SlideShare a Scribd company logo
Adaptive Restore algorithm
Christian P. Robert
U. Paris Dauphine & Warwick U.
0 10 20 30 40 50
−2
0
2
4
x
t
Joint work with H Krimm, M Pollock, AQ Wang
and GO Roberts
arXiv:2210.09901
EaRly Call
Incoming 2023-2030 ERC funding for postdoctoral
collaborations with
I Michael Jordan (Berkeley)
I Eric Moulines (Paris)
I Gareth Roberts (Warwick)
I myself (Paris)
MCMC regeneration
Reversibility attached with (sub-efficient) random walk
behaviour
Recent achievements in non-reversible MCMC with PDMPs
[Bouchard-Côté et al., 2018; Bierkens et al., 2019]
Linked with regeneration
[Nummelin, 1978; Mykland et al., X, 1995]
Restore process
Take {Yt}t≥0 diffusion / jump process on Rd with infinitesimal
generator LY and Y0 ∼ µ —e..g., LY = ∆/2 for Brownian—
Regeneration rate κ with associated tour length
τ = inf t ≥ 0 :
Zt
0
κ(Ys)ds ≥ ξ ξ ∼ Exp(1)
Take {Y
(i)
t }t≥0, τ(i)
∞
i=0
iid realisations
Regeneration times Tj =
Pj−1
i=0 τ(i)
Restore process {Xt}t≥0 given by:
Xt =
∞
X
i=0
I[Ti,Ti+1)(t)Y
(i)
t−Ti
[Wang et al., 2021]
Restore process
0 2 4 6 8
−2
−1
0
1
2
x
t
Path of 5 tours of Brownian Restore with
π ≡ N(0, 12), µ ≡ N(0, 22) and C chosen so that
minx∈R κ(x) = 0, with K = 200 and Λ0 = 1000. First and last
output states shown by green dots and red crosses
Stationarity of Restore
Infinitesimal generator of {Xt}t≥0
LXf(x) = LYf(x) + κ(x)
Z
[f(y) − f(x)]µ(y)dy
Regeneration rate κ chosen as1
κ(x) =
L†
Yπ(x)
π(x)
+ C
µ(x)
π(x)
makes {Xt}t≥0 π-invariant
Z
Rd
LXf(x)π(x)dx = 0
1
L†
Y formal adjoint
Restore sampler
Rewrite
κ(x) =
L†
Yπ(x)
π(x)
+ C
µ(x)
π(x)
= κ̃(x) + C
µ(x)
π(x)
with
I κ̃ partial regeneration rate
I C  0 regeneration constant and
I Cµ regeneration measure, large enough so that κ(·)  0
Resulting Monte Carlo method called Restore Sampler
[Wang et al., 2021]
Restore sampler
Given π-invariance of {Xt}t≥0
Eπ[f] = EX0∼µ
h Zτ(0)
0
f(Xs)dx
i.
EX0∼µ[τ(0)
]
and a.s. convergence of ergodic averages:
1
t
Zt
0
f(Xs)ds → Eπ[f]
For iid Zi :=
RTi+1
Ti
f(Xs)ds, CLT
√
n
ZTn
0
f(Xs)dx

Tn − Eπ[f]
!
→ N(0, σ2
f)
where
σ2
f := EX0∼µ

Z0 − τ(0)
Eπ[f]
2
.
EX0∼µ[τ(0)
]
2
Restore sampler
Given π-invariance of {Xt}t≥0
Eπ[f] = EX0∼µ
h Zτ(0)
0
f(Xs)dx
i.
EX0∼µ[τ(0)
]
and a.s. convergence of ergodic averages:
1
t
Zt
0
f(Xs)ds → Eπ[f]
Estimator variance depends on expected tour length: choose µ
towards long tours
[Wang et al., 2020]
Minimal regeneration
Minimal regeneration measure, C+µ+ corresponding to smallest
possible rate
κ+
(x) := κ̃(x) ∨ 0 = κ̃(x) + C+ µ+(x)
π(x)
leading to
µ+
(x) =
1
C+
[0 ∨ −κ̃(x)]π(x)
Frequent regeneration not necessarily detrimental, except when
when µ is not well-aligned to π, leading to wasted computation
Minimal Restore maximizes expected tour length / minimizes
asymptotic variance
[Wang et al., 2021]
Constant approximation
When target misses normalizing constant
π(x) = π̃(x)

Z,
take energy U(x) := − log π(x) = log Z − log π̃(x)
When {Yt}t≥0 Brownian, κ̃ function of ∇U(x) and ∆U(x)
Constant approximation
When target misses normalizing constant
π(x) = π̃(x)

Z,
take energy U(x) := − log π(x) = log Z − log π̃(x)
When {Yt}t≥0 Brownian, κ̃ function of ∇U(x) and ∆U(x)
In regeneration rate, Z “absorbed” into C
κ(x) = κ̃(x) + C
µ(x)

