Vous êtes sur la page 1sur 21

Mezić Research Group

TIME-AVERAGES OF OBSERVABLES

Marko Budišić
mbudisic@engr.ucsb.edu
SIAM Conference on Applications of Dynamical Systems
Snowbird, May 2011
The minisymposium that follows focuses on time averaging in analysis and design of dynamical systems. This presentation aims to introduce
the recent advances in analyzing dynamical systems by averaging functions, or observables, along the trajectories of the system.

First of all, let me briefly say what I will NOT be talking about. This work should not be confused with averaging of the model where the order
of the model is reduced by substituting dynamics evolving on fast time scales by their averaged effect.

The time-averages of observables, on the other hand, involve averaging observables, or measurements, along trajectories of the system.
Common names used are: time-averages, trajectory-averages, dynamical averages, Cesàro means, ergodic averages, lagrangian averages,
cumulative moving averages, statistics...
Mezić Research Group

Problem: Trajectories are not robust descriptions of dynamics in


chaotic systems, due to exponential sensitivity to errors.

Motivation: Time-averages are robust under time-discretization


and round-off error even when trajectory is not.
[Sigurgeirsson, Stuart, 2001]

Time-averages provide macroscopic descriptions of


geometry of dynamics: Lyapunov exponents, residence
times, power spectrum,...
[Cvitanović et al., Chaos Book]

Use time-averages to extend the library of robust indicators of


dynamical behaviors for analysis and design.

Monday, May 23, 2011 Marko Budišić: Time Averages of Observables 2

In chaotic systems, trajectories are non-robust. Even the smallest perturbations, common in numerical simulations, result in completely
different trajectories. In Chaos Book, Cvitanović argues that trajectories are wrong objects to look at in such cases, and that the best we can
do is to analyze features of the set in state space that the trajectory explores. Despite trajectories being sensitive to perturbations,
simulations and numerics show that averaging over them still results in reliable time-averages, as noticed by Sigurgeirsson and Stuart.

Time averages provide a macroscopic description of features of the set a trajectory is exploring in the state space. Many common dynamical
indicators (Lyapunov coefficients, residence times, power spectra...) can be formulated as time averages of particular observables along
trajectories. This body of work aims to systematically extend the library of those indicators by a careful choice of observables that are
averaged.
Mezić Research Group

Compact manifold M
Flow map Φt : M → M
Observable f :M→C
� T
˜ 1
f (x) = lim f [Φt (x)]dt
T →∞ T 0

Birkhoff: Limits exist a.e.


wrt any measure preserved by
flow.

Monday, May 23, 2011 Marko Budišić: Time Averages of Observables 3

Dynamics are modeled by an flow map on a COMPACT manifold. Observables are scalar functions on that manifold. In most of the cases,
we will focus on continuous observables. Tilde above a function will indicate that we are taking a time average of
the function along the trajectory. Of course, there exist analogous formulations for maps and PDEs.

Since the definition of the time averages includes taking infinite-time limits, it is a valid question to ask whether these limits exist or not. A large
class of dynamical systems preserve SOME measure. Birkhoffʼs theorem asserts that time averages converge almost anywhere with respect
to any measure preserved. A common examples are area-preserving systems, e.g., advection in divergence-free flows and Hamiltonian
systems. While Birkhoffʼs result provides some comfort, it might still include cases where the exceptional set, i.e., the set where the time-
averages are ill-defined, includes a big portion of the state space.
<transition>
However, there is a known structure that is a good example of non-existence, an attracting heteroclinic loop. As trajectories are attracted to it,
they linger for longer and longer periods of time close to each fixed point in the loop, which prevents the time average from settling down. Most
other configurations in state space have well defined time averages and they can be computed by numerical simulations.
Mezić Research Group

Compact manifold M Non-existence: Attracting heteroclinic loop

Flow map Φt : M → M
Observable f :M→C
� T Trajectories slow down more and more
˜ 1
f (x) = lim f [Φt (x)]dt as they pass equilibria, preventing
T →∞ T 0 convergence.

Birkhoff: Limits exist a.e.


wrt any measure preserved by
flow.

Monday, May 23, 2011 Marko Budišić: Time Averages of Observables 3

Dynamics are modeled by an flow map on a COMPACT manifold. Observables are scalar functions on that manifold. In most of the cases,
we will focus on continuous observables. Tilde above a function will indicate that we are taking a time average of
the function along the trajectory. Of course, there exist analogous formulations for maps and PDEs.

Since the definition of the time averages includes taking infinite-time limits, it is a valid question to ask whether these limits exist or not. A large
class of dynamical systems preserve SOME measure. Birkhoffʼs theorem asserts that time averages converge almost anywhere with respect
to any measure preserved. A common examples are area-preserving systems, e.g., advection in divergence-free flows and Hamiltonian
systems. While Birkhoffʼs result provides some comfort, it might still include cases where the exceptional set, i.e., the set where the time-
averages are ill-defined, includes a big portion of the state space.
<transition>
However, there is a known structure that is a good example of non-existence, an attracting heteroclinic loop. As trajectories are attracted to it,
they linger for longer and longer periods of time close to each fixed point in the loop, which prevents the time average from settling down. Most
other configurations in state space have well defined time averages and they can be computed by numerical simulations.
Mezić Research Group

Convergence of time-averages
Orbits in regular zones Orbit in mixing zone “Sticky” orbit

images: [Levnajić, Mezić, 2010]

Standard map convergence error (log)

Intermittency: trajectories get temporarily


trapped around marginally stable periodic
islands.
[Cvitanović et al., Chaos Book]

Coping with slow convergence:


• cycle expansions
• comp. benefits for adaptive sampling
• finite-time analysis
Difference in averages at 0.6×105 and 1.0×105 steps
Monday, May 23, 2011 Marko Budišić: Time Averages of Observables 4

Even when time averages are well defined, the phenomenon of trajectories lingering close to marginally stable for extended period of time
and then moving away can (theoretically) cause arbitrarily slow convergence rates.

This is demonstrated on an example of a Chirikov standard map – a Hamiltonian iterated map with a state space that contains both
regular and mixing zones that come in contact in a fractal region. Theoretically, the convergence can be arbitrarily slow even when the
time averages of observables exist. The graphs on the top show log-log plots difference between finite time averages and infinite time
averages. Convergence of time averages in regular zones is dominated by geometric slope of -1. In mixing zones, convergence is
dominated by slope of -1/2. In zones where intermittency occurs, convergence is characterized by intermittent periods of large errors as
trajectories get trapped around marginally stable objects. Such zones are known to occur in KAM systems when tori break down.

Predrag Cvitanovic championed an analytical method of cycle expansions to compute time-averages in intermittent zones. He proposed
that, instead of averaging an observables along chaotic trajectories, we could obtain the result by averaging along periodic orbits that are
embedded within the chaotic zones. This is a successful strategy, but it involves determining where unstable periodic orbits lie, which is not
an easy task. Therefore cycle expansions do not provide a general-enough replacement for direct averaging.

