PWS: Path Weight Sampling Techniques

• Path Weight Sampling (PWS) is a family of probabilistic methods that assigns weights to complete paths—be they stochastic trajectories or network paths—to facilitate estimation, compression, or variance reduction.
• PWS techniques span exact Monte Carlo estimation for mutual information, transfer entropy computations, and sparse approximants in deep networks through importance sampling and thermodynamic integration.
• In neural architecture search, PWS is applied to reduce supernet gradient variance and optimize training by sampling diverse architecture paths based on gradient-driven or uniform distributions.

Path Weight Sampling (PWS) denotes a family of path-based probabilistic methods rather than a single universally standardized technique. In one prominent usage, PWS is an exact Monte Carlo framework for computing mutual information between stochastic input and output trajectories by evaluating conditional path probabilities from a master equation and averaging in trajectory space (Reinhardt et al., 2022). In another, it is a path-based probabilistic sparsification method for deep ReLU and, more generally, positive homogeneous networks, where path weights define sampling distributions used to construct sparse approximants and derive covering-number and generalization bounds (Theisen et al., 2019). Closely related work in weight-sharing neural architecture search uses non-uniform, gradient-based path sampling as a principled path-weighted sampling mechanism, even when the acronym itself is not explicit (Lu et al., 2023). Across these literatures, the common structural idea is to assign weights to complete paths—trajectories, network paths, or architecture paths—and to use those weights for estimation, compression, or variance reduction.

1. Terminological scope and domain-specific meanings

The term “Path Weight Sampling” is polysemous in the recent arXiv literature. In stochastic information processing, it names a method for exact computation of mutual information between time-varying signals by sampling whole trajectories and evaluating their path probabilities (Reinhardt et al., 2022). In deep learning theory, it refers to sampling input–output paths of positive homogeneous networks according to distributions induced by path weights, with the goal of constructing sparse approximants and controlling capacity by global norms of products of weights (Theisen et al., 2019). In one-shot NAS, the same phrase is not always used, but the underlying mechanism is present whenever paths through a supernet are sampled according to explicitly assigned path weights, whether those weights are uniform as in Single Path One-Shot or proportional to gradient-based importance indicators as in PA&DA (Guo et al., 2019, Lu et al., 2023).

This heterogeneity is substantive rather than merely lexical. The information-theoretic lineage treats a path as a stochastic trajectory in continuous time or discrete time. The deep-network lineage treats a path as a sequence of units or edges from an input coordinate to an output unit. NAS treats a path as a sampled architecture or subgraph inside a weight-sharing supernet. A plausible implication is that “PWS” is best understood as a methodological schema—sampling over whole paths with nontrivial weights—rather than as a single algorithmic object.

The acronym is also domain-dependent. In differential privacy, “PWS” denotes “Private Weighted Sampling,” not Path Weight Sampling, and concerns differentially private weighted samples of keys by frequency rather than trajectory or architecture path measures (Cohen et al., 2020). Any encyclopedia treatment therefore has to distinguish the path-based usages from unrelated expansions of the same acronym.

2. Exact Monte Carlo PWS for stochastic trajectories

In the information-theoretic formulation, PWS addresses the mutual information between input and output trajectories of a stochastic dynamical system described by a master equation (Reinhardt et al., 2022). If the input is a time-varying signal trajectory $\bm{s}$ and the output is a stochastic trajectory $\bm{x}$ of a continuous-time Markov jump process, the core object is the conditional path probability

$$\mathcal{P}[\bm{x}|\bm{s}] = \mathrm{P}(x_0|s_0) \left( \prod_{i=1}^{n-1} Q_{t_i}(x_i,x_{i-1};\bm{s}) \right) \left( \prod_{i=1}^{n} \exp\!\left[\int_{t_{i-1}}^{t_i} dt\, Q_t(x_{i-1},x_{i-1};\bm{s}) \right] \right),$$

where the jump factors contribute instantaneous transition rates and the waiting-time factors contribute survival probabilities. In log form, this becomes a path action built from the initial-state term, jump-rate terms, and time integrals of the diagonal generator.

