Simulated Maximum Likelihood for Continuous

Simulated Maximum Likelihood for
Continuous-Discrete State Space Models
using
Langevin Importance Sampling.
Simulated Maximum
Likelihood for
Continuous-Discrete
State Space Models
using
Langevin Importance
Sampling.
Hermann Singer
State Space Models
Parameter
Estimation
Langevin Sampling
Simulated Likelihood
Examples
Hermann Singer
FernUniversität in Hagen,
[email protected]
References
presented at the
9th International Conference on Social Science
Methodology (RC33)
11.-16. September 2016, Leicester, UK
Simulated Maximum
Likelihood for
Continuous-Discrete
State Space Models
using
Langevin Importance
Sampling.
Hermann Singer
State Space Models
Parameter
Estimation
Langevin Sampling
Simulated Likelihood
Examples
Session:
Estimation of Stochastic Differential Equations with Time
Series, Panel and Spatial Data.
References
Nonlinear continuous/discrete
state space model
Simulated Maximum
Likelihood for
Continuous-Discrete
State Space Models
using
Langevin Importance
Sampling.
Hermann Singer
dY (t) = f (Y (t), t)dt + g (Y (t), t)dW (t)
Zi
= h(Yi , ti ) + i
State Space Models
Nonlinear
continuous/discrete
state space model
Linear stochastic
differential equations
(LSDE)
Parameter
Estimation
i
= 0, ..., T
Langevin Sampling
Simulated Likelihood
nonlinear drift and diffusion functions f , g
f = f (Y (t), t, x(t), ψ)
nonlinear diffusion g (Y ): Itô calculus
Spatial models: Yn (t) = Y (xn , t), xn ∈ Rd :
Random field Y (x, t, ω)
Examples
References
Linear stochastic differential equations (LSDE)
(exact ML, Singer; 1990, Kalman filter)
Simulated Maximum
Likelihood for
Continuous-Discrete
State Space Models
using
Langevin Importance
Sampling.
Hermann Singer
dY (t) = AY (t)dt + GdW (t)
Z t
Y (t) = e A(t−t0 ) Y (t0 ) +
e A(t−s) GdW (s)
t0
State Space Models
Nonlinear
continuous/discrete
state space model
Linear stochastic
differential equations
(LSDE)
Parameter
Estimation
W (t, ω): Wiener process:
continuous time random walk (Brownian motion)
Langevin Sampling
Itô stochastic differential equations:
dW /dt = ζ(t) = Gaussian white noise
Examples
A: drift matrix
G : diffusion matrix
Simulated Likelihood
References
Simulated Maximum
Likelihood for
Continuous-Discrete
State Space Models
using
Langevin Importance
Sampling.
Wiener process and stock index
DAX
DAX-Zuwachs
8000
400
7000
6000
Out[467]=
200
5000
Out[454]=
Hermann Singer
0
4000
-200
State Space Models
3000
Nonlinear
continuous/discrete
state space model
Linear stochastic
differential equations
(LSDE)
-400
2000
1995
2000
1995
2005
2000
2005
Figure : German stock index (DAX)
Simulierter Wiener-Prozeß
Parameter
Estimation
Langevin Sampling
Zuwächse: Simulierter Wiener-Prozeß
0.2
3
Simulated Likelihood
2
Out[503]=
Examples
0.1
1
Out[504]=
References
0.0
0
-0.1
-1
0
1000
2000
3000
4000
5000
6000
7000
0
1000
2000
3000
4000
5000
Figure : Simulated Wiener process (random walk)
6000
7000
Simulated Maximum
Likelihood for
Continuous-Discrete
State Space Models
using
Langevin Importance
Sampling.
Simulated Wiener processes
Simulierte Wiener-Prozesse
4
3
0.5
0.0
2
3
1
2
Hermann Singer
-0.5
-1.0
0
-1.5
-2.0
1
State Space Models
-1
0
-2.5
0
1000
2000
3000
4000
5000
6000
7000
0
1000
2000
3000
4000
5000
6000
7000
0
5
7
4
6
1000
2000
3000
4000
5000
6000
7000
0
-1
5
3
Out[501]=
4
-2
Nonlinear
continuous/discrete
state space model
Linear stochastic
differential equations
(LSDE)
2
3
-3
Parameter
Estimation
1
2
-4
0
1000
2000
3000
4000
5000
6000
0
1
-1
0
7000
0
1000
2000
3000
4000
5000
6000
7000
0
1000
2000
3000
4000
5000
6000
7000
Langevin Sampling
3
2
Simulated Likelihood
1
2
1
Examples
0
1
0
-1
0
-1
References
-2
-1
-2
-3
-2
0
1000
2000
3000
4000
5000
6000
7000
0
1000
2000
3000
4000
5000
6000
7000
0
1000
2000
Figure : Simulated Wiener processes
3000
4000
5000
6000
7000
Simulated Maximum
Likelihood for
Continuous-Discrete
State Space Models
using
Langevin Importance
Sampling.
