Physics-Based Non-Uniform Random Variate Generators

• Physics-Based Non-Uniform Random Variate Generators (PPRVGs) are systems that exploit physical stochasticity to directly sample non-uniform distributions, bypassing traditional uniform-to-nonuniform transforms.
• They utilize hardware elements such as MEMS accelerometers and GFETs to perform real-time, programmable transformations that enhance the speed and accuracy of Monte Carlo simulations.
• Recent implementations like Spot and Grappa demonstrate efficient sampling for complex distributions—including Gaussian, lognormal, and heavy-tailed—while reducing computational overhead.

Physics-based non-uniform random variate generators are systems that generate samples directly from non-uniform probability laws by exploiting physical stochasticity, physical transfer characteristics, or physical invariances rather than relying exclusively on algorithmic transformations of uniform pseudo-random numbers. In recent usage, the acronym PPRVG denotes “physics-based programmable non-uniform random variate generators,” with recent work highlighting hardware such as Spot and Grappa as ways to overcome common limitations of traditional Monte Carlo sampling (Petangoda et al., 10 Aug 2025). The broader literature includes sensor-noise Gaussian generators, graphene-based programmable circuits, mathematically physics-based normal generators, and simulation-oriented procedures for relativistic, phononic, and heavy-tailed distributions (Meech et al., 2020, Tye et al., 2020, Brent, 2010, Swisdak, 2013, Maurer et al., 2016, Tabbara et al., 26 Jun 2026).

1. Conceptual basis and scope

The conventional pipeline for non-uniform sampling begins with a uniform PRNG and then applies distribution-specific transforms such as Box-Muller, Ziggurat, ICDF-based methods, or acceptance-rejection. The framework summarized in recent work contrasts this with PPRVGs that are “truly stochastic,” “directly programmable,” and capable of “large batches” in a “single shot,” with time per sample that can be independent of distribution complexity when the distribution is realized physically rather than algorithmically (Petangoda et al., 10 Aug 2025).

Two mathematical patterns recur across the literature. The first is change of variables:

$X = F_X^{-1}(U),$

which underlies implementations that physically approximate an inverse CDF. The second is direct sampling from a physically realized native law followed by a simple transformation, as in Gaussian sources where a sampled Gaussian can be shifted and scaled to another Gaussian. Spot, as summarized in later work, uses

$Y = \sigma Y_0 + \mu$

to obtain any other univariate Gaussian from a sampled standard Gaussian, while Grappa uses a hardware approximation to the ICDF of the target distribution (Petangoda et al., 10 Aug 2025).

The scope of the field is not limited to hardware devices. Some work treats physically derived target laws—relativistic Maxwellians, Bose–Einstein-related phonon distributions, or $α$-stable laws—as the central problem, with inversion, rejection, CMS, or LePage methods furnishing the variate generator. This suggests that PPRVGs sit at the intersection of device physics, stochastic modeling, and Monte Carlo architecture rather than in a single methodological niche (Swisdak, 2013, Maurer et al., 2016, Tabbara et al., 26 Jun 2026).

2. Device mechanisms and representative architectures

Direct hardware realizations differ chiefly in the physical mechanism from which the non-uniformity arises. Some systems sample intrinsic noise that is already close to a target law; others exploit nonlinear transfer characteristics to transform an easier source into the desired output.

System	Physical principle	Reported capability
Sensor-noise PRVA	Intrinsic noise of the z-axis of a commercial MEMS accelerometer (Bosch BMX055) sampled via ADC	Output is well-approximated by a Gaussian; proof-of-concept reduced Monte Carlo integration error by 1068 times and doubled speed
GFET non-uniform generator	Nonlinear, ambipolar transfer characteristics of GFETs	Requires as few as two transistors and a resistor; targets arbitrary univariate distributions via configurable biases
Spot	Electron tunneling noise sampled by ADC	Direct Gaussian source; programmable Gaussian outputs and Gaussian mixtures
Grappa	GFET nonlinear transfer characteristics implementing ICDF approximation	Digitally reconfigurable for lognormal, exponential, generalized Pareto, and Gaussian mixtures

The sensor-noise accelerator samples the intrinsic z-axis noise of a commercial MEMS accelerometer through the ADC within a microcontroller at an I2C rate of 1154 Hz, and measurements consistently showed that the sensor noise is well-approximated by a Gaussian distribution (Meech et al., 2020). Its architecture couples an analog MEMS sensor, ADC, and microcontroller or FPGA, with the CPU specifying target mean and variance and the hardware performing an affine transformation before placing results in a cache (Meech et al., 2020).

