asymptotics of ABC

Asymptotics of ABC
Christian P. Robert
U. Paris Dauphine PSL & Warwick U. & CREST
ENSAE
KSS Meeting, Seoul, 9 Nov. 2019

Co-authors
Jean-Marie Cornuet, Arnaud Estoup (INRA
Montpellier)
David Frazier, Gael Martin (Monash U)
Jean-Michel Marin (U Montpellier)
Kerrie Mengersen (QUT)
Espen Bernton, Pierre Jacob, Natesh Pillai
(Harvard U)
Judith Rousseau (U Oxford)
Gr´egoire Clart´e, Robin Ryder, Julien Stoehr (U
Paris Dauphine)

Outline
1 Motivation
2 Concept
3 Implementation
4 Convergence
5 Advanced topics

Posterior distribution
Central concept of Bayesian inference:
π( θ
parameter
| xobs
observation
)
proportional
∝ π(θ)
prior
× L(θ | xobs
)
likelihood
Drives
derivation of optimal decisions
assessment of uncertainty
model selection
prediction
[McElreath, 2015]

Monte Carlo representation
Exploration of Bayesian posterior π(θ|xobs) may (!) require to
produce sample
θ1, . . . , θT
distributed from π(θ|xobs) (or asymptotically by Markov chain
Monte Carlo aka MCMC)
[McElreath, 2015]

Diﬃculties
MCMC = workhorse of practical Bayesian analysis (BUGS,
JAGS, Stan, &tc.), except when product
π(θ) × L(θ | xobs
)
well-deﬁned but numerically unavailable or too costly to
compute
Only partial solutions are available:
demarginalisation (latent variables)
exchange algorithm (auxiliary variables)
pseudo-marginal (unbiased estimator)

Example 1: Dynamic mixture
Mixture model
{1 − wµ,τ(x)}fβ,λ(x) + wµ,τ(x)gε,σ(x) x > 0
where
fβ,λ Weibull density,
gε,σ generalised Pareto density, and
wµ,τ Cauchy (arctan) cdf
Intractable normalising constant
C(µ, τ, β, λ, ε, σ) =
∞
0
{(1 − wµ,τ(x))fβ,λ(x) + wµ,τ(x)gε,σ(x)} dx
[Frigessi, Haug & Rue, 2002]

Example 2: truncated Normal
Given set A ⊂ Rk (k large), truncated Normal model
f(x | µ, Σ, A) ∝ exp{−(x − µ)T
Σ−1
(x − µ)/2} IA(x)
with intractable normalising constant
C(µ, Σ, A) =
A
exp{−(x − µ)T
Σ−1
(x − µ)/2}dx

Example 3: robust Normal statistics
Normal sample
x1, . . . , xn ∼ N(µ, σ2
)
summarised into (insuﬃcient)
^µn = med(x1, . . . , xn)
and
^σn = mad(x1, . . . , xn)
= med |xi − ^µn|
Under a conjugate prior π(µ, σ2), posterior close to intractable.
but simulation of (^µn, ^σn) straightforward

Example 4: exponential random graph
ERGM: binary random vector x indexed
by all edges on set of nodes plus graph
f(x | θ) =
1
C(θ)
exp(θT
S(x))
with S(x) vector of statistics and C(θ)
intractable normalising constant
[Grelaud & al., 2009; Everitt, 2012; Bouranis & al., 2017]

Realistic[er] applications
Kingman’s coalescent in population genetics
[Tavar´e et al., 1997; Beaumont et al., 2003]
α-stable distributions
[Peters et al, 2012]
complex networks
[Dutta et al., 2018]
astrostatistics & cosmostatistics
[Cameron & Pettitt, 2012; Ishida et al., 2015]

Concept
2 Concept
algorithm
summary statistics
random forests
calibration of tolerance
3 Implementation
4 Convergence
5 Advanced topics

A?B?C?
A stands for approximate [wrong
likelihood]
B stands for Bayesian [right prior]
C stands for computation [producing
a parameter sample]

A?B?C?
Rough version of the data [from dot
to ball]
Non-parametric approximation of
the likelihood [near actual
observation]
Use of non-suﬃcient statistics
[dimension reduction]
Monte Carlo error [and no
unbiasedness]

A seemingly na¨ıve representation
When likelihood f(x|θ) not in closed form, likelihood-free
rejection technique:
ABC-AR algorithm
For an observation xobs ∼ f(x|θ), under the prior π(θ), keep
jointly simulating
θ ∼ π(θ) , z ∼ f(z|θ ) ,
until the auxiliary variable z is equal to the observed value,
z = xobs
[Diggle & Gratton, 1984; Rubin, 1984; Tavar´e et al., 1997]

Why does it work?
The mathematical proof is trivial:
f(θi) ∝
z∈D
π(θi)f(z|θi)Iy(z)
∝ π(θi)f(y|θi)
= π(θi|y)
[Accept–Reject 101]
But very impractical when
Pθ(Z = xobs
) ≈ 0

