Online Monotone Density Estimation

• Online monotone density estimation is a framework for sequentially predicting a nonincreasing probability density on [0,1] using log-loss performance metrics.
• The online Grenander estimator adapts the classical isotonic MLE to a sequential setting, achieving an O(n^(1/3)) excess KL risk under well-specified i.i.d. conditions.
• The expert aggregation approach discretizes the density space and applies exponential weighting to secure an O(√(n log n)) adversarial regret bound while quickly adapting to change-points.

Online monotone density estimation addresses the sequential prediction of an unknown probability density that is known a priori to be monotone nonincreasing on $[0,1]$. At each time $t$, the estimator observes a real-valued data stream $x_1, x_2, \ldots$ and outputs a measurable function $\hat{f}_t$ of past data, so that for every $t$, $\hat{f}_t$ estimates the underlying density $q \in \mathcal{D}$, where

$$\mathcal{D} = \left\{ f : [0,1] \to [0, \infty) \,\Bigg|\, \int_0^1 f(u)\,du = 1,\, f\ \text{nonincreasing}\ \right\}.$$

Performance is measured by sequential log-loss $\ell_t(f) = -\log f(X_t)$ and its cumulative version $$L(\hat{f}, n) = \sum_{t=1}^n \ell_t(\hat{f}_t) = -\sum_{t=1}^n \log \hat{f}_t(X_t)$$. The theoretical framework considers two benchmarks: the well-specified stochastic setting (oracle risk under i.i.d. draws from $t$0) and an adversarial setting (pathwise regret compared to the best hindsight monotone density with amplitude bounds).

1. Formal Problem Statement

The data consists of sequentially observed values $t$1 assumed to be in $t$2 (by linear rescaling). At each $t$3, the algorithm outputs an estimator $t$4 based only on $t$5. The performance of a sequence of estimators $t$6 is evaluated in two principal regimes:

• Stochastic benchmark: $t$7 are i.i.d. from some $t$8. The cumulative excess Kullback-Leibler risk is

$t$9

• Adversarial benchmark (pathwise regret): For arbitrary (well-spaced) sequences, regret is measured against the best monotone density in hindsight within amplitude bounds $x_1, x_2, \ldots$0

$x_1, x_2, \ldots$1

where $x_1, x_2, \ldots$2.

The adversarial regime requires minimal regularity on the input: points are well-spaced (no two closer than $x_1, x_2, \ldots$3), and not close to the upper boundary ($x_1, x_2, \ldots$4).

2. The Online Grenander (OG) Estimator

The classical Grenander estimator provides the offline (batch) maximum likelihood estimator (MLE) over $x_1, x_2, \ldots$5, resulting in a piecewise-constant histogram supported at the order statistics. The online analogue (OG) operates as follows at each $x_1, x_2, \ldots$6:

• Input: Amplitude bounds $x_1, x_2, \ldots$7; initialize $x_1, x_2, \ldots$8.
• For $x_1, x_2, \ldots$9: Solve

$\hat{f}_t$0

Theorem 2.1 (Excess KL risk of OG).

If $\hat{f}_t$1 are i.i.d. $\hat{f}_t$2 with $\hat{f}_t$3,

$\hat{f}_t$4

The derivation utilizes classical entropy arguments: the (offline) Grenander estimator achieves $\hat{f}_t$5; summing over $\hat{f}_t$6 gives an aggregate $\hat{f}_t$7 cumulative risk.

3. The Expert Aggregation (EA) Estimator

To bypass the computational burden of solving isotonic MLEs at each step, the Expert Aggregation (EA) estimator discretizes $\hat{f}_t$8 into a finite net of monotone histograms, and employs exponential weighting:

• Expert set construction:
  • Grid breakpoints $\hat{f}_t$9, $t$0.
  • Log-heights grid $t$1, $t$2.
  • $t$3: All monotone histograms with at most $t$4 bins, breakpoints from $t$5, heights from $t$6, normalized to integrate to 1 and lying in $t$7.
  • Cardinality: $t$8.
• Exponential weighting over $t$9 experts $\hat{f}_t$0, $\hat{f}_t$1, and for $\hat{f}_t$2

$\hat{f}_t$3

• Mixability Lemma: For log-loss, $\hat{f}_t$4.