Mixing zones and intermittent zones are sufficiently explored by a single trajectory, due to local ergodicity with respect to measures
supported on sets containing such zones. If our goal is to compute time averages on a sufficiently dense set of points in the state space, we
might use this fact in an adaptive sampling of state space. While waiting for the time average to converge, the trajectory explores a fat set in
state space, which does not have to be sampled by additional initial conditions later, as all points on the trajectory will have the same time-
average and can be identified with the initial condition of the trajectory. Therefore, even if we have to wait for a long period of time, if our goal
is coverage of state space, we get additional computational benefit for waiting for time averages to converge. This benefit is especially
important when the space is high-dimensional

Finally, we might just have to accept that finite-time averages are the best that we can do. I will revisit this topic briefly at the end of this
presentation, but understanding of how to harness finite-time averages is one of the burning issues that users of time averages will have to
face.
Mezić Research Group

Computing averages for continuous systems

Trajectory-based Statistics-based
numerical consistency ? numerical consistency

Monday, May 23, 2011 Marko Budišić: Time Averages of Observables 5

When dynamical system is an iterative map, we have no issues of computing it on a computer. When simulating ODEs and PDEs, we use
discretization schemes to approximate continuous trajectories. Most discretization schemes are validated by providing error estimates
compared to trajectories of continuous systems. The work of Paul Tupper, Jonathan Mattingly, Andrew Stuart, Xiaoming Wang and others
demonstrates that trajectory-consistency is not always enough to ensure that time-averages will be correctly computed when systems
are high-dimensional, or have a special structure that needs to be conserved.
*** slide transition ***
A common observation is that a numerical scheme shouldnʼt introduce additional attractors into dynamics. Both works by Tupper and
Wang show that some commonly used schemes fail to satisfy this requirement. Tupper analyzed molecular dynamics (high-DOF
Hamiltonian) integrators that conserve energy. If they are not symplectic, they might stabilize periodic orbits that are otherwise unstable for
continuous systems. If the continuous system was ergodic, such stable periodic islands might destroy the ergodicity of the numerical scheme
resulting in incorrectly computed time-averages.

Wang showed that in PDE simulations of heat convection, time-averaged indicators, such as Nusselt number, can be incorrectly computed
if the scheme introduces attractors that arenʼt there for continuous system. This occurs in some commonly used numerical schemes. His
theoretical result about precompactness of the set of attractors generated by numerical schemes is essentially the same as Tupperʼs, implying
that the culprit for non-consistency of time averages is the possible set of attractors that numerical methods might introduce, even at small step
sizes.

In conclusion, while the work of Sigurgeirsson and Stuart provides hope that time-averages can be computed reliably for maps, we still need to
take care in using integrators for continuous systems, paying special attention to systems with symplectic structure and infinite-dimensional
systems.
Mezić Research Group

Computing averages for continuous systems


Trajectory-based Statistics-based
numerical consistency ? numerical consistency

[P. Tupper, 2007] Numerical ergodicity of Hamiltonian systems


Time averages are correct if numerical scheme:
1. ... is ergodic, e.g., has no artificial limit cycles.
2. ... conserves energy.
3. ... conserves phase space volume (symplectic). Symplectic schemes perform better
than energy-conserving ones.

[X. Wang, 2010] Simulation of cts. infinite-dimensional dissipative systems


Time averages are correct if system is uniformly cts. and:
Proof-of-concept
1. ... numerical global attractors for small numerical scheme for
time steps form a pre-compact set. computation of Nusselt
2. ... numerical scheme is finite-time no. in a convection
model.
uniformly convergent.
Monday, May 23, 2011 Marko Budišić: Time Averages of Observables 5

When dynamical system is an iterative map, we have no issues of computing it on a computer. When simulating ODEs and PDEs, we use
discretization schemes to approximate continuous trajectories. Most discretization schemes are validated by providing error estimates
compared to trajectories of continuous systems. The work of Paul Tupper, Jonathan Mattingly, Andrew Stuart, Xiaoming Wang and others
demonstrates that trajectory-consistency is not always enough to ensure that time-averages will be correctly computed when systems
are high-dimensional, or have a special structure that needs to be conserved.
*** slide transition ***
A common observation is that a numerical scheme shouldnʼt introduce additional attractors into dynamics. Both works by Tupper and
Wang show that some commonly used schemes fail to satisfy this requirement. Tupper analyzed molecular dynamics (high-DOF
Hamiltonian) integrators that conserve energy. If they are not symplectic, they might stabilize periodic orbits that are otherwise unstable for
continuous systems. If the continuous system was ergodic, such stable periodic islands might destroy the ergodicity of the numerical scheme
resulting in incorrectly computed time-averages.

Wang showed that in PDE simulations of heat convection, time-averaged indicators, such as Nusselt number, can be incorrectly computed
if the scheme introduces attractors that arenʼt there for continuous system. This occurs in some commonly used numerical schemes. His
theoretical result about precompactness of the set of attractors generated by numerical schemes is essentially the same as Tupperʼs, implying
that the culprit for non-consistency of time averages is the possible set of attractors that numerical methods might introduce, even at small step
sizes.

In conclusion, while the work of Sigurgeirsson and Stuart provides hope that time-averages can be computed reliably for maps, we still need to
take care in using integrators for continuous systems, paying special attention to systems with symplectic structure and infinite-dimensional
systems.
Chaos 20, 033114 !2010" Mezić Research Group

f˜ ◦ Φt ≡ f˜ Time-averages are invariants of motion.

Time averaged
Trajectories Φt (x) 033114-6 An observable
033114-6 Z.Z.Levnajić
Levnajićand
andI.I.Mezić
Mezić f (x) observable
Chaos
Chaos f˜(x)
20,20, 033114
033114 !2010"
!2010"

M = T2

Level-sets of time-averaged observables partition the


state space into invariant sets.