A as approximative
When y is a continuous random variable, strict equality
z = xobs
is replaced with a tolerance zone
ρ(xobs
, z) ≤ ε
where ρ is a distance
Output distributed from
π(θ) Pθ{ρ(xobs
, z) < ε}
def
∝ π(θ|ρ(xobs
, z) < ε)
[Pritchard et al., 1999]

ABC algorithm
Algorithm 1 Likelihood-free rejection sampler
for i = 1 to N do
repeat
generate θ from prior π(·)
generate z from sampling density f(·|θ )
until ρ{η(z), η(xobs)} ≤ ε
set θi = θ
end for
where η(xobs) deﬁnes a (not necessarily suﬃcient) statistic
Custom: η(xobs) called summary statistic

ABC algorithm
Algorithm 2 Likelihood-free rejection sampler
for i = 1 to N do
repeat
generate θ from prior π(·)
generate z from sampling density f(·|θ )
until ρ{η(z), η(xobs)} ≤ ε
set θi = θ
end for
where η(xobs) deﬁnes a (not necessarily suﬃcient) statistic
Custom: η(xobs) called summary statistic

Example 3: robust Normal statistics
mu=rnorm(N<-1e6) #prior
sig=sqrt(rgamma(N,2,2))
medobs=median(obs)
madobs=mad(obs) #summary
for(t in diz<-1:N){
psud=rnorm(1e2)/sig[t]+mu[t]
medpsu=median(psud)-medobs
madpsu=mad(psud)-madobs
diz[t]=medpsu^2+madpsu^2}
#ABC subsample
subz=which(diz<quantile(diz,.1))

Exact ABC posterior
Algorithm samples from marginal in z of [exact] posterior
πABC
ε (θ, z|xobs
) =
π(θ)f(z|θ)IAε,xobs
(z)
Aε,xobs ×Θ π(θ)f(z|θ)dzdθ
,
where Aε,xobs = {z ∈ D|ρ{η(z), η(xobs)} < ε}.
Intuition that proper summary statistics coupled with small
tolerance ε = εη should provide good approximation of the
posterior distribution:
πABC
ε (θ|xobs
) = πABC
ε (θ, z|xobs
)dz ≈ π{θ|η(xobs
)}

Why summaries?
reduction of dimension
improvement of signal to noise ratio
reduce tolerance ε considerably
whole data may be unavailable (as in Example 3)
medobs=median(obs)
madobs=mad(obs) #summary

Example 6: MA inference
Moving average model MA(2):
xt = εt + θ1εt−1 + θ2εt−2 εt
iid
∼ N(0, 1)
Comparison of raw series:

Moving average model MA(2):
xt = εt + θ1εt−1 + θ2εt−2 εt
iid
∼ N(0, 1)
[Feller, 1970]
Comparison of acf’s:

Summary vs. raw:

Why not summaries?
loss of suﬃcient information when πABC(θ|xobs) replaced
with πABC(θ|η(xobs))
arbitrariness of summaries
uncalibrated approximation
whole data may be available (at same cost as summaries)
(empirical) distributions may be compared (Wasserstein
distances)
[Bernton et al., 2019]

Summary selection strategies
Fundamental difficulty of selecting summary statistics when
there is no non-trivial sufficient statistic [except when done by
experimenters from the field]

1. Starting from large collection of summary statistics, Joyce
and Marjoram (2008) consider the sequential inclusion into the
ABC target, with a stopping rule based on a likelihood ratio test

2. Based on decision-theoretic principles, Fearnhead and
Prangle (2012) end up with E[θ|xobs] as the optimal summary
statistic

3. Use of indirect inference by Drovandi, Pettit, & Paddy
(2011) with estimators of parameters of auxiliary model as
summary statistics

4. Starting from large collection of summary statistics, Raynal
& al. (2018, 2019) rely on random forests to build estimators
and select summaries

5. Starting from large collection of summary statistics, Sedki &
Pudlo (2012) use the Lasso to eliminate summaries

Semi-automated ABC
Use of summary statistic η(·), importance proposal g(·), kernel
K(·) ≤ 1 with bandwidth h ↓ 0 such that
(θ, z) ∼ g(θ)f(z|θ)
accepted with probability (hence the bound)
K[{η(z) − η(xobs
)}/h]
and the corresponding importance weight deﬁned by
π(θ) g(θ)
Theorem Optimality of posterior expectation E[θ|xobs] of
parameter of interest as summary statistics η(xobs)
[Fearnhead & Prangle, 2012; Sisson et al., 2019]

Random forests
Technique that stemmed from Leo Breiman’s bagging (or
bootstrap aggregating) machine learning algorithm for both
classiﬁcation [testing] and regression [estimation]
[Breiman, 1996]
Improved performances by averaging over classiﬁcation schemes
of randomly generated training sets, creating a “forest” of
(CART) decision trees, inspired by Amit and Geman (1997)
ensemble learning
[Breiman, 2001]

Growing the forest
Breiman’s solution for inducing random features in the trees of
the forest:
boostrap resampling of the dataset and
random subset-ing [of size
√
t] of the covariates driving the
classiﬁcation or regression at every node of each tree
Covariate (summary) xτ that drives the node separation
xτ cτ
and the separation bound cτ chosen by minimising entropy or
Gini index