π̃(x)
Z
 = κ̃(x) + CZ
µ(x)
π̃(x)
= κ̃(x) + C̃
µ(x)
π̃(x)
,
where C̃ = CZ set by user. Since C = 1/Eµ[τ] Using n tours
with simulation time T, Z ≈ C̃T

n.
Adapting Restore
Adaptive Restore process defined by enriching underlying
continuous-time Markov process with regenerations at rate κ+
from distribution µt at time t
Convergence of (µt, πt) to (µ+, π): a.s. convergence of stochastic
approximation algorithms for discrete-time processes on
compact spaces
[Benaı̈m et al., 2018]
Adapting Restore
Initial regeneration distribution µ0 and updates by addition of
point masses
µt(x) =

µ0(x), if N(t) = 0,
t
a+t
1
N(t)
PN(t)
i=1 δXζi
(x) + a
a+t µ0(x), if N(t)  0,
where a  0 constant2 and ζi arrival times of inhomogeneous
Poisson process (N(t) : t ≥ 0) with rate κ−(Xt)
κ−
(x) := [0 ∨ −κ̃(x)]
Poisson process simulated by Poisson thinning, under
assumption
K−
:= sup
x∈X
κ−
(x)  0
2
Discrete measure dominance time: when regeneration most likely to be
from discrete measure in mixture distribution
Adaptive Brownian Restore Algorithm
t ← 0, E ← ∅, i ← 0, X ∼ µ0.
while i  n do
τ̃ ∼ Exp(K+
), s ∼ Exp(Λ0), ζ̃ ∼ Exp(K−
).
if τ̃  s and τ̃  ζ̃ then
X ∼ N(X, τ̃), t ← t + τ̃, u ∼ U[0, 1].
if u  κ+
(X)/K+
then
if |E| = 0 then
X ∼ µ0.
else
X ∼ U(E) with probability t/(a + t), else X ∼ µ0.
end
i ← i + 1.
end
else if s  τ̃ and s  ζ̃ then
X ∼ N(X, s), t ← t + s, record X, t, i.
else
X ∼ N(X, ζ̃), t ← t + ζ̃, u ∼ U[0, 1].
If u  κ−
(X)/K−
then E ← E ∪ {X}.
end
end
Adaptive Brownian Restore Algorithm
0 10 20 30 40 50
−2
0
2
4
x
t
Path of an Adaptive Brownian Restore process with
π ≡ N(0, 1), µ0 ≡ N(2, 1), a = 10. Green dots and red crosses
show the first and last output states of each tour.
Calibrating initial regeneration and parameters
I µ0 as initial approximate π, e.g., µ0 ≡ Nd(0, I) with π
pre-transformed
I Tradeoff in choosing a discrete measure dominance time:
smaller choices of a lead to faster convergence while larger
value of a produces more regenerations from µ0, hence
better exploration (range of a between 1,000 and 10,000)
I K+ and K− based on quantiles of κ̃, from preliminary
MCMC run
I or, assuming π close to Gaussian, initial guess of K− is d/2
and initial estimate for K+ based on chi-squared
approximation
Final remarks
I Adaptive Restore benefits from global moves, for targets
hard to approximate with a parametric distribution, with
large moves across the space
I Use of minimal regeneration rate makes simulation
computationally feasible and more likely in areas where π
has significant mass
I In comparison to e.g. Random Walk Metropolis, process
can be slow but higher efficiency in sampling distributions
with skewed tails

More Related Content

Similar to restore.pdf

002 ray modeling dynamic systems
002 ray modeling dynamic systems002 ray modeling dynamic systems
002 ray modeling dynamic systems
Institute of Technology Telkom
 
Sampling strategies for Sequential Monte Carlo (SMC) methods
Sampling strategies for Sequential Monte Carlo (SMC) methodsSampling strategies for Sequential Monte Carlo (SMC) methods
Sampling strategies for Sequential Monte Carlo (SMC) methodsStephane Senecal
 
International journal of engineering and mathematical modelling vol2 no1_2015_1
International journal of engineering and mathematical modelling vol2 no1_2015_1International journal of engineering and mathematical modelling vol2 no1_2015_1
International journal of engineering and mathematical modelling vol2 no1_2015_1
IJEMM
 
Mit2 092 f09_lec20
Mit2 092 f09_lec20Mit2 092 f09_lec20
Mit2 092 f09_lec20
Rahman Hakim
 
Statistics Homework Help
Statistics Homework HelpStatistics Homework Help
Statistics Homework Help
Statistics Homework Helper
 
Multiple Linear Regression Homework Help
Multiple Linear Regression Homework HelpMultiple Linear Regression Homework Help
Multiple Linear Regression Homework Help
Statistics Homework Helper
 
Program on Quasi-Monte Carlo and High-Dimensional Sampling Methods for Applie...
Program on Quasi-Monte Carlo and High-Dimensional Sampling Methods for Applie...Program on Quasi-Monte Carlo and High-Dimensional Sampling Methods for Applie...
Program on Quasi-Monte Carlo and High-Dimensional Sampling Methods for Applie...
The Statistical and Applied Mathematical Sciences Institute
 
Btech admission in india
Btech admission in indiaBtech admission in india
Btech admission in india
Edhole.com
 
Iterative procedure for uniform continuous mapping.
Iterative procedure for uniform continuous mapping.Iterative procedure for uniform continuous mapping.
Iterative procedure for uniform continuous mapping.
Alexander Decker
 