Theorem 2.2 (Pathwise regret of EA).

Under mild regularity (the data in $\hat{f}_t$5), choosing $\hat{f}_t$6 for $\hat{f}_t$7 yields

$\hat{f}_t$8

The proof uses histogram compression (bin-merge for $\hat{f}_t$9-bin approximation, error $q \in \mathcal{D}$0), bin/height rounding (error $q \in \mathcal{D}$1), the bound $q \in \mathcal{D}$2, and exponential-weights log-loss mixability. Selecting $q \in \mathcal{D}$3 produces the displayed regret rate.

4. Log-Optimal Sequential $q \in \mathcal{D}$4-to-$q \in \mathcal{D}$5 Calibration

An application of online monotone density estimation appears in sequential hypothesis testing for calibration of $q \in \mathcal{D}$6-values to $q \in \mathcal{D}$7-values. Under the null hypothesis, the $q \in \mathcal{D}$8-values $q \in \mathcal{D}$9 are sequentially super-uniform: $$\mathcal{D} = \left\{ f : [0,1] \to [0, \infty) \,\Bigg|\, \int_0^1 f(u)\,du = 1,\, f\ \text{nonincreasing}\ \right\}.$$0. A $$\mathcal{D} = \left\{ f : [0,1] \to [0, \infty) \,\Bigg|\, \int_0^1 f(u)\,du = 1,\, f\ \text{nonincreasing}\ \right\}.$$1-to-$$\mathcal{D} = \left\{ f : [0,1] \to [0, \infty) \,\Bigg|\, \int_0^1 f(u)\,du = 1,\, f\ \text{nonincreasing}\ \right\}.$$2 calibrator is a nonincreasing function $$\mathcal{D} = \left\{ f : [0,1] \to [0, \infty) \,\Bigg|\, \int_0^1 f(u)\,du = 1,\, f\ \text{nonincreasing}\ \right\}.$$3 with $$\mathcal{D} = \left\{ f : [0,1] \to [0, \infty) \,\Bigg|\, \int_0^1 f(u)\,du = 1,\, f\ \text{nonincreasing}\ \right\}.$$4, so that $$\mathcal{D} = \left\{ f : [0,1] \to [0, \infty) \,\Bigg|\, \int_0^1 f(u)\,du = 1,\, f\ \text{nonincreasing}\ \right\}.$$5 is a valid $$\mathcal{D} = \left\{ f : [0,1] \to [0, \infty) \,\Bigg|\, \int_0^1 f(u)\,du = 1,\, f\ \text{nonincreasing}\ \right\}.$$6-value. Admissible calibrators correspond exactly to monotone densities on $$\mathcal{D} = \left\{ f : [0,1] \to [0, \infty) \,\Bigg|\, \int_0^1 f(u)\,du = 1,\, f\ \text{nonincreasing}\ \right\}.$$7.

Under an alternative $$\mathcal{D} = \left\{ f : [0,1] \to [0, \infty) \,\Bigg|\, \int_0^1 f(u)\,du = 1,\, f\ \text{nonincreasing}\ \right\}.$$8 with density $$\mathcal{D} = \left\{ f : [0,1] \to [0, \infty) \,\Bigg|\, \int_0^1 f(u)\,du = 1,\, f\ \text{nonincreasing}\ \right\}.$$9, the log-optimal calibrator is the true density: $\ell_t(f) = -\log f(X_t)$0. Thus, estimating the log-optimal $\ell_t(f) = -\log f(X_t)$1-to-$\ell_t(f) = -\log f(X_t)$2 calibrator corresponds to online monotone density estimation of $\ell_t(f) = -\log f(X_t)$3.

An empirical adaptive procedure: at each $\ell_t(f) = -\log f(X_t)$4, set $\ell_t(f) = -\log f(X_t)$5 (OG or EA) applied to the observed $\ell_t(f) = -\log f(X_t)$6, and define the $\ell_t(f) = -\log f(X_t)$7-process $\ell_t(f) = -\log f(X_t)$8.