ABC with random forests
Idea: Starting with
possibly large collection of summary statistics (η1, . . . , ηp)
(from scientiﬁc theory input to available statistical
softwares, methodologies, to machine-learning alternatives)
ABC reference table involving model index, parameter
values and summary statistics for the associated simulated
pseudo-data
run R randomforest to infer M or θ from (η1i, . . . , ηpi)

Idea: Starting with
pseudo-data
Average of the trees is resulting summary statistics, highly
non-linear predictor of the model index

Idea: Starting with
pseudo-data
Potential selection of most active summaries, calibrated against
pure noise

Classiﬁcation of summaries by random forests
Given large collection of summary statistics, rather than
selecting a subset and excluding the others, estimate each
parameter by random forests
handles thousands of predictors
ignores useless components
fast estimation with good local properties
automatised with few calibration steps
substitute to Fearnhead and Prangle (2012) preliminary
estimation of ^θ(yobs)
includes a natural (classiﬁcation) distance measure that
avoids choice of either distance or tolerance
[Marin et al., 2016, 2018]

Calibration of tolerance
Calibration of threshold ε
from scratch [how small is small?]
from k-nearest neighbour perspective [quantile of prior
predictive]
subz=which(diz<quantile(diz,.1))
from asymptotics [convergence speed]
related with choice of distance [automated selection by
random forests]
[Fearnhead & Prangle, 2012; Biau et al., 2013; Liu & Fearnhead 2018]

Implementation
3 Implementation
ABC-MCMC
ABC-PMC
post-processed ABC
4 Convergence
5 Advanced topics

Sofware
Several ABC R packages for performing parameter
estimation and model selection
[Nunes & Prangle, 2017]

Sofware
abctools R package tuning ABC analyses

Sofware
abcrf R package ABC via random forests

Sofware
EasyABC R package several algorithms for performing
eﬃcient ABC sampling schemes, including four sequential
sampling schemes and 3 MCMC schemes

Sofware
DIYABC non R software for population genetics

ABC-IS
Basic ABC algorithm limitations
blind [no learning]
ineﬃcient [curse of dimension]
inapplicable to improper priors
Importance sampling version
importance density g(θ)
bounded kernel function Kh with bandwidth h
acceptance probability of
Kh{ρ[η(xobs
), η(x{θ})]} π(θ) g(θ) max
θ
Aθ
[Fearnhead & Prangle, 2012]

ABC-MCMC
Markov chain (θ(t)) created via transition function
θ(t+1)
=



θ ∼ Kω(θ |θ(t)) if x ∼ f(x|θ ) is such that x ≈ y
and u ∼ U(0, 1) ≤ π(θ )Kω(θ(t)|θ )
π(θ(t))Kω(θ |θ(t))
,
θ(t) otherwise,
has the posterior π(θ|y) as stationary distribution
[Marjoram et al, 2003]

ABC-MCMC
Algorithm 3 Likelihood-free MCMC sampler
get (θ(0), z(0)) by Algorithm 1
for t = 1 to N do
generate θ from Kω ·|θ(t−1) , z from f(·|θ ), u from U[0,1],
if u ≤ π(θ )Kω(θ(t−1)|θ )
π(θ(t−1)Kω(θ |θ(t−1))
IAε,xobs
(z ) then
set (θ(t), z(t)) = (θ , z )
else
(θ(t), z(t))) = (θ(t−1), z(t−1)),
end if
end for

ABC-PMC
Generate a sample at iteration t by
^πt(θ(t)
) ∝
N
j=1
ω
(t−1)
j Kt(θ(t)
|θ
(t−1)
j )
modulo acceptance of the associated xt, with tolerance εt ↓, and
use importance weight associated with accepted simulation θ
(t)
i
ω
(t)
i ∝ π(θ
(t)
i ) ^πt(θ
(t)
i )
c Still likelihood free
[Sisson et al., 2007; Beaumont et al., 2009]

ABC-SMC
Use of a kernel Kt associated with target πεt and derivation of
the backward kernel
Lt−1(z, z ) =
πεt (z )Kt(z , z)
πεt (z)
Update of the weights
ω
(t)
i ∝ ω
(t−1)
i
M
m=1 IAεt
(x
(t)
im)
M
m=1 IAεt−1
(x
(t−1)
im )
when x
(t)
im ∼ Kt(x
(t−1)
i , ·)
[Del Moral, Doucet & Jasra, 2009]

ABC-NP
Better usage of [prior] simulations by
adjustement: instead of throwing
away θ such that ρ(η(z), η(xobs)) > ε,
replace θ’s with locally regressed
transforms
θ∗
= θ − {η(z) − η(xobs
)}T ^β [Csill´ery et al., TEE, 2010]
where ^β is obtained by [NP] weighted least square regression on
(η(z) − η(xobs)) with weights
Kδ ρ(η(z), η(xobs
))
[Beaumont et al., 2002, Genetics]

ABC-NN
Incorporating non-linearities and heterocedasticities:
θ∗
= ^m(η(xobs
)) + [θ − ^m(η(z))]
^σ(η(xobs))
^σ(η(z))
where
^m(η) estimated by non-linear regression (e.g., neural
network)
^σ(η) estimated by non-linear regression on residuals
log{θi − ^m(ηi)}2
= log σ2
(ηi) + ξi
[Blum & Fran¸cois, 2009]

