the generalized correlated sampling approach: toward an...
TRANSCRIPT
Plan
Introduction
The generalized correlated sampling approach 9 avril 2008 1 de 44
The generalized correlated sampling approach:toward an exact calculation of energy
derivatives in Diffusion Monte Carlo
Roland Assarafa,Alexander Kolliasb and Michel Caffarelc
a) Laboratoire de Chimie Theorique, CNRS-UMR 7616, Universite Pierre et Marie Curie Paris VI, Case137, 4, place Jussieu 75252 PARIS Cedex 05, France
b) Carnegie Institution of Washington, Geophysical Laboratory 5251 Broad Branch Rd., N.W.,Washington, DC 20015, USA
c) Laboratoire de Chimie et Physique Quantiques, CNRS-UMR 5626, IRSAMC Universite PaulSabatier, 118 route de Narbonne 31062 Toulouse Cedex, France
Plan
Introduction
The generalized correlated sampling approach 9 avril 2008 2 de 44
General perspective
Quantum Monte Carlo (QMC) : stochastic technics usedto solve the Schroedinger equation
In principle adapted to the many body problem (weaklimitation in system sizes)
In Practice
Reference methods for groundstate energies of largesystems.
Less successful for other quantities.
Plan
Introduction
The generalized correlated sampling approach 9 avril 2008 3 de 44
Quantities of physical interest ?
Most of them can be obtained from total energies
Binding energies. Transition state energies. One, two particlegaps (electron affinities, ionization energies) . . .
First order derivatives of the energy : Any observable (force,dipole, moment, densities...).
Higher order derivatives : spectroscopic constants . . .
Differences of energies of very close systems(small energy differences)
Plan
Introduction
The generalized correlated sampling approach 9 avril 2008 4 de 44
Paradigm : Calculation of an observable O
Hλ = H + λO => O =dEλ
dλ≃ Eλ − E0
λ(1)
Direct calculation
Eλ, E0 computed independently.
δ(Eλ −E0
λ) ∼ δE0
λ−→∞λ→ 0
Plan
Introduction
The generalized correlated sampling approach 9 avril 2008 5 de 44
System size dependency in a direct calculation
Example : a one particle gap ∆ = E (N + 1)− E (N )
δE ∼ N and ∆ ∼ 1 (best case)
Independent calculation of energies :
=⇒ δ∆
∆∼ N
1/N plays the role of the small parameter λ
In practice limiting factor on system sizes
Plan
Introduction
The generalized correlated sampling approach 9 avril 2008 6 de 44
Exploiting accurate QMC total energies to obtain accurate smalldifferences is not as simple as in a deterministic method.
=> At the heart of practical limitations regarding properties one can
compute and system sizes one can reach in QMC.
Objective
Eλ − E0 ∼ λ =⇒ δ(Eλ − E0) ∼ λ
⇐⇒Finite statistical error on energy derivatives.
Plan
Introduction
The generalized correlated sampling approach 9 avril 2008 6 de 44
Exploiting accurate QMC total energies to obtain accurate smalldifferences is not as simple as in a deterministic method.
=> At the heart of practical limitations regarding properties one can
compute and system sizes one can reach in QMC.
Objective
Eλ − E0 ∼ λ =⇒ δ(Eλ − E0) ∼ λ
⇐⇒Finite statistical error on energy derivatives.
Plan
Introduction
The generalized correlated sampling approach 9 avril 2008 6 de 44
Exploiting accurate QMC total energies to obtain accurate smalldifferences is not as simple as in a deterministic method.
=> At the heart of practical limitations regarding properties one can
compute and system sizes one can reach in QMC.
Objective
Eλ − E0 ∼ λ =⇒ δ(Eλ − E0) ∼ λ
⇐⇒Finite statistical error on energy derivatives.