The mutual information between input and output trajectories can then be written as

$$\mathrm{I}(\mathcal{S},\mathcal{X}) = \left\langle \ln \frac{\mathcal{P}[\bm{x}|\bm{s}]}{\mathcal{P}[\bm{x}]} \right\rangle_{\mathcal{P}[\bm{s},\bm{x}]},$$

so the computational problem reduces to evaluating $\ln \mathcal{P}[\bm{x}|\bm{s}]$ and the marginal output path probability $\mathcal{P}[\bm{x}]$ for sampled pairs $(\bm{s},\bm{x})$. The conditional term is exact once the master equation is known; the difficulty is the path-space marginalization

$$\mathcal{P}[\bm{x}] = \int \mathcal{D}[\bm{s}]\, \mathcal{P}[\bm{s}]\,\mathcal{P}[\bm{x}|\bm{s}].$$

Three variants of PWS are given. Direct PWS estimates $\mathcal{P}[\bm{x}]$ by brute-force Monte Carlo averaging over independently sampled input trajectories. Rosenbluth–Rosenbluth PWS interprets the marginalization as a partition-function or free-energy computation in trajectory space, segments time, and uses sequential importance sampling with pruning and enrichment of partial signal trajectories. Thermodynamic Integration PWS introduces an auxiliary interpolation parameter $\theta$ between a reference path ensemble and the full interacting ensemble, and computes $\bm{x}$0 by trajectory-space MCMC and numerical integration over $\bm{x}$1. All three generate trajectories with the standard stochastic simulation algorithm and differ only in how the marginal path weight is computed.

Within this formulation, PWS is “exact” in the Monte Carlo sense: the only approximation is statistical error from finite sampling, not a closure approximation or a time discretization of the underlying master equation dynamics. The method also extends to systems with hidden internal states and to systems with feedback from output to input, provided the relevant joint or conditional path probabilities can be evaluated from the model (Reinhardt et al., 2022).

3. Transfer entropy, learned generative models, and later extensions

The same path-weight formalism was extended from path mutual information to directed information measures in “Exact computation of Transfer Entropy with Path Weight Sampling” (Das et al., 2024). In discrete time, the transfer entropy from $\bm{x}$2 to $\bm{x}$3 is written as

$\bm{x}$4

and equivalently as an average of log-likelihood ratios of conditional transition probabilities. For jump processes, the paper expresses transfer-entropy increments in terms of marginal jump propensities $\bm{x}$5 and $\bm{x}$6 and corresponding escape propensities $\bm{x}$7 and $\bm{x}$8, all obtained by path-space marginalization over hidden trajectories. TE-PWS then uses importance sampling and Rosenbluth–Rosenbluth resampling to generate conditional ensembles of hidden trajectories and to compute these marginalized quantities exactly for the specified stochastic model. The paper states that TE-PWS makes it possible, for the first time, to quantify transfer entropy and its variants exactly for any stochastic model, including those with multiple hidden variables, nonlinearity, transient conditions, and feedback (Das et al., 2024).

A second extension, “ML-PWS: Estimating the Mutual Information Between Experimental Time Series Using Neural Networks,” addresses the main limitation of classical PWS, namely its dependence on a known mechanistic model (Reinhardt et al., 22 Aug 2025). There the information transmission rate is defined as

$\bm{x}$9

and the conditional path distribution is factorized autoregressively as

$$\mathcal{P}[\bm{x}|\bm{s}] = \mathrm{P}(x_0|s_0) \left( \prod_{i=1}^{n-1} Q_{t_i}(x_i,x_{i-1};\bm{s}) \right) \left( \prod_{i=1}^{n} \exp\!\left[\int_{t_{i-1}}^{t_i} dt\, Q_t(x_{i-1},x_{i-1};\bm{s}) \right] \right),$$0