Convergence
4 Convergence
ABC inference consistency
ABC model choice consistency
ABC under misspeciﬁcation
5 Advanced topics

Asymptotics of ABC
Since πABC(· | xobs) is an approximation of π(· | xobs) or
π(· | η(xobs)) coherence of ABC-based inference need be
established on its own
[Li & Fearnhead, 2018a,b; Frazier et al., 2018,2020]
Meaning
establishing large sample (n) properties of ABC posteriors
and ABC procedures
ﬁnding suﬃcient conditions and checks on summary
statistics η(·)
determining proper rate ε = εn of convergence of tolerance
to 0
[mostly] ignoring Monte Carlo errors

Consistency of ABC posteriors
ABC algorithm Bayesian consistent at θ0 if for any δ > 0,
Π θ − θ0 > δ | ρ{η(xobs
), η(Z)} ≤ ε → 0
as n → +∞, ε → 0
Bayesian consistency implies that sets containing θ0 have
posterior probability tending to one as n → +∞, with
implication being the existence of a speciﬁc rate of
concentration

ABC algorithm Bayesian consistent at θ0 if for any δ > 0,
Π θ − θ0 > δ | ρ{η(xobs
), η(Z)} ≤ ε → 0
as n → +∞, ε → 0
Concentration around true value and Bayesian consistency
impose less stringent conditions on the convergence speed
of tolerance εn to zero, when compared with asymptotic
normality of ABC posterior
asymptotic normality of ABC posterior mean does not
require asymptotic normality of ABC posterior

Asymptotic setup
Assumptions:
asymptotic: xobs = xobs(n)
∼ Pn
θ0
and ε = εn, n → +∞
parametric: θ ∈ Rk, k ﬁxed concentration of summary
statistic η(zn):
∃b : θ → b(θ) η(zn
) − b(θ) = oPθ
(1), ∀θ
identiﬁability of parameter b(θ) = b(θ ) when θ = θ

Concentration of summary η(z): there exists b(θ) such that
η(z) − b(θ) = oPθ
(1)
Consistency:
Πεn θ − θ0 ≤ δ | η(xobs
) = 1 + op(1)
Convergence rate: there exists δn = o(1) such that
Πεn θ − θ0 ≤ δn | η(xobs
) = 1 + op(1)

Consistency:
Πεn θ − θ0 ≤ δ | η(xobs
) = 1 + op(1)
Convergence rate: there exists δn = o(1) such that
Πεn θ − θ0 ≤ δn | η(xobs
) = 1 + op(1)
Point estimator consistency
^θε = EABC[θ|η(xobs(n)
)], EABC[θ | η(xobs(n)
)] − θ0 = op(1)
vn(EABC[θ | η(xobs(n)
)] − θ0) ⇒ N(0, v)

Asymptotic shape of posterior distribution
Shape of
Π · | η(xobs
), η(z) ≤ εn
depending on relation between εn and rate σn at which η(xobsn
)
satisfy CLT
Three diﬀerent regimes:
1. σn = o(εn) −→ Uniform limit
2. σn εn −→ perturbated Gaussian limit
3. σn εn −→ Gaussian limit

Asymptotic behaviour of posterior mean
When kη = dim(η(xobs)) = kθ = dim(θ) and εn = o(n−3/10)
EABC[νn(θ − θ0) | xobs
] ⇒ N(0, ( bo
)T
Σ−1
bo
−1
[Li & Fearnhead (2018a)]
In fact, if εβ+1
n
√
n = o(1), with β H¨older-smoothness of π
EABC[(θ−θ0) | xobs
] =
( bo)−1Zo
√
n
+
k
j=1
hj(θ0)ε2j
n +op(1), 2k = β

Asymptotic behaviour of posterior mean
When kη = dim(η(xobs)) = kθ = dim(θ) and εn = o(n−3/10)
EABC[νn(θ − θ0) | xobs
] ⇒ N(0, ( bo
)T
Σ−1
bo
−1
[Li & Fearnhead (2018a)]
Iterating for ﬁxed kθ mildly interesting: if
˜η(xobs
) = EABC[θ | xobs
]
then
EABC[θ|˜η(xobs
)] = θ0 +
( bo)−1Zo
√
n
+
π (θ0)
π(θ0)
ε2
n + o()

Curse of dimension
for reasonable statistical behavior, decline of acceptance αn
the faster the larger the dimension of θ, kθ, but unaﬀected
by dimension of η, kη
theoretical justiﬁcation for dimension reduction methods
that process parameter components individually and
independently of other components
[Fearnhead & Prangle, 2012; Martin & al., 2016]
importance sampling approach of Li & Fearnhead (2018a)
yields acceptance rates αn = O(1), when εn = O(1/vn)