Plan
Introduction
The generalized correlated sampling approach 9 avril 2008 7 de 44
Plan
1 QMC : A reference method for the total energyVariational energyDiffusion Monte Carlo
2 Small energy differences in VMCObservable in VMC
3 Generalization to DMCForward walkingGeneralized correlated samplingCalculation of the DMC energy derivative in H2 and Li2Preliminary results on other moleculesCriteria of stabilty
4 Conclusion and perspectives
Plan
Introduction
The generalized correlated sampling approach 9 avril 2008 7 de 44
Plan
1 QMC : A reference method for the total energyVariational energyDiffusion Monte Carlo
2 Small energy differences in VMCObservable in VMC
3 Generalization to DMCForward walkingGeneralized correlated samplingCalculation of the DMC energy derivative in H2 and Li2Preliminary results on other moleculesCriteria of stabilty
4 Conclusion and perspectives
Plan
Introduction
The generalized correlated sampling approach 9 avril 2008 7 de 44
Plan
1 QMC : A reference method for the total energyVariational energyDiffusion Monte Carlo
2 Small energy differences in VMCObservable in VMC
3 Generalization to DMCForward walkingGeneralized correlated samplingCalculation of the DMC energy derivative in H2 and Li2Preliminary results on other moleculesCriteria of stabilty
4 Conclusion and perspectives
Plan
Introduction
The generalized correlated sampling approach 9 avril 2008 7 de 44
Plan
1 QMC : A reference method for the total energyVariational energyDiffusion Monte Carlo
2 Small energy differences in VMCObservable in VMC
3 Generalization to DMCForward walkingGeneralized correlated samplingCalculation of the DMC energy derivative in H2 and Li2Preliminary results on other moleculesCriteria of stabilty
4 Conclusion and perspectives
Plan
I. Energy calculation
Variational energy
Diffusion Monte Carlo
Small energydifferences in VMC
Observable in VMC
Generalization toDMC
Forward walking
Generalized correlatedsampling
Calculation of the DMC energy
derivative in H2 and Li2
Preliminary results on othermolecules
Criteria of stability
Conclusion andperspectives
The generalized correlated sampling approach 9 avril 2008 8 de 44
Variational energy
EV ≡ 〈Ψ|H |Ψ〉
Average on a probability distribution
〈Ψ|H |Ψ〉 =
∫
dRΨ2(R)HΨ
Ψ(R)
= 〈HΨΨ (R)〉Ψ2 = 〈EL(R)〉Ψ2
Ev =1
N
N∑
k=1
EL(Rk )
Plan
I. Energy calculation
Variational energy
Diffusion Monte Carlo
Small energydifferences in VMC
Observable in VMC
Generalization toDMC
Forward walking
Generalized correlatedsampling
Calculation of the DMC energy
derivative in H2 and Li2
Preliminary results on othermolecules
Criteria of stability
Conclusion andperspectives
The generalized correlated sampling approach 9 avril 2008 8 de 44
Variational energy
EV ≡ 〈Ψ|H |Ψ〉
Average on a probability distribution
〈Ψ|H |Ψ〉 =
∫
dRΨ2(R)HΨ
Ψ(R)
= 〈HΨΨ (R)〉Ψ2 = 〈EL(R)〉Ψ2
Ev =1
N
N∑
k=1
EL(Rk )
Plan
I. Energy calculation
Variational energy
Diffusion Monte Carlo
Small energydifferences in VMC
Observable in VMC
Generalization toDMC
Forward walking
Generalized correlatedsampling
Calculation of the DMC energy
derivative in H2 and Li2
Preliminary results on othermolecules
Criteria of stability
Conclusion andperspectives
The generalized correlated sampling approach 9 avril 2008 9 de 44
Sampling of Ψ2
Dynamic over the configurations
R(t + dt) = R(t) + bdt + dW (2)
b(t) ≡ ∇ΨΨ (drift)
dW gaussian random numbers (diffusion).〈dWidWj 〉 =
√dtδij
A trajectory R(t) <=> Sample of Ψ2.
Plan
I. Energy calculation
Variational energy
Diffusion Monte Carlo
Small energydifferences in VMC
Observable in VMC
Generalization toDMC
Forward walking
Generalized correlatedsampling
Calculation of the DMC energy
derivative in H2 and Li2
Preliminary results on othermolecules
Criteria of stability
Conclusion andperspectives
The generalized correlated sampling approach 9 avril 2008 10 de 44
Why VMC can be a very accurate method ?
Zero-variance - zero-bias property (ZVZB)
Bias EV − E0 |Ψ− Φ|2Variance σ2(EL) |Ψ− Φ|2
φ Exact groundstate
Important sinceAccuracy in QMC <=> EV − E0 systematic error
+ σ(EL)√N
statistical error.
Plan
I. Energy calculation
Variational energy
Diffusion Monte Carlo
Small energydifferences in VMC
Observable in VMC
Generalization toDMC
Forward walking
Generalized correlatedsampling
Calculation of the DMC energy
derivative in H2 and Li2
Preliminary results on othermolecules
Criteria of stability
Conclusion andperspectives
The generalized correlated sampling approach 9 avril 2008 11 de 44
Diffusion Monte Carlo (DMC)
Sampling the exact groundstate
e−tH |Ψ〉 = Φ (3)
Trotter formula : e−tH = e−δtH e−δtH . . . e−δtH
〈R′|e−δtH|R〉 = P(R→ R′)W(R)
Overdamped Langevin Weight
In practice for fermions, Fixed node approximation :H −→ HFN =⇒ Φ −→ ΦFN (variational solution in the spaceof functions having the same nodes as Ψ).
Plan
I. Energy calculation
Variational energy
Diffusion Monte Carlo
Small energydifferences in VMC
Observable in VMC
Generalization toDMC
Forward walking
Generalized correlatedsampling
Calculation of the DMC energy
derivative in H2 and Li2
Preliminary results on othermolecules
Criteria of stability
Conclusion andperspectives
The generalized correlated sampling approach 9 avril 2008 12 de 44
Conclusion
We can sample something better than Ψ, ΦFN , even if itsanalytic form is unknown.
Illustration
-460.8
-460.6
-460.4
-460.2
-460
-459.8
1 2 3 4 5 6 7 8 9 10
En
ergie
(H
art
ree)
R (Bohr)
RHF
MCSCF CAS(8,5)
MRCISD CAS(8,5)
VMC J × CAS(8,5)
DMC J × CAS(8,5) Exact
Dissociation curve for HCl (collaboration J. Toulouse)
Can we compute accurate small differences or derivativeson the VMC curve, on the DMC curve ?