The GFET-based architecture begins from a uniform random voltage applied to the gate of a GFET. Because graphene is a zero-bandgap semi-metal with ambipolar transfer characteristics and a tunable nonlinear current–voltage relation, a uniform random gate voltage yields a non-uniform random drain current. Cascading GFET stages and adjusting $V_{DS}$ biases further shapes the output distribution. The reported minimal circuit requires as few as two transistors and a resistor, or optionally a transimpedance amplifier for lower noise (Tye et al., 2020).

The later architectural synthesis names two PPRVGs explicitly. Spot uses electron tunneling noise as a Gaussian source sampled by an ADC, while Grappa exploits GFET nonlinearities to implement a transfer function corresponding to the ICDF of a desired distribution. In this account, Spot and Grappa are presented as device-level exemplars of direct physical sampling and direct physical transformation, respectively (Petangoda et al., 10 Aug 2025).

3. Programmability and distribution synthesis

Programmability is central to the contemporary notion of a PPRVG. In the MEMS-based accelerator, the native output is already close to Gaussian, so the principal programming mechanism is an affine transform from one univariate Gaussian to another. The system is described as adaptable to different target means and variances by simple configuration, and future FPGA-based logic is proposed for low-latency conversion to arbitrary Gaussian variates (Meech et al., 2020).

The GFET work pursues a more general objective: approximation of arbitrary univariate PDFs. Its key mechanism is wavelet decomposition,

$g(x) \approx \sum_k c_k \psi_k(x),$

with a GFET at a particular bias playing the role of a basis function. The CPU computes the discrete wavelet transform of the desired distribution, selects significant coefficients, and sets the GFET biases so that summed outputs reconstruct an approximation of the target law. The reported trade-off is explicit: more GFETs correspond to more wavelet coefficients and therefore higher approximation accuracy (Tye et al., 2020).

Grappa, as summarized later, adopts a related but distinct synthesis strategy. The target ICDF is approximated through the Galerkin method with orthonormal basis functions selected and fitted using linear least squares and the Gram-Schmidt process to the GFET’s measured response. This makes the chip digitally reconfigurable for different target laws, including lognormal, exponential, generalized Pareto, and Gaussian mixtures (Petangoda et al., 10 Aug 2025).

A mathematically different route to programmability appears in Chris Wallace’s normal generators. Wallace’s “maximum entropy” method does not emulate a device but exploits the invariance of the Gaussian law under orthogonal transforms. If $x$ is a vector of independent standard normals and $Q$ is orthogonal, then $y = Qx$ has the same joint Gaussian law. Implementations therefore maintain a pool of normal variates, apply orthogonal transformations and strong mixing, and then restore the correct $\chi^2$ fluctuation by chi-squared correction. The data explicitly characterizes Wallace’s method as a prime example of a PPRVG because it applies the mathematical symmetry directly for computational speed and efficiency (Brent, 2010).

4. Statistical quality and Monte Carlo consequences

The principal empirical motivation for PPRVGs is reduction of Monte Carlo cost when the target distribution is itself non-uniform. The MEMS-based proof of concept reported that Monte Carlo integration of a univariate Gaussian achieved an error reduction by 1068 times while doubling the speed of the simulation (Meech et al., 2020). In the detailed evaluation, the KL divergence between the sensor-derived distribution and a fitted Gaussian was 0.00263, compared to 0.0392 for MATLAB-generated discrete Gaussians after quantization, and increasing quantization decreased KL divergence (Meech et al., 2020).

That same work reported that the sensor-based PRVA outperformed the standard C++ Gaussian generator in both speed and accuracy for Monte Carlo integration tasks. Using single-threaded execution it was about 2× faster, and with eight threads it was up to 1.4× faster. A limitation was slightly fewer “unique numbers in the tails” than synthetic Gaussians, although the reported practical effect on integration was minimal (Meech et al., 2020).

The GFET study reported simulated Monte Carlo integration speedups of up to $2\times$, under the assumption that the ADCs reading the outputs from the circuit can produce samples in the same amount of time that it takes to perform memory accesses. The paper also states that hardware-based univariate non-uniform sampling leads to lower error for the same sample count than uniform sampling with accept-reject or bounded uniform proposals, and that generation speed is limited by ADC throughput rather than intrinsic device speed (Tye et al., 2020).

The later synthesis situates these results within a broader acceleration narrative. Spot is summarized as being 260× faster for Gaussian random variable generation than an ARM Cortex-M0+ running Box-Muller and dissipating lower power than the microcontroller. Grappa is summarized as at least 2× faster than MATLAB for common non-uniform distributions and at least 1.26× faster than standard C++ lognormal generators. The significance assigned to these devices is that they remove or vastly speed up the sampling step in Monte Carlo pipelines (Petangoda et al., 10 Aug 2025).