Monte Carlo error
Link the choice of εn to Monte Carlo error associated with Nn
draws in ABC Algorithm
Conditions (on εn) under which
^αn = αn{1 + op(1)}
where ^αn = Nn
i=1 I [d{η(y), η(z)} ≤ εn] /Nn proportion of
accepted draws from Nn simulated draws of θ
Either
(i) εn = o(v−1
n ) and (vnεn)−kη ε−kθ
n ≤ MNn
or
(ii) εn v−1
n and ε−kθ
n ≤ MNn
for M large enough

Bayesian model choice
Model candidates M1, M2, . . . to be compared for dataset xobs
making model index M part of inference
Use of a prior distribution. π(M = m), plus a prior distribution
on the parameter conditional on the value m of the model
index, πm(θm)
Goal to derive the posterior distribution of M, challenging
computational target when models are complex
[Savage, 1964; Berger, 1980]

Generic ABC for model choice
Algorithm 4 Likelihood-free model choice sampler (ABC-MC)
for t = 1 to T do
repeat
Generate m from the prior π(M = m)
Generate θm from the prior πm(θm)
Generate z from the model fm(z|θm)
until ρ{η(z), η(xobs)} < ε
Set m(t) = m and θ(t) = θm
end for
[Cornuet et al., DIYABC, 2009]

ABC model choice consistency
Leaving approximations aside, limiting ABC procedure is Bayes
factor based on η(xobs)
B12(η(xobs
))
Potential loss of information at the testing level
[Robert et al., 2010]
When is Bayes factor based on insuﬃcient statistic η(xobs)
consistent?
[Marin et al., 2013]

Example 7: Gauss versus Laplace
Model M1: xobs ∼ N(θ1, 1)⊗n opposed to model M2:
xobs ∼ L(θ2, 1/
√
2)⊗n, Laplace distribution with mean θ2 and
variance one
Four possible statistics η(xobs)
1. sample mean xobs (suﬃcient for M1 if not M);
2. sample median med(xobs) (insuﬃcient);
3. sample variance var(xobs) (ancillary);
4. median absolute deviation
mad(xobs) = med(|xobs − med(xobs)|);

Example 7: Gauss versus Laplace
Model M1: xobs ∼ N(θ1, 1)⊗n opposed to model M2:
xobs ∼ L(θ2, 1/
√
2)⊗n, Laplace distribution with mean θ2 and
variance one
q
q
q
q
q
q
q
q
q
q
q
Gauss Laplace
0.00.10.20.30.40.50.60.7
n=100
q
q
q
q
q
q
q
q
q
q
q
q
q
q
q
q
q
q
Gauss Laplace
0.00.20.40.60.81.0
n=100

Consistency
Summary statistics
η(xobs
) = (τ1(xobs
), τ2(xobs
), · · · , τd(xobs
)) ∈ Rd
with
true distribution η ∼ Gn, true mean µ0,
distribution Gi,n under model Mi, corresponding posteriors
πi(· | ηn)
Assumptions of central limit theorem and large deviations for
η(xobs) under true, plus usual Bayesian asymptotics with di
eﬀective dimension of the parameter)
[Pillai et al., 2013]

Asymptotic marginals
Asymptotically
mi,n(t) =
Θi
gi,n(t|θi) πi(θi) dθi
such that
(i) if inf{|µi(θi) − µ0|; θi ∈ Θi} = 0,
Cl
√
n
d−di
≤ mi,n(ηn
) ≤ Cu
√
n
d−di
and
(ii) if inf{|µi(θi) − µ0|; θi ∈ Θi} > 0
mi,n(ηn
) = oPn [
√
n
d−τi
+
√
n
d−αi
].

Between-model consistency
Consequence of above is that asymptotic behaviour of the Bayes
factor is driven by the asymptotic mean value µ(θ) of ηn under
both models. And only by this mean value!

Indeed, if
inf{|µ0 − µ2(θ2)|; θ2 ∈ Θ2} = inf{|µ0 − µ1(θ1)|; θ1 ∈ Θ1} = 0
then
Cl
√
n
−(d1−d2)
≤ m1,n(ηn
) m2(ηn
) ≤ Cu
√
n
−(d1−d2)
,
where Cl, Cu = OPn (1), irrespective of the true model.
c Only depends on the diﬀerence d1 − d2: no consistency

Else, if
inf{|µ0 − µ2(θ2)|; θ2 ∈ Θ2} > inf{|µ0 − µ1(θ1)|; θ1 ∈ Θ1} = 0
then
m1,n(ηn)
m2,n(ηn)
≥ Cu min
√
n
−(d1−α2)
,
√
n
−(d1−τ2)

Checking for adequate statistics
Run a practical check of the relevance (or non-relevance) of ηn
null hypothesis that both models are compatible with the
statistic ηn
H0 : inf{|µ2(θ2) − µ0|; θ2 ∈ Θ2} = 0
against
H1 : inf{|µ2(θ2) − µ0|; θ2 ∈ Θ2} > 0
testing procedure provides estimates of mean of ηn under each
model and checks for equality

ABC under misspeciﬁcation
ABC methods rely on simulations z(θ) from the model to
identify those close to xobs
What is happening when the model is wrong?
for some tolerance sequences εn ↓ ε∗, well-behaved ABC
posteriors that concentrate posterior mass on pseudo-true
value
if εn too large, asymptotic limit of ABC posterior uniform
with radius of order εn − ε∗
even if
√
n{εn − ε∗} → 2c ∈ R, limiting distribution no
longer Gaussian
ABC credible sets invalid conﬁdence sets
[Frazier et al., 2020]