Plan
I. Energy calculation
Variational energy
Diffusion Monte Carlo
Small energydifferences in VMC
Observable in VMC
Generalization toDMC
Forward walking
Generalized correlatedsampling
Calculation of the DMC energy
derivative in H2 and Li2
Preliminary results on othermolecules
Criteria of stability
Conclusion andperspectives
The generalized correlated sampling approach 9 avril 2008 13 de 44
Observable and energy on a similar footing inVMC
ZVZB improved estimators (R. Assaraf, M. Caffarel, 2003)
Hλ = H + λO => 〈O〉Φ2 =dEλ
dλ≃ dEV [Ψλ]
dλ
Ψλ = Ψ + λΨ
dEV [Ψλ]
dλ= 〈O〉Ψ2
O =⇒ O [Ψ, Ψ] = O +(H − EL)Ψ
Ψ+ 2(EL − EV )
Ψ
Ψ
Plan
I. Energy calculation
Variational energy
Diffusion Monte Carlo
Small energydifferences in VMC
Observable in VMC
Generalization toDMC
Forward walking
Generalized correlatedsampling
Calculation of the DMC energy
derivative in H2 and Li2
Preliminary results on othermolecules
Criteria of stability
Conclusion andperspectives
The generalized correlated sampling approach 9 avril 2008 13 de 44
Observable and energy on a similar footing inVMC
ZVZB improved estimators (R. Assaraf, M. Caffarel, 2003)
Hλ = H + λO => 〈O〉Φ2 =dEλ
dλ≃ dEV [Ψλ]
dλ
Ψλ = Ψ + λΨ
dEV [Ψλ]
dλ= 〈O〉Ψ2
O =⇒ O [Ψ, Ψ] = O +(H − EL)Ψ
Ψ+ 2(EL − EV )
Ψ
Ψ
Plan
I. Energy calculation
Variational energy
Diffusion Monte Carlo
Small energydifferences in VMC
Observable in VMC
Generalization toDMC
Forward walking
Generalized correlatedsampling
Calculation of the DMC energy
derivative in H2 and Li2
Preliminary results on othermolecules
Criteria of stability
Conclusion andperspectives
The generalized correlated sampling approach 9 avril 2008 14 de 44
Energy/observable in VMC
Bias Variance Choice of
EV [Ψ] δΨ2 Idem Ψ
EV′λ = 〈O [Ψ, Ψ]〉Ψ2 δΨ2 + δΨδΨ′ Idem (Ψ, Ψ)
〈O〉Ψ2 δΨ O(1) (Ψ, Ψ = 0)
δΨ = Ψ−Φ
δΨ′ = Ψ− dΦλ
dλ
Plan
I. Energy calculation
Variational energy
Diffusion Monte Carlo
Small energydifferences in VMC
Observable in VMC
Generalization toDMC
Forward walking
Generalized correlatedsampling
Calculation of the DMC energy
derivative in H2 and Li2
Preliminary results on othermolecules
Criteria of stability
Conclusion andperspectives
The generalized correlated sampling approach 9 avril 2008 15 de 44
Electronic densityR. Assaraf, M. Caffarel, and A. Scemama, Phys. Rev. E. 75 035701, (2007)
ρ(r) =∑
i
δ(ri − r)
Regularized bare estimator :
Counting electrons in small boxes of size ∆R around r :=> Bias O(∆R)
=> Variance O( 1∆R
3).
Poorest, in regions rarely or never visited by electrons
Plan
I. Energy calculation
Variational energy
Diffusion Monte Carlo
Small energydifferences in VMC
Observable in VMC
Generalization toDMC
Forward walking
Generalized correlatedsampling
Calculation of the DMC energy
derivative in H2 and Li2
Preliminary results on othermolecules
Criteria of stability
Conclusion andperspectives
The generalized correlated sampling approach 9 avril 2008 16 de 44
Simplest improved estimator
δ(ri − r) = −14π
∆ 1|ri−r| + Green formula
ρZV (r) =
⟨
− 1
4π
N∑
i=1
1
|ri − r|∇2
i (Ψ2)
Ψ2
⟩
Ψ2
.
Used by P. Langfelder 1et al for the density at a nucleus, in
combination with a modified important sampling (to cure the still
infinite variance).
