Diffusion Maps as Invariant Functions of Dynamical Systems

Embed Size (px)

Citation preview

  • 7/28/2019 Diffusion Maps as Invariant Functions of Dynamical Systems

    1/11

    Mezi Research Group

    SIAM Conference on Applications of Dynamical Systems

    Snowbird, UT, May 2013

    Marko Budii

    Igor Mezi

    DIFFUSION MAPS AS

    INVARIANT FUNCTIONS OF

    DYNAMICAL SYSTEMS

    Funding:

    ONR MURI

    Ocean 3D+1

  • 7/28/2019 Diffusion Maps as Invariant Functions of Dynamical Systems

    2/11

    Marko Budii: Diffusion Maps as Invariant Functions of Dynamical SystemsMay 23, 2013 2

    Invariant functions define barriers to dynamical transport

    and simplify dynamics.

    Classical mechanics:

    Integrals of motion foliate the state space.

    We can reduce system to each of the leaves of the

    foliation and reduce order of the system.

    Fluid dynamics:

    Boundaries of (sub-)level sets of

    invariant functions arebarriers to material transport.

    Invariant functions: functions on the state space whose values are constant along trajectories.

  • 7/28/2019 Diffusion Maps as Invariant Functions of Dynamical Systems

    3/11

    xp = u(xp), xp(0) = p

    (p, t) 7! xp(t)

    f(p) := limT!1

    1

    T

    ZT0

    f(xp())d

    : !

    p! g(xp(t)) p! g p

    p! f(xp(t)) p! f(p)

    f(xp(t)) f(p)

    Marko Budii: Diffusion Maps as Invariant Functions of Dynamical SystemsMay 23, 2013

    Every invariant function is an

    average of a scalar function along trajectories.

    3

    xp(t)p

    PTrajectory:

    Time-average:

    Vector field Initial state

    Time averages are invariant functions:

    Different initial functions may not yield independent

    time-averages.

    Level sets of time-averages are invariant sets.

  • 7/28/2019 Diffusion Maps as Invariant Functions of Dynamical Systems

    4/11

    R

    Marko Budii: Diffusion Maps as Invariant Functions of Dynamical SystemsMay 23, 2013

    Ergodic Quotient is a static representation of

    dynamical invariants in the space of averaged functions.

    4

    p 7!

    264f1(p)

    f2(p)...

    375

    Functions selected from a basis

    xp = u(t, xp)

    Entire trajectories are represented

    by single points.

    Axes are averagedfunctions.

    f2(p)

    f1(p)

    f(p) := limT!1

    1

    T

    ZT0

    f(xp())d

    Ergodic Quotient (EQ)

    maps initial conditions to

    infinite vectors

    of time averages.

    Ergodic Quotient

    Any function with the domain in EQ

    is again an invariant function.

  • 7/28/2019 Diffusion Maps as Invariant Functions of Dynamical Systems

    5/11

    Marko Budii: Diffusion Maps as Invariant Functions of Dynamical SystemsMay 23, 2013

    Sobolev space topology on Ergodic Quotient

    captures families of orbits as continuous segments.

    5

    State space

    portrait

    Reeb Graphs

    (ideal EQ structure)

    Fixed point onseparatrices preventsegments fromconnecting.

    d(p1, p2)2 =

    X

    k2Zd

    fk(p1) fk(p2)2

    (1 + |k|2)s Acts as a low-pass filter:

    de-emphasizes

    small scale differences.

    Wavevectors

    Ergodic Quotientf2(p)

    f1(p)

    Desired topology:

    Time-averages of Fourier harmonics are

    spatial Fourier coefficients of

    averaging distributions, supported on orbits.

    P

    fk(x) = ex

    Required choice of observables:

    Sobolev space norm does the trick:

  • 7/28/2019 Diffusion Maps as Invariant Functions of Dynamical Systems

    6/11

    Marko Budii: Diffusion Maps as Invariant Functions of Dynamical SystemsMay 23, 2013

    Diffusion coordinates reduce ambient dimension,

    while preserving intrinsic geometry of the EQ.

    6

    Ergodic Quotient

    f2(p)

    f1(p)

    The ambient dimension is the number of

    functions averaged (axes).

    It is ideally large, to achieve

    a high resolution.

    It is known a priori.

    The intrinsic dimension depends on the

    complexity of dynamics, but can be

    very small, even zero (ergodicity).

    It is unknown a priori.

    To disentangle the wire (EQ),sort the points by time it takes them to heat up.

    Heat sources are placed onan entangled wire (EQ).Diffusion Maps:

    [Coifman, Lafon,

    ACHA, 2006]

    Diffusion coordinates

    pulled back onto the state space

    are invariant functions,

    sensitive to similarity in dynamics.

    p

    !

    264

    f1(p)

    f2(p)...

    375!

    264

    1((p))2((p))

    ..

    .