Litvinenko low-rank kriging +FFT poster
Litvinenko low-rank kriging +FFT  posterLitvinenko low-rank kriging +FFT  poster
Litvinenko low-rank kriging +FFT poster
Alexander Litvinenko
 
ch9.pdf
ch9.pdfch9.pdf
ch9.pdf
KavS14
 
Chm3410hwk02 soln.252145237
Chm3410hwk02 soln.252145237Chm3410hwk02 soln.252145237
Chm3410hwk02 soln.252145237
Samandar Yuldoshev
 
somenath_fixedpoint_dasguptaIMF17-20-2013
somenath_fixedpoint_dasguptaIMF17-20-2013somenath_fixedpoint_dasguptaIMF17-20-2013
somenath_fixedpoint_dasguptaIMF17-20-2013Somenath Bandyopadhyay
 
3.pdf
3.pdf3.pdf
3.pdf
AlaaOdeh18
 
SIAM - Minisymposium on Guaranteed numerical algorithms
SIAM - Minisymposium on Guaranteed numerical algorithmsSIAM - Minisymposium on Guaranteed numerical algorithms
SIAM - Minisymposium on Guaranteed numerical algorithms
Jagadeeswaran Rathinavel
 
Computing f-Divergences and Distances of\\ High-Dimensional Probability Densi...
Computing f-Divergences and Distances of\\ High-Dimensional Probability Densi...Computing f-Divergences and Distances of\\ High-Dimensional Probability Densi...
Computing f-Divergences and Distances of\\ High-Dimensional Probability Densi...
Alexander Litvinenko
 
QMC: Operator Splitting Workshop, Compactness Estimates for Nonlinear PDEs - ...
QMC: Operator Splitting Workshop, Compactness Estimates for Nonlinear PDEs - ...QMC: Operator Splitting Workshop, Compactness Estimates for Nonlinear PDEs - ...
QMC: Operator Splitting Workshop, Compactness Estimates for Nonlinear PDEs - ...
The Statistical and Applied Mathematical Sciences Institute
 
Program on Quasi-Monte Carlo and High-Dimensional Sampling Methods for Applie...
Program on Quasi-Monte Carlo and High-Dimensional Sampling Methods for Applie...Program on Quasi-Monte Carlo and High-Dimensional Sampling Methods for Applie...
Program on Quasi-Monte Carlo and High-Dimensional Sampling Methods for Applie...
The Statistical and Applied Mathematical Sciences Institute
 

Similar to restore.pdf (20)

002 ray modeling dynamic systems
002 ray modeling dynamic systems002 ray modeling dynamic systems
002 ray modeling dynamic systems
 
Sampling strategies for Sequential Monte Carlo (SMC) methods
Sampling strategies for Sequential Monte Carlo (SMC) methodsSampling strategies for Sequential Monte Carlo (SMC) methods
Sampling strategies for Sequential Monte Carlo (SMC) methods
 
International journal of engineering and mathematical modelling vol2 no1_2015_1
International journal of engineering and mathematical modelling vol2 no1_2015_1International journal of engineering and mathematical modelling vol2 no1_2015_1
International journal of engineering and mathematical modelling vol2 no1_2015_1
 
Mit2 092 f09_lec20
Mit2 092 f09_lec20Mit2 092 f09_lec20
Mit2 092 f09_lec20
 
Statistics Homework Help
Statistics Homework HelpStatistics Homework Help
Statistics Homework Help
 
Multiple Linear Regression Homework Help
Multiple Linear Regression Homework HelpMultiple Linear Regression Homework Help
Multiple Linear Regression Homework Help
 
Program on Quasi-Monte Carlo and High-Dimensional Sampling Methods for Applie...
Program on Quasi-Monte Carlo and High-Dimensional Sampling Methods for Applie...Program on Quasi-Monte Carlo and High-Dimensional Sampling Methods for Applie...
Program on Quasi-Monte Carlo and High-Dimensional Sampling Methods for Applie...
 
Btech admission in india
Btech admission in indiaBtech admission in india
Btech admission in india
 
Iterative procedure for uniform continuous mapping.
Iterative procedure for uniform continuous mapping.Iterative procedure for uniform continuous mapping.
Iterative procedure for uniform continuous mapping.
 
Ch06 3
Ch06 3Ch06 3
Ch06 3
 
Litvinenko low-rank kriging +FFT poster
Litvinenko low-rank kriging +FFT  posterLitvinenko low-rank kriging +FFT  poster
Litvinenko low-rank kriging +FFT poster
 
ch9.pdf
ch9.pdfch9.pdf
ch9.pdf
 
Chm3410hwk02 soln.252145237
Chm3410hwk02 soln.252145237Chm3410hwk02 soln.252145237
Chm3410hwk02 soln.252145237
 
somenath_fixedpoint_dasguptaIMF17-20-2013
somenath_fixedpoint_dasguptaIMF17-20-2013somenath_fixedpoint_dasguptaIMF17-20-2013
somenath_fixedpoint_dasguptaIMF17-20-2013
 
maa_talk
maa_talkmaa_talk
maa_talk
 
3.pdf
3.pdf3.pdf
3.pdf
 
SIAM - Minisymposium on Guaranteed numerical algorithms
SIAM - Minisymposium on Guaranteed numerical algorithmsSIAM - Minisymposium on Guaranteed numerical algorithms
SIAM - Minisymposium on Guaranteed numerical algorithms
 
Computing f-Divergences and Distances of\\ High-Dimensional Probability Densi...
Computing f-Divergences and Distances of\\ High-Dimensional Probability Densi...Computing f-Divergences and Distances of\\ High-Dimensional Probability Densi...
Computing f-Divergences and Distances of\\ High-Dimensional Probability Densi...
 