1P. Langfelder, S.M. Rothstein, J. Vrbik, J. Chem. Phys. 107 8525 (1997)
Plan
I. Energy calculation
Variational energy
Diffusion Monte Carlo
Small energydifferences in VMC
Observable in VMC
Generalization toDMC
Forward walking
Generalized correlatedsampling
Calculation of the DMC energy
derivative in H2 and Li2
Preliminary results on othermolecules
Criteria of stability
Conclusion andperspectives
The generalized correlated sampling approach 9 avril 2008 17 de 44
Illustration : Helium atom
0
0.5
1
1.5
2
2.5
3
3.5
0 0.5 1 1.5 2 2.5 3
Usual EstimatorSimple Improved Estimator
Best Improved EstimatorExact
2.4
2.5
2.6
2.7
2.8
2.9
3
3.1
3.2
3.3
0 0.02 0.04 0.06
-0.001
-0.0005
0
0.0005
0.001
0.0015
0.002
0.0025
0.003
0.0035
2.5 2.6 2.7 2.8 2.9 3
ρ
r
(r)
Plan
I. Energy calculation
Variational energy
Diffusion Monte Carlo
Small energydifferences in VMC
Observable in VMC
Generalization toDMC
Forward walking
Generalized correlatedsampling
Calculation of the DMC energy
derivative in H2 and Li2
Preliminary results on othermolecules
Criteria of stability
Conclusion andperspectives
The generalized correlated sampling approach 9 avril 2008 18 de 44
Application : water dimer
Fig.: Isodensity surfaces of the water dime
Bare regularized estimator Improved estimator
Plan
I. Energy calculation
Variational energy
Diffusion Monte Carlo
Small energydifferences in VMC
Observable in VMC
Generalization toDMC
Forward walking
Generalized correlatedsampling
Calculation of the DMC energy
derivative in H2 and Li2
Preliminary results on othermolecules
Criteria of stability
Conclusion andperspectives
The generalized correlated sampling approach 9 avril 2008 19 de 44
Pair correlation function
[J. Toulouse, R. Assaraf, C. J. Umrigar. J. Chem. Phys. 126 244112 (2007)]
I (u) =∑
i<j δ(rij − u)Probability density to find a pair of electrons at distance u
0
0.05
0.1
0.15
0.2
0.25
0.3
0 0.5 1 1.5 2
I(u
) (a
.u.)
u (a.u.)
histogram estimator with HF wave functionZV1 estimator with HF wave function
ZV1ZB1 estimator with HF wave functionaccurate intracule
He atom
Plan
I. Energy calculation
Variational energy
Diffusion Monte Carlo
Small energydifferences in VMC
Observable in VMC
Generalization toDMC
Forward walking
Generalized correlatedsampling
Calculation of the DMC energy
derivative in H2 and Li2
Preliminary results on othermolecules
Criteria of stability
Conclusion andperspectives
The generalized correlated sampling approach 9 avril 2008 20 de 44
Impact of the optimisation of Ψ
Illustration : Intracule calculations (collaboration with J.
Toulouse et C. Umrigar)
-0.4
-0.2
0
0.2
0.4
0.6
0 1 2 3 4 5
4 π
u2 [
I (u
) - IH
F (u
) ]
(a.u
.)
u (a.u.)
Jastrow × HFJastrow × SD
Jastrow × CAS(10,8)
N2 molecule
VMC with ZVZB2 estimator
Fig.: Correlation part of the radial intracule in function of thedistance electron-electron u for the molecule N2.
Plan
I. Energy calculation
Variational energy
Diffusion Monte Carlo
Small energydifferences in VMC
Observable in VMC
Generalization toDMC
Forward walking
Generalized correlatedsampling
Calculation of the DMC energy
derivative in H2 and Li2
Preliminary results on othermolecules
Criteria of stability
Conclusion andperspectives
The generalized correlated sampling approach 9 avril 2008 21 de 44
In summary
Derivatives (or small differences) of VMC energies : onthe same footing as total VMC energies
Strategy : Using improved estimators OZVZB [Ψ, Ψ].
=> control of the statistical error and the systematic error, ata reasonable computational cost.
Current work :
Improving estimators for forces (geometry optimization).
Application to a variety of other properties (dipolemoments...).
Plan
I. Energy calculation
Variational energy
Diffusion Monte Carlo
Small energydifferences in VMC
Observable in VMC
Generalization toDMC
Forward walking
Generalized correlatedsampling
Calculation of the DMC energy
derivative in H2 and Li2
Preliminary results on othermolecules
Criteria of stability
Conclusion andperspectives
The generalized correlated sampling approach 9 avril 2008 22 de 44
Generalization to DMCR. Assaraf, A. Kollias, M. Caffarel
The DMC energy
Trial function Groundstate of H
EDMC = 〈Ψ | H | Φ〉 = 〈EL〉ΨΦ
ΨΦ = Ψe−tH | Ψ〉 =
⟨
e−
R
t
0EL(R(s))ds | R(t)〉
⟩
[R(s)](2)
[R(s)] Overdamped Langevin process with the drift b = ∇ΨΨ
Plan
I. Energy calculation
Variational energy
Diffusion Monte Carlo
Small energydifferences in VMC
Observable in VMC
