A random walk or Lévy process (a random path built from independent increments, possibly with jumps) has a characteristic exponent, a single function that encodes its whole law. The Wiener–Hopf factorisation splits this function into two pieces. One piece describes how the path climbs to new all-time highs, and the other describes how it falls to new lows. These pieces govern practical quantities such as the probability that a noisy baseline ever exceeds a trigger threshold. A recent theorem shows that for processes that run forever, with no random "killing" time, this split is unique up to a single positive scale constant. This note proposes to turn that theorem into a learning problem. We would fit the two factors from data using small networks whose rational activations are constrained to produce valid "ladder" exponents. We would then use bagging, meaning training many copies on bootstrap resamples of the data, to measure empirically whether the learned split is identifiable.
The idea has two uses. The practical one is estimating threshold-exceedance and overshoot laws from finite noisy data, which matters for background rejection in rare-event detectors. The mathematical one is probing a question that is open: whether the matrix version of the factorisation is unique for Markov-modulated processes that run forever. Wherever uniqueness is proven, bootstrap ensembles of constrained factor networks should agree after scale normalisation. Where they disagree while fitting the exponent equally well, we have a numerical counterexample candidate that can then be checked by hand.
Core Claim
For a Lévy process with characteristic exponent $\Psi(\theta)$, the spatial factorisation is
$$\Psi(\theta) = \kappa(-i\theta)\,\hat{\kappa}(i\theta),$$
where $\kappa$ and $\hat\kappa$ are the Laplace exponents of the ascending and descending ladder height subordinators. A subordinator is an increasing Lévy process; here it records the successive record highs, or record lows. The uniqueness of the Wiener-Hopf factorisation of Lévy processes and random walks proves that without killing, any pair of subordinator exponents satisfying this identity equals the ladder pair up to $\kappa \to c\kappa$, $\hat\kappa \to \hat\kappa/c$. The proof uses tempered distributions rather than Liouville's theorem, and treats lattice and non-lattice cases separately. That paper is purely analytical. It motivates our work but supplies no estimator.
The claim is that uniqueness is exactly an identifiability statement for a learning problem. It holds only if the model class is restricted to genuine subordinator exponents, which are Bernstein functions. Unconstrained function approximators admit infinitely many spurious splits, because one can multiply one factor by any non-vanishing function and divide the other by it.
We therefore predict three things. Constrained factor networks trained on bootstrap resamples will converge to a common scale-normalised factor, with spread shrinking as sample size grows. Unconstrained networks will fit $\Psi$ equally well yet disagree. And this contrast gives a working diagnostic that can be pointed at the unproven Markov-additive case.
Research Program
Parameterisation. Rational functions supply the method. Padé Activation Units learn activations as ratios of trainable polynomials, with a "safe" denominator to avoid poles. Rational Neural Networks have Expressivity Advantages proves that rational networks approximate kinks and non-smooth targets with exponentially fewer parameters than smooth ones. We specialise this idea rather than reuse the generic activation. Complete Bernstein functions, an important subclass of subordinator exponents, have Stieltjes-type representations. A finite version is a rational function,
$$\kappa(\lambda) = a + b\lambda + \sum_{k=1}^{K} w_k \frac{\lambda}{\lambda + t_k},$$
where $a, b, w_k \ge 0$ and $t_k > 0$ are learned and $K$ is the number of poles. By construction every pole sits on the negative real axis.
Unlike PAUs, this rational unit is not an activation inside a deep network. It is the entire model of each factor, and its positivity constraints are what make the learning problem identifiable. For general Bernstein functions outside the complete class, we add a monotone mixture of $1-e^{-s\lambda}$ terms with non-negative weights. We report separately how much each component is used.
Objective. From observed increments we form the empirical characteristic exponent $\hat\Psi(\theta)$ on a frequency window where the empirical characteristic function is well estimated. For random walks we use the analogous factorisation of $1-\hat\phi(\theta)$, with $\hat\phi$ the empirical characteristic function of one step. We then minimise a weighted squared residual between $\hat\Psi$ and $\kappa(-i\theta)\hat\kappa(i\theta)$. The scale ambiguity is fixed by normalising $\kappa(1)=1$ after training, so the theorem's constant $c$ is removed before any comparison.