Monday, May 23, 2011 Marko Budišić: Time Averages of Observables 6

The first important property of time-averages is their invariance with respect to dynamics. It is common knowledge that invariants of motion,
such as energy and momenta in some mechanical systems, play a strong role in understanding the dynamics of systems. However, as said
before, we can compute time-averages for most dynamical systems. What role do time-averages play there?
uation !7" !left" and the corresponding phase space portraits !right" done
f = Dynamical
cos!2"y". Top row:system I will
# = 0, middle: bebottom:
# = 0.09, using throughout
this presentation is a Chirikov standard map. It is a Hamiltonian system evolving on a 2-torus,
# = 0.18. The grid
ations.
with a parameter epsilon that takes it from a
completely regular discrete shear, to global chaos. Letʼs visualize time-averages of an observable
that alternates signs on a checkerboard pattern. On the rightmost image I am coloring the level sets of the time-averaged observable.
quency. In the context of standard map it is convenient to
FIG. 2. !Color online" The graph of the wavelet W as defined by Eq. !8" in !a". Its time average under the dynamics of the standard map equation !7" for
e a slightly bigger frequency for y coordinate—since the
As a consequence of invariance, level-sets of # =time-averaged observables
0.09 in !b", # = 0.16 in !c", are
and # = 0.18 in !d". invariant
Grid: sets.
800$ 800 and Therefore,
tfinal = 50 000 iterations. by looking at level-sets of an averaged
nsport in x direction is much faster, the function averages FIG. 2. !Color online" The graph of the wavelet W as defined by Eq. !8" in !a". Its time average under the dynamics of the standard map equation !7" for
t toobservable, we can
zero rather quickly immediately
in this establish# =between
direction. Time-average 0.09 in !b", # which regions
= 0.16 in !c", and # = 0.18of state
in !d". Grid:space
800$ 800 there
and tfinal =is
50 no transport. Just from looking at the
000 iterations.
ots rightmost figure,basis
for several Fourier wefunctions
can immediately see
with increasing fre-that there is no transport between regions colored light-blue 31 and red.
quency are shown in Fig. 3. Note the relationship between obtained, but in our context it is relevant to study how
the used frequency and the scale of the best-visualized phase convergence properties depend on the type of trajectory of f.
quency are shown in Fig. 3. Note the relationship between obtained,31 but in our context it is relevant to study how
space details. Thus, in this section we study numerically the convergence
the used frequency and the scale of the best-visualized phase convergence properties depend on the type of trajectory of f.
While the Fourier functions give smoothly colored plots, properties of time averages in order to estimate the precision
space details. Thus, in this section we study numerically the convergence
wavelets produce sharper and more detailed plots with a bet- of the values obtained and relate them to the different types
While the Fourier functions give smoothly colored plots,
ter distinction between the independent invariant sets. How-
properties of time averages in order to estimate the precision
of orbits. For simplicity we use only Fourier functions !the
wavelets produceinvariant
ever, different sharper sets
and still
morehappen
detailed to plots with a bet-
be assigned the of the for
results values obtained
wavelets and relate
are similar". them to
Consider thethe
tth different types
partial time
tersame
distinction
colors. between
To remedy thethis,
independent
we use the invariant sets.ofHow-
full power the of orbits.
average forFor simplicity
a function by only Fourier functions !the
we use
f given
ever, different
ergodic invariant
partition conceptsets still happen
in Sec. V. to be assigned the results for wavelets are similar". Consider the tth partial time
same colors. To remedy this, we use the full power of the average for a functiont−1
f given by
ergodic partition concept in Sec. V. 1
IV. THE CONVERGENCE PROPERTIES f t!x0,y 0" = # f!Tk!x0,y 0"", !9"
t k=0t−1
1
IV. THEIt isCONVERGENCE PROPERTIES f t!x0,y 0"t = # f!T! k!x0,y 0"", !9"
well known that time averages of functions under with limt→! f !x0t, k=0 y 0" = f !x0 , y 0" $which we assume exists for
dynamics of measure-preserving maps can converge arbi- all grid-points !x0 , y 0"%. The difference
It isslowly
trarily well known that
!see Ref. 30 time averages and
for discussion of functions under
related results". with limt→! f t!x0 , y 0" = f !!x0 , y 0" $which we assume exists for
dynamics
However,offormeasure-preserving maps canbounds
specific continuous functions converge arbi-
on rate of all grid-points
"!t" = &f !!x ,y !x0", y−0f"%.
t The difference
!x ,y "& !10"
31 0 0 0 0
trarily slowly !see
convergence Ref. 30 for discussion
are computable. and depend
Bounds that related only
results".
on
the maximum
However, absolute
for specific value of afunctions
continuous continuous function
bounds on can
ratebe
of is a sequence !whose asymptotic behavior is to be studied in
31 "!t" = &f !x0,y 0" − f t!x0,y 0"& !10"
convergence are computable. Bounds that depend only on
the maximum absolute value of a continuous function can be is a sequence whose asymptotic behavior is to be studied in
Mezić Research Group

Intersections of invariant partitions are again invariant partitions.


Intersecting increases the resolution of information.
033114-7 Ergodic theory and visualization. I
Mezić
[f˜ ∩
f˜(x)
g̃](x) Chaos 20, 033114 !2010"
[f˜ ∩ g̃g̃(x)
∩ . . . ](x)

Transport?

Transport?

Repeating the procedure for linearly independent observables


limits to the ergodic partition – collection of ergodic sets.

[Susuki, Mezić, 2009] [Levnajić, Mezić, 2010]


Monday, May 23, 2011 Marko Budišić: Time Averages of Observables 7

There are two questions that are left by looking at a single averaged observable:
1. Is there a transport within any region of uniform color?
2. Does choice of the observables matter?

Different observables will result in different level-sets of time-averaged observables. Each of these level-set partitions will be an invariant
partition. Single invariant partition does not answer whether there is transport inside of individual level sets.
*** slide transition ***
However,
the wavelet W asby intersecting
defined by Eq. !8" in invariant partitions
!a". Its time average generated
under the dynamics of from different
the standard observables
map equation !7" for we can increase the resolution of information and get finer
= 0.18 in !d". Grid: 800$
information about800 the
and tfinal
level= 50 000
sets.iterations.
Continuing this process for linearly independent observables refines the invariant partition more and more,
limiting to the ergodic partition. In it, every set is an ergodic set. Within any ergodic set, any two points are connected by transport. In this
sense, ergodic partition extends the notion of foliation of the state space into invariant tori for completely integrable dynamical systems.
Note the relationship between obtained,31 but in our context it is relevant to study how
aleTherefore,
of the best-visualized
combining phase
information convergencefrom properties
differentdepend on the typecan
time-averages of trajectory
provideofveryf. detailed information about global dynamical features of the
state space. Yoshihiko Susukiʼs Thus,paper
in this is
section
aboutweapplications
study numerically the convergence
of ergodic partition visualization to analysis of continuous power systems.
ns give smoothly colored plots, properties of time averages in order to estimate the precision
more detailed plots with a bet- of the values obtained and relate them to the different types
dependent invariant sets. How- of orbits. For simplicity we use only Fourier functions !the
still happen to be assigned the results for wavelets are similar". Consider the tth partial time
, we use the full power of the average for a function f given by
Sec. V.
t−1
1 FIG. 3. !Color online" Plots for various time-averaged functions for the standard map equation !7" w
ROPERTIES f !x0,y 0" = # f!Tk!x0,y 0"",
t
!9"islands" are revealed when the frequency of the function is increased. Grid: 800$ 800 and
resonant
t k=0
me averages of functions under with limt→! f t!x0 , y 0" = f !!x0 , y 0" $which we assume exists for
ving maps can converge arbi- all grid-points !x0 , y 0"%. The difference
discussion and related results". relation to the initial point !x0 , y 0" and the !-value, depend- A. The regular
ous functions bounds on rate of ing on
31 "!t" = &f !!x0,y 0" − f t!x0,y 0"& !10"the phase space regions with different dynamical be-
Bounds that depend only on haviors. We consider the function f = cos!2"y" !cf. Fig. 1", For all the
of a continuous function can be is a sequence whose asymptotic behavior is to be studied
define inf ! = f t for t = 108, and consider the first 106 iterations. averages of con
Mezić Research Group

Intersections of invariant partitions are again invariant partitions.


Intersecting increases the resolution of information.
033114-15
[f˜ ∩
f˜(x)
g̃](x)
Ergodic theory and visualization. I [f˜ ∩ g̃g̃(x)
∩ . . . ](x) Chaos 20, 033114 !2010"

Transport?

Transport?

FIG. 11. !Color online" Approximations of the ergodic partition for the standard map equation !7" with "=0.12. Left: three-function approximation using two
functions from Fig. 10 and the function sin!4$x"sin!4$y". Right: four-function approximation using these three functions in addition to sin!10$x"sin!10$y".
The grid 800#800 is used for all time averages and the cell division for both approximations is L=50.