Generalization toDMC
Forward walking
Generalized correlatedsampling
Calculation of the DMC energy
derivative in H2 and Li2
Preliminary results on othermolecules
Criteria of stability
Conclusion andperspectives
The generalized correlated sampling approach 9 avril 2008 22 de 44
Generalization to DMCR. Assaraf, A. Kollias, M. Caffarel
The DMC energy
Trial function Groundstate of H
EDMC = 〈Ψ | H | Φ〉 = 〈EL〉ΨΦ
ΨΦ = Ψe−tH | Ψ〉 =
⟨
e−
R
t
0EL(R(s))ds | R(t)〉
⟩
[R(s)](2)
[R(s)] Overdamped Langevin process with the drift b = ∇ΨΨ
Plan
I. Energy calculation
Variational energy
Diffusion Monte Carlo
Small energydifferences in VMC
Observable in VMC
Generalization toDMC
Forward walking
Generalized correlatedsampling
Calculation of the DMC energy
derivative in H2 and Li2
Preliminary results on othermolecules
Criteria of stability
Conclusion andperspectives
The generalized correlated sampling approach 9 avril 2008 22 de 44
Generalization to DMCR. Assaraf, A. Kollias, M. Caffarel
The DMC energy
Trial function Groundstate of H
EDMC = 〈Ψ | H | Φ〉 = 〈EL〉ΨΦ
ΨΦ = Ψe−tH | Ψ〉 =
⟨
e−
R
t
0EL(R(s))ds | R(t)〉
⟩
[R(s)](2)
[R(s)] Overdamped Langevin process with the drift b = ∇ΨΨ
Plan
I. Energy calculation
Variational energy
Diffusion Monte Carlo
Small energydifferences in VMC
Observable in VMC
Generalization toDMC
Forward walking
Generalized correlatedsampling
Calculation of the DMC energy
derivative in H2 and Li2
Preliminary results on othermolecules
Criteria of stability
Conclusion andperspectives
The generalized correlated sampling approach 9 avril 2008 23 de 44
Forward walking
Hλ = H + λO => HλΦλ = EλΦλ
Computation ofdEλ
dλ=
d〈ELλ〉ΦλΨ
dλ
Forward walking
Same trial function for H and Hλ (Ψ = Ψλ).
dΨΦλ
dλ=
⟨[
−∫
t
0
dsO[R(s)]
]
e−
R
t
0EL(R(s))ds | R(t)
⟩
[R(s)]
In principle exact, but tractable only if O has a small variance.Example for a force at the nucleus, variance infinite.
Plan
I. Energy calculation
Variational energy
Diffusion Monte Carlo
Small energydifferences in VMC
Observable in VMC
Generalization toDMC
Forward walking
Generalized correlatedsampling
Calculation of the DMC energy
derivative in H2 and Li2
Preliminary results on othermolecules
Criteria of stability
Conclusion andperspectives
The generalized correlated sampling approach 9 avril 2008 24 de 44
Approximations
Example : Hybrid improved estimator, in a FNapproximation
〈O〉h ≡ 2〈O〉DMC − 〈O〉VMC
Systematic error : O(|Ψ− Φ|2) + O(|Ψ− Φ1|.|ΦFN − Φ|)
O(|ΦFN − Φ|) for 〈O〉hVariance O(|Ψ− Φ|2)
O(1) for〈O〉hApplication to forces (R. Assaraf, M. Caffarel, 2003)
Plan
I. Energy calculation
Variational energy
Diffusion Monte Carlo
Small energydifferences in VMC
Observable in VMC
Generalization toDMC
Forward walking
Generalized correlatedsampling
Calculation of the DMC energy
derivative in H2 and Li2
Preliminary results on othermolecules
Criteria of stability
Conclusion andperspectives
The generalized correlated sampling approach 9 avril 2008 25 de 44
Generalized correlated sampling
First ingredient : two different guiding functions
Ψλ = Ψ + λΨ => bλ = b + λb
Eλ − E0 = 〈ELλ〉ΨλΦλ− 〈EL〉ΨΦ
Second ingredient correlating the Overamped Langevinprocess
R(t + dt) = R(t) + bdt + dW
Rλ(t + dt) = Rλ(t) + bλdt + dWλ
Wiener processes related by a unitary transform
dWλ = UλdW (3)
Plan
I. Energy calculation
Variational energy
Diffusion Monte Carlo
Small energydifferences in VMC
Observable in VMC
Generalization toDMC
Forward walking
Generalized correlatedsampling
Calculation of the DMC energy
derivative in H2 and Li2
Preliminary results on othermolecules
Criteria of stability
Conclusion andperspectives
The generalized correlated sampling approach 9 avril 2008 26 de 44
Difference of energies : average on the process [R,Rλ].
Eλ − E0 =⟨
ELλ(Rλ)e−R
t
0dsELλ(Rλ) − EL(R)e−
R
t
0dsEL(R)
⟩
[R,Rλ]
Why a hope of improvement over usual forward walking ?
∫ t
0dsO [R(s)] −→
∫ t
0dsO [R(s)]
O =dELλ(Rλ)
dλ= λOZV (R)+ ~T .~∇EL(R)
~T ≡ Rλ −R
λ
ZV principle on O (choice of ψ) ! !Small fluctuations if 〈~T 2〉 finite !
Plan
I. Energy calculation
Variational energy
Diffusion Monte Carlo
Small energydifferences in VMC
Observable in VMC
Generalization toDMC
Forward walking
Generalized correlatedsampling
Calculation of the DMC energy
derivative in H2 and Li2
Preliminary results on othermolecules
Criteria of stability
Conclusion andperspectives
The generalized correlated sampling approach 9 avril 2008 27 de 44
Why can we hope 〈T 2〉 finite ?