QMC: Operator Splitting Workshop, Compactness Estimates for Nonlinear PDEs - ...
QMC: Operator Splitting Workshop, Compactness Estimates for Nonlinear PDEs - ...QMC: Operator Splitting Workshop, Compactness Estimates for Nonlinear PDEs - ...
QMC: Operator Splitting Workshop, Compactness Estimates for Nonlinear PDEs - ...
 
Program on Quasi-Monte Carlo and High-Dimensional Sampling Methods for Applie...
Program on Quasi-Monte Carlo and High-Dimensional Sampling Methods for Applie...Program on Quasi-Monte Carlo and High-Dimensional Sampling Methods for Applie...
Program on Quasi-Monte Carlo and High-Dimensional Sampling Methods for Applie...
 

More from Christian Robert

Asymptotics of ABC, lecture, Collège de France
Asymptotics of ABC, lecture, Collège de FranceAsymptotics of ABC, lecture, Collège de France
Asymptotics of ABC, lecture, Collège de France
Christian Robert
 
Workshop in honour of Don Poskitt and Gael Martin
Workshop in honour of Don Poskitt and Gael MartinWorkshop in honour of Don Poskitt and Gael Martin
Workshop in honour of Don Poskitt and Gael Martin
Christian Robert
 
discussion of ICML23.pdf
discussion of ICML23.pdfdiscussion of ICML23.pdf
discussion of ICML23.pdf
Christian Robert
 
How many components in a mixture?
How many components in a mixture?How many components in a mixture?
How many components in a mixture?
Christian Robert
 
Testing for mixtures at BNP 13
Testing for mixtures at BNP 13Testing for mixtures at BNP 13
Testing for mixtures at BNP 13
Christian Robert
 
Inferring the number of components: dream or reality?
Inferring the number of components: dream or reality?Inferring the number of components: dream or reality?
Inferring the number of components: dream or reality?
Christian Robert
 
CDT 22 slides.pdf
CDT 22 slides.pdfCDT 22 slides.pdf
CDT 22 slides.pdf
Christian Robert
 
Testing for mixtures by seeking components
Testing for mixtures by seeking componentsTesting for mixtures by seeking components
Testing for mixtures by seeking components
Christian Robert
 
discussion on Bayesian restricted likelihood
discussion on Bayesian restricted likelihooddiscussion on Bayesian restricted likelihood
discussion on Bayesian restricted likelihood
Christian Robert
 
NCE, GANs & VAEs (and maybe BAC)
NCE, GANs & VAEs (and maybe BAC)NCE, GANs & VAEs (and maybe BAC)
NCE, GANs & VAEs (and maybe BAC)
Christian Robert
 
ABC-Gibbs
ABC-GibbsABC-Gibbs
ABC-Gibbs
Christian Robert
 
Coordinate sampler : A non-reversible Gibbs-like sampler
Coordinate sampler : A non-reversible Gibbs-like samplerCoordinate sampler : A non-reversible Gibbs-like sampler
Coordinate sampler : A non-reversible Gibbs-like sampler
Christian Robert
 
eugenics and statistics
eugenics and statisticseugenics and statistics
eugenics and statistics
Christian Robert
 
Laplace's Demon: seminar #1
Laplace's Demon: seminar #1Laplace's Demon: seminar #1
Laplace's Demon: seminar #1
Christian Robert
 
ABC-Gibbs
ABC-GibbsABC-Gibbs
ABC-Gibbs
Christian Robert
 
asymptotics of ABC
asymptotics of ABCasymptotics of ABC
asymptotics of ABC
Christian Robert
 
ABC-Gibbs
ABC-GibbsABC-Gibbs
ABC-Gibbs
Christian Robert
 
Likelihood-free Design: a discussion
Likelihood-free Design: a discussionLikelihood-free Design: a discussion
Likelihood-free Design: a discussion
Christian Robert
 
the ABC of ABC
the ABC of ABCthe ABC of ABC
the ABC of ABC
Christian Robert
 
CISEA 2019: ABC consistency and convergence
CISEA 2019: ABC consistency and convergenceCISEA 2019: ABC consistency and convergence
CISEA 2019: ABC consistency and convergence
Christian Robert
 

More from Christian Robert (20)

Asymptotics of ABC, lecture, Collège de France
Asymptotics of ABC, lecture, Collège de FranceAsymptotics of ABC, lecture, Collège de France
Asymptotics of ABC, lecture, Collège de France
 
Workshop in honour of Don Poskitt and Gael Martin
Workshop in honour of Don Poskitt and Gael MartinWorkshop in honour of Don Poskitt and Gael Martin
Workshop in honour of Don Poskitt and Gael Martin
 
discussion of ICML23.pdf
discussion of ICML23.pdfdiscussion of ICML23.pdf
discussion of ICML23.pdf
 
How many components in a mixture?
How many components in a mixture?How many components in a mixture?
How many components in a mixture?
 