The method learns this conditional trajectory model from data using neural generative models—GRU-based autoregressive Gaussian models in the synthetic nonlinear benchmark and causal convolutional Poisson models for neuronal spike counts—and then applies PWS on the learned model. A second neural network approximates the posterior $$\mathcal{P}[\bm{x}|\bm{s}] = \mathrm{P}(x_0|s_0) \left( \prod_{i=1}^{n-1} Q_{t_i}(x_i,x_{i-1};\bm{s}) \right) \left( \prod_{i=1}^{n} \exp\!\left[\int_{t_{i-1}}^{t_i} dt\, Q_t(x_{i-1},x_{i-1};\bm{s}) \right] \right),$$1 to reduce the variance of the importance-sampling estimator for the marginal path probability. On synthetic time series generated from a nonlinear model, ML-PWS closely matches ground-truth results obtained by applying PWS directly to the model; on salamander retinal ganglion cell recordings, the inferred information transmission rates are about $$\mathcal{P}[\bm{x}|\bm{s}] = \mathrm{P}(x_0|s_0) \left( \prod_{i=1}^{n-1} Q_{t_i}(x_i,x_{i-1};\bm{s}) \right) \left( \prod_{i=1}^{n} \exp\!\left[\int_{t_{i-1}}^{t_i} dt\, Q_t(x_{i-1},x_{i-1};\bm{s}) \right] \right),$$2 higher than previous decoding-based estimates, which are lower bounds (Reinhardt et al., 22 Aug 2025).

These extensions preserve the core PWS architecture: whole-trajectory sampling, exact or learned conditional path likelihoods, and explicit path-weight averaging in trajectory space. They also broaden the scope from mutual information between trajectories to transfer entropy, directed information, filtered transfer entropy, and empirical time-series settings.

4. Path distributions, sparse approximants, and capacity control in deep networks

In deep-network theory, PWS takes a very different form. For an $$\mathcal{P}[\bm{x}|\bm{s}] = \mathrm{P}(x_0|s_0) \left( \prod_{i=1}^{n-1} Q_{t_i}(x_i,x_{i-1};\bm{s}) \right) \left( \prod_{i=1}^{n} \exp\!\left[\int_{t_{i-1}}^{t_i} dt\, Q_t(x_{i-1},x_{i-1};\bm{s}) \right] \right),$$3-layer positive homogeneous network with path $$\mathcal{P}[\bm{x}|\bm{s}] = \mathrm{P}(x_0|s_0) \left( \prod_{i=1}^{n-1} Q_{t_i}(x_i,x_{i-1};\bm{s}) \right) \left( \prod_{i=1}^{n} \exp\!\left[\int_{t_{i-1}}^{t_i} dt\, Q_t(x_{i-1},x_{i-1};\bm{s}) \right] \right),$$4, the path weight is the product of weights along that path,

$$\mathcal{P}[\bm{x}|\bm{s}] = \mathrm{P}(x_0|s_0) \left( \prod_{i=1}^{n-1} Q_{t_i}(x_i,x_{i-1};\bm{s}) \right) \left( \prod_{i=1}^{n} \exp\!\left[\int_{t_{i-1}}^{t_i} dt\, Q_t(x_{i-1},x_{i-1};\bm{s}) \right] \right),$$5

and the network can be rewritten as a path-sum once signs are absorbed by the positive-homogeneous representation (Theisen et al., 2019). The basic 1-path variation is

$$\mathcal{P}[\bm{x}|\bm{s}] = \mathrm{P}(x_0|s_0) \left( \prod_{i=1}^{n-1} Q_{t_i}(x_i,x_{i-1};\bm{s}) \right) \left( \prod_{i=1}^{n} \exp\!\left[\int_{t_{i-1}}^{t_i} dt\, Q_t(x_{i-1},x_{i-1};\bm{s}) \right] \right),$$6

which induces the 1-path distribution

$$\mathcal{P}[\bm{x}|\bm{s}] = \mathrm{P}(x_0|s_0) \left( \prod_{i=1}^{n-1} Q_{t_i}(x_i,x_{i-1};\bm{s}) \right) \left( \prod_{i=1}^{n} \exp\!\left[\int_{t_{i-1}}^{t_i} dt\, Q_t(x_{i-1},x_{i-1};\bm{s}) \right] \right),$$7