Analysis in 1D U = Id
dT
dt=
∂b
∂xT Growth rate term (4)
+ b Deviation term
If b turned off at time t0
The accumulated deviation T (t0) will evolve as
T (t) = e
R
t
t0
∂b
∂x T (t0)→ 0
Since 〈 ∂b∂x〉 < 0 (the system lies in a finite region).
Plan
I. Energy calculation
Variational energy
Diffusion Monte Carlo
Small energydifferences in VMC
Observable in VMC
Generalization toDMC
Forward walking
Generalized correlatedsampling
Calculation of the DMC energy
derivative in H2 and Li2
Preliminary results on othermolecules
Criteria of stability
Conclusion andperspectives
The generalized correlated sampling approach 9 avril 2008 28 de 44
Illustration in many dimensions
1e-14
1e-12
1e-10
1e-08
1e-06
0.0001
0.01
1
100
0 100 200 300 400 500 600 700
(R2-
R1)
2 (t)
Simulation time (au)
Convergence of the joint process, molecule Li8
Fig.: Quadratic deviation of two correlated simulations on the Li8molecule
Plan
I. Energy calculation
Variational energy
Diffusion Monte Carlo
Small energydifferences in VMC
Observable in VMC
Generalization toDMC
Forward walking
Generalized correlatedsampling
Calculation of the DMC energy
derivative in H2 and Li2
Preliminary results on othermolecules
Criteria of stability
Conclusion andperspectives
The generalized correlated sampling approach 9 avril 2008 29 de 44
Stability for the H2 molecule
1e-11
1e-10
1e-09
0 50 100 150 200 250
<(R
2-R
1)2
(t)>
Simulation time (au)
H2 molecule at equilibrium Delta=1e-5
Fig.: Quadratic distance between the two different processes
Plan
I. Energy calculation
Variational energy
Diffusion Monte Carlo
Small energydifferences in VMC
Observable in VMC
Generalization toDMC
Forward walking
Generalized correlatedsampling
Calculation of the DMC energy
derivative in H2 and Li2
Preliminary results on othermolecules
Criteria of stability
Conclusion andperspectives
The generalized correlated sampling approach 9 avril 2008 30 de 44
DMC energy derivative in H2
Crude Ψ : Unoptimized CSF × Minimal jastrow (e-e cusp)
-0.01
0
0.01
0.02
0.03
0.04
0.05
0.06
0.07
0.08
0 2 4 6 8 10 12 14 16-0.1
-0.05
0
0.05
0.1
0.15
Ene
rgy
(au)
Ene
rgy
deriv
ativ
e (a
u)
Projecton time (au)
PDMC convergence for H2 (R=1.6 au, ∆R=1e-4)
E(t)-Eexact
dE/dR(t)-dEexact/dR
Plan
I. Energy calculation
Variational energy
Diffusion Monte Carlo
Small energydifferences in VMC
Observable in VMC
Generalization toDMC
Forward walking
Generalized correlatedsampling
Calculation of the DMC energy
derivative in H2 and Li2
Preliminary results on othermolecules
Criteria of stability
Conclusion andperspectives
The generalized correlated sampling approach 9 avril 2008 31 de 44
-0.02
0
0.02
0.04
0.06
0.08
0.1
0.12
0.14
0.8 1 1.2 1.4 1.6 1.8 2 2.2 2.4
[dE
/dR
](t)
-[dE
/dR
] exa
ct
d(H-H) (au)
PDMC convergences for H2
t=0
t=0.05
t=2.5
t=5
t=1
t=2
t=4
t=8
Plan
I. Energy calculation
Variational energy
Diffusion Monte Carlo
Small energydifferences in VMC
Observable in VMC
Generalization toDMC
Forward walking
Generalized correlatedsampling
Calculation of the DMC energy
derivative in H2 and Li2
Preliminary results on othermolecules
Criteria of stability
Conclusion andperspectives
The generalized correlated sampling approach 9 avril 2008 32 de 44
Finite time step error
-1.0205
-1.02
-1.0195
-1.019
-1.0185
-1.018
-1.0175
-1.017
-1.0165
-1.016
0 0.005 0.01 0.015 0.02 0.025 0.03
E (
au)
Time step (au)
Time step error on the energy
Exact energy (Kolos)
VMC (Projection time t=0)
DMC (t=10)
-0.79
-0.785
-0.78
-0.775
-0.77
-0.765
0 0.005 0.01 0.015 0.02 0.025 0.03
∆E
/∆R
(au),
∆R
=1e-5
Time step (au)
Time step error on the energy derivative
Exact derivative (Kolos)
VMC (Projection time t=0)
DMC (t=10)
Fig.: Finite step error for the energy its derivative. H2 molecule,R = 0.8. Ψ = simple CSF × Full optimized jastrow
Plan
I. Energy calculation
Variational energy
Diffusion Monte Carlo
Small energydifferences in VMC
Observable in VMC
Generalization toDMC
Forward walking
Generalized correlatedsampling
Calculation of the DMC energy
derivative in H2 and Li2
Preliminary results on othermolecules
Criteria of stability
Conclusion andperspectives
The generalized correlated sampling approach 9 avril 2008 33 de 44
Calculation on Li2
2 VMC with same pseudo random numbers (U = Id)
-0.02
0
0.02
0.04
0.06
0.08
0.1
0.12
0.14
0 2000 4000 6000 8000 10000 12000 14000 16000 18000 20000 0
20
40
60
80
100
120
140
< d
EL/
dR >
(au
)
<T
2 > (
au)
Simulation time (au)
Naive calculation on Li2 (R=5.051, ∆R=1e-3), time step=0.05 au
Li2 molecule: R=5.051, λ=∆R = 1e-3
< dEL/dR >
<T2>
Plan
I. Energy calculation
Variational energy
Diffusion Monte Carlo
Small energydifferences in VMC
Observable in VMC
Generalization toDMC
Forward walking
Generalized correlatedsampling
Calculation of the DMC energy
derivative in H2 and Li2
Preliminary results on othermolecules
Criteria of stability
Conclusion andperspectives
The generalized correlated sampling approach 9 avril 2008 34 de 44
Origin of the instabilities for Li2 ?
