anomalous diffusion in a finite world: time-scale dependent trajectory analysis
TRANSCRIPT
Anomalous diffusion in a finite world: time-scale
dependent trajectory analysis
Konrad Hinsen
Centre de Biophysique Moléculaire, Orléans, Franceand
Synchrotron SOLEIL, Saint Aubin, France
13 December 2016
Konrad Hinsen (CBM/SOLEIL) Anomalous diffusion in a finite world: time-scale dependent trajectory analysis13 December 2016 1 / 21
Anomalous diffusion
Defined by the asymptotic MSD for t →∞:
W(t) = 2Dαtα, 0 < α < 2
Study of different models that have this asymptotic behavior.
Asymptotic diffusion not accessible in experiment or simulation.
Konrad Hinsen (CBM/SOLEIL) Anomalous diffusion in a finite world: time-scale dependent trajectory analysis13 December 2016 2 / 21
Diffusion in homogeneous molecular systems
Continuous process X(t)
W(t)→ kBTm t2 for t → 0 (ballistic motion)
No trapping, no rare events
Idea:
analyze finite-length trajectories at different time scales
use fractional Brownian Motion (fBM) as an analysis tool
use Bayesian inference to get error bars on the fBM parameters
K. Hinsen & G. Kneller, J. Chem. Phys. 145, 151101 (2016)
Konrad Hinsen (CBM/SOLEIL) Anomalous diffusion in a finite world: time-scale dependent trajectory analysis13 December 2016 3 / 21
Trajectory sampling at different time scales
1 Choose the number of samples L
2 Pick a lag time value ∆t
3 Analyze the discrete and finite series
X(∆t)i = X(i∆t), i = 1, . . . , L
4 Vary ∆t keeping L fixed
Konrad Hinsen (CBM/SOLEIL) Anomalous diffusion in a finite world: time-scale dependent trajectory analysis13 December 2016 4 / 21
Fractional Brownian Motion
Probability density for a Gaussian L-step process X = (X0, . . . ,XL):
P(X) =1√
(2π)L|Σ|exp
(−1
2X ·Σ−1 · X
)
Fractional Brownian Motion:
Σi,j = Dα∆tα (iα + jα + |i− j|α)
Parameters: α,Dα
Konrad Hinsen (CBM/SOLEIL) Anomalous diffusion in a finite world: time-scale dependent trajectory analysis13 December 2016 5 / 21
Bayesian inference
Goal: infer parameter distribution P(α,Dα) from observed L-stepprocesses X.Interpretation: P(α,Dα) represents the knowledge we have aboutthe parameters after exploiting the observations.
1 Assume a prior distribution P0(α,Dα)
2 Add information from independent observations X(j), j = 1, . . . ,N:
Pj(α,Dα) =P(X(j)|α,Dα)∫
dα∫dDαP(X(j)|α,Dα)
Pj−1(α,Dα),
3 Interpret the posterior distribution PN(α,Dα)
Konrad Hinsen (CBM/SOLEIL) Anomalous diffusion in a finite world: time-scale dependent trajectory analysis13 December 2016 6 / 21
Bayesian inference for fBM
Normalization doesn’t matter for us, so let’s be sloppy!
P0(α,Dα) =
{1 if 0 < α < 2 and 0 < Dα∆tα < Dmax
0 otherwise
Dmax: upper limit introduced for normalizability
Pj(α,Dα) =1√
|Σ(α,Dα)|exp
(−1
2X(j) ·Σ−1(α,Dα) · X(j)
)Pj−1(α,Dα)
For numerical computations, use log P instead of P.
Konrad Hinsen (CBM/SOLEIL) Anomalous diffusion in a finite world: time-scale dependent trajectory analysis13 December 2016 7 / 21
Pros and cons
Direct analysis of trajectories - no intermediate quantities
Estimation of the parameters and their uncertainties
We can know if our data is good enough to support conclusions
No check for model quality - if fBM is a bad match for the data,we can still get nice parameter distributions
Limited to L ≤ 500 with double-precision floats
Konrad Hinsen (CBM/SOLEIL) Anomalous diffusion in a finite world: time-scale dependent trajectory analysis13 December 2016 8 / 21
Parameter distributions
Probability distributions of two variables are difficult to analyze
What we care most about is α
Dα turns out to be sharply localized around
(2Dα∆tα)est =⟨
(X(k)j+1 − X
(k)j )2
⟩j,k
(1)
I will show mostly 1D distributions P(α).