Bagging as a measurement instrument. An Empirical Study of Bagging Predictors for Different Learning Algorithms establishes bagging as a variance reducer for unstable learners. We use it mainly to separate two sources of disagreement.
The first source is optimisation variance: the same data with different initialisations. Rational fits are known to be brittle in exactly this way (Padé Activation Units). The second source is data variance: different bootstrap resamples. We borrow the self-consistency idea of Arbitrariness and Social Prediction, which measured how often independently trained models agree on an instance in tabular fairness data. We carry it over from instance labels to whole functions. Our self-consistency score is the fraction of bags whose normalised $\kappa$ lies within a tolerance band of the ensemble median on a grid. The median-aggregation choice follows the "bragging" used by Ensemble-SINDy. That work used bootstrap inclusion probabilities to separate real terms from noise artefacts in sparse equation discovery over fixed libraries. We apply the same logic to a continuous function class whose identifiability is guaranteed by a theorem.
Markov-additive probe. Markov additive friendships and its companion Markov additive friendships extend factorisation and inverse construction to Markov additive processes. These are Lévy-type processes whose parameters switch according to a hidden Markov chain. The papers introduce $\pi$-friendship and $\pi$-fellowship conditions involving the chain's stationary distribution $\pi$. They state that uniqueness for unkilled processes of this kind requires additional conditions and remains open in general. The same is true for non-irreducible modulating chains.
We replace each scalar factor by a matrix of rational Bernstein-type entries, following the matrix form in those papers. We then run the same bootstrap-and-seed protocol on unkilled examples. A persistent split into two or more tight clusters of equally good fits would flag a candidate non-uniqueness, which we would then hand-verify. A single cluster would count as evidence, though not proof, for uniqueness in that regime.
Detector application. Signal-to-background discrimination supplies the downstream stakes. The probability that a negatively drifting noise trace ever exceeds a level $x$ is determined by the ascending factor. This is one ingredient of spurious-trigger and anomaly rates of the kind that Machine learning-powered data cleaning for LEGEND reports as leakage, up to 15.034±0.224% on high-rate test-stand data. We treat this as a secondary case study on baseline segments of germanium detector traces, and only if such data can be obtained. If those traces are not well described by a random walk (for instance, if they have strong autocorrelation), we report that as a limitation rather than force the fit.
Baselines
The naive reference point is an unconstrained factorisation. Two small multilayer perceptrons with fixed activations are trained to reproduce $\hat\Psi$ as a product, with no positivity or Bernstein structure. They receive the same data, optimiser, parameter count within 20%, and tuning budget of an equal number of hyperparameter trials. This baseline should match our residual on $\Psi$ while failing to agree with itself across bags. Beating it means our method has comparable residual but substantially higher self-consistency and lower error against the true factor.
The strongest existing competitor is the plug-in Spitzer–Baxter estimator. This is the classical fluctuation identity that expresses the ladder exponent as an exponential of an integral against the marginal laws of the partial sums. We evaluate it with the empirical distribution of partial sums from the same sample. It is consistent, needs no training, and uses all the data, which is why it is the method to beat. We give it the same data and choose its truncation and smoothing parameters on the same validation split. No reported result found on Lacuna for any estimator of Wiener–Hopf factors from data, so its accuracy must be established here. Beating it means lower mean sup-norm error of the normalised $\kappa$, and of the tail of the supremum, at small and moderate sample sizes, with the gap closing as data grows.
A third baseline reads ladder heights directly off the observed path. It records each new maximum and fits the jumps between successive records with the same rational Bernstein family. This is also given matched data and tuning budget.
Two ablations isolate the key ingredients. The first swaps the rational Stieltjes parameterisation for a monotone neural Bernstein parameterisation with fixed smooth activations and the same constraints, which tests whether rationality rather than constraints drives accuracy. The second compares a single model against the bagged median at equal total compute, which tests whether bagging adds accuracy or only diagnostics.
Why not just…?