    375

    Averaging DiffusionMaps

    Practical matters:Bubacarr Bah,MSc Thesis,Oxford, 2008.

    http://eprints.maths.ox.ac.uk/740/http://eprints.maths.ox.ac.uk/740/http://eprints.maths.ox.ac.uk/740/http://eprints.maths.ox.ac.uk/740/http://eprints.maths.ox.ac.uk/740/http://eprints.maths.ox.ac.uk/740/
  • 7/28/2019 Diffusion Maps as Invariant Functions of Dynamical Systems

    7/11

    1

    Marko Budii: Diffusion Maps as Invariant Functions of Dynamical SystemsMay 23, 2013

    Coloring the state space by values of dominant

    diffusion coordinates reveals large scale features.

    7

    21 3

    Diffusion pseudocolors identify invariant sets

    Different colors indicate there is no

    material transport between regions.

    Coordinates of higher order

    distinguish between finer features.

    State space portraitEQ in Diffusion Coordinates Coloring of the state space

    1

    2

    Averaging

    +Diffusion Maps

  • 7/28/2019 Diffusion Maps as Invariant Functions of Dynamical Systems

    8/11

    R

    z

    0

    2

    0.2

    0.3

    0.1

    0.1

    Cluster 3

    Cluster 2

    Cluster 1

    Cluster 0

    2

    1

    1 0.5 0 0.5 11

    0.8

    0.6

    0.4

    0.2

    0

    0.2

    0.4

    0.

    6

    0.8

    1

    Marko Budii: Diffusion Maps as Invariant Functions of Dynamical SystemsMay 23, 2013

    Clustering in diffusion coord. identifies large-scale invariant sets.

    8

    0.00

    0.50

    1.00

    0

    2

    4

    Cluster

    x

    z

    y

    3

    1

    0.00

    0.50

    1.00 0.00

    0.50

    1.00

    6

    5

    Index

    6

    Index 4 Index 20.15

    0.1

    0.05

    00.

    05

    0.1

    0.08

    0.04

    0

    0.

    04

    0.080.03

    0.01

    0.01

    0.03

    0.05

    Vortices identified as clusters ofpoints in diff. coord. space.

    z

    R

    Secondary vortex appears asa bifurcation in the primary segment

    Steady ABC flow:

    Time-periodic Hills vortex ring:

    Time-periodic perturbation splits

    the primary vertex core.

    State space

    State space

    Ergodic Quotient

    Ergodic Quotient

  • 7/28/2019 Diffusion Maps as Invariant Functions of Dynamical Systems

    9/11Marko Budii: Diffusion Maps as Invariant Functions of Dynamical SystemsMay 23, 2013

    Future: method resolves structures in finite-time, unsteady flows.

    9by Drew Poje (CUNY)

    HYCOM simulation of 2D Gulf Stream (along surface of equal potential density of water)

    Averaging Time = 20 days

    Averaging Time = 10 days

    1 2 3

    1 2 3

    12

    3

    1Color:

    12

    3

    1Color:

  • 7/28/2019 Diffusion Maps as Invariant Functions of Dynamical Systems

    10/11Marko Budii: Diffusion Maps as Invariant Functions of Dynamical SystemsMay 23, 2013

    To achieve automatic analysis,

    there are several issues that need resolving:

    10

    Tuning the diffusion bandwidth.

    Incorrect bandwidth can artificially

    connect or disconnect regions.

    Tuning is heuristic: Lafon (min. distance),

    Lee (neighborhood size), Singer (linear sensitivity).

    Detection of intrinsic dimensionand independent coordinates.

    Cluster 3

    Cluster 2

    Cluster 1

    Cluster 0

    2

    1

    1 0.5 0 0.5 11

    0.8

    0.6

    0.4

    0.

    2

    0

    0.2

    0.4

    0.6

    0.8

    1

    Due to highly

    non-convex structures,

    K-means clustering

    is not a good choice.

    Alternative segmenting

    would be preferable.

    Diffusion eigenvalues depend on the tuning

    Where is the

    spectral gap?

    k = 5

    k = 4

    k = 3

    k = 2

    0.3 0.35 0.4 0.45 0.51

    0.8

    0.6

    0.4

    0.2

    0

    0.2

    0.4

    0.6

    0.8

    1

    k

    Dependence of diffusion coordinates

    Post-processing:

    K-means, proximity graphs, etc.

    What is

    relevant for

    dynamical

    systems?

    Ergodic quotient for Hills ring vortex

  • 7/28/2019 Diffusion Maps as Invariant Functions of Dynamical Systems

    11/11Marko Budii: Diffusion Maps as Invariant Functions of Dynamical SystemsMay 23, 2013 11

    Take away points:

    Budii and Mezi, Geometry of the ergodic quotient reveals coherent structures in flows,Physica D, 241, (2012).

    Budii,Ergodic Quotients in Analysis of Dynamical Systems, PhD Thesis,UC Santa Barbara, (2012).

    Budii, Mohr, and Mezi,Applied Koopmanism, Chaos 22, (2012).

    Ergodic Quotient (EQ) is a geometric representation

    of the space of invariant functions.

    Time-averaged functions provide computable axes for EQ.

    Diffusion Maps provide efficient axes for EQ,

    while preserving intrinsic geometry.

    EQ was successfully used to extract invariant sets

    in the state space.

    Even in finite-time, method recovers significant structures.