Exact discrete model (EDM)
Bergstrom (1976, 1988)
Hermann Singer
Yi+1 = e
A(ti +1 −ti )
Z
Yi +
ti +1
e A(ti +1 −s) GdW (s)
ti
restricted VAR(1) model
State Space Models
Nonlinear
continuous/discrete
state space model
Linear stochastic
differential equations
(LSDE)
Parameter
Estimation
Langevin Sampling
Yi+1 = Φ(ti+1 , ti )Yi + ui
Simulated Likelihood
Examples
References
Φ: fundamental matrix of the system
Yi := Y (ti ): sampled measurements
dt = 0.5
Dt = 2
Simulated Maximum
Likelihood for
Continuous-Discrete
State Space Models
using
Langevin Importance
Sampling.
Hermann Singer
accumulated interaction Dt = 2
State Space Models
Nonlinear
continuous/discrete
state space model
Linear stochastic
differential equations
(LSDE)
accumulated interaction Dt = 4
Parameter
Estimation
Langevin Sampling
Simulated Likelihood
Examples
References
Figure : 3 variable model:
Product representation of matrix exponential within the
measurement interval ∆t = 2. Latent states ηj , discretization
interval δt = 2/4 = 0.5 (Singer; 2012)
Maximum Likelihood Parameter Estimation
likelihood function of all observations
Simulated Maximum
Likelihood for
Continuous-Discrete
State Space Models
using
Langevin Importance
Sampling.
Hermann Singer
Z
p(zT , ..., z0 ) =
p(zT , ..., z0 |yT , ..., y0 )p(yT , ..., y0 )dy
State Space Models
Parameter
Estimation
Langevin Sampling
High dimensional integration over latent variables
smooth dependence on parameter vector ψ:
L(ψ) = p(z; ψ)
Problem: p(yT , ..., y0 ) not known
Sampling interval ∆ti
Transition density p(yi+1 , ∆ti |yi ) difficult to compute
Use additional latent variables
yT = ηJ , ..., η0 = y0 , y (ti ) = yi = ηji
Simulated Likelihood
Examples
References
Simulated Maximum
Likelihood for
Continuous-Discrete
State Space Models
using
Langevin Importance
Sampling.
Integration
latent states ηj
Z
p(zT , ..., z0 |ηJ , ..., η0 )p(ηJ , ..., η0 )d η
h
i
= E p(zT , ..., z0 |ηJ , ..., η0 )
p(zT , ..., z0 ) =
Hermann Singer
State Space Models
Parameter
Estimation
Langevin Sampling
ηji = y (ti ) = yi , ti = measurement times
even higher (infinite) dimensional integration over
latent variables:
Markov Chain Monte Carlo: simulate η
Euler density (discretization interval δt)
p(ηj+1 , δt|ηj+1 ) ≈ φ(ηj+1 ; ηj + fj δt, Ωj δt)
fj = f (ηj , τj ), Ωj = (gg 0 )(ηj , τj )
Simulated Likelihood
Examples
References
Simulated Maximum
Likelihood for
Continuous-Discrete
State Space Models
using
Langevin Importance
Sampling.
Importance Sampling
Z
p(zT , ..., z0 ) =
p(η)
p(z|η)
p2 (η)d η
p2 (η)
Hermann Singer
State Space Models
Parameter
Estimation
p2 : importance density
Langevin Sampling
Simulated Likelihood
p2,optimal =
p(z|η)p(η)
p(z)
= p(η|z)
however, p(z) is unknown
η → η(t, u): random field, u = simulation time
Examples
References
Langevin Sampling
Langevin (1908); Roberts and Stramer (2002)
Langevin equation
d η(u) = ∂η log p(η(u)|z)du +
√
Simulated Maximum
Likelihood for
Continuous-Discrete
State Space Models
using
Langevin Importance
Sampling.