More generally, for $$\mathcal{P}[\bm{x}|\bm{s}] = \mathrm{P}(x_0|s_0) \left( \prod_{i=1}^{n-1} Q_{t_i}(x_i,x_{i-1};\bm{s}) \right) \left( \prod_{i=1}^{n} \exp\!\left[\int_{t_{i-1}}^{t_i} dt\, Q_t(x_{i-1},x_{i-1};\bm{s}) \right] \right),$$8, the paper defines data-dependent $$\mathcal{P}[\bm{x}|\bm{s}] = \mathrm{P}(x_0|s_0) \left( \prod_{i=1}^{n-1} Q_{t_i}(x_i,x_{i-1};\bm{s}) \right) \left( \prod_{i=1}^{n} \exp\!\left[\int_{t_{i-1}}^{t_i} dt\, Q_t(x_{i-1},x_{i-1};\bm{s}) \right] \right),$$9-path variations $$\mathrm{I}(\mathcal{S},\mathcal{X}) = \left\langle \ln \frac{\mathcal{P}[\bm{x}|\bm{s}]}{\mathcal{P}[\bm{x}]} \right\rangle_{\mathcal{P}[\bm{s},\bm{x}]},$$0 and distributions

$$\mathrm{I}(\mathcal{S},\mathcal{X}) = \left\langle \ln \frac{\mathcal{P}[\bm{x}|\bm{s}]}{\mathcal{P}[\bm{x}]} \right\rangle_{\mathcal{P}[\bm{s},\bm{x}]},$$1

where $$\mathrm{I}(\mathcal{S},\mathcal{X}) = \left\langle \ln \frac{\mathcal{P}[\bm{x}|\bm{s}]}{\mathcal{P}[\bm{x}]} \right\rangle_{\mathcal{P}[\bm{s},\bm{x}]},$$2 depends on the dataset through an input-coordinate norm.

A key structural result is that $$\mathrm{I}(\mathcal{S},\mathcal{X}) = \left\langle \ln \frac{\mathcal{P}[\bm{x}|\bm{s}]}{\mathcal{P}[\bm{x}]} \right\rangle_{\mathcal{P}[\bm{s},\bm{x}]},$$3 has a Markov factorization across layers. This permits sampling $$\mathrm{I}(\mathcal{S},\mathcal{X}) = \left\langle \ln \frac{\mathcal{P}[\bm{x}|\bm{s}]}{\mathcal{P}[\bm{x}]} \right\rangle_{\mathcal{P}[\bm{s},\bm{x}]},$$4 paths from the path distribution, forming empirical marginals and conditionals, and constructing an empirical Markov chain $$\mathrm{I}(\mathcal{S},\mathcal{X}) = \left\langle \ln \frac{\mathcal{P}[\bm{x}|\bm{s}]}{\mathcal{P}[\bm{x}]} \right\rangle_{\mathcal{P}[\bm{s},\bm{x}]},$$5 that defines a sparse approximant network $$\mathrm{I}(\mathcal{S},\mathcal{X}) = \left\langle \ln \frac{\mathcal{P}[\bm{x}|\bm{s}]}{\mathcal{P}[\bm{x}]} \right\rangle_{\mathcal{P}[\bm{s},\bm{x}]},$$6. The resulting approximant has at most $$\mathrm{I}(\mathcal{S},\mathcal{X}) = \left\langle \ln \frac{\mathcal{P}[\bm{x}|\bm{s}]}{\mathcal{P}[\bm{x}]} \right\rangle_{\mathcal{P}[\bm{s},\bm{x}]},$$7 nonzero transition probabilities, hence at most $$\mathrm{I}(\mathcal{S},\mathcal{X}) = \left\langle \ln \frac{\mathcal{P}[\bm{x}|\bm{s}]}{\mathcal{P}[\bm{x}]} \right\rangle_{\mathcal{P}[\bm{s},\bm{x}]},$$8 nonzero weights. The approximation theorem states that for $$\mathrm{I}(\mathcal{S},\mathcal{X}) = \left\langle \ln \frac{\mathcal{P}[\bm{x}|\bm{s}]}{\mathcal{P}[\bm{x}]} \right\rangle_{\mathcal{P}[\bm{s},\bm{x}]},$$9,

$\ln \mathcal{P}[\bm{x}|\bm{s}]$0

where $\ln \mathcal{P}[\bm{x}|\bm{s}]$1 is a path-complexity term defined from the $\ln \mathcal{P}[\bm{x}|\bm{s}]$2-Rényi entropies of layerwise marginals. The same machinery yields a representer-set cardinality bound $\ln \mathcal{P}[\bm{x}|\bm{s}]$3, a covering-number bound, and a margin-based generalization bound of order

$\ln \mathcal{P}[\bm{x}|\bm{s}]$4

with $\ln \mathcal{P}[\bm{x}|\bm{s}]$5 itself controlled by norms of the product of absolute-value weight matrices rather than products of layerwise norms (Theisen et al., 2019).

