Kirill Neklyudov
Langevin Dynamics
for sampling and
global optimization
Goals of this talk
Langevin Equation
Ito Stochastic Differential Equation (SDE):
Force
Random fluctuations
Discrete approximation:
1-d simulation
Langevin equation
Fokker-Planck equation
Derivation of the Fokker-Planck equation
Langevin equation:
Increments of the Brownian motion:
Consider a small increment of :
Derivation of the Fokker-Planck equation
Using the change of variables formula, we obtain:
Density of particle distribution:
Unknown density!
Changing variables
The change of variables:
We can’t invert this
Changing variables
Changing variables
From the previous slide:
Equation for the increment
(Langevin equation)
Note that
since
Changing variables
Finally!
With a little more efforts (homework):
Reminder:
Derivation of the Fokker-Planck equation
Change of variables
Increment of the density
???
Derivation of the Fokker-Planck equation
0th order:
1st order:
2nd order:
Taking the expectation
0th order:
1st order:
Taking the expectation (2nd order)
Derivation of the Fokker-Planck equation
Increment of the density
Taylor series
0
5 min break
Fokker-Planck equation
Stationary distribution of the Langevin dynamics
With a little efforts (homework), we obtain:
when
(density does not change anymore)
Let
Gibbs distribution
Sampling via the Langevin dynamics
Particles have the stationary distribution:
Gibbs distribution
Langevin equation
We want to sample from
Note!
Sampling via the Langevin dynamics
Stochastic Differential Equation for sampling:
Discrete approximation:
More popular way:
Langevin dynamics for the Bayesian inference
Predictive distribution
We need samples
Langevin equation
Discrete approximation
???
???
Borkar, Mitter, 1999
Consider a SDE:
Theorem
Discrete approximation:
Stationary distribution
Stationary distribution
Sketch of the proof
Lemma
What happened to the noise?
Lemma says
Original dynamics
The same but integrated
Our approximation
Free speed-up?
What happened to the noise?
Computational efforts are hidden here
Sketch of the proof
Lemma 1
Stationary distribution
Theorem
Stationary distribution
Lemma 2
Global optimization
Temperature annealing
Langevin equation
Particles have the stationary distribution:
Gibbs distribution
Concentrates on the global minima
Annealing example
Distribution of particles
Langevin equation
???
Annealing schedule
Theorem:
Distribution of particles
Annealing schedule
Theorem:
Continuous process
Discrete approximation
Stationary distribution
Discretization error
Convergence speed
Further reading
The end