Testing for mixtures at BNP 13
Testing for mixtures at BNP 13Testing for mixtures at BNP 13
Testing for mixtures at BNP 13
 
Inferring the number of components: dream or reality?
Inferring the number of components: dream or reality?Inferring the number of components: dream or reality?
Inferring the number of components: dream or reality?
 
CDT 22 slides.pdf
CDT 22 slides.pdfCDT 22 slides.pdf
CDT 22 slides.pdf
 
Testing for mixtures by seeking components
Testing for mixtures by seeking componentsTesting for mixtures by seeking components
Testing for mixtures by seeking components
 
discussion on Bayesian restricted likelihood
discussion on Bayesian restricted likelihooddiscussion on Bayesian restricted likelihood
discussion on Bayesian restricted likelihood
 
NCE, GANs & VAEs (and maybe BAC)
NCE, GANs & VAEs (and maybe BAC)NCE, GANs & VAEs (and maybe BAC)
NCE, GANs & VAEs (and maybe BAC)
 
ABC-Gibbs
ABC-GibbsABC-Gibbs
ABC-Gibbs
 
Coordinate sampler : A non-reversible Gibbs-like sampler
Coordinate sampler : A non-reversible Gibbs-like samplerCoordinate sampler : A non-reversible Gibbs-like sampler
Coordinate sampler : A non-reversible Gibbs-like sampler
 
eugenics and statistics
eugenics and statisticseugenics and statistics
eugenics and statistics
 
Laplace's Demon: seminar #1
Laplace's Demon: seminar #1Laplace's Demon: seminar #1
Laplace's Demon: seminar #1
 
ABC-Gibbs
ABC-GibbsABC-Gibbs
ABC-Gibbs
 
asymptotics of ABC
asymptotics of ABCasymptotics of ABC
asymptotics of ABC
 
ABC-Gibbs
ABC-GibbsABC-Gibbs
ABC-Gibbs
 
Likelihood-free Design: a discussion
Likelihood-free Design: a discussionLikelihood-free Design: a discussion
Likelihood-free Design: a discussion
 
the ABC of ABC
the ABC of ABCthe ABC of ABC
the ABC of ABC
 
CISEA 2019: ABC consistency and convergence
CISEA 2019: ABC consistency and convergenceCISEA 2019: ABC consistency and convergence
CISEA 2019: ABC consistency and convergence
 

Recently uploaded

Nutraceutical market, scope and growth: Herbal drug technology
Nutraceutical market, scope and growth: Herbal drug technologyNutraceutical market, scope and growth: Herbal drug technology
Nutraceutical market, scope and growth: Herbal drug technology
Lokesh Patil
 
justice-and-fairness-ethics with example
justice-and-fairness-ethics with examplejustice-and-fairness-ethics with example
justice-and-fairness-ethics with example
azzyixes
 
GBSN- Microbiology (Lab 3) Gram Staining
GBSN- Microbiology (Lab 3) Gram StainingGBSN- Microbiology (Lab 3) Gram Staining
GBSN- Microbiology (Lab 3) Gram Staining
Areesha Ahmad
 
plant biotechnology Lecture note ppt.pptx
plant biotechnology Lecture note ppt.pptxplant biotechnology Lecture note ppt.pptx
plant biotechnology Lecture note ppt.pptx
yusufzako14
 
ESR_factors_affect-clinic significance-Pathysiology.pptx
ESR_factors_affect-clinic significance-Pathysiology.pptxESR_factors_affect-clinic significance-Pathysiology.pptx
ESR_factors_affect-clinic significance-Pathysiology.pptx
muralinath2
 
insect morphology and physiology of insect
insect morphology and physiology of insectinsect morphology and physiology of insect
insect morphology and physiology of insect
anitaento25
 
Unveiling the Energy Potential of Marshmallow Deposits.pdf
Unveiling the Energy Potential of Marshmallow Deposits.pdfUnveiling the Energy Potential of Marshmallow Deposits.pdf
Unveiling the Energy Potential of Marshmallow Deposits.pdf
Erdal Coalmaker
 
Structures and textures of metamorphic rocks
Structures and textures of metamorphic rocksStructures and textures of metamorphic rocks
Structures and textures of metamorphic rocks
kumarmathi863
 
Comparative structure of adrenal gland in vertebrates
Comparative structure of adrenal gland in vertebratesComparative structure of adrenal gland in vertebrates
Comparative structure of adrenal gland in vertebrates
sachin783648
 
RNA INTERFERENCE: UNRAVELING GENETIC SILENCING
RNA INTERFERENCE: UNRAVELING GENETIC SILENCINGRNA INTERFERENCE: UNRAVELING GENETIC SILENCING
RNA INTERFERENCE: UNRAVELING GENETIC SILENCING
AADYARAJPANDEY1
 
erythropoiesis-I_mechanism& clinical significance.pptx
erythropoiesis-I_mechanism& clinical significance.pptxerythropoiesis-I_mechanism& clinical significance.pptx
erythropoiesis-I_mechanism& clinical significance.pptx
muralinath2
 