Example 8: Normal model with wrong variance
Assumed data generating process (DGP) is z ∼ N(θ, 1) but
actual DGP is xobs ∼ N(θ, ˜σ2)
Use of summaries
sample mean η1(xobs) = 1
n
n
i=1 xi
centered summary η2(xobs) = 1
n−1
n
i=1(xi − η1(xobs)2 − 1
Three ABC:
ABC-AR: accept/reject approach with
Kε(d{η(z), η(xobs)}) = I d{η(z), η(xobs)} ≤ ε and
d{x, y} = x − y
ABC-K: smooth rejection approach, with
Kε(d{η(z), η(xobs)}) univariate Gaussian kernel
ABC-Reg: post-processing ABC approach with weighted
linear regression adjustment

Example 8: Normal model with wrong variance
posterior means for ABC-AR, ABC-K and ABC-Reg as σ2
increases (N = 50, 000 simulated data sets)
αn = n−5/9 quantile for ABC-AR
ABC-K and ABC-Reg bandwidth of n−5/9
[Frazier et al., 2020]

ABC misspecification
data xobs with true distribution P0 assumed issued from
model Pθ θ ∈ Θ ⊂ Rkθ and summary statistic
η(xobs) = (η1(xobs), ..., ηkη (xobs))
misspecification
inf
θ∈Θ
D(P0||Pθ) = inf
θ∈Θ
log
dP0(x)
dPθ(x)
dP0(y) > 0,
with
θ∗
= arg inf
θ∈Θ
D(P0||Pθ)
[Muller, 2013]
ABC misspecification:
for b0 (resp. b(θ)) limit of η(xobs) (resp. η(z))
inf
θ∈Θ
d{b0, b(θ)} > 0

ABC misspeciﬁcation
data xobs with true distribution P0 assumed issued from
model Pθ θ ∈ Θ ⊂ Rkθ and summary statistic
η(xobs) = (η1(xobs), ..., ηkη (xobs))
ABC misspeciﬁcation:
for b0 (resp. b(θ)) limit of η(xobs) (resp. η(z))
inf
θ∈Θ
d{b0, b(θ)} > 0
ABC pseudo-true value:
θ∗
= arg inf
θ∈Θ
d{b0, b(θ)}.

Minimum tolerance
Under identiﬁcation conditions on b(·) ∈ Rkη , there exists ε∗
such that
ε∗
= inf
θ∈Θ
d{b0, b(θ)} > 0
Once εn < ε∗ no draw of θ to be selected and posterior
Πε[A|η(xobs)] ill-conditioned
But appropriately chosen tolerance sequence (εn)n allows
ABC-based posterior to concentrate on distance-dependent
pseudo-true value θ∗

ABC concentration under misspeciﬁcation
Assumptions
[A0] Existence of unique b0 such that d(η(xobs), b0) = oP0
(1)
and of sequence v0,n → +∞ such that
lim inf
n→+∞
P0 d(η(xobs
n ), b0) ≥ v−1
0,n = 1.

Assumptions
[A1] Existence of injective map b : Θ → B ⊂ Rkη and function
ρn with ρn(·) ↓ 0 as n → +∞, and ρn(·) non-increasing,
such that
Pθ [d(η(Z), b(θ)) > u] ≤ c(θ)ρn(u),
Θ
c(θ)dΠ(θ) < ∞
and assume either
(i) Polynomial deviations: existence of vn ↑ +∞ and
u0, κ > 0 such that ρn(u) = v−κ
n u−κ, for u ≤ u0
(ii) Exponential deviations:

Assumptions
[A1] Existence of injective map b : Θ → B ⊂ Rkη and function
ρn with ρn(·) ↓ 0 as n → +∞, and ρn(·) non-increasing,
such that
Pθ [d(η(Z), b(θ)) > u] ≤ c(θ)ρn(u),
Θ
c(θ)dΠ(θ) < ∞
and assume either
(i) Polynomial deviations:
(ii) Exponential deviations: existence of hθ(·) > 0 such
that Pθ[d(η(z), b(θ)) > u] ≤ c(θ)e−hθ(uvn) and
existence of m, C > 0 such that
Θ
c(θ)e−hθ(uvn)
dΠ(θ) ≤ Ce−m·(uvn)τ
, for u ≤ u0.

Assumptions
[A2] existence of D > 0 and M0, δ0 > 0 such that, for all
δ0 ≥ δ > 0 and M ≥ M0, existence of
Sδ ⊂ {θ ∈ Θ : d(b(θ), b0) − ε∗ ≤ δ} for which
(i) In case (i), D < κ and Sδ
1 − c(θ)
M dΠ(θ) δD.
(ii) In case (ii), Sδ
1 − c(θ)e−hθ(M) dΠ(θ) δD.