5. Physics-derived target laws and algorithmic relatives

A large part of the relevant literature concerns non-uniform variates whose target distributions arise directly from physical law. In the relativistic Maxwellian case, the Jüttner or Synge distribution does not separate into products of 1D distributions, so the paper adopts Devroye’s log-concave rejection sampling algorithm for arbitrary temperature and arbitrary drift velocity. The reported efficiency is about 88–90% in the isotropic case and about 77–80% in the drifting case, with sample Fortran codes generating $Y = \sigma Y_0 + \mu$0 variates in approximately 0.1 seconds unoptimized (Swisdak, 2013).

Phonon Monte Carlo provides another systematic treatment of physically grounded non-uniform generation. There the inversion and rejection methods are applied to Bose–Einstein thermal distributions, full or isotropic phonon dispersion, Lambert’s cosine law for diffuse scattering, Soffer-type momentum-dependent specularity, and contact injection. A central methodological claim is that the correct physical distribution should be sampled even when exact energy matching per time step is inconvenient, because forcing the last phonon’s energy to close a local budget can bias the simulation (Maurer et al., 2016).

Heavy-tailed models furnish a further extension of the landscape. For $Y = \sigma Y_0 + \mu$1-stable laws, the AUB-HTP package implements the Chambers–Mallows–Stuck algorithm for univariate sampling and the LePage series representation for multivariate sampling with flexible spectral measures. The package supports all $Y = \sigma Y_0 + \mu$2, arbitrary skewness, scale, shift, isotropic or elliptical or discrete or mixture or custom spectral measures, and automatic truncation control for the multivariate LePage series (Tabbara et al., 26 Jun 2026). Although this is software rather than direct hardware sampling, it illustrates the continuing importance of non-uniform variate generation where the target law is physically or empirically heavy-tailed.

Wallace’s normal generator occupies a distinctive position among these methods. Unlike rejection or inversion methods, it seeks to generate new normal random numbers directly from previously generated normal numbers by orthogonal transforms, mixing, and chi-squared rescaling. Reported implementations can be approximately as fast as good uniform generators and 3–10× faster than classical normal generators, while still carrying caveats about finite-pool correlations and rare-event clustering (Brent, 2010).

6. Limitations, misconceptions, and emerging directions

A recurrent limitation of physical generators is that raw physical outputs are not automatically stationary. In the MEMS sensor system, both the mean and standard deviation of the noise distribution were sensitive to temperature and supply voltage, with temperature the primary factor and voltage a secondary but non-negligible effect. The paper states that these parameters must be measured and compensated for to prevent drift, and that current throughput is limited by the on-board ADC sample rate of less than 2 kHz unless custom hardware ADCs are used (Meech et al., 2020).

Another recurring issue is that physical unpredictability, empirical randomness, and direct non-uniformity are distinct properties. The double-pendulum PRNG uses deterministic chaotic dynamics, ordinary differential equations, and binary extraction from final states; it is therefore a physics-based PRNG rather than a direct hardware non-uniform variate generator. Its raw outputs are not guaranteed to be uniformly distributed, the paper does not specify elaborate post-processing to enforce strict uniformity, and the non-damped version merely passed the NIST suite more reliably than Java’s basic random class while remaining less robust than Java’s SecureRandom. The damped version performed worse because energy loss reduced unpredictability over time and led to eventual convergence (Sigit, 2024). A plausible implication is that chaos-based generators and direct PPRVGs solve different problems even when both are “physics-based.”

GFET systems also expose a distinction between device nonlinearity and exact distributional fidelity. Their flexibility depends on measured transfer characteristics, bias control, and ADC readout assumptions; output approximation quality is assessed through KL divergence or task-level Monte Carlo error rather than by a universal exactness guarantee (Tye et al., 2020). Likewise, native Gaussian sources are less suitable when other distributions are desired natively, because additional transformations are needed (Meech et al., 2020).

Research directions in the cited literature are correspondingly diverse. Proposed hardware extensions include analog-out accelerometers such as ADXL335, high-speed ADCs such as ADS54J60, static proof-mass configurations, amplified signal ranges, and removal of low-pass filtering to maximize entropy in the sampled signal (Meech et al., 2020). GFET work points toward dynamically selectable non-uniform distributions through integrated switching matrices, CPU control, and reconfigurable biasing (Tye et al., 2020). The chaos-based PRNG work suggests exploring more complex pendulum systems, additional test suites such as DIEHARDER, and more sophisticated sampling or post-processing (Sigit, 2024). Finally, the architectural perspective in later work goes beyond PPRVGs entirely by proposing “uncertainty-tracking” architectures such as Signaloid’s UxHw, which represent and manipulate full probability distributions in microarchitectural state and are described as “convergence-oblivious” (Petangoda et al., 10 Aug 2025). This suggests a bifurcation in the field: one line accelerates sampling from physically realized non-uniform laws, while another seeks to eliminate repeated sampling for many uncertainty-propagation workloads.