Earliest Galaxies in the JADES Origins Field: Luminosity Function and Cosmic ...
Earliest Galaxies in the JADES Origins Field: Luminosity Function and Cosmic ...Earliest Galaxies in the JADES Origins Field: Luminosity Function and Cosmic ...
Earliest Galaxies in the JADES Origins Field: Luminosity Function and Cosmic ...
Sérgio Sacani
 
Astronomy Update- Curiosity’s exploration of Mars _ Local Briefs _ leadertele...
Astronomy Update- Curiosity’s exploration of Mars _ Local Briefs _ leadertele...Astronomy Update- Curiosity’s exploration of Mars _ Local Briefs _ leadertele...
Astronomy Update- Curiosity’s exploration of Mars _ Local Briefs _ leadertele...
NathanBaughman3
 
THE IMPORTANCE OF MARTIAN ATMOSPHERE SAMPLE RETURN.
THE IMPORTANCE OF MARTIAN ATMOSPHERE SAMPLE RETURN.THE IMPORTANCE OF MARTIAN ATMOSPHERE SAMPLE RETURN.
THE IMPORTANCE OF MARTIAN ATMOSPHERE SAMPLE RETURN.
Sérgio Sacani
 
Penicillin...........................pptx
Penicillin...........................pptxPenicillin...........................pptx
Penicillin...........................pptx
Cherry
 
platelets_clotting_biogenesis.clot retractionpptx
platelets_clotting_biogenesis.clot retractionpptxplatelets_clotting_biogenesis.clot retractionpptx
platelets_clotting_biogenesis.clot retractionpptx
muralinath2
 
Orion Air Quality Monitoring Systems - CWS
Orion Air Quality Monitoring Systems - CWSOrion Air Quality Monitoring Systems - CWS
Orion Air Quality Monitoring Systems - CWS
Columbia Weather Systems
 
Multi-source connectivity as the driver of solar wind variability in the heli...
Multi-source connectivity as the driver of solar wind variability in the heli...Multi-source connectivity as the driver of solar wind variability in the heli...
Multi-source connectivity as the driver of solar wind variability in the heli...
Sérgio Sacani
 
Viksit bharat till 2047 India@2047.pptx
Viksit bharat till 2047  India@2047.pptxViksit bharat till 2047  India@2047.pptx
Viksit bharat till 2047 India@2047.pptx
rakeshsharma20142015
 
NuGOweek 2024 Ghent - programme - final version
NuGOweek 2024 Ghent - programme - final versionNuGOweek 2024 Ghent - programme - final version
NuGOweek 2024 Ghent - programme - final version
pablovgd
 

Recently uploaded (20)

Nutraceutical market, scope and growth: Herbal drug technology
Nutraceutical market, scope and growth: Herbal drug technologyNutraceutical market, scope and growth: Herbal drug technology
Nutraceutical market, scope and growth: Herbal drug technology
 
justice-and-fairness-ethics with example
justice-and-fairness-ethics with examplejustice-and-fairness-ethics with example
justice-and-fairness-ethics with example
 
GBSN- Microbiology (Lab 3) Gram Staining
GBSN- Microbiology (Lab 3) Gram StainingGBSN- Microbiology (Lab 3) Gram Staining
GBSN- Microbiology (Lab 3) Gram Staining
 
plant biotechnology Lecture note ppt.pptx
plant biotechnology Lecture note ppt.pptxplant biotechnology Lecture note ppt.pptx
plant biotechnology Lecture note ppt.pptx
 
ESR_factors_affect-clinic significance-Pathysiology.pptx
ESR_factors_affect-clinic significance-Pathysiology.pptxESR_factors_affect-clinic significance-Pathysiology.pptx
ESR_factors_affect-clinic significance-Pathysiology.pptx
 
insect morphology and physiology of insect
insect morphology and physiology of insectinsect morphology and physiology of insect
insect morphology and physiology of insect
 
Unveiling the Energy Potential of Marshmallow Deposits.pdf
Unveiling the Energy Potential of Marshmallow Deposits.pdfUnveiling the Energy Potential of Marshmallow Deposits.pdf
Unveiling the Energy Potential of Marshmallow Deposits.pdf
 
Structures and textures of metamorphic rocks
Structures and textures of metamorphic rocksStructures and textures of metamorphic rocks
Structures and textures of metamorphic rocks
 
Comparative structure of adrenal gland in vertebrates
Comparative structure of adrenal gland in vertebratesComparative structure of adrenal gland in vertebrates
Comparative structure of adrenal gland in vertebrates
 
RNA INTERFERENCE: UNRAVELING GENETIC SILENCING
RNA INTERFERENCE: UNRAVELING GENETIC SILENCINGRNA INTERFERENCE: UNRAVELING GENETIC SILENCING
RNA INTERFERENCE: UNRAVELING GENETIC SILENCING
 
erythropoiesis-I_mechanism& clinical significance.pptx
erythropoiesis-I_mechanism& clinical significance.pptxerythropoiesis-I_mechanism& clinical significance.pptx
erythropoiesis-I_mechanism& clinical significance.pptx
 
Earliest Galaxies in the JADES Origins Field: Luminosity Function and Cosmic ...
Earliest Galaxies in the JADES Origins Field: Luminosity Function and Cosmic ...Earliest Galaxies in the JADES Origins Field: Luminosity Function and Cosmic ...
Earliest Galaxies in the JADES Origins Field: Luminosity Function and Cosmic ...
 
