0 i
= max
q0
hq0 , q0 i
hAH (t)A(t)q0 , q0 i
= max
q0
hq0 , q0 i
for Gme-dependent system
hAH (t)A(t)q0 , q0 i
G(t) = max
q0
hq0 , q0 i
AH A is a normal matrix
Maximum amplification for largest eigenvector of AH A

Principal eingenvector and eigenvalue can be
found by power iterations
(n+1)
q0
(n)
(n)
A q0
(n+1)
q0
updaGng
(n)
q0
(n)
(n)
A q0
A
direct problem
(n)
scaling
(n)
Aq0
(n)
Aq0
AH
adjoint problem
updaGng
(n+1)
q0
(n)
= (n) AH A q0
(n)
q0
direct problem
(n)
scaling
(n)
Aq0
A
(n)
Aq0
AH
adjoint problem
This technique (adjoint looping) can be applied to any general

time-dependent stability problem
Propagation of initial condition forward in time and of adjoint
initial condition backward in time
q0 (n+1) = (n) PH Pq0 (n)

updaGng q (n)
0
(n)
scaling
direct
problem
Pq0 (n)
PH Pq0 (n)
adjoint
adjoint
problem
problem
VariaGonal formulaGon of the problem

VariaGonal formulaGon of the opGmal growth problem
more general
We wish to maximize
with the constraint
||q||2
J=
! max
2
||q0 ||
d
q = L(t)q
dt
Listen Carlo Cossu on Friday!
Global modes: how to

Most problems with inhomogeneous direcGons,
complex geometry
Cannot use Fourier transform in two direcGons
EigenfuncGons and opGmals depend on more than
one direcGon u+ (x, y)
ComputaGonal issues
One vs two inhomogeneous directions:
state vector, matrix, operation count
0
q1
B q2 C
B C
C
q=B
B C
@ A
qN
L 2 CN N O(N 2 )
3
O(N )
State vector
Matrix size
Operation count
q1,1
B q1,2 C
B
C
B
q=B C
C
@ A
qN,N
L2C
N 2 N 2
O(N 4 )
O(N 6 )
ComputaGonal issues
One vs two inhomogeneous directions:
state vector, matrix, operation count
L2C
N 2 N 2
O(N 4 )
storage
O(N 6 )
CPU time
Direct eigenvalue algorithms become too expensive

Iterative algorithms, Arnoldi technique
Arnoldi algorithm
AcGon of the linear operator within an orthonormal basis V
H
L
System matrix
Orthogonal
basis
System matrix:
Hessenberg
matrix
Orthogonal
basis
L, exp(tL), LLH , exp(tL) exp(tLH )
USE THIS APPROACH TO COMPUTE MODAL AND NON-MODAL STABILITY
Arnoldi algorithm
AcGon of the linear operator within an orthonormal basis V
Represent stability matrix by a low-order approximaGon
based on V
H
L
System matrix
Hessenberg
matrix
Orthogonal
basis
Hessenberg matrix
Only multiplication by L are necessary
System eigenvalues approximated

by eigenvalues of H
Eig(L) Eig(H)
Example in 2D: ow in X-juncGon
Base flow
Lashgari et al, in preparation

Spectrum
Steady two-dimensional bifurcation


Eigenfunction


New asymmetric state

X-juncGon: Example in 2,5D

Asymmetric state unstable to 3D periodic disturbances
Snapshot method based on linear DNS
Example in 3D: Jet in cross ow

Use DNS and compute spectrum of matrix exponential q = exp(tL)q0
Bagheri et al, JFM, 2011, Schlatter et al.
Summary
non-normal operators are ubiquitous in uid ow problems
non-modal (mulG-modal) eects therefore dominant
non-modal analysis computaGonally more costly, many
extensions possible though

RecepGvity and sensiGvity

Use of adjoint modes
Structural sensiGvity, wavemaker
Base-ow sensiGvity, passive control
Tutorials
Consider Poiseuille or Couette flow
Orr-Sommerfeld and Squire system for
3D disturbances
Parallel shear ows
Parallel shear ows, cont
Orr-Sommerfeld and Squire equaGons
i
u
= 2 (D
v
)
k
i
w
= 2 ( D
v + )
k
Tutorials
TransientGrowth.m"
G(t) for selected ,; Re and given max time
Resolvent.m
R(omega) for selected ,; Re
NumRange.m
Numerical range for selected ,; Re
Tutorials
Neutral_a_Re.m"
Eigenvalues and G(t) for selected and given max time,
for a range of and Re
Neutral_alpha_beta.m
Eigenvalues and G(t) for given Re and max time, for a
range of ,