Finite time is a destabilization factor
Proposition
Pacc(R,Rλ) = min[Pacc(R),Pacc(Rλ)]
Adaptative time step τ (smaller time steps in regions withlarge local kinetic energy)
Note : Finite time step error O(〈τ〉).
Plan
I. Energy calculation
Variational energy
Diffusion Monte Carlo
Small energydifferences in VMC
Observable in VMC
Generalization toDMC
Forward walking
Generalized correlatedsampling
Calculation of the DMC energy
derivative in H2 and Li2
Preliminary results on othermolecules
Criteria of stability
Conclusion andperspectives
The generalized correlated sampling approach 9 avril 2008 35 de 44
Effect on 〈T 2〉
0
20
40
60
80
100
120
140
0 2000 4000 6000 8000 10000 12000 14000 16000 18000 20000
< T
2 > (
au)
Simulation time (au)
Li2 molecule: R=5.051, λ=∆R = 1e-3
Fixed time step, unmodified detailed balance
Adaptative time step, modified detailed balance
Plan
I. Energy calculation
Variational energy
Diffusion Monte Carlo
Small energydifferences in VMC
Observable in VMC
Generalization toDMC
Forward walking
Generalized correlatedsampling
Calculation of the DMC energy
derivative in H2 and Li2
Preliminary results on othermolecules
Criteria of stability
Conclusion andperspectives
The generalized correlated sampling approach 9 avril 2008 36 de 44
Effect on dEL
dR
-0.02
-0.01
0
0.01
0.02
0.03
0.04
0.05
2000 4000 6000 8000 10000 12000 14000 16000 18000 20000
< d
EL/
dR >
(au
)
Simulation time (au)
Li2 molecule: R=5.051, λ=∆R = 1e-3
Fixed time step, unmodified detailed balance
Adaptative time step, modified detailed balance
σ(E ) = 0.16e−02
σ(∆E∆R
) = 0.43e−03 => σ(∆E ) = 0.43e−06 << σ(E ) ! !
Plan
I. Energy calculation
Variational energy
Diffusion Monte Carlo
Small energydifferences in VMC
Observable in VMC
Generalization toDMC
Forward walking
Generalized correlatedsampling
Calculation of the DMC energy
derivative in H2 and Li2
Preliminary results on othermolecules
Criteria of stability
Conclusion andperspectives
The generalized correlated sampling approach 9 avril 2008 37 de 44
Application : Force calculation in thedissociation of Li2
0.02
0.
-0.02
87654
For
ce (
a.u.
)
R (a.u.)
Li2
Exact
Force RHF
FN derivative
Plan
I. Energy calculation
Variational energy
Diffusion Monte Carlo
Small energydifferences in VMC
Observable in VMC
Generalization toDMC
Forward walking
Generalized correlatedsampling
Calculation of the DMC energy
derivative in H2 and Li2
Preliminary results on othermolecules
Criteria of stability
Conclusion andperspectives
The generalized correlated sampling approach 9 avril 2008 38 de 44
Impact of the unitary transform
0.01
0.1
1
10
100
1000
0 100 200 300 400 500 600 700 800
(R2-
R1)
2 (t)
Simulation time (au)
<(R2-R1)2>, U=Id / U = Pe
F2H2 molecule, ∆ dFH-FH=0.1
U=PeU=Id
Fig.: F2H2, 2 geometries ∆dFH−FH = 0.1, U = Id / U = Pe
permutation in the space of electrons (the closest same spin electronshave the same random numbers).
Plan
I. Energy calculation
Variational energy
Diffusion Monte Carlo
Small energydifferences in VMC
Observable in VMC
Generalization toDMC
Forward walking
Generalized correlatedsampling
Calculation of the DMC energy
derivative in H2 and Li2
Preliminary results on othermolecules
Criteria of stability
Conclusion andperspectives
The generalized correlated sampling approach 9 avril 2008 39 de 44
0
0.05
0.1
0.15
0.2
0.25
0.3
100 200 300 400 500 600
Sta
tistic
al e
rror
Simulation time (au)
Error on the energy difference U=Id / U = Pe
F2H2 molecule, ∆ dFH-FH=0.1
sqrt(2)σ(E1)
σ(E2-E1), U=Id
σ(E2-E1), U=Pe
Fig.: F2H2, 2 geometries ∆dFH−FH = 0.1, U = Id / U = Pe
permutation in the space of electrons (the closest same spin electronshave the same random numbers).