This line of work therefore treats PWS as a path-based compression and capacity-analysis tool. Empirically, it reports that networks trained on easier tasks are much more compressible by path sampling than networks trained on random labels: with $\ln \mathcal{P}[\bm{x}|\bm{s}]$6 sampled paths, the easy four-layer network’s approximant achieves $\ln \mathcal{P}[\bm{x}|\bm{s}]$7 training accuracy, the medium network about $\ln \mathcal{P}[\bm{x}|\bm{s}]$8, and the random-label network only about $\ln \mathcal{P}[\bm{x}|\bm{s}]$9 (Theisen et al., 2019). The paper interprets this as evidence that path variation and normalized margins track both compressibility and generalization.

5. Path-weighted sampling in weight-sharing NAS

In NAS, PWS appears as a supernet-training strategy over architecture paths. Single Path One-Shot (SPOS) builds a supernet in which each architecture is a single path and trains it by sampling one path per SGD step from a fixed prior $\mathcal{P}[\bm{x}]$0, usually uniform (Guo et al., 2019). The training objective is

$\mathcal{P}[\bm{x}]$1

implemented by activating only the sampled path, computing the loss, and updating only the corresponding weights. Because $\mathcal{P}[\bm{x}]$2 is fixed and uniform, every architecture is trained equally in expectation. This is a path-weight sampling scheme with uniform path weights. In the ImageNet building-block supernet described in the paper, the search space has $\mathcal{P}[\bm{x}]$3 choice blocks and size $\mathcal{P}[\bm{x}]$4; on NAS-Bench-201, the resulting supernet rankings have Kendall’s $\mathcal{P}[\bm{x}]$5 roughly in the range $\mathcal{P}[\bm{x}]$6–$\mathcal{P}[\bm{x}]$7 on CIFAR-10, $\mathcal{P}[\bm{x}]$8–$\mathcal{P}[\bm{x}]$9 on CIFAR-100, and $(\bm{s},\bm{x})$0–$(\bm{s},\bm{x})$1 on ImageNet-16-120 (Guo et al., 2019).

PA&DA generalizes this idea by explicitly optimizing path and data sampling distributions to minimize supernet gradient variance (Lu et al., 2023). With supernet objective

$(\bm{s},\bm{x})$2

the paper derives the optimal path-sampling probability

$(\bm{s},\bm{x})$3

where $(\bm{s},\bm{x})$4 is the gradient of the loss with respect to supernet weights under the sampled path and data. For data, it uses the analogous importance-sampling principle and approximates the expensive per-sample gradient norm by the norm of the gradient with respect to the pre-softmax outputs of the last layer. In implementation, PA&DA records candidate-operation gradient norms during backpropagation, updates path and data distributions once per epoch, and mixes the resulting importance distributions with the uniform distribution through smoothing parameters that increase linearly from $(\bm{s},\bm{x})$5 to $(\bm{s},\bm{x})$6. The method is therefore an explicit path-weighted sampling strategy in the sense that paths are sampled in proportion to normalized gradient norms and the stochastic gradient is corrected by the usual importance-sampling factor.

Empirically, PA&DA reports lower supernet gradient variance and better ranking consistency than uniform sampling. On NAS-Bench-201 for CIFAR-10, it achieves Kendall’s Tau $(\bm{s},\bm{x})$7, Precision@Top5% $(\bm{s},\bm{x})$8, and training cost about $(\bm{s},\bm{x})$9 GPU hours; the SPOS baseline reports $$\mathcal{P}[\bm{x}] = \int \mathcal{D}[\bm{s}]\, \mathcal{P}[\bm{s}]\,\mathcal{P}[\bm{x}|\bm{s}].$$0, $$\mathcal{P}[\bm{x}] = \int \mathcal{D}[\bm{s}]\, \mathcal{P}[\bm{s}]\,\mathcal{P}[\bm{x}|\bm{s}].$$1, and cost about $$\mathcal{P}[\bm{x}] = \int \mathcal{D}[\bm{s}]\, \mathcal{P}[\bm{s}]\,\mathcal{P}[\bm{x}|\bm{s}].$$2 GPU hours (Lu et al., 2023). A different NAS variant, Farthest Greedy Path Sampling, addresses the same supernet-ranking problem in recommender search by first filtering a large pool of well-performing paths using a predictor and evolutionary search, then selecting a diverse subset by maximizing total Hamming distance in an architecture encoding space; Theorem 3.1 in that paper links pairwise Hamming-distance maximization to approximate maximization of Shannon entropy over path encodings (Cao et al., 2023). This suggests that the NAS literature now contains at least two distinct PWS-style principles: variance-minimizing importance sampling and diversity-aware quality-constrained path selection.