Consistency
Assume [A0] – [A2], with εn ↓ ε∗ with
εn ≥ ε∗
+ Mv−1
n + v−1
0,n,
for M large enough. Let Mn ↑ ∞ and δn ≥ Mn(εn − ε∗), then
Πε d(b(θ), b0) ≥ ε∗
+ δn | η(xobs
) = oP0
(1),
1. if δn ≥ Mnv−1
n u
−D/κ
n = o(1) in case (i)
2. if δn ≥ Mnv−1
n | log(un)|1/τ = o(1) in case (ii)
with un = εn − (ε∗ + Mv−1
n + v−1
0,n) ≥ 0 .
[Bernton et al., 2017; Frazier et al., 2020]

Regression adjustement under misspeciﬁcation
Accepted value θ artiﬁcially related to η(xobs) and η(z) through
local linear regression model
θ = µ + β {η(xobs
) − η(z)} + ν,
where νi model residual
[Beaumont et al., 2003]
Asymptotic behavior of ABC-Reg posterior
Πε[· | η(xobs
)]
determined by behavior of
Πε[· | η(xobs
)], ^β, and {η(xobs
) − η(z)}

Regression adjustement under misspeciﬁcation
ABC-Reg takes draws of asymptotically optimal θ,
perturbed in a manner that need not preserve original
optimality
for β0 large, pseudo-true value ˜θ∗ possibly outside Θ
extends to nonlinear regression adjustments [Blum
& Fran¸cois, 2010]
potential correction of the adjustment [Frazier et al., 2020]
local regression adjustments with smaller posterior
variability than ABC-AR but fake precision

Example 9: misspeciﬁed g-&-k
Quantile function of Tukey’s g-&-k model:
F−1
(q) = a + b 1 + 0.8
1 − exp(−gz(q))
1 + exp(−gz(q))
1 + z(q)2
k
z(q),
where z(q) q-th N(0, 1) quantile
But data generated from a mixture distribution with minor
bi-modality

Example 9: misspeciﬁed g-&-k

Advanced topics
5 Advanced topics
big data issues
big parameter issues

Computational bottleneck
Time per iteration increases with sample size n of the data:
cost of sampling O(n1+?) associated with a reasonable
acceptance probability makes ABC infeasible for large datasets
surrogate models to get samples (e.g., using copulas)
direct sampling of summary statistics (e.g., synthetic
likelihood)
[Wood, 2010]
borrow from proposals for scalable MCMC (e.g., divide
& conquer)

Approximate ABC [AABC]
Idea approximations on both parameter and model spaces by
resorting to bootstrap techniques.
[Buzbas & Rosenberg, 2015]
Procedure scheme
1. Sample (θi, xi), i = 1, . . . , m, from prior predictive
2. Simulate θ∗ ∼ π(·) and assign weight wi to dataset x(i)
simulated under k-closest θi to θ∗
3. Generate dataset x∗ as bootstrap weighted sample from
(x(1), . . . , x(k))
Drawbacks
If m too small, prior predictive sample may miss
informative parameters
large error and misleading representation of true posterior
only suited for models with very few parameters

Approximate ABC [AABC]
Procedure scheme
1. Sample (θi, xi), i = 1, . . . , m, from prior predictive
2. Simulate θ∗ ∼ π(·) and assign weight wi to dataset x(i)
simulated under k-closest θi to θ∗
3. Generate dataset x∗ as bootstrap weighted sample from
(x(1), . . . , x(k))
Drawbacks
If m too small, prior predictive sample may miss
informative parameters
large error and misleading representation of true posterior
only suited for models with very few parameters

Divide-and-conquer perspectives
1. divide the large dataset into smaller batches
2. sample from the batch posterior
3. combine the result to get a sample from the targeted
posterior
Alternative via ABC-EP
[Barthelm´e & Chopin, 2014]

Geometric combination: WASP
Subset posteriors given partition xobs
[1] , . . . , xobs
[B] of observed
data xobs, let deﬁne
π(θ | xobs
[b] ) ∝ π(θ)
j∈[b]
f(xobs
j | θ)B
.
[Srivastava et al., 2015]
Subset posteriors are combined via Wasserstein barycenter
[Cuturi, 2014]

[1] , . . . , xobs
[B] of observed
π(θ | xobs
[b] ) ∝ π(θ)
j∈[b]
f(xobs
j | θ)B
.
[Cuturi, 2014]
Drawback require sampling from f(· | θ)B by ABC means.
Should be feasible for latent variable (z) representations when
f(x | z, θ) available in closed form
[Doucet & Robert, 2001]

[1] , . . . , xobs
[B] of observed
π(θ | xobs
[b] ) ∝ π(θ)
j∈[b]
f(xobs
j | θ)B
.
[Cuturi, 2014]
Alternative backfeed subset posteriors as priors to other
subsets, partitioning summaries

Consensus ABC
Na¨ıve scheme
For each data batch b = 1, . . . , B
1. Sample (θ
[b]
1 , . . . , θ
[b]
n ) from diﬀused prior ˜π(·) ∝ π(·)1/B
2. Run ABC to sample from batch posterior
^π(· | d(S(xobs
[b] ), S(x[b])) < ε)
3. Compute sample posterior variance Σ−1
b
Combine batch posterior approximations
θj =
B
b=1
Σbθ
[b]
j
B
b=1
Σb
Remark Diﬀuse prior ˜π(·) non informative calls for
ABC-MCMC steps