Repeating the procedure for linearly independent observables


u =x −x , v =y −y , However, since we cannot graphically represent the 5D space
limits towhichthe ergodic partition –results,
2 1 2 2 1 2

reduce the system to two maps: a standard map in


collection of ergodic
we consider here three-function MSP visualized insets.
3D embedding. An example with "= ! =0 is shown in Fig.
coordinates !u1 , v1" with the parameter 2! and an integrable
13, obtained for 4D grid 13#13#13#13 initial points as
twist map in !u2 , v2". Given this structure, an interesting way
follows:
of varying parameters of the map equation !13" is by letting
"=2!. We slice the 4D phase space by a 2D section with • The lattice-grid is set in full 4D phase space, and iterated
fixed !x2 , y 2"=!0,0", and consider a grid of 500#500 initial using the Froeschlé map equation !13" with "= ! =1 for ten
grid-points in !x1 , y 1"-space, plotting the time averages of a [Susuki, Mezić, 2009]
iterations, in order to randomize the points !starting from [Levnajić, Mezić, 2010]
Monday, May 23, 2011 single function f 2 =2 cos!2$y 1"+cos!2 $y 1"cos!2
Marko Budišić: Time 2"
$yAverages the uniform lattice-grid in the case of "= ! =0 might create
of Observables 7
+cos!12$x2"+cos!12$x1" obtained by running the dynamics trajectories that strongly overlap".
There are two questions thatfor are200left
000 by
iterations.
looking The at results are shown
a single in Fig. 12. Note
averaged • From thus obtained 13#13#13#13=28 561 initial 4D
observable:
1. Is there a transport withinthe anydeparture
region from
of the known color?
uniform standard map’s phase space points, set 13#13#13=2197 to have !y 1 =0, y 2 =0", an-
due to increasing interaction between the maps. For small " other 13#13#13 to have !y 1 = 21 , y 2 = 21 ", and so on, de-
2. Does choice of the observables matter?
values, the phase space maintains its regular structure with pending on how many resonances are to be visualized.
only small and localized chaos similar to standard map for • Run the dynamics for tfinal =100 000 iteration starting from
Different observables will result
smallin".different
The coloring level-sets
here indicates of two-dimensional
time-averaged inter-observables.
thus createdEachensemble of these
of initiallevel-set
4D points. partitions will be an invariant
partition. Single invariant partition
sections of doesthe realnot answer
4D invariant setswhether there2D
with the selected is transport
• As the result,inside of individual
majority level for
of points will account sets. quasip-
*** slide transition *** phase space section. It appears that the global chaotic transi- eriodic orbits, while the selected points will visualize the
tion in the
However, by intersecting invariant sense of vertical
partitions transport through
generated the phase space
from different observables we can increase the resolution of information and get finer
chosen resonances
occurs around "=2 ! =0.05.
information about the level sets. Continuing this process for linearly independent observables refines the invariant partition more and more,
Theit,structure Note that in Fig. 13, all the mentioned phase space fea-
limiting to the ergodic partition. In everyofset Qe foris the
an Froeschlé
ergodicmap set.forWithin
"= ! =0 is any ergodic set, any two points are connected by transport. In this
clearly a product of two “rational combs” discussed earlier. tures are visualized as expected. The chosen resonances for
sense, ergodic partition extends the notion of foliation of the state space into invariant tori! for1 completely integrable dynamical systems.
This product is topologically a torus to which this case are !y 1 =0, y 2 =0", y 1 = 2 , y 2 = 21 ", !y 1 = 31 , y 2 = 31 ",
and !y 1 = 32 , y 2 = 32 ".
• a line of length 1/q is
Therefore, combining information from different time-averages can attached at every point !y , y " such
1 2provide very detailed information about global dynamical features of the
We examine the three-function MSPs for the Froeschlé
state space. Yoshihiko Susukiʼs thatpaper
one of the y i’s is irrational
is about and the other
applications rational, p/q
of ergodic partition visualization
map with "=2! %0 byto analysis
employing 4D of
gridcontinuous
of 12#12#12power systems.
!with p/q an irreducible fraction",
• a rectangle of sides q1 ,q2 is attached at every point #12 initial grid-points and the same number of total itera-
!y 1 = p1 /q1 , y 2 = p2 /q2", where p1 /q1 and p2 /q2 are irre- tions. The ensemble of initial points with selected resonance
ducible fractions. points is constructed as above. We consider structural
changes in the MSPs by changing the "-value, as shown in
Since two rectangles generically do not intersect in a five Fig. 14. As the coupling intensity grows, the structure from
dimensional !5D" Euclidean space, the appropriate embed- the previous figure is destroyed, in a way similar to what
ding dimension for the Froeschlé map is five. To avoid self- observed for the standard map !cf. Fig. 8". In particular, note
intersections in the embedding, we would thus need to con- the different mechanisms of destruction of resonances and
sider at least five independent functions. This information is irrational orbits. The MSP’s structure reports a given level of
important when choosing the number of functions that are regularity in the phase space persisting for a certain range of
used in clustering and approximation of the ergodic partition. coupling parameter strengths. This is a manifestation of
Mezić Research Group

Ergodic Quotient (EQ) and empirical measures


State space portrait Averaging Ergodic quotient
� �
Φt (x) ⊂ M ˜ ˜
f1 (x), f2 (x), . . . ⊂ l ∞
∀f from cts. basis

Empirical measures represent time


µx
averaging through spatial integration.

f˜(x) = f dµx
M x Φt(x)
EQ is the set of weak-representatives of
empirical measures for all initial cond.

Choice of topology on EQ defines coherent structures,


e.g., discrete metric topology corresponds to ergodic partition.
Monday, May 23, 2011 Marko Budišić: Time Averages of Observables 8

Going from state space portrait to time-averages quotients the state space by ergodic partition – we lose the ability to resolve features inside of
ergodic sets. That is OK, because we only need one initial condition to understand any ergodic set since the dynamics restricted to any
ergodic set is a system ergodic with respect to a measure supported on the ergodic set. What do we gain? We gain the ability to compare
dynamics between different parts of the systems in a robust way.

Averaging is a linear, bounded functional on the space of continuous functions that can be represented by delta-like invariant measures,
supported along trajectories. Empirical measures are weak-limits (in infinite time) of such invariant measures. Averaging a basis of continuous
functions produces a weak representation of empirical measures. Ergodic quotient is then the set of weak-representatives of empirical
measures.

The coherent structures that we can identify in state space are directly related to topologies imposed on the ergodic quotient.
What do I mean by that? Example: assume we have sourced a lot of initial conditions on the state space and computed their time averages. A
discrete metric topology is the one where objects are at distance zero only to themselves, and infinity to everybody else. Taking an infinite
intersection of level sets of observables achieves the same thing - any element of the limiting partition contains initial conditions which had the
same averages for all functions. Therefore, the ergodic partition, mentioned on the previous slide, is equivalent to ergodic quotient viewed
through the lens of discrete metric topology.

But we can consider other, continuous, topologies.


Mezić Research Group

Coherent structures: similarity in residence times


[MB, Mezić, 2009] Elliptic Duffing Gyres

State space portrait


Φt (x) ⊂ M

Ergodic quotient
� �
˜ ˜
f1 (x), f2 (x), . . . ⊂ l∞

Negative Sobolev space norms


compare residence times of
trajectories. [Mathew, Mezić, 2010]

Easily computed on the ergodic


quotient – low pass filter!

Intrinsic coordinates recovered by


diffusion modes on data.
Monday, May 23, 2011 Marko Budišić: Time Averages of Observables 9

Endowing the empirical measure space with a negative-index Sobolev space norm topology (proposed by G. Mathew) results defines a
particular type of coherent features. In that topology, trajectories are deemed to be a part of the same coherent structure if their residence
times sets in the state space are similar. Moreover, by selecting a proper index of the Sobolev space norm we can make sure that differences
across larger spatial scales matter more than differences over small scales. These types of norms amount to using Fourier basis as
observables, and comparing the distance between trajectories is comparing the energy of the sequences with appropriate discounting of the
higher wave numbers.