Plan
I. Energy calculation
Variational energy
Diffusion Monte Carlo
Small energydifferences in VMC
Observable in VMC
Generalization toDMC
Forward walking
Generalized correlatedsampling
Calculation of the DMC energy
derivative in H2 and Li2
Preliminary results on othermolecules
Criteria of stability
Conclusion andperspectives
The generalized correlated sampling approach 9 avril 2008 40 de 44
Stable case
1e-11
1e-10
1e-09
0 50 100 150 200 250
<(R
2-R
1)2
(t)>
Simulation time (au)
H2 molecule at equilibrium Delta=1e-5
1e-25
1e-20
1e-15
1e-10
1e-05
1
0 50 100 150 200 250
<(R
2-R
1)2
(t)>
Simulation time (au)
H2 molecule at equilibrium Delta=0
Fig.: Quadratic distance between the processes / sensitivity of theunperturbed process to initial conditions (H2 molecule).
Plan
I. Energy calculation
Variational energy
Diffusion Monte Carlo
Small energydifferences in VMC
Observable in VMC
Generalization toDMC
Forward walking
Generalized correlatedsampling
Calculation of the DMC energy
derivative in H2 and Li2
Preliminary results on othermolecules
Criteria of stability
Conclusion andperspectives
The generalized correlated sampling approach 9 avril 2008 41 de 44
Unstable case
1e-09
1e-08
1e-07
1e-06
1e-05
0.0001
0.001
0.01
0.1
50 100 150 200 250 300 350 400
<(R
2-R
1)2
(t)>
Simulation time (au)
CO molecule at equilibrium Delta=1e-5
1e-09
1e-08
1e-07
1e-06
1e-05
0.0001
0.001
0.01
0.1
50 100 150 200 250 300 350 400
<(R
2-R
1)2
(t)>
Simulation time (au)
CO molecule at equilibrium Delta=0
Fig.: Quadratic distance between the two different processes /sensitivity of the unperturbed process to initial conditions (COmolecule.
Plan
I. Energy calculation
Variational energy
Diffusion Monte Carlo
Small energydifferences in VMC
Observable in VMC
Generalization toDMC
Forward walking
Generalized correlatedsampling
Calculation of the DMC energy
derivative in H2 and Li2
Preliminary results on othermolecules
Criteria of stability
Conclusion andperspectives
The generalized correlated sampling approach 9 avril 2008 42 de 44
Some comparisons DMC energies differences correlated /independent
Calculation of ∆E = E (dA−B + ∆R)− E (dA−B)
Molecule ∆R (B) Geometry gain on σ(∆) cpu gain
Li8 1e-5 eq 7e4 0.5e10Li4 1e-5 eq 1e5 1e10Li2 1e-5 5.051 (eq) 4e5 16e10Li2 1e-5 3.0 1.7e5 2.9e10Li2 0.1 3 42 1600Li2 0.1 1 3.6 13CO 0.1 2.175 B (eq) 3.9 15CO 0.1 20 15 225Li4 0.1 eq 14 196F2H2 0.1 eq 10.3 106Li4 1 eq 5.6 31Li2 1 eq 5.25 26
Plan
I. Energy calculation
Variational energy
Diffusion Monte Carlo
Small energydifferences in VMC
Observable in VMC
Generalization toDMC
Forward walking
Generalized correlatedsampling
Calculation of the DMC energy
derivative in H2 and Li2
Preliminary results on othermolecules
Criteria of stability
Conclusion andperspectives
The generalized correlated sampling approach 9 avril 2008 43 de 44
Conclusion and perspectives
Small differences, derivatives of VMC energies
Basically a solved problem, though a lot of work has yet to bedone (improving estimators).
Generalized correlated sampling method for DMC smalldifferences and derivatives
Different guiding functions for the two systems.
Pseudo-random numbers (Wiener process) related by aunitary transform.
Finite time step stabilization (joint acceptation probability,adaptative time step).
Plan
I. Energy calculation
Variational energy
Diffusion Monte Carlo
Small energydifferences in VMC
Observable in VMC
Generalization toDMC
Forward walking
Generalized correlatedsampling
Calculation of the DMC energy
derivative in H2 and Li2
Preliminary results on othermolecules
Criteria of stability
Conclusion andperspectives
The generalized correlated sampling approach 9 avril 2008 44 de 44
Main results
Non trivial DMC energy derivatives obtained.
DMC derivatives ← chaotic properties of the dynamics.
Substantial gain for small but finite differences (∼ 0.1)even for chaotic systems.
Work still to be done to have 〈T 2〉 always finite !
Open questions
Better insight needed into the chaotic properties of theOverdamped Langevin process.
Stabilization techniques to be searched for.