Big parameter issues
Curse of dimension: as dim(Θ) = kθ increases
exploration of parameter space gets harder
summary statistic η forced to increase, since at least of
dimension kη ≥ dim(Θ)
Some solutions
adopt more local algorithms like ABC-MCMC or
ABC-SMC
aim at posterior marginals and approximate joint posterior
by copula
[Li et al., 2016]
run ABC-Gibbs
[Clart´e et al., 2016]

Example 11: Hierarchical MA(2)
xi
iid
∼ MA2(µi, σi)
µi = (βi,1 − βi,2, 2(βi,1 + βi,2) − 1) with
(βi,1, βi,2, βi,3)
iid
∼ Dir(α1, α2, α3)
σi
iid
∼ IG(σ1, σ2)
α = (α1, α2, α3), with prior E(1)⊗3
σ = (σ1, σ2) with prior C(1)+⊗2
α
µ1 µ2 µn. . .
x1 x2 xn. . .
σ
σ1 σ2 σn. . .

c Based on 106 prior and 103 posteriors simulations, 3n
summary statistics, and series of length 100, ABC-Rej posterior
hardly distinguishable from prior!

ABC-Gibbs
When parameter decomposed into θ = (θ1, . . . , θn)
Algorithm 5 ABC-Gibbs sampler
starting point θ(0) = (θ
(0)
1 , . . . , θ
(0)
n ), observation xobs
for i = 1, . . . , N do
for j = 1, . . . , n do
θ
(i)
j ∼ πεj
(· | x , sj, θ
(i)
1 , . . . , θ
(i)
j−1, θ
(i−1)
j+1 , . . . , θ
(i−1)
n )
end for
end for
Divide & conquer:
one tolerance εj for each parameter θj
one statistic sj for each parameter θj

ABC-Gibbs
When parameter decomposed into θ = (θ1, . . . , θn)
Algorithm 6 ABC-Gibbs sampler
starting point θ(0) = (θ
(0)
1 , . . . , θ
(0)
n ), observation xobs
for i = 1, . . . , N do
for j = 1, . . . , n do
θ
(i)
j ∼ πεj
(· | x , sj, θ
(i)
1 , . . . , θ
(i)
j−1, θ
(i−1)
j+1 , . . . , θ
(i−1)
n )
end for
end for
Divide & conquer:
one tolerance εj for each parameter θj
one statistic sj for each parameter θj

Compatibility
When using ABC-Gibbs conditionals with different acceptance
events, e.g., different statistics
π(α)π(sα(µ) | α) and π(µ)f(sµ(x ) | α, µ).
conditionals are incompatible
ABC-Gibbs does not necessarily converge (even for
tolerance equal to zero)
potential limiting distribution
not a genuine posterior (double use of data)
unknown [except for a specific version]
possibly far from genuine posterior(s)

Convergence
In hierarchical case n = 2,
Theorem If there exists 0 < κ < 1/2 such that
sup
θ1,˜θ1
πε2
(· | x , s2, θ1) − πε2
(· | x , s2, ˜θ1) TV = κ
ABC-Gibbs Markov chain geometrically converges in total
variation to stationary distribution νε, with geometric rate
1 − 2κ.

Separation from the prior for identical number of simulations

For model
i.i.d.
∼ π(µj | α) , α ∼ π(α)
that is,
prior simulations of α ∼ π(α) and of ˜µ ∼ π(˜µ | α) until
η(sα(µ), sα(˜µ)) < εα
simulation of µ from instrumental q(µ) and of auxiliary
variables ˜µ and ˜x until both constraints satisﬁed

For model
i.i.d.
∼ π(µj | α) , α ∼ π(α)
Resulting Gibbs sampler stationary for posterior proportional to
π(α, µ) q(sα(µ))
projection
f(sµ(x ) | µ)
projection
that is, for likelihood associated with sµ(x ) and prior
distribution proportional to π(α, µ)q(sα(µ)) [exact!]

Incoming ABC workshops
[A]BayesComp, Gainesville, Florida, Jan 7-10 2020
ABC in Grenoble, France, March 18-19 2020
ISBA(BC), Kunming, China, June 26-30 2020
ABC in Longyearbyen, Svalbard, April 11-13 2021

ABC postdoc positions
2 post-doc positions with the ABSint research grant:
Focus on approximate Bayesian techniques like ABC,
variational Bayes, PAC-Bayes, Bayesian non-parametrics,
scalable MCMC, and related topics. A potential direction
of research would be the derivation of new Bayesian tools
for model checking in such complex environments.
Terms: up to 24 months, no teaching duty attached,
primarily located in Universit´e Paris-Dauphine, with
supported periods in Oxford (J. Rousseau) and visits to
Montpellier (J.-M. Marin). No hard deadline.
If interested, send application to me:
bayesianstatistics@gmail.com

asymptotics of ABC

More Related Content

What's hot (20)

Similar to asymptotics of ABC (20)

More from Christian Robert (19)

Recently uploaded (20)

asymptotics of ABC