In ergodic quotient, that lives in a high-dimensional space, certain state space features will be mapped to low-dimensional structures,
even if the state space itself is fairly high-dimensiona.
1. The first image depicts an elliptic well that maps to a string in the ergodic quotient. This signifies that ergodic sets in an elliptic zone
(periodic orbits in two-dimensional state space) can be parametrized by a single conserved quantity.
2. The second image shows a Duffing-type of behavior, or a double-well, where the inside “libration” orbits each map to separate strings which
end up joining when we cross the lip of wells into the “revolution” orbits.
3. Conversely, convection cells do not have revolution orbits that would join two strings together. Instead they have hyperbolic fixed points in
their corners that ensure that the time averages stay apart.

These low-dimensional structures inhabit an infinitely-dimensional sequence space. Using a single averaged observable, we are coloring
the state space according to projections of such features to a one-dimensional space. In my own research, I extract the intrinsic
coordinates of such structures and use them to visualize features of the state space. This provides us with low-dimensional embeddings of
ergodic quotient that maximize the detail shown.
*** slide transition ***
If we know which trajectories mapped to certain features in EQ, we can assign colors to different features of EQ, and use these colors to color
the state space portrait in order to visualize coherent structures.
*** slide transition ***
For example, take the Arnold-Beltrami-Childress flow: steady state flow characterized by prominent vortices. We visualized the vortices
extracting features in the ergodic quotient and using them to color trajectories. The major elliptic region was colored by a gradient, while side
vortices were picked out as disconnected components and each given its own color. However, we have done that by studying only results of
simulation, which could be potentially done in a model-free setting, e.g., for experimental data, but that is consistent with “physical” properties,
such as vorticity of the velocity field.
Mezić Research Group

Coherent structures: similarity in residence times


[MB, Mezić, 2009] Elliptic Duffing Gyres

State space portrait


Φt (x) ⊂ M

Ergodic quotient
� �
˜ ˜
f1 (x), f2 (x), . . . ⊂ l∞

Negative Sobolev space norms


compare residence times of
trajectories. [Mathew, Mezić, 2010]

Easily computed on the ergodic


quotient – low pass filter!

Intrinsic coordinates recovered by


diffusion modes on data.
Monday, May 23, 2011 Marko Budišić: Time Averages of Observables 9

Endowing the empirical measure space with a negative-index Sobolev space norm topology (proposed by G. Mathew) results defines a
particular type of coherent features. In that topology, trajectories are deemed to be a part of the same coherent structure if their residence
times sets in the state space are similar. Moreover, by selecting a proper index of the Sobolev space norm we can make sure that differences
across larger spatial scales matter more than differences over small scales. These types of norms amount to using Fourier basis as
observables, and comparing the distance between trajectories is comparing the energy of the sequences with appropriate discounting of the
higher wave numbers.

In ergodic quotient, that lives in a high-dimensional space, certain state space features will be mapped to low-dimensional structures,
even if the state space itself is fairly high-dimensiona.
1. The first image depicts an elliptic well that maps to a string in the ergodic quotient. This signifies that ergodic sets in an elliptic zone
(periodic orbits in two-dimensional state space) can be parametrized by a single conserved quantity.
2. The second image shows a Duffing-type of behavior, or a double-well, where the inside “libration” orbits each map to separate strings which
end up joining when we cross the lip of wells into the “revolution” orbits.
3. Conversely, convection cells do not have revolution orbits that would join two strings together. Instead they have hyperbolic fixed points in
their corners that ensure that the time averages stay apart.

These low-dimensional structures inhabit an infinitely-dimensional sequence space. Using a single averaged observable, we are coloring
the state space according to projections of such features to a one-dimensional space. In my own research, I extract the intrinsic
coordinates of such structures and use them to visualize features of the state space. This provides us with low-dimensional embeddings of
ergodic quotient that maximize the detail shown.
*** slide transition ***
If we know which trajectories mapped to certain features in EQ, we can assign colors to different features of EQ, and use these colors to color
the state space portrait in order to visualize coherent structures.
*** slide transition ***
For example, take the Arnold-Beltrami-Childress flow: steady state flow characterized by prominent vortices. We visualized the vortices
extracting features in the ergodic quotient and using them to color trajectories. The major elliptic region was colored by a gradient, while side
vortices were picked out as disconnected components and each given its own color. However, we have done that by studying only results of
simulation, which could be potentially done in a model-free setting, e.g., for experimental data, but that is consistent with “physical” properties,
such as vorticity of the velocity field.
Mezić Research Group

Coherent structures: similarity in residence times


[MB, Mezić, 2009] Elliptic Duffing Gyres

State space portrait


Φt (x) ⊂ M

Ergodic quotient
� �
˜ ˜
f1 (x), f2 (x), . . . ⊂ l∞

Negative Sobolev space norms


compare residence times of
trajectories. [Mathew, Mezić, 2010]

Easily computed on the ergodic


quotient – low pass filter!

Intrinsic coordinates recovered by


diffusion modes on data. Primary/secondary vortices of an ABC flow
Monday, May 23, 2011 Marko Budišić: Time Averages of Observables 9

Endowing the empirical measure space with a negative-index Sobolev space norm topology (proposed by G. Mathew) results defines a
particular type of coherent features. In that topology, trajectories are deemed to be a part of the same coherent structure if their residence
times sets in the state space are similar. Moreover, by selecting a proper index of the Sobolev space norm we can make sure that differences
across larger spatial scales matter more than differences over small scales. These types of norms amount to using Fourier basis as
observables, and comparing the distance between trajectories is comparing the energy of the sequences with appropriate discounting of the
higher wave numbers.

In ergodic quotient, that lives in a high-dimensional space, certain state space features will be mapped to low-dimensional structures,
even if the state space itself is fairly high-dimensiona.
1. The first image depicts an elliptic well that maps to a string in the ergodic quotient. This signifies that ergodic sets in an elliptic zone
(periodic orbits in two-dimensional state space) can be parametrized by a single conserved quantity.
2. The second image shows a Duffing-type of behavior, or a double-well, where the inside “libration” orbits each map to separate strings which
end up joining when we cross the lip of wells into the “revolution” orbits.
3. Conversely, convection cells do not have revolution orbits that would join two strings together. Instead they have hyperbolic fixed points in
their corners that ensure that the time averages stay apart.

These low-dimensional structures inhabit an infinitely-dimensional sequence space. Using a single averaged observable, we are coloring
the state space according to projections of such features to a one-dimensional space. In my own research, I extract the intrinsic
coordinates of such structures and use them to visualize features of the state space. This provides us with low-dimensional embeddings of
ergodic quotient that maximize the detail shown.
*** slide transition ***
If we know which trajectories mapped to certain features in EQ, we can assign colors to different features of EQ, and use these colors to color
the state space portrait in order to visualize coherent structures.
*** slide transition ***
For example, take the Arnold-Beltrami-Childress flow: steady state flow characterized by prominent vortices. We visualized the vortices
extracting features in the ergodic quotient and using them to color trajectories. The major elliptic region was colored by a gradient, while side
vortices were picked out as disconnected components and each given its own color. However, we have done that by studying only results of
simulation, which could be potentially done in a model-free setting, e.g., for experimental data, but that is consistent with “physical” properties,
such as vorticity of the velocity field.
Mezić Research Group

Continuous Dynamical Indicators

Ergodicity defect Quality of mixing and sampling


[Scott et al., 2010] [Mathew, Mezić, 2010]

Haar wavelets as observables. Fourier basis as observables.