Astronomy Update- Curiosity’s exploration of Mars _ Local Briefs _ leadertele...
Astronomy Update- Curiosity’s exploration of Mars _ Local Briefs _ leadertele...Astronomy Update- Curiosity’s exploration of Mars _ Local Briefs _ leadertele...
Astronomy Update- Curiosity’s exploration of Mars _ Local Briefs _ leadertele...
 
THE IMPORTANCE OF MARTIAN ATMOSPHERE SAMPLE RETURN.
THE IMPORTANCE OF MARTIAN ATMOSPHERE SAMPLE RETURN.THE IMPORTANCE OF MARTIAN ATMOSPHERE SAMPLE RETURN.
THE IMPORTANCE OF MARTIAN ATMOSPHERE SAMPLE RETURN.
 
Penicillin...........................pptx
Penicillin...........................pptxPenicillin...........................pptx
Penicillin...........................pptx
 
platelets_clotting_biogenesis.clot retractionpptx
platelets_clotting_biogenesis.clot retractionpptxplatelets_clotting_biogenesis.clot retractionpptx
platelets_clotting_biogenesis.clot retractionpptx
 
Orion Air Quality Monitoring Systems - CWS
Orion Air Quality Monitoring Systems - CWSOrion Air Quality Monitoring Systems - CWS
Orion Air Quality Monitoring Systems - CWS
 
Multi-source connectivity as the driver of solar wind variability in the heli...
Multi-source connectivity as the driver of solar wind variability in the heli...Multi-source connectivity as the driver of solar wind variability in the heli...
Multi-source connectivity as the driver of solar wind variability in the heli...
 
Viksit bharat till 2047 India@2047.pptx
Viksit bharat till 2047  India@2047.pptxViksit bharat till 2047  India@2047.pptx
Viksit bharat till 2047 India@2047.pptx
 
NuGOweek 2024 Ghent - programme - final version
NuGOweek 2024 Ghent - programme - final versionNuGOweek 2024 Ghent - programme - final version
NuGOweek 2024 Ghent - programme - final version
 