Hermann Singer
2dW (u)
State Space Models
Parameter
Estimation
stationary distribution of conditional latent states
pstat (η) = e −Φ(η) = p(η|z)
drift = −gradient of potential Φ(η)
−∂η Φ(η) = ∂η [log p(z|η) + log p(η) − log p(z)]
sample from p(η|z), p(z) not needed!
optimal nonlinear smoothing
η(u) ∼ p(η|z) in equilibrium u → ∞
Langevin Sampling
Simulated Likelihood
Examples
References
Simulated Likelihood
Variance reduced MC-integration
p̂(zT , ..., z0 ) = L−1
X
p(ηl )
p(z|ηl )
p2 (ηl )
ηl ∼ p(η|z) in equilibrium
Simulated Maximum
Likelihood for
Continuous-Discrete
State Space Models
using
Langevin Importance
Sampling.
Hermann Singer
State Space Models
Parameter
Estimation
Langevin Sampling
Simulated Likelihood
draw optimal ηl = η(ul ) from discretized Langevin
equation (including Metropolis-Hastings mechanism)
p2,optimal =
p(z|η)p(η)
p(z)
= p(η|z) is unknown
Idea: estimate p2 = p(η|z) from ηl ∼ p(η|z)
Examples
References
Estimation of importance density
use known (suboptimal) reference density
p2 = p0 (η|z) = p0 (z|η)p0 (η)/p0 (z)
Hermann Singer
State Space Models
Parameter
Estimation
kernel density estimate
p̂(η|z) = L−1
Simulated Maximum
Likelihood for
Continuous-Discrete
State Space Models
using
Langevin Importance
Sampling.
Langevin Sampling
X
k(η − ηl ; smooth)
l
Problem:
– high dimensional state,
– no structure imposed on p(η|z)
Simulated Likelihood
Examples
References
Estimation of importance density
use Markov structure of state space model
ηj+1 = f (ηj )δt + g (ηj )δWj
zi
Simulated Maximum
Likelihood for
Continuous-Discrete
State Space Models
using
Langevin Importance
Sampling.
Hermann Singer
= h(yi ) + i
p(η|z) = p(ηJ |ηJ−1 , ..., η0 , z) ∗ p(ηJ−1 , ..., η0 |z)
State Space Models
Parameter
Estimation
Langevin Sampling
Conditional Markov process
Simulated Likelihood
Examples
p(ηj+1 |ηj , ..., η0 , z) = p(ηj+1 |ηj , z)
References
use conditional independence of past z i = (z0 , ..., zi ) and
future z̄ i = (zi+1 , ...zT ) given η j :
j
i
i
p(ηj +1 |η , z , z̄ )
i
i
p(ηj +1 |ηj , z , z̄ )
ji ≤
j
i
i
=
p(ηj +1 |η , z̄ ) = p(ηj +1 |ηj , z̄ )
=
p(ηj +1 |ηj , z̄ )
j
< ji +1
i
Euler transition kernel
Euler density (discretization interval δt)
Simulated Maximum
Likelihood for
Continuous-Discrete
State Space Models
using
Langevin Importance
Sampling.
Hermann Singer
p(ηj+1 , δt|ηj ) ≈ φ(ηj+1 ; ηj + fj δt, Ωj δt)
modified drift and diffusion matrix
conditional Euler density
p(ηj+1 , δt|ηj , z) ≈
φ(ηj+1 ; ηj + (fj + δfj )δt, (Ωj + δΩj )δt)
nonlinear regression for δfj and δΩj
(parametric and nonparametric)
draw data ηjl = η(τj , ul ) ∼ p(η|z)
State Space Models
Parameter
Estimation
Langevin Sampling
Simulated Likelihood
Examples
References
Kernel density
Simulated Maximum
Likelihood for
Continuous-Discrete
State Space Models
using
Langevin Importance
Sampling.
Hermann Singer
conditional transition density
p(ηj+1 , ηj |z)
p(ηj+1 , δt|ηj , z) =
p(ηj |z)
State Space Models
Parameter
Estimation
Langevin Sampling
Simulated Likelihood
estimate joint density p(ηj+1 , ηj |z) and p(ηj |z)
with kernel density estimates
variant: use φ(ηj+1 , ηj |z) and φ(ηj |z)
Examples
References
Examples: Geometrical Brownian motion
dy (t) = µy (t)dt + σy (t) dW (t)
Simulated Maximum
Likelihood for
Continuous-Discrete
State Space Models
using
Langevin Importance
Sampling.
Hermann Singer
State Space Models
nonlinear model (multiplicative noise) y ∗ dW
Parameter
Estimation
exact solution: set x = log y , use Itô’s lemma
Langevin Sampling
Simulated Likelihood
dx = dy /y + 1/2(−y −2 )dy 2 = (µ − σ 2 /2)dt + σdW
Examples
References
exact discrete model
y (t) = y (t0 )e (µ−σ
2 /2)(t−t
0 )+σ[W (t)−W (t0 )]
���� μ = �����σ = ����
���(���)
����
����
����
����
����
����
����
����
Simulated Maximum
Likelihood for
Continuous-Discrete
State Space Models
using
Langevin Importance
Sampling.