OptimalDisturbance.m
Optimal disturbance and response for selected ,; Re
Tutorials
Sens_OptDist.m"
Structural sensitivity of least stable mode and
q + Aq
= +
q Bq
of optimal disturbance to base flow modification

for selected ,; Re
Extra slides
Stability
Stable
Unstable
Neutral
Nonlinearly
unstable
Stability deniGons
E=
1
2
ui ui dV
V
E(t)
lim
t E(0)
Stable:
CondiGonally stable:
: E(0) <
Globally stable:
Linearly unstable:
Monotonically stable:
Stable
CondiGonally Stable with
0
dE
<0
dt
t>0
CriGcal Reynolds numbers

Flow monotonically stable

Flow globally stable

Flow linearly unstable
Re < ReE
Re < ReG
Re > ReL
CriGcal values for shear ows
Stability analysis
Search for ReE: monotonically stable ows
Linear analysis: ReL
Amplitude ? ReG ?

Disturbance equaGons
Energy equaGon
Linear growth mechanisms
Energy theory: ReE

1
= max
ui
ReE
V
V
ui uj
ui
xj
Ui
xj dV
ui
xj dV
VariaGonal problem: very conservaGve esGmate,

need for constraints!
Stability analysis
Search for ReE: monotonically stable ows
Linear analysis: ReL
Amplitude ? ReG ?

Linear analysis
Departure from equilibrium: slope

IdenGfy relevant mechanisms

Examine recepGvity and sensiGvity

Linear analysis
u
= L(U, t; Re) u
t
Time-independent problem: eigenvalue problem in Gme. Ex?
u(xi )e ),
u = L(U ; Re) u
u(xi , t) = (

Classic 1d problem:
Orr-Sommerfeld, Squire system for parallel ows
t
Parallel shear ows
Parallel shear ows, cont
Orr-Sommerfeld and Squire equaGons
Nonmodal stability analysis

Input-output approach
Nonnormal operators
Non-orthogonal eigenvectors
Matrix norm G = max Gw

w =1
Euclidean scalar product
2
2
= wjH wj
Matrix as transformaGon with associated amplicaGon

H
Gw 2
w G Gw
=
w 2
wH w
1/2
Eigenvalues of GH G or singular values of G

:
GH Gvi =
Largest amplicaGon for the largest singular value:

1
= max
w
2 =1
Gw
Gv1 2
=
,
v1 2
Input and output basis vi , u i : G = U V H
1,
i vi
v1
Gv1 = u1

Slides KTH Stability and Transition

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

Slides KTH Stability and Transition

Uploaded by

Copyright:

Available Formats

Non-modal stability

Luca Brandt & Peter Schmid

an example of instability and transiGon that is going to disappear...

Numerical simulaGon of a jet combustor: mixing of fuel and air

An example of Kelvin-Helmoltz shear instability

Reynolds pipe ow experiment (1883)

Reynolds pipe ow experiment (1883)

TransiGon on a swept wing

turbulent boundary layer

Low levels of freestream turbulence

turbulent boundary layer

Oscillators vs. noise ampliers

Globally transient growth of perturba4ons!

Two concepts of stability

Energy stability: we are interested in the maximum

Can we observe them in the lab?

CoueVe ow, Re=400

Turbulent chaoGc moGons

We will talk about linear stability mostly

Benards problem, natural convecGon

3. Derive ekv for the disturabnce evoluGon u = F (u; U)

4. Assume small amplitudes, linear ekv.

Linear stability: normal modes

Homogeneous base ow in x och z U = (U (y), 0, 0) ,

Thermal instability: Benards problem

Steady soluGon with linear temperature prole and

Warmer lighter parGcle below

Termiskt instabilitet: Benards problem

Destabilising force: Buoyancy

Termiskt instabilitet: Benards problem

Destabilising force: Buoyancy

Thermal expansion coecient

Linear stability theory: above a critical Rayleigh number of

Linear stability theory: above a critical Rayleigh number of

Experiments show the onset of convec>ve instabili>es at a

Linear stability theory: above a critical Reynolds number of

Linear stability theory: above a critical Reynolds number of

Experiments show turbulent patches at a cri>cal Reynolds

Two opposite behaviors

Linear growth mechanisms

Linear stability problem

With soluGon in the form of matrix exponenGal

Norm of matrix exponenGal

Input output analysis:

Norm of matrix exponenGal

Input output analysis:

Definition of the norm of a matrix

Matrix norm G = max Gw

Euclidean scalar product

Matrix as transformaGon with associated amplicaGon

Eigenvalues of GH G or singular values of G

Largest amplicaGon for the largest singular value:

Input and output basis vi , u i : G = U V H

Eigenvalues vs Propagator Norm

Matrix exponenGal dicult to compute

)||2 = ||S exp(t)S

Traditional stability analysis:

Eigenvalues vs Propagator Norm

The energy cannot decay at a faster rate than that given by

Bounds of the matrix exponenGal

|| exp(tL)||2 ||S||2 ||S

(S) = ||S||2 ||S

Two distinct cases: