Endpoint-Only Estimation of Occupation Densities for Self-Similar Markov Processes via a Mellin Poisson Identity

Suppose we observe where a random process started and where it was stopped, and nothing in between. We do not know the rule that stopped it, and that rule may depend on the path, for example "stop the first time the process falls 30% below its running maximum." Can we still learn how much time the process spent at each level before stopping? For processes that look the same under rescaling of space and time, we argue the answer is yes, and that this is easier than it sounds. The Mellin transform, the multiplicative analogue of the Fourier transform that turns scalings into simple factors, converts the governing equation into a division by a known function. That division is well-posed. The quantity we propose to recover is the expected occupation density $H$: the expected time spent near each level before stopping. Recovering $H$ from endpoint samples should achieve parametric $n^{-1/2}$ rates for self-similar diffusions. It should hold whether or not the stopping time is independent of the path.

This matters because the existing "statistical Skorokhod embedding" methods have two limitations. They solve a harder and different problem, recovering the full law of a random clock. They also assume that clock is independent of the process, which is exactly what a real stopping time (a rule that decides when to stop using only the past) is not. Statistical Skorohod embedding problem: Optimality and asymptotic normality recovers the density of $T$ from $X = \sqrt{T}B_1$ by Mellin deconvolution and obtains only $n^{-2\beta/(\pi+2\beta)}$ or $\log^{-\beta} n$ minimax rates. Statistical Skorohod embedding problem and its generalizations extends this to Lévy processes with combined Mellin and Laplace transforms. Both require $T$ to be independent of the process. We change the estimand rather than the transform. The occupation density is what the Poisson-equation theory of embeddings actually characterizes, and it remains identifiable under path-dependent stopping.

Core Claim

The setting is a positive self-similar Markov process. This is a process on $(0,\infty)$ whose law is unchanged when space is multiplied by $c$ and time by $c^{\alpha}$. Bessel processes, which describe the distance of a Brownian particle from the origin, are examples, as are geometric Brownian motion and Lamperti-stable processes with two-sided jumps. Every such process is the exponential of a Lévy process $\xi$ run on a random clock. Let $\psi$ be the Laplace exponent of $\xi$, defined by $\mathbb{E}e^{z\xi_t} = e^{t\psi(z)}$. The generator $A$ then acts on powers as $A x^{z} = \psi(z)x^{z-\alpha}$. Combined with Dynkin's formula, this gives a "Mellin Dynkin identity" valid for any integrable stopping time $\tau$, path-dependent or not, for all $z$ in a strip determined by $\psi$:

$$\mathbb{E}[X_\tau^{z}] - \mathbb{E}[X_0^{z}] = \psi(z)\,\mathbb{E}\int_0^\tau X_u^{z-\alpha}\,du .$$

Write $h_0, h_1$ for the start and stop densities, $f^*(s)=\int_0^\infty x^{s-1}f(x)\,dx$ for the Mellin transform, and $H$ for the expected occupation density. The identity then becomes an algebraic solution of the adjoint Poisson equation $A^*H = h_1 - h_0$ from On Skorokhod embeddings and Poisson equations:

$$H^*(w) = \frac{h_1^*(w+\alpha) - h_0^*(w+\alpha)}{\psi(w+\alpha-1)} .$$

That paper proves that a finite-mean embedding exists if and only if a nonnegative integrable $H$ solves this equation. It gives explicit criteria via Fourier transforms only for Lévy processes, and it leaves general Markov processes non-constructive. Our formula is the self-similar counterpart of its Fourier condition. Two consequences make the formula informative rather than cosmetic.

First, the zeros of $\psi$ must be zeros of the numerator, which yields necessary moment-matching conditions. For a Bessel process of dimension $\delta$, $\psi(z)=z^2/2+(\delta/2-1)z$. The root at $z = 2-\delta$ forces $\mathbb{E}X_\tau^{2-\delta}=\mathbb{E}X_0^{2-\delta}$, which is the martingale condition on the scale function. This plays the role that the mean-matching condition plays for Brownian motion.

Second, along a vertical line, the empirical Mellin transform of a positive sample has variance bounded uniformly in the imaginary part, while $|\psi|$ grows like $|t|^{\beta_\psi}$, with $\beta_\psi = 2$ for diffusions. Mellin inversion of the estimated $H^*$ therefore needs no regularization when $\beta_\psi>1$. In that case the pointwise error is $O(n^{-1/2})$, because the integral of $n^{-1/2}|t|^{-\beta_\psi}$ converges. For jump processes with $\beta_\psi\le 1$, a mild frequency cutoff is required. This is the reverse of the severely ill-posed division by a gamma function in the clock-recovery problem.

Research Program

The theoretical part has three pieces. The first is a precise strip of validity for the identity, together with integrability conditions under which the inverted $H$ is the true occupation density. The analysis must include processes that hit zero, such as Bessel processes with $\delta<2$, and target laws with an atom at zero, where log-based reductions break. The second is rates for the estimator $\hat H$, obtained by plugging empirical Mellin transforms of the start and stop samples into the second display and inverting along a line $\mathrm{Re}\,w=c$. We will prove asymptotic normality in the spirit of the confidence-interval results of Statistical Skorohod embedding problem: Optimality and asymptotic normality. Error bounds will follow the Mellin Berry–Esseen route of Statistical Inference for Scale Mixture Models via Mellin Transform Approach, which we reuse as method. The third is a data-driven test of embeddability, built on the condition $H\ge 0$ plus the root conditions. Its null is "these start and stop samples are consistent with some finite-mean stopping rule for this process."

