Metropolis-adjusted Langevin algorithm
Mala算法/Langevin Monte Carlo (LMC)v by wikipedia
Algorithm
1.Initialization:set X 0 = x 0 X_0=x_0 X0=x0 (and recursively define an app X k X_k Xk to the true solution X ( k τ ) X(k\tau) X(kτ))
2. (proposal dis)the transition probability density
Note that X k + 1 X_{k+1} Xk+1 is normally distributed with mean X k + τ ∇ log π ( X k ) X_{k}+\tau \nabla \log \pi (X_{k}) Xk+τ∇logπ(Xk) and covariance equal to 2 τ 2\tau 2τ times the d × d d\times d d×d identity matrix.
3. This proposal is accepted or rejected according to the Metropolis–Hastings algorithm:
注意:
- Compared to naive Metropolis–Hastings, MALA has the advantage that it usually proposes moves into regions of higher π \pi π probability, which are then more likely to be accepted.
- On the other hand, when
π
\pi
π is strongly anisotropic (i.e. it varies much more quickly in some directions than others), it is necessary to take
0
<
τ
≪
1
0<\tau \ll 1
0<τ≪1 in order to properly capture the Langevin dynamics;
the use of a positive-definite preconditioning matrix A ∈ R d × d A\in \mathbb {R} ^{d\times d} A∈Rd×d can help to alleviate this problem, by generating proposals according to
X ~ k + 1 : = X k + τ A ∇ log π ( X k ) + 2 τ A ξ k \tilde {X}_{k+1}:=X_{k}+\tau A\nabla \log \pi (X_{k})+ \sqrt {2\tau A} \xi _{k} X~k+1:=Xk+τA∇logπ(Xk)+2τAξk
操作guide:
In practical applications, the optimal acceptance rate for this algorithm is
0.574
0.574
0.574.