Why not just use the Spitzer–Baxter plug-in formula? It is a real threat. For large samples it may be as good as anything we build, and we expect it to win asymptotically in some regimes. Its weakness is structural. The output is not guaranteed to be a valid Bernstein function at finite sample size, it extrapolates poorly beyond the frequencies the data resolves, and it offers no way to probe uniqueness in the Markov-additive case where no analogous formula is settled. The sample-size sweep in the experiments settles the accuracy question directly. If the plug-in estimator matches us at all sample sizes on scalar processes, the contribution narrows to the identifiability diagnostic and the matrix probe, and we will say so.
Why not just read ladder heights off the path? For a walk drifting downward there are only a handful of record highs. This estimator therefore throws away almost every increment, and its variance should be large exactly in the regime relevant to threshold exceedance. It also yields the ascending factor alone, not both jointly. The ladder baseline tests whether this intuition holds.
Why not just fit a parametric process, such as Brownian motion plus two-sided exponential jumps, and factor it analytically? When the parametric family is correct this is optimal, and our synthetic benchmarks include such cases so the comparison is fair. When the family is wrong, the factors inherit the misspecification. We will include misspecified cases, such as heavier-tailed jumps, to measure how much this costs.
Why not just add rational activations and bagging to a pulse-shape classifier? This is the obvious combination suggested by the literature. Deep learning based pulse shape discrimination for germanium detectors already reaches 90% background rejection at 90% signal acceptance with a 141-parameter classifier on an autoencoder code, matching the hand-crafted current-to-energy cut. The hand-crafted cuts themselves are strong: 100% surface alpha rejection at 0.2% bulk loss in α-event characterization, and 6% single-escape-peak survival in Multisite event discrimination. The headroom is small and the result would say nothing about factorisation. We therefore deliberately do not make this our core study.
Proposed Experiments
The primary benchmarks are synthetic, because they have known answers. Brownian motion with drift and the Laplace-increment random walk have closed-form factors. Brownian motion with two-sided exponential jumps has a rational exponent whose factors follow from its roots. The simple ±1 random walk tests the lattice case, which the uniqueness proof treats separately. Two-state Markov-modulated Brownian motion, with factors computed numerically by standard matrix-analytic routines, covers the matrix case.
For each process we sweep sample sizes over roughly two orders of magnitude and use at least 50 bootstrap bags crossed with 5 seeds. We report four quantities: residual on $\Psi$, sup-norm error of the normalised $\kappa$, error in $P(\sup_t X_t > x)$ over a range of $x$, and functional self-consistency.
The decisive check is the contrast between the constrained and unconstrained factorisations. For the constrained model, the spread across bags should shrink roughly as $n^{-1/2}$, the rate inherited from the empirical characteristic function. Across seeds on fixed data it should be near zero once the scale is normalised. The unconstrained model should show a residual on $\Psi$ indistinguishable from ours with no such shrinkage. If the constrained model also fails to agree across seeds on scalar processes where uniqueness is proven, then the optimisation is the bottleneck, not the mathematics. In that case the Markov-additive probe is uninterpretable and we stop there.
If the check passes, we run the probe on unkilled Markov-additive examples. These include cases with jumps at modulator switches and a reducible modulating chain, both flagged as hard in Markov additive friendships. We report any cluster structure in the learned factor matrices.
The detector case study comes last. It asks whether the learned ascending factor predicts observed exceedance rates on held-out baseline segments better than the plug-in estimator and than direct Monte Carlo counting at matched data size.
References
- The uniqueness of the Wiener–Hopf factorisation of Lévy processes and random walks
- Padé Activation Units: End-to-end Learning of Flexible Activation Functions in Deep Networks
- Rational Neural Networks have Expressivity Advantages
- An Empirical Study of Bagging Predictors for Different Learning Algorithms
- Arbitrariness and Social Prediction: The Confounding Role of Variance in Fair Classification
- Ensemble-SINDy: Robust sparse model discovery in the low-data, high-noise limit, with active learning and control
- Markov additive friendships
- Machine learning-powered data cleaning for LEGEND: a semi-supervised approach using affinity propagation and support vector machines
- Deep learning based pulse shape discrimination for germanium detectors
- $$\alpha $$-event characterization and rejection in point-contact HPGe detectors
- Multisite event discrimination for the majorana demonstrator