The payoff we target is concrete. From endpoint samples alone we can estimate $\mathbb{E}\tau = H^*(1)$ and $\mathbb{E}\int_0^\tau g(X_u)\,du$ for any $g$, such as expected time spent below a threshold. The estimates apply to processes with two-sided jumps, where the excursion constructions of On an Explicit Skorokhod Embedding for Spectrally Negative Lévy Processes and An explicit Skorokhod embedding for functionals of Markovian excursions do not apply. The method does not recover the law of $\tau$. We state this limit openly: $H$ is exactly what the endpoint laws identify, while the law of $\tau$ is not identified once the stopping rule is allowed to depend on the path.

Baselines

The naive reference point is the single-moment Wald-type identity. For Brownian-like cases one estimates $\mathbb{E}\tau$ from $\mathbb{E}X_\tau^2-\mathbb{E}X_0^2$, and in general from our first display at the one value of $z$ that isolates $\tau$. This baseline uses the same samples and is exact for $\mathbb{E}\tau$. Our method should match it on $\mathbb{E}\tau$ and add the whole occupation profile. Beating it means accurate recovery of $H$ and of threshold occupation times, which the single moment cannot provide.

The strongest existing competitor is the Mellin clock-deconvolution estimator of Statistical Skorohod embedding problem: Optimality and asymptotic normality for Bessel and Brownian cases, and of Statistical Skorohod embedding problem and its generalizations for jump cases. We run it on identical samples. We tune its bandwidth by the same held-out criterion and grid size we allow ourselves, and convert its estimated clock law into $H$ via $H(x)=\int p_t(x)\,\mathbb{P}(T>t)\,dt$. The file reports rates for it, $n^{-2\beta/(\pi+2\beta)}$ and $O(\log^{-\beta} n)$, and qualitative Gamma and generalized-hyperbolic simulations. No reported numerical error values were found on Lacuna for this estimator. We therefore regenerate its numbers ourselves. Beating it means lower error on $H$ under independent clocks at matched $n$, and a widening gap under path-dependent stopping, where its model is misspecified.

The ablation that isolates our key ingredient is the Lamperti route. We map to $\log X$, apply the Fourier criterion of On Skorokhod embeddings and Poisson equations to $\xi$, and in one variant drop the $x^{\alpha}$ weight, i.e. the shift by $\alpha$. The variant without the weight estimates occupation in the wrong clock. The weighted variant evaluated on the shifted complex line is algebraically equivalent to ours. We compare both, including on targets with an atom at zero. Finally, the oracle is the empirical occupation density computed from full simulated paths. It sets the ceiling.

Why not just…?

Why not just take logarithms and use the Lévy-process Fourier result? On the interior of the strip, the correctly weighted version is the same computation in different coordinates. If our experiments show it matches our estimator everywhere, we will say so. In that case the contribution is the identity, the choice of estimand, and the rates, not the Mellin transform per se. The routes differ in practice in two places. One is processes that hit zero and targets with mass at zero, where $\log X$ is undefined. The other is the shifted line needed to count time in the original clock, which naive log-Fourier implementations omit. In the multiplicative-deconvolution setting, Statistical Inference for Scale Mixture Models via Mellin Transform Approach reports the Mellin approach to be 60 to 1,000 times more accurate than a Fourier alternative under two-point discrete noise. That supports our expectation but does not prove it in our setting.

Why not just use the existing statistical Skorokhod estimator? It targets the clock law. It assumes independence between clock and path, and its rates are slow because it divides by a rapidly decaying Mellin transform. Under barrier or drawdown stopping its model is wrong. Our path-dependent experiments are designed to show the resulting bias directly.

Why not just learn the generator from data? Resolvent regression in Laplace Transform Based Low-Complexity Learning of Continuous Markov Semigroups needs sampled trajectories. In our regime only endpoints are observed and the generator family is known up to $\psi$, so it does not apply. It does supply a useful oracle variant when paths are available.

Why not just solve the free-boundary PDE? PDE methods for optimal Skorokhod embeddings computes optimal barriers given known measures with $L^2$ densities on bounded domains. Our problem is inference of $H$ from samples with unknown stopping rules. We will still use PDE-computed Root barriers to generate test stopping rules.

Proposed Experiments

We simulate Bessel processes with $\delta=3$ and $\delta=1.5$ (the latter hits zero), geometric Brownian motion, and a Lamperti-stable process with two-sided jumps. Each is stopped by four rules. The first is an independent Gamma(2,2) clock, matching the benchmark used in the cited statistical embedding papers. The others are a Root-type space-time barrier, an Azéma–Yor-style drawdown stop in the spirit of An iterated Azéma-Yor type embedding for finitely many marginals, and a mixture of the two. Sample sizes run from $n=100$ to $10^5$. Metrics are $L^2$ and sup error of $\hat H$ against the full-path oracle, error in $\mathbb{E}\tau$ and in occupation below fixed thresholds, coverage of asymptotic confidence intervals, and size and power of the embeddability test against deliberately non-embeddable start/stop pairs.

Three checks are decisive. The first is whether $\hat H$ attains slope $-1/2$ on a log-log error plot for the Bessel cases without regularization, while the clock-deconvolution baseline does not. The second is whether our error is unchanged when switching from independent clocks to barrier stopping, while the baseline's error stops decreasing with $n$. The third is whether the Lamperti log-Fourier ablation is indistinguishable from our estimator for $\delta=3$ but degrades for $\delta=1.5$ with mass at zero. If the third check shows no difference anywhere, we will reframe the transform choice as convenience and rest the contribution on the first two.