Konrad Hinsen (CBM/SOLEIL) Anomalous diffusion in a finite world: time-scale dependent trajectory analysis13 December 2016 9 / 21
Analysis of fBM trajectories
α = 0.6, Dα = 1, L = 100, N = 500
Konrad Hinsen (CBM/SOLEIL) Anomalous diffusion in a finite world: time-scale dependent trajectory analysis13 December 2016 10 / 21
Analysis of perfect fBM trajectories, α = 0.6, L = 10
0 10 20 30 40 50Number of trajectories
0.0
0.5
1.0
1.5
2.0
α
0 200 400 600 800 1000Number of trajectories
0.0
0.5
1.0
1.5
2.0
α
fBM trajectories, L= 10
Konrad Hinsen (CBM/SOLEIL) Anomalous diffusion in a finite world: time-scale dependent trajectory analysis13 December 2016 11 / 21
Analysis of perfect fBM trajectories, α = 0.6, L = 100
0 10 20 30 40 50Number of trajectories
0.0
0.2
0.4
0.6
0.8
1.0
α
0 200 400 600 800 1000Number of trajectories
0.0
0.2
0.4
0.6
0.8
1.0
α
fBM trajectories, L= 100
With a few hundred trajectories, a maximum-entropy estimate is OK.
Konrad Hinsen (CBM/SOLEIL) Anomalous diffusion in a finite world: time-scale dependent trajectory analysis13 December 2016 12 / 21
Imperfect fBM trajectories
ck =< ∆X(t)∆X(t + kδt)) > /(Dαδtα)
0 2 4 6 8 10
k
0.4
0.2
0.0
0.2
0.4
0.6
0.8
1.0
1.2
c k
fBM α= 0. 6
fBM∗ α= 0. 6
cfBM0 = 1
cfBMk = (k + 1)α − 2kα + (k − 1)α for k > 0
Konrad Hinsen (CBM/SOLEIL) Anomalous diffusion in a finite world: time-scale dependent trajectory analysis13 December 2016 13 / 21
Imperfect fBM trajectories: MSD
100 101 102 103
τ/δt
100
101
102
<W
(τ)>/(D·δtα
)fBM α= 0. 6
fBM∗ α= 0. 6
Konrad Hinsen (CBM/SOLEIL) Anomalous diffusion in a finite world: time-scale dependent trajectory analysis13 December 2016 14 / 21
Analysis of imperfect fBM trajectories, α = 0.6
0 10 20 30 40 50Number of trajectories
0.0
0.2
0.4
0.6
0.8
1.0
α
0 200 400 600 800 1000Number of trajectories
0.0
0.2
0.4
0.6
0.8
1.0
α
fBM with modified short-time behavior, L= 100, s= 10
Good convergence... to a wrong α.
Konrad Hinsen (CBM/SOLEIL) Anomalous diffusion in a finite world: time-scale dependent trajectory analysis13 December 2016 15 / 21
Multiscale Bayesian inference as an analysis tool
Bayesian inference cannot say if fBM is a good model...
... but it always yields a number α ...
... that says something about the data.
Multiscale analysis: look at α as a function of time scale
For a scale-free model (fBM): no time scale dependence.
Konrad Hinsen (CBM/SOLEIL) Anomalous diffusion in a finite world: time-scale dependent trajectory analysis13 December 2016 16 / 21
Multiscale analysis of fBM∗, α = 0.6, L = 100
0 20 40 60 80 100
∆t/δt
0.56
0.58
0.60
0.62
0.64
0.66αML
Konrad Hinsen (CBM/SOLEIL) Anomalous diffusion in a finite world: time-scale dependent trajectory analysis13 December 2016 17 / 21
Lipid bilayer simulation
Coarse-grained MARTINI model
2033 POPC molecules, 57952 water beads
600 ns production run
NVT ensemble, T=320 K
Simulations run by Sławomir Stachura for his thesis
Lipid center-of-mass trajectories available from Zenodo:short-time, long-time
Preprocessing:
Use only in-plane coordinates x and y
Shift initial position to 0
Konrad Hinsen (CBM/SOLEIL) Anomalous diffusion in a finite world: time-scale dependent trajectory analysis13 December 2016 18 / 21
Convergence of Bayesian inference, L = 100
0 10 20 30 40 50Number of trajectories
0.0
0.2
0.4
0.6
0.8
1.0
α
0 500 1000 1500 2000 2500 3000 3500 4000 4500Number of trajectories
0.0
0.2
0.4
0.6
0.8
1.0
α
lipid trajectories, ∆t= 18. 00 ps
Konrad Hinsen (CBM/SOLEIL) Anomalous diffusion in a finite world: time-scale dependent trajectory analysis13 December 2016 19 / 21
Multiscale analysis, L = 100
0.4
0.6
0.8
1.0
1.2
1.4
1.6
1.8
2.0
αML
α= 0. 55
ballisticMSDWDFTMEE
10-2 10-1 100 101 102 103 104
∆t [ps]
10-7
10-6
10-5
10-4
10-3
10-2
10-1
100
2Dα∆tα
[nm
2]
τ= 0. 59 ps
α= 0. 55
ballisticMSDWDFTMEE
fBM behavior for 10ps < ∆t < 10ns
Ballistic behavior for ∆t < 0.1ps
Konrad Hinsen (CBM/SOLEIL) Anomalous diffusion in a finite world: time-scale dependent trajectory analysis13 December 2016 20 / 21