436 G. Mathew, I. Mezić / Physica D 240 (2011) 432–442
436 G. Mathew, I. Mezić / Physica D 240 (2011) 432–442

(a) Time, t =
(a)0.Time,
0. t = 0.0. (b) Time, t = (b)
6.0.Time, t = 6.0. (c) Time, t = 12(c)
.0. Time, t = 1

Replaces “ergodic/non-ergodic”
Applied to design of mixing, optimal
with continuous measure of
sampling of a distribution.
area-ergodicity.

(d) Time, t = 18.0.


Monday, May 23, 2011 Marko Budišić: Time Averages of Observables (d) Time, t = 18.0. 10
Fig. 1. Snapshots at various times of the agent trajectories generated by the SMC algorithm with first-order dynamics. One can observe the multisca
The spacingFig. 1. Snapshots
between at various
the trajectories times of
becomes the agent
smaller trajectories
and smaller generated
as time by the SMC algorithm with first-order dynamics. One can obse
proceeds.
The application of analysis of topology of empirical measures goes beyond visualization of features. Sherry Scott and George Mathew have
The spacing between the trajectories becomes smaller and smaller as time proceeds.
both applied this concept to convert binary descriptions – ergodic/non-ergodic and mixing/not-mixing into continuous
Now, µ is defined as quantifications
generates trajectories such that the fraction o
Now, µ is defined as generates
of ergodicity and mixing. Continuous quantifications can then be used as design the foliage regions
Ter(x) parameters, for example to drive theFig.
system
2 the
shows a plot
tooftrajectories
is close
foliage the
to zero. such that t
ergodicity
regions
decayis of
close to zer
the covera
µ(x) = � .Ter(x) (34)
with respect to a particular measure. Ter( y
U x) = �
µ( )dy . time.
(34)The decayFig.
is 2
notshows a
monotonicplot of
and the
in decay
partic
Ter(y)dy the decay is time. irregular.The This
decay is partly
is not due to the f
monotonic
The snapshots in U Fig. 1 were generated with a simulation of
functions used and the fixed time-stepping
the decay is irregular. This is partl
3 agents andThe with the feedback
snapshots in Fig.law in (26).
1 were The domain
generated withUaissimulation
a of
differential equations. Fig. 2 shows a plot of tim
th
The goal will often determine the appropriate choice of observables. Sherry Scott looked at a multiscale measure of ergodicity by taking functions used and the fixed
unit square domain
3 agents and [0,with
1] ×the[0, feedback
1], the total
lawsimulation
in (26). Thetime is
domain U
spent aby thedifferential
is agents outside the foliage
equations. Fig.regions
2 show
T = 18, unit
and usquare
max = 5
Haarʼs wavelets as observables, which allow precise tuning of scale and position of features we are interested in. .0. The basis
domain × [0, 1]f,k the
[0, 1]functions usedtotal
are as in (9)
simulation time is
this fraction approaches
5 spent by the one,agents as time
outside proceeds
the fo
for K1 , K2T == 018,
, 1, and
. . . , u50. The initial positions of the agents are
max = 5.0. The basis functions fk used are asagents in (9) spend very little time in the foliage
this fraction approaches one, as tr
chosen randomly.
for K1 , K2 = 0, 1, . . . , 50.5 The initial positions of the agents a plot areof theagents norm spend
of the very vectors B (t ). As on
little time in
j
The differential equations for the closed-loop dynamics of the
George Mathew, on the other hand, applied this concept to design of trajectories. In his work, he needed to plan trajectories for searchers,
chosen randomly. �Bj (t )�2 comes close to zero often, it never sta
a plot of the norm of the vector
extended system as described
The differential by (19) for
equations arethe
solved using a dynamics
closed-loop fixed of the
e.g., unmanned vehicles (represented as dots), such that they collaboratively sample a particular known prior distribution on the state space.
time-stepextended
4th order system
Runge–Kutta method. Note thatare
thesolved
Neumann
�Bj (t )�2 comes close to zero ofte
as described by (19) using a3.2. fixedSecond-order dynamics
boundary condition of the basis functions fk guarantees that the
By comparing weak representations of empirical measures of searchers and the weak representation of the target distribution, he
time-step 4th order Runge–Kutta method. Note that the Neumann
velocity component normal to the boundary at the boundaries of 3.2. Second-order dynamics
boundary condition of the basis functions fk guarantees that The the dynamics of the extended system tha
designed optimal controllers that drove searchers to quickly sample the desired distributions. The same idea was applied to design of
the domain is always zero and therefore the agents never escape
velocity component normal to the boundary at the boundaries tionsof and velocities of the agents and the varia
the domain. The dynamics of the extende
mixing trajectories. the domain is always zero and therefore the agents never escape is described as:
From the snapshots in Fig. 1, one can see the multiscale nature tions and velocities of the agents
the domain. ẋ (τ ) = v (τ is
) described as:
of the algorithm. The spacing between the trajectories becomes j j
From the snapshots in Fig. 1, one can
smaller and smaller as time proceeds. One should note that the see the multiscale nature
v̇ (τ ) = u (τ )
j j ẋ (τ ) = v (τ )
of the
trajectories are algorithm.
not such that Thethey
spacing between
completely the going
avoid trajectories
over becomes
j j

smaller Ṡ (τ ) = W (τ ) = u
regionsand smaller as avoidance.
time proceeds. One
theshould note that the
k k v̇ j (τ ) j (τ )
the foliage as in collision Rather, algorithm
trajectories are not such that they completely avoid going over � N
Ṡk (τ ) = Wk (τ )
the foliage regions as in collision avoidance. Rather, the algorithm Ẇk (τ ) = ∇ fk (xj (τ )) · vj (τ ).
j=1 � N
5 For numerical purposes, we need to have a cutoff for the number of Fourier Ẇk (τ ) = ∇ f (xj (τ )) · vj (τ ).
The forces on the agents areksubject to the co
efficients used. The exact effect of this cutoff on the performance of the algorithm j=1
5 For
is a challenging andnumerical
interestingpurposes,
problem. Roughly
F max . We are going to use the same model p
we needspeaking,
to have aforcutoff
a lower
forcutoff, the
the number of Fourier
gaps left between the trajectories of the agents will be bigger than those obtained Thefirst-order
proach as with forces on dynamics.
the agentsThe are cost
sub
efficients used. The exact effect of this cutoff on the performance of the algorithm
with a higher cutoff. Fmax .isWe
ing to use here
is a challenging and interesting problem. Roughly speaking, for a lower cutoff, the
are goingsum
a weighted to use thefirs
of the sa
gaps left between the trajectories of the agents will be bigger than those obtained proach as with first-order dynam
with a higher cutoff. ing to use here is a weighted su
Mezić Research Group

� f˜
f→ Focus on dynamics of T : M → M, T ≡ Φτ
“on the attractor”, systems with recurrent dynamics.

linear, unitary on attractor


Koopman Operator Uf = f ◦ T captures full nonlinear dynamics
[Mezić, Banaszuk, 2004.] n

U= ei2πωk PTωk + Uc
8 C.W. Rowley, I. Mezić, S. Bagheri, P. Schlatter, and D.S. Henningson
k=1
400
1
(a) (b)
300
0.5 Time-averaging projects
�vj � U f˜ = f˜ onto eigenspace at 1
Im{λj } 200
0 (invariant functions).

