Perron-Mengersen (2001) Bayesian Nonparametric Modeling Using Mixtures of Triangular Distributions

Summary

Perron and Mengersen (2001) propose Bayesian nonparametric estimation of a monotone nondecreasing function $F:[0,1]\to[0,1]$ by approximating it as a cumulative distribution function (CDF) whose derivative is a mixture of triangular densities — equivalently, a quadratic spline with piecewise-linear derivative. The paper proves three theorems establishing that the variable-partition triangular mixture achieves a uniform approximation error bound of $\|F-H\|\leq 3/2r$ (Theorem 3), strictly better than the $O(r^{-1/2})$ Bernstein/beta-mixture bound (Theorem 1) and the $O(r^{-1})$ fixed-partition triangular bound (Theorem 2). The key insight is that allowing the partition to vary with $F$ — rather than working on a fixed convex hull — yields guaranteed uniform approximation over all nondecreasing $F$. A reversible jump Markov chain Monte Carlo (RJMCMC) algorithm with birth, death, move, and $\sigma^2$-renewal operations jointly estimates the partition $\{x_1,\ldots,x_{r-1}\}$ and dimension $r$ (truncated Poisson prior). Applications include a cubic spline comparison on Eppright height/weight data and a survival analysis extending the Gelfand-Mallick (1995) proportional-hazards framework to unknown link and hazard functions. Published in Biometrics 57(2): 518–528.

Key Claims

• Theorem 1 (Beta mixture, fixed partition): $\|H-F\| \leq \delta(r)(1+\sqrt{\pi r}/2)/2r$ — $O(r^{-1/2})$; cannot uniformly cover $\mathcal{F}$ (convex hull limitation).
• Theorem 2 (Triangular mixture, fixed partition): $\|H-F\| \leq 3\delta(r)/2r$ — $O(r^{-1})$; still a convex hull, same coverage failure.
• Theorem 3 (Triangular mixture, variable partition): there exists a partition $\{x_i\}$ (with $F(x_i)=i/r$ optimal) and $H\in\mathcal{F}_{r-1}$ such that $\|F-H\|\leq 3/2r$, uniformly over all nondecreasing $F$ onto $[0,1]$. For $r\geq 3/2\varepsilon$, $\lambda(\mathcal{F}_{r-1},\mathcal{F})\leq\varepsilon$.
• Derivative representation: $h(x)=\sum_{k=0}^r w_k h_k(x)$ where $h_k$ is a triangular density with support $[x_{k-1},x_{k+1}]$, mode $x_k$, and $w_k=((x_{k+1}-x_{k-1})/2)\,h(x_k)$.
• RJMCMC algorithm: four operations — (0) Gibbs renewal of $\sigma^2$; (1) death (remove node $x_j$, decrease $r$); (2) move (perturb $x_j$ within $[x_{j-1},x_{j+1}]$, Metropolis-Hastings); (3) birth (insert node $z\in[x_{j-1},x_j]$, increase $r$) — with Poisson prior $\pi(r)\propto P(\lambda)$ truncated at $K_{\max}$.
• Simulation results: posterior for $r$ is unimodal, centered 10–30; triangular mixture beats beta mixture for step functions (better discontinuity handling); comparable for smooth functions.
• Triangular vs. beta efficiency: evaluating a triangular CDF mixture is computationally cheaper than a beta mixture, especially for inverse $F^{-1}$.
• Survival analysis: reanalysis of Lawless (1982) 40-patient lung cancer data via Gibbs updates of $H$ and $G$ jointly; nonparametric hazard $h$ recovers a decreasing shape consistent with clinical prior.

Concepts Introduced or Extended

• Reversible Jump MCMC

Entities Mentioned

• François Perron
• Kerrie Mengersen

Quotes

"The triangular mixture model is smoother and superior to the beta mixture model in accommodating the discontinuities in this function."

"Here we obtain the desired result, an upper bound on the quality of the approximation that is smaller than one and is uniformly good on $\mathcal{F}$."

My Take

Theoretically elegant — the Theorem 3 result genuinely solves the coverage failure of convex-hull approximations. The connection to the time-series wiki is thin: the paper targets monotone regression and survival analysis, not financial or macroeconomic time series. The primary value for this wiki is methodological: RJMCMC for dimension-selection and the insight that variable-partition mixtures outperform fixed-basis approximations. The survival analysis application closely mirrors the Gelfand-Mallick (1995) framework (not in this wiki) rather than GARCH or VAR applications. Worth knowing as a toolbox entry; not a core source.