6. Related path-weight constructions, limitations, and conceptual boundaries

Several adjacent literatures do not always use the term “Path Weight Sampling” but are structurally relevant. In Monte Carlo light transport, “Path Throughput Importance Weights” proposes to compute multiple-importance-sampling weights from inverse path throughput rather than path-space PDFs, with

$$\mathcal{P}[\bm{x}] = \int \mathcal{D}[\bm{s}]\, \mathcal{P}[\bm{s}]\,\mathcal{P}[\bm{x}|\bm{s}].$$3

and proves this formulation equivalent to the standard balance heuristic (Jendersie, 2018). Because path throughput already includes BSDF factors, cosines, geometry terms, emitted radiance, sensor response, and all inverse sampling probabilities, this perspective makes explicit that Russian roulette factors, shading-normal corrections, and random connection probabilities must enter MIS weights consistently.

In geometric data analysis, power-weighted shortest-path distances define path costs

$$\mathcal{P}[\bm{x}] = \int \mathcal{D}[\bm{s}]\, \mathcal{P}[\bm{s}]\,\mathcal{P}[\bm{x}|\bm{s}].$$4

and a continuum analogue

$$\mathcal{P}[\bm{x}] = \int \mathcal{D}[\bm{s}]\, \mathcal{P}[\bm{s}]\,\mathcal{P}[\bm{x}|\bm{s}].$$5

That work does not define PWS, but it provides a density-sensitive path-weighting theory in which $$\mathcal{P}[\bm{x}] = \int \mathcal{D}[\bm{s}]\, \mathcal{P}[\bm{s}]\,\mathcal{P}[\bm{x}|\bm{s}].$$6 recovers geodesic distance and $$\mathcal{P}[\bm{x}] = \int \mathcal{D}[\bm{s}]\, \mathcal{P}[\bm{s}]\,\mathcal{P}[\bm{x}|\bm{s}].$$7 increasingly favors paths through high-density regions (Little et al., 2020). A plausible implication is that PWSPD supplies a principled metric substrate for path-sampling methods that seek to balance geometry and density on manifolds or graphs.

A central limitation of re-weighted PWS appears in the analysis of re-weighted sampling for high-order path integration (Ceriotti et al., 2011). If target expectations are estimated from reference samples with weights $$\mathcal{P}[\bm{x}] = \int \mathcal{D}[\bm{s}]\, \mathcal{P}[\bm{s}]\,\mathcal{P}[\bm{x}|\bm{s}].$$8 and one writes $$\mathcal{P}[\bm{x}] = \int \mathcal{D}[\bm{s}]\, \mathcal{P}[\bm{s}]\,\mathcal{P}[\bm{x}|\bm{s}].$$9, then under a jointly Gaussian model for the observable $\mathcal{P}[\bm{x}]$0 and $\mathcal{P}[\bm{x}]$1,

$\mathcal{P}[\bm{x}]$2

The paper’s main conclusion is that re-weighted sampling is bound to fail as the size of the system increases because the statistical uncertainty in the re-weighted average grows exponentially. This is a general warning for any PWS variant based on extensive path-weight ratios: once the variance of the log-weight becomes large, the effective sample size collapses.

Taken together, these related constructions show both the breadth and the limits of the PWS idea. Across information theory, deep-network theory, NAS, rendering, and geometric learning, path weighting is used to transform a global path-space problem into tractable sampling, compression, or variance-control procedures. At the same time, the efficacy of those procedures depends sharply on how concentrated or diffuse the induced path-weight distribution becomes.