restore.pdf

  • 1. Adaptive Restore algorithm Christian P. Robert U. Paris Dauphine & Warwick U. 0 10 20 30 40 50 −2 0 2 4 x t Joint work with H Krimm, M Pollock, AQ Wang and GO Roberts arXiv:2210.09901
  • 2. EaRly Call Incoming 2023-2030 ERC funding for postdoctoral collaborations with I Michael Jordan (Berkeley) I Eric Moulines (Paris) I Gareth Roberts (Warwick) I myself (Paris)
  • 3. MCMC regeneration Reversibility attached with (sub-efficient) random walk behaviour Recent achievements in non-reversible MCMC with PDMPs [Bouchard-Côté et al., 2018; Bierkens et al., 2019] Linked with regeneration [Nummelin, 1978; Mykland et al., X, 1995]
  • 4. Restore process Take {Yt}t≥0 diffusion / jump process on Rd with infinitesimal generator LY and Y0 ∼ µ —e..g., LY = ∆/2 for Brownian— Regeneration rate κ with associated tour length τ = inf t ≥ 0 : Zt 0 κ(Ys)ds ≥ ξ ξ ∼ Exp(1) Take {Y (i) t }t≥0, τ(i) ∞ i=0 iid realisations Regeneration times Tj = Pj−1 i=0 τ(i) Restore process {Xt}t≥0 given by: Xt = ∞ X i=0 I[Ti,Ti+1)(t)Y (i) t−Ti [Wang et al., 2021]
  • 5. Restore process 0 2 4 6 8 −2 −1 0 1 2 x t Path of 5 tours of Brownian Restore with π ≡ N(0, 12), µ ≡ N(0, 22) and C chosen so that minx∈R κ(x) = 0, with K = 200 and Λ0 = 1000. First and last output states shown by green dots and red crosses
  • 6. Stationarity of Restore Infinitesimal generator of {Xt}t≥0 LXf(x) = LYf(x) + κ(x) Z [f(y) − f(x)]µ(y)dy Regeneration rate κ chosen as1 κ(x) = L† Yπ(x) π(x) + C µ(x) π(x) makes {Xt}t≥0 π-invariant Z Rd LXf(x)π(x)dx = 0 1 L† Y formal adjoint
  • 7. Restore sampler Rewrite κ(x) = L† Yπ(x) π(x) + C µ(x) π(x) = κ̃(x) + C µ(x) π(x) with I κ̃ partial regeneration rate I C 0 regeneration constant and I Cµ regeneration measure, large enough so that κ(·) 0 Resulting Monte Carlo method called Restore Sampler [Wang et al., 2021]
  • 8. Restore sampler Given π-invariance of {Xt}t≥0 Eπ[f] = EX0∼µ h Zτ(0) 0 f(Xs)dx i. EX0∼µ[τ(0) ] and a.s. convergence of ergodic averages: 1 t Zt 0 f(Xs)ds → Eπ[f] For iid Zi := RTi+1 Ti f(Xs)ds, CLT √ n ZTn 0 f(Xs)dx Tn − Eπ[f] ! → N(0, σ2 f) where σ2 f := EX0∼µ Z0 − τ(0) Eπ[f] 2 . EX0∼µ[τ(0) ] 2
  • 9. Restore sampler Given π-invariance of {Xt}t≥0 Eπ[f] = EX0∼µ h Zτ(0) 0 f(Xs)dx i. EX0∼µ[τ(0) ] and a.s. convergence of ergodic averages: 1 t Zt 0 f(Xs)ds → Eπ[f] Estimator variance depends on expected tour length: choose µ towards long tours [Wang et al., 2020]
  • 10. Minimal regeneration Minimal regeneration measure, C+µ+ corresponding to smallest possible rate κ+ (x) := κ̃(x) ∨ 0 = κ̃(x) + C+ µ+(x) π(x) leading to µ+ (x) = 1 C+ [0 ∨ −κ̃(x)]π(x) Frequent regeneration not necessarily detrimental, except when when µ is not well-aligned to π, leading to wasted computation Minimal Restore maximizes expected tour length / minimizes asymptotic variance [Wang et al., 2021]
  • 11. Constant approximation When target misses normalizing constant π(x) = π̃(x) Z, take energy U(x) := − log π(x) = log Z − log π̃(x) When {Yt}t≥0 Brownian, κ̃ function of ∇U(x) and ∆U(x)
  • 12. Constant approximation When target misses normalizing constant π(x) = π̃(x) Z, take energy U(x) := − log π(x) = log Z − log π̃(x) When {Yt}t≥0 Brownian, κ̃ function of ∇U(x) and ∆U(x) In regeneration rate, Z “absorbed” into C κ(x) = κ̃(x) + C µ(x) π̃(x) Z = κ̃(x) + CZ µ(x) π̃(x) = κ̃(x) + C̃ µ(x) π̃(x) , where C̃ = CZ set by user. Since C = 1/Eµ[τ] Using n tours with simulation time T, Z ≈ C̃T n.
  • 13. Adapting Restore Adaptive Restore process defined by enriching underlying continuous-time Markov process with regenerations at rate κ+ from distribution µt at time t Convergence of (µt, πt) to (µ+, π): a.s. convergence of stochastic approximation algorithms for discrete-time processes on compact spaces [Benaı̈m et al., 2018]
  • 14. Adapting Restore Initial regeneration distribution µ0 and updates by addition of point masses µt(x) = µ0(x), if N(t) = 0, t a+t 1 N(t) PN(t) i=1 δXζi (x) + a a+t µ0(x), if N(t) 0, where a 0 constant2 and ζi arrival times of inhomogeneous Poisson process (N(t) : t ≥ 0) with rate κ−(Xt) κ− (x) := [0 ∨ −κ̃(x)] Poisson process simulated by Poisson thinning, under assumption K− := sup x∈X κ− (x) 0 2 Discrete measure dominance time: when regeneration most likely to be from discrete measure in mixture distribution
  • 15. Adaptive Brownian Restore Algorithm t ← 0, E ← ∅, i ← 0, X ∼ µ0. while i n do τ̃ ∼ Exp(K+ ), s ∼ Exp(Λ0), ζ̃ ∼ Exp(K− ). if τ̃ s and τ̃ ζ̃ then X ∼ N(X, τ̃), t ← t + τ̃, u ∼ U[0, 1]. if u κ+ (X)/K+ then if |E| = 0 then X ∼ µ0. else X ∼ U(E) with probability t/(a + t), else X ∼ µ0. end i ← i + 1. end else if s τ̃ and s ζ̃ then X ∼ N(X, s), t ← t + s, record X, t, i. else X ∼ N(X, ζ̃), t ← t + ζ̃, u ∼ U[0, 1]. If u κ− (X)/K− then E ← E ∪ {X}. end end
  • 16. Adaptive Brownian Restore Algorithm 0 10 20 30 40 50 −2 0 2 4 x t Path of an Adaptive Brownian Restore process with π ≡ N(0, 1), µ0 ≡ N(2, 1), a = 10. Green dots and red crosses show the first and last output states of each tour.
  • 17. Calibrating initial regeneration and parameters I µ0 as initial approximate π, e.g., µ0 ≡ Nd(0, I) with π pre-transformed I Tradeoff in choosing a discrete measure dominance time: smaller choices of a lead to faster convergence while larger value of a produces more regenerations from µ0, hence better exploration (range of a between 1,000 and 10,000) I K+ and K− based on quantiles of κ̃, from preliminary MCMC run I or, assuming π close to Gaussian, initial guess of K− is d/2 and initial estimate for K+ based on chi-squared approximation
  • 18. Final remarks I Adaptive Restore benefits from global moves, for targets hard to approximate with a parametric distribution, with large moves across the space I Use of minimal regeneration rate makes simulation computationally feasible and more likely in areas where π has significant mass I In comparison to e.g. Random Walk Metropolis, process can be slow but higher efficiency in sampling distributions with skewed tails