100 Complex eigenspaces


U f˜ω = ei2πω f˜ω
!0.5
reveal periodicity of
0 dynamics.
!1 0 0.03 0.06 0.09 0.12 0.15 0.18 0.21 0.24
!1 !0.5 0 0.5 1
Re{λj } St

Figure 2. (a) The empirical Ritz values λj . The value corresponding to the first Koopman mode
is shown with the blue symbol. (b) The magnitudes of the Koopman modes (expect the first
Koopman one) at each spectrum ofInaboth
frequency. jet in the crossflow.
figures, the colors vary [Rowley
smoothlyetfrom al., 2009]
red to white, depending
on the magnitude of the corresponding mode.
Monday, May 23, 2011 Marko Budišić: Time Averages of Observables 11

So far I have talked about the relationship between state space features
1 and the time-averaged observables. What is the relationship between
spectrum shows the frequency content û 1
(ω) of u (t). The peak frequency corresponds
observable and its time-average? At every step of the averaging process, we are composing an observable with the evolution of the
system.to a vortex
Operator thatshedding
corresponds oftowake vorticesiswith
that operation the Strouhal
the Koopman number
operator. St ≡and
It is a linear, f D/V jet =
it is dual to 0.0174.
the Perron-Frobenius operator
that evolvesIndensities
figure of1(d,f), a secondOn
initial conditions. probe located
the attractor, a few operator
Koopman jet diameters
is unitaryalong the jet
with respect trajectory
to the L2 function space over any
= (12,
invariantxmeasure
2
P
on 6,
the2), shows a second oscillation that can be identified with the shedding of
attractor.
the shear-layer vortices. The peak frequency beats with St = 0.141 which is nearly one
Let me stress that the operator itself is LINEAR in observables but that it captures full dynamics of a NONLINEAR dynamical system.
order
The unitarity of of
themagniture larger
operator enables than
us to studythe low-frequency
its spectrum mode.
in an effort Note that
to understand the peak
its behavior. Thefrequencies
image shows a numerically computed
spectrumofofthe power spectra
a Koopman vary
operator for slightly
a fluid flow. depending on the location of the probe.

Time-averaging projects the observable4.1.


onto Koopman modes
the eigenspace and
at 1 for thefrequencies
Koopan operator, which contains all invariant functions for dynamics.
The complex eigenspaces reveal periodicity in dynamics of the system corresponding to the frequency of the eigenvalue associated with the
In this section we compute the Koopman modes and show that they directly allow an
eigenspace. We can access the projection to periodic eigenspaces by harmonic averages, which are introduced on the next slide.
identification of the various shedding frequencies. The empirical Ritz values λj and the
Koopmanempirical
operator isvectors vj of a sequence
an instantaneous object; its of flow-fields
definition 0 , u1a, finite
{uonly
contains ...,u time
m−1 } = {u(t =of200),
advancement u(t =
dynamics. The infinite-time
202),
averaging is just
. . . ,au(t
TOOL= through
700)} with
whichmwe= 251 are
analyze the computed using
eigenspace at 1 – thethe algorithm
space described
of invariant functions.earlier.
If we want to access other
eigenspaces, we can modify simple time averages to form harmonic averages.
Thus, the transient time (t < 200) is not sampled and only the asymptotic motion in
phase space is considered.
Figure 2(a) shows that nearly all the Ritz values are on the unit circle |λj | = 1
indicating that the sample points ui lie on or near an attracting set. The Koopman
eigenvalue corresponding to the first Koopman mode is the time-averaged flow and is
depicted with blue symbol in figure 2(a). This mode, shown in figure 1(b), captures the
steady flow structures as discussed previously. In figure 2(a), the other (unsteady) Ritz
values vary smoothly in color from red to white, depending on the magnitude of the
corresponding Koopman mode. The magnitudes defined by the global energy norm �vj �,
and are shown in figure 2(b) with the same coloring as the spectrum. In figure 2(b) each
mode is displayed with a vertical line scaled with its magnitude at its corresponding
frequency ωj = Im{log(λj )}/∆t (with ∆t = 2 in our case). Only the ωj ≥ 0 are shown,
since the eigenvalues come in complex conjugate pairs. Ordering the modes with respect
to their magnitude, the first (2-3) and second (4-5) pair of modes oscillate with St2 =
0.141 and St4 = 0.136 respectively, whereas the third pair of modes (6-7) oscillate with
St6 = 0.017. All linear combinations of the frequencies excite higher modes, for instance,
the nonlinear interaction of the first and third pair results in the fourth pair, i.e. St8 =
0.157 and so on.
In figures 1(e) and (f) the power spectra of the two DNS time signals (black lines)
Mezić Research Group

Harmonic averages
N
� −1
˜ 1
fω (x) = lim e−i2πωn f [T n (x)]
N →∞ N
n=0

U f˜ω = ei2πω f˜ω


[Mezić, Banaszuk, 2004.]

f˜ω1 g̃ω1 h̃ω1 ...


Harmonic
f˜ω2 g̃ω2 h̃ω2 ...
Quotient f˜ω3 g̃ω3 h̃ω3 ...
at ω
3 .. .. .. ..
. . . .
Discrete
Fourier
Transform

R. Mohr, I Mezić: Applications to analysis of computer networks (in progress)


Monday, May 23, 2011 Marko Budišić: Time Averages of Observables 12

Harmonic averages are formed from time averages by adding a complex term that rotates with frequency of the eigenvalue we are interested
in. Harmonic averages span the complex eigenspaces in the same fashion that time averaged spanning. Harmonic averaging projects an
observable to the complex eigenspace.

For any initial condition studies, there are two different parameters to choose: an observable and the frequency of averaging. While this might
seem daunting, fixing any of those parameters results in familiar objects:
1. Fixing an observable, and computing all frequencies, results in a DFT of an observable.
2. Fixing a frequency, and computing averages of all continuous observables results in harmonic quotients: analogues to ergodic quotients
discussed earlier, except oscillating at a certain frequency.
Ryan Mohr, also at UC Santa Barbara, is working on applications of these methods to analysis of computer TCP networks.
*** slide transition ***
To demonstrate what harmonic averages look like, Iʼll turn again to the standard map. Looking at harmonic averages at period 5, we see that
averages over features that arenʼt periodic with period 5 are equal to zero. The level sets at non-zero values correspond to periodic sets which
return to themseves after 5 time steps. As harmonic averages are complex functions, we can use both real and imaginary part to distinguish
the phase within these sets.
Mezić Research Group

Harmonic averages Standard map ε = 0.12 f (x, y) = sin(8πx + 2πy)

N
� −1
˜ 1
fω (x) = lim e−i2πωn f [T n (x)]
N →∞ N
n=0

U f˜ω = ei2πω f˜ω


[Mezić, Banaszuk, 2004.]