��� �������
�����
�����
�����
�����
-�����
-�����
����
����
�
��
���
���
���
���
���
���
�
��
���
���
���
���
���
���
-�����
�
��
���
���
���
���
���
���
Hermann Singer
Figure : Trajectory and log returns.
State Space Models
Parameter
Estimation
��������
����
Langevin Sampling
����
���
����
����
���
Simulated Likelihood
����
����
���
����
����
���
����
����
���
����
�
���
���
���
���
���
���
���
���
���
���
���
���
���
�
���
��
���
���
���
���
���
���
���
���
���
���
���
���
����
����
References
����
�
����
����
Examples
����
���
����
���
����
���
����
����
����
����
����
���
����
���
����
�
��
���
���
���
���
���
���
���
�
���
���
���
���
Figure : Langevin sampler, p̂2 =
���
Q
���
j
���
�
���
φ(ηj+1 , ηj |z)/φ(ηj |z) .
���
Simulated Maximum
Likelihood for
Continuous-Discrete
State Space Models
using
Langevin Importance
Sampling.
��
���
��
���
���
��
���
��
���
��
Hermann Singer
���
��
- �����
- ���
�����
�����
�����
�����
�����
-�����
�����
�����
�����
�����
Figure : likelihood and score, p̂2 = conditional kernel density.
State Space Models
Parameter
Estimation
Langevin Sampling
Simulated Likelihood
Examples
�
References
���
-�
- ��
���
- ��
- ��
���
- ��
- ��
��
-�����
�����
�����
�����
�����
�����
- �����
�����
�����
�����
Figure : likelihood and score, p̂2 = full kernel density.
�����
Simulated Maximum
Likelihood for
Continuous-Discrete
State Space Models
using
Langevin Importance
Sampling.
��
���
��
���
���
��
���
��
���
��
Hermann Singer
���
��
-�����
- ���
�����
�����
�����
�����
�����
-�����
�����
�����
�����
�����
Figure : likelihood and score, p̂2 = conditionally Gaussian.
State Space Models
Parameter
Estimation
Langevin Sampling
Simulated Likelihood
Examples
��
���
��
References
���
���
��
���
��
���
��
��
- �����
���
- ���
�����
�����
�����
�����
�����
-�����
�����
�����
�����
Figure : likelihood and score, p̂2 = linear GLS.
�����
Simulated Maximum
Likelihood for
Continuous-Discrete
State Space Models
using
Langevin Importance
Sampling.
Hermann Singer
��
�
��
�
State Space Models
�
Parameter
Estimation
��
�
Langevin Sampling
�
Simulated Likelihood
��
- �����
�����
�����
�����
�����
�����
-�����
�����
�����
�����
�����
Examples
Figure : likelihood and score, p̂2 = linear GLS,
constant diffusion matrix.
References
Simulated Maximum
Likelihood for
Continuous-Discrete
State Space Models
using
Langevin Importance
Sampling.
Examples: Cameron-Martin formula
p
λ2 R T
2
E e − 2 0 W (t) dt = 1/ cosh(T λ)
Cameron and Martin (1945); Gelfand and Yaglom (1960)
Hermann Singer
State Space Models
��������
���
�
�
���
�
Parameter
Estimation
�
���
�
�
���
-�
-�
���
�
���
���
���
���
���
Langevin Sampling
-�
���
-�
�
���
���
���
Simulated Likelihood
���
�
�
�
�
��
Examples
���
�
���
���
�
References
���
���
-�
���
���
���
-��
���
���
�
���
���
���
���
�
���
���
���
���
�
���
���
���
���
Figure : Cameron-Martin formula.
Simulation using a conditionally gaussian importance density.
Tp
= 1, λ = 1, dt = 0.1 and L = 500 replications. Exact value
1/ cosh(1) = 0.805018
Simulated Maximum
Likelihood for
Continuous-Discrete
State Space Models
using
Langevin Importance
Sampling.
Examples: Cameron-Martin formula
Hermann Singer
�������-������ �������� �������� ��������� �� = ����
�������-������ �������� �������� ��������� �� = ����
���
���
���
���
���
���
���
State Space Models
Parameter
Estimation
���
��������
Langevin Sampling
��������
���
���
���
���
���
���
���
Simulated Likelihood
Examples
���
�
�
�
�
�
�
�
�
�
�
�
�
Figure : Expectation value as a function of T . Right: Ω = fix.