Theorem 3.1 (Asymptotic log-optimality).

If $\ell_t(f) = -\log f(X_t)$9 are i.i.d. $$L(\hat{f}, n) = \sum_{t=1}^n \ell_t(\hat{f}_t) = -\sum_{t=1}^n \log \hat{f}_t(X_t)$$0 with $$L(\hat{f}, n) = \sum_{t=1}^n \ell_t(\hat{f}_t) = -\sum_{t=1}^n \log \hat{f}_t(X_t)$$1, then for either OG- or EA-based calibrator,

$$L(\hat{f}, n) = \sum_{t=1}^n \ell_t(\hat{f}_t) = -\sum_{t=1}^n \log \hat{f}_t(X_t)$$2

Corollary (Consistency of $$L(\hat{f}, n) = \sum_{t=1}^n \ell_t(\hat{f}_t) = -\sum_{t=1}^n \log \hat{f}_t(X_t)$$3-process).

If $$L(\hat{f}, n) = \sum_{t=1}^n \ell_t(\hat{f}_t) = -\sum_{t=1}^n \log \hat{f}_t(X_t)$$4 and $$L(\hat{f}, n) = \sum_{t=1}^n \ell_t(\hat{f}_t) = -\sum_{t=1}^n \log \hat{f}_t(X_t)$$5, then $$L(\hat{f}, n) = \sum_{t=1}^n \ell_t(\hat{f}_t) = -\sum_{t=1}^n \log \hat{f}_t(X_t)$$6 grows linearly at rate $$L(\hat{f}, n) = \sum_{t=1}^n \ell_t(\hat{f}_t) = -\sum_{t=1}^n \log \hat{f}_t(X_t)$$7, implying that the sequential test rejects $$L(\hat{f}, n) = \sum_{t=1}^n \ell_t(\hat{f}_t) = -\sum_{t=1}^n \log \hat{f}_t(X_t)$$8 in finite time with probability 1.

5. Empirical Investigation

Numerical assessments are performed in both well-specified and mis-specified regimes:

• Stochastic setting ($$L(\hat{f}, n) = \sum_{t=1}^n \ell_t(\hat{f}_t) = -\sum_{t=1}^n \log \hat{f}_t(X_t)$$9, $t$00 replicates):
  • Densities: linear ($t$01), quadratic ($t$02), and a 4-bin piecewise constant.
  • Both OG and EA estimators closely track the oracle log-likelihood $t$03; EA demonstrates smaller finite-sample regret, particularly when $t$04 is milder (e.g., linear).
• Change-point (mis-specified) setting:
  • Data follows one linear monotone form for $t$05, then switches to another.
  • EA quickly adapts to the new regime via weight reallocation among experts—incurring substantially lower post-change regret—whereas OG, which relies on the full data history, adapts more slowly.

These experiments corroborate the theoretical $t$06 excess-risk for OG and $t$07 pathwise regret for EA. A plausible implication is that expert aggregation offers practical advantages in nonstationary environments.

6. Summary Table: Algorithmic and Statistical Properties

Estimator	Stochastic (i.i.d.) Excess Risk	Adversarial Regret Bound
Online Grenander	$t$08 (Theorem 2.1)	Not minimized; recomputes MLE
Expert Aggregation	$t$09 (via expert approximation)	$t$10 (Theorem 2.2)

Both algorithms support online monotone density estimation, but EA provides tighter adversarial guarantees and superior adaptivity in nonstationary or change-point settings.

7. Context and Significance

Online monotone density estimation extends classical nonparametric MLEs to the sequential, potentially adversarial context, providing both theoretical guarantees in KL-risk and worst-case regret. Its connection to sequential calibration for $t$11-to-$t$12 procedures emphasizes the foundational role of monotonicity constraints in adaptive hypothesis testing. The link to exponential weighting and mixture-of-experts methods also places this problem at the intersection of shape-constrained inference, online learning, and sequential testing theory. The approach synthesizes classical statistical theory (Grenander estimator, entropy arguments) with modern online learning methods (expert aggregation, mixability) (Hore et al., 9 Feb 2026).