New Insights Into Approximate Bayesian Computation

1 Introduction

• ABC is motivated by settings where posterior computation is intractable or suitable likelihoods are unavailable, while theoretical guarantees for the approach remain limited.The paper analyzes ABC through nearest-neighbor theory.Source: §1 · pp. 2, 4, 5
• Algorithm 2 fixes the number kN of accepted points and lets their distance from s0 vary, dual to the fixed-tolerance algorithm’s random accepted count.In practice, N is typically of order 10^6, and kN is commonly specified as a percentile.Source: §1 · pp. 3–4
• Conditioned on the distance to the (kN+1)th nearest summary, Algorithm 2’s outputs are distributed as kN i.i.d. draws from the joint density restricted to a cylinder around s0.This gives a precise description of the algorithm’s output distribution.Source: §1 · p. 5
• The posterior estimate practitioners use smooths the parameter values of the kN nearest neighbors around θ0, combining k-nearest-neighbor and kernel density estimation.Its bandwidth hN and kernel K control the smoothing.Source: §1 · p. 5
• The paper establishes consistency properties and convergence rates for this genuine conditional-density estimate.The estimate targets g(θ0|s0).Source: §1 · pp. 5–6
• The paper leaves automatic summary-statistic selection, efficiency improvements, and ABC model choice outside its scope.The authors note that theoretical arguments for model choice are still missing.Source: §1 · p. 6

• Conditionally on d(kN+1), the accepted pairs behave as i.i.d. draws from the joint density restricted to summaries within that distance of s0.This gives a precise description of the distribution of ABC Algorithm 2 outputs.Source: §2 · pp. 7–8
• The distribution result extends an order-statistics representation to the Bayesian setting not covered by Kaufmann and Reiss (1992).The paper notes that its proof uses simpler arguments.Source: §2 · p. 7
• For bounded ϕ(Θ), the ABC estimate of E[ϕ(Θ)|S=s0] is consistent for almost all s0 when kN/log log N→∞ and kN/N→0.The result follows by transferring nearest-neighbor regression theory to ABC.Source: §2 · pp. 8–9

• The estimate combines retaining the kN nearest summaries to s0 with kernel smoothing of their corresponding parameter values.Unlike kernel-type competitors, it avoids evaluating a ratio with a potentially small denominator, making it easy to implement.Source: §3 · p. 10
• The analysis targets convergence under minimal conditions on the unknown densities, using real-analysis arguments to reduce assumptions.The paper treats pointwise and integrated mean-square error consistency as separate results.Source: §3 · pp. 10, 14
• Theorem 3.2 addresses a gap in results for general product kernels by specializing to ABC’s setting, where the second kernel is naive.A broader result for two general kernels is given in the appendix.Source: §3 · p. 12
• Theorem 3.2 allows h and δ to approach zero without restrictions on their relative behavior, at the cost of an integrability condition on f.The condition holds in particular when f belongs to Lq for q>1.Source: §3 · p. 13
• Theorem 3.3 establishes pointwise mean-square consistency, and Theorem 3.4 establishes mean integrated square error consistency under their respective stated assumptions.For Theorem 3.4, the stated conditions include f and π in L2, kN→∞, kN/N→0, hN→0, and kN hN^p N→∞.Source: §3 · pp. 13–14

4 Rates of convergence

• The analysis studies convergence rates while aiming to keep assumptions on the unknown probability densities as mild as possible.The results rely on regularity and integrability assumptions on the densities and kernel.Source: §4 · pp. 15–16
• Theorem 4.1 gives mean integrated square error rates for the estimate, with terms reflecting smoothing in parameter and summary-statistic dimensions.The theorem’s displayed rate expressions are incomplete in the supplied passages.Source: §4 · pp. 16–17
• Balancing the terms in Theorem 4.1 yields dimension-specific rate regimes for m ∈ {1, 2, 3}, m = 4, and m > 4.The corollary states that a suitable sequence kN exists in each regime.Source: §4 · pp. 17–18
• Selecting kN and hN jointly is left open because existing bandwidth rules are difficult to adapt to the nearest-neighbor setting and target local performance at fixed s0.The authors state that devising such a method is beyond the paper’s scope.Source: §4 · p. 18
• For m > 4, a rough rule of thumb sets kN ≈ N^(p+4)/(m+p+4), with an acceptance fraction of about kN/N ≈ N^-m/(m+p+4).The paper presents this as a practical implication of Corollary 4.1.Source: §4 · p. 19

5.1 Proof of Proposition 2.1

• The proof of Proposition 2.1 establishes the conditional i.i.d. representation of ABC outputs by adapting arguments for order statistics and replacing combinations with permutations.The normalizing constant is positive when s0 belongs to the support of the marginal density.Source: §5.1 · p. 19

5.2 Proof of Theorem 3.3

• Theorem 3.3 is proved through a variance–bias decomposition: kN/N → 0 shrinks the neighbor radius, while kN hN^p → ∞ controls the variance term.The argument assumes s0 is a positive-density Lebesgue point.Source: §5.2 · pp. 20–22

5.3 Proof of Theorem 3.4

• Theorem 3.4’s integrated-error proof uses Fubini’s theorem and dominated convergence, with L2 assumptions supporting the required integrability.The proof also invokes maximal-function bounds.Source: §5.3 · pp. 22–23

5.4 Proof of Theorem 4.1

• Theorem 4.1’s rate proof expands the density by multivariate Taylor’s theorem and uses bounds on the second and fourth moments of the neighbor distance.The argument relies on symmetry and the stated smoothness assumptions.Source: §5.4 · pp. 23, 25, 27

6 Some technical results

• Lemma 6.1 applies at a Lebesgue point s0 where the marginal density is positive and establishes a boundedness property for a local volume-to-mass ratio.The displayed definition and full statement are incomplete in the supplied passages.Source: §6 · p. 27
• Lemma 6.2 assumes a three-times continuously differentiable joint density and bounds a parameterized quantity using generalized Minkowski’s inequality.The bound is finite under Assumption [K2] and independent of h and δ.Source: §6 · pp. 27–28
• Propositions 6.1 and 6.2 each assume that the support of ¯f is compact with diameter L > 0, and that s0 is a Lebesgue point of ¯f with ¯f(s0) > 0.Source: §6 · pp. 29–30

A Complements on singular integrals

• Theorem A.1 establishes that (f ⋆ ϕε1,ε2)(x) converges to f(x) for almost all x when f and the product kernel satisfy suitable restrictions.The two parameters ε1 and ε2 may tend to zero independently.Source: A Complements on singular integrals · pp. 31–32
• The appendix’s stated novelty is allowing an approximation-of-identity family to depend on two independent parameters, a question the authors describe as fundamental to multivariate nonparametric estimates.They note that real-analysis literature offers little help on this question.Source: A Complements on singular integrals · p. 31
• The maximal function with two degrees of freedom is bounded on Lq for 1 < q ≤ ∞, with a bound depending on q and dimension.The appendix defines it using separate radii for the two component spaces.Source: A Complements on singular integrals · pp. 32, 37
• The proof handles the continuous compact-support case first, then establishes almost-everywhere existence and extends the convergence result using maximal-function bounds.The argument uses approximation by continuous functions with compact support.Source: A Complements on singular integrals · pp. 35–36