References
Examples: Feynman-Kac-formula
Schrödinger Equation (imaginary time t = iτ )
ut
=
1
2 uxx
− φ(x)u
Simulated Maximum
Likelihood for
Continuous-Discrete
State Space Models
using
Langevin Importance
Sampling.
Hermann Singer
State Space Models
u(x, t = 0) = δ(x − z)
φ(x) =
1 2 2
2γ x
Parameter
Estimation
Langevin Sampling
Simulated Likelihood
Examples
Feynman-Kac-formula
h γ2 R t
i
2
u(x, t) = Ex e − 2 0 W (u) du δ(W (t) − z) =
r
γ
×
2π sinh(γt)
γ
2
2
exp
2xz − (x + z ) cosh(γt)
2 sinh(γt)
References
Simulated Maximum
Likelihood for
Continuous-Discrete
State Space Models
using
Langevin Importance
Sampling.
Feynman-Kac-formula
�������
���
���
���
���
���
���
���
���
���
���
���
���
���
���
���
���
���
���
���
���
- ���
���
���
���
���
�
���
���
���
����
�
���
���
����
��
��
��
��
���
�
State Space Models
���
�
����
���
����
Parameter
Estimation
�
����
��
����
�
��
����
����
Hermann Singer
-���
�
Langevin Sampling
�
���
���
���
���
�
���
���
���
���
�
���
���
���
���
Simulated Likelihood
�������-������ �������� �������� ��������� �� = ����
Examples
����
References
����
����
����
����
-���
-���
���
���
���
���
���
�
Figure : Expectation value as a function of z, x = 1, t = 1,
γ = 1.
Conclusion
Simulated Maximum
Likelihood for
Continuous-Discrete
State Space Models
using
Langevin Importance
Sampling.
Hermann Singer
smooth likelihood simulation for nonlinear
continuous-discrete state space models.
State Space Models
Parameter
Estimation
Nonlinear smoothing of latent variables between
measurements.
Langevin Sampling
Simulated Likelihood
Examples
References
Variance reduced MC estimation of functional
integrals in finance, statistics and quantum theory
(Feynman-Kac formula).
Bergstrom, A. (1988). The history of continuous-time econometric
models, Econometric Theory 4: 365–383.
Bergstrom, A. (ed.) (1976). Statistical Inference in Continuous Time
Models, North Holland, Amsterdam.
Borodin, A. and Salminen, P. (2002). Handbook of Brownian Motion
– Facts and Formulae, second edn, Birkhäuser-Verlag, Basel.
Cameron, R. H. and Martin, W. T. (1945). Transformations of
Wiener Integrals Under a General Class of Linear
Transformations, Transactions of the American Mathematical
Society 58(2): 184–219.
Feynman, R. and Hibbs, A. (1965). Quantum Mechanics and Path
Integrals, McGraw-Hill, New York.
Gelfand, I. and Yaglom, A. (1960). Integration in Functional Spaces
and its Application in Quantum Physics, Journal of Mathematical
Physics 1(1): 48–69.
Langevin, P. (1908). Sur la théorie du mouvement brownien [on the
theory of brownian motion], Comptes Rendus de l’Academie des
Sciences (Paris) 146: 530–533.
Oud, J. and Singer, H. (2008). Special issue: Continuous time
modeling of panel data. Editorial introduction, Statistica
Neerlandica 62, 1: 1–3.
Roberts, G. O. and Stramer, O. (2002). Langevin Diffusions and
Metropolis-Hastings Algorithms, Methodology And Computing In
Applied Probability 4(4): 337–357.
Simulated Maximum
Likelihood for
Continuous-Discrete
State Space Models
using
Langevin Importance
Sampling.
Hermann Singer
State Space Models
Parameter
Estimation
Langevin Sampling
Simulated Likelihood
Examples
References
Singer, H. (1990). Parameterschätzung in zeitkontinuierlichen
dynamischen Systemen [Parameter estimation in continuous time
dynamical systems; Ph.D. thesis, University of Konstanz, in
German], Hartung-Gorre-Verlag, Konstanz.
Singer, H. (2012). SEM modeling with singular moment matrices.
Part II: ML-Estimation of Sampled Stochastic Differential
Equations., Journal of Mathematical Sociology 36(1): 22–43.
Simulated Maximum
Likelihood for
Continuous-Discrete
State Space Models
using
Langevin Importance
Sampling.
Hermann Singer
State Space Models
Parameter
Estimation
Langevin Sampling
Simulated Likelihood
Examples
References