�f˜ω (x, y) �f˜ω (x, y), ω = 1/5


f˜ω1 g̃ω1 h̃ω1 ...
Harmonic
f˜ω2 g̃ω2 h̃ω2 ...
Quotient f˜ω3 g̃ω3 h̃ω3 ...
at ω
3 .. .. .. ..
. . . .
Discrete
FIG. 1. !Color online" Single-function plots of time averages for the standard map equation !7" !left" and the corresponding phase space portraits !right" done
Fourier
as 100 iterations of 11! 11 random trajectories picked from #0 , 1$2 for the function f = cos!2"y". Top row: # = 0, middle: # = 0.09, bottom: # = 0.18. The grid
of 800! 800 initial points was used, and the dynamics was run for tfinal = 30 000 iterations.
Transform
curve, a different coloration occurs there. This type of cha- frequency. In the context of standard map it is convenient to
otic transport in Hamiltonian maps can be approximated by use a slightly bigger frequency for y coordinate—since the
the Markov tree model.27–29 transport in x direction is much faster, the function averages
R. Mohr, I Mezić: Applications to analysis of computer networks (in progress)
In order to investigate the phase space structure at larger
!smaller" scale, one takes smaller !bigger" Fourier/wavelet
out to zero rather quickly in this direction. Time-average
plots for several Fourier basis functions with increasing fre-
Monday, May 23, 2011 Marko Budišić: Time Averages of Observables 12

Harmonic averages are formed from time averages by adding a complex term that rotates with frequency of the eigenvalue we are interested
in. Harmonic averages span the complex eigenspaces in the same fashion that time averaged spanning. Harmonic averaging projects an
observable to the complex eigenspace.

For any initial condition studies, there are two different parameters to choose: an observable and the frequency of averaging. While this might
seem daunting, fixing any of those parameters results in familiar objects:
1. Fixing an observable, and computing all frequencies, results in a DFT of an observable.
2. Fixing a frequency, and computing averages of all continuous observables results in harmonic quotients: analogues to ergodic quotients
discussed earlier, except oscillating at a certain frequency.
Ryan Mohr, also at UC Santa Barbara, is working on applications of these methods to analysis of computer TCP networks.
*** slide transition ***
To demonstrate what harmonic averages look like, Iʼll turn again to the standard map. Looking at harmonic averages at period 5, we see that
averages over features that arenʼt periodic with period 5 are equal to zero. The level sets at non-zero values correspond to periodic sets which
return to themseves after 5 time steps. As harmonic averages are complex functions, we can use both real and imaginary part to distinguish
the phase within these sets.
Mezić Research Group

Finite-time averages

Koopman Operator is an instantaneous object. Do we really need the


infinite-time formalism to analyze it?

• finite-time Lyapunov exponents [Haller, 2010] and more...


• finite-time turbulence quantifiers [Foias, Jolly, 2005]
• time-averaged velocity [Poje, Haller, Mezić, 1999], [Mezić et al. 2010]

� T � τ
1 1
f [Φt (x)]dt = lim f [Φ(t mod T ) (x)]dt
T 0 τ →∞ τ 0

Finite time averaging is equivalent to periodizing the dynamics.


Monday, May 23, 2011 Marko Budišić: Time Averages of Observables 13

As mentioned, Koopman operator is instantaneous object, in case of iterated maps, it composes observable with a single step dynamics.
This gives hope that analyzing it through infinite time averages is not necessary.

There have been several research efforts to convert infinite time indicators to finite time.

As a glimpse of one path to understand finite time averages, this identity shows that finite-time averages are equivalent to infinite time
averages for modified flow map. The modification consists of periodizing dynamics of the original time map with the period equal to the time
over which finite-time average was computed.

While this might seem artificial and useless at first, this approach is readily used in signal processing, when Discrete Fourier Transform is
used to analyze and filter measured signals that are not necessarily periodic. However Koopman operator frames this approach within the
context of full nonlinear dynamics.
Mezić Research Group

Conclusions
Numerical methods should be designed with time-averaging in mind.
Trajectory-based consistency does not always imply consistency of
time-averages for Hamiltonian and infinite-dimensional systems.

Wealth of information is waiting to be managed.


Fully nonlinear data can be analyzed by linear tools.

Choosing observables: can we do better than the entire function


basis? Compressed sensing? Empirical measure reconstruction?

Finite-time averages are here to stay.


Infinite averaging is just a tool to access the instantaneous
Koopman op. What can finite-time averages tell us about it?
Physical data, nonrecurrent systems require finite-time averages.
Monday, May 23, 2011 Marko Budišić: Time Averages of Observables 14

These are the main take-aways from this talk:

1. Numerical integrators that were built around trajectories of ODEs might not be adequate for computation of statistics. On the other hand,
since time averages show increased robustness to noise and numerical errors, as compared to trajectories, we might be able to build
averaging-integrators that would be more efficient at computing time averages of trajectories, while ensuring the statistical consistency.

2. Moving the full nonlinear data to the linear world results in a wealth of information stored in averages of observables. At this point, unless
we are physically motivated to choose a particular observable, we store the information in a large number of observables. A more selective
approach to choosing observable could enable us to reduce computational requirements of the method. It might be possible to use ideas
from compressed sensing to do this. A step further would involve full empirical measure reconstruction, effectively viewing the ergodic quotient
as input data to the inverse problem. Due to singular nature of a lot of empirical measures, with respect to Lebesgue, makes it difficult to use
current inverse problem solutions without encountering large reconstruction artifacts.

3. A requirement for viable methods is the ability to handle experimental data, and truly aperiodic data. It seems that to do this, we have
to abandon the infinity and embrace finite-time averages. Understanding how finite-time features relate to properties of the system is an
important step in acceptance of time-averaging as a standard tool in analysis and design.
Mezić Research Group

Main references
• Budišić and Mezić. An approximate parametrization of the ergodic partition using time averaged
observables. Proceedings of the 48th IEEE Conference on Decision and Control, Shanghai, China
(2010) pp. 3162-3168
• Cvitanović et al. Chaos: Classical and Quantum, ChaosBook.org, Niels Bohr Institute, Copenhagen,
2011
• Levnajić and Mezić. Ergodic theory and visualization. I. Mesochronic plots for visualization of ergodic
partition and invariant sets. Chaos (2010) vol. 20 (3) pp. 19
• Mathew and Mezić. Metrics for ergodicity and design of ergodic dynamics for multi-agent systems.
Physica D (2010) vol. 240 pp. 432-442
• Mattingly et al. Convergence of numerical time-averaging and stationary measures via Poisson
equations. SIAM J. Numer. Anal. (2010) vol. 48 (2) pp. 552-577
• Mezić and Banaszuk. Comparison of systems with complex behavior. Physica D (2004) vol. 197 (1-2)
pp. 101-133
• Rowley et al. Spectral analysis of nonlinear flows. J. Fluid Mech. (2009)
• Scott et al. Capturing deviation from ergodicity at different scales. Physica D (2009) vol. 238 (16) pp.
1668-1679
• Sigurgeirsson and Stuart. Statistics From Computations. Foundations of Computational Mathematics,
LMS Lecture Note Series, Cambridge University Press (2001) vol. 284 pp. 323-344
• Tupper. Computing statistics for Hamiltonian systems: A case study. Journal of Computational and
Applied Mathematics (2007)
• Wang. Approximation of stationary statistical properties of dissipative dynamical systems: time
discretization. Math. Comp (2010)

Monday, May 23, 2011 Marko Budišić: Time Averages of Observables 15

For additional information, please e-mail: mbudisic@engr.ucsb.edu

Vous aimerez peut-être aussi