Deterministic Distribution Conversion

• Deterministic distribution conversion refers to methods that deterministically transform probability distributions into target forms using invertible mappings, enabling reproducible outcomes.
• Techniques such as inverse CDF transforms, Monge–Ampère mass transport, shell mapping in communication, and Fourier-based convolution in bootstrapping optimize divergence and energy efficiency.
• These methods support derandomization in simulation and learning, underpinning applications in signal processing, quantum analog-to-digital conversion, and robust statistical inference.

Deterministic distribution conversion refers to methodologies that transform random variables or entire probability distributions into new forms with tractable or desirable properties, such that the transformation itself is a deterministic map—there is no intrinsic randomness in the conversion procedure once the data or initial random variable is fixed. Deterministic distribution conversion is fundamental in areas including signal processing, communication, learning theory, quantum information, numerical analysis, and statistical inference. Major frameworks range from invertible distribution matching in communication to the solution of Monge–Ampère partial differential equations for mass transport, to deterministic analog-to-digital encoding in quantum computation, as well as derandomization in learning and simulation.

1. Mathematical Foundations and Definitions

The essence of deterministic distribution conversion is the construction of a mapping $T: \mathcal{X} \to \mathcal{Y}$ such that the pushforward measure, $T_{\#}P_X$, matches a target distribution on $\mathcal{Y}$. In one dimension, this is classically achieved by the inverse cumulative distribution function (CDF) method: if $P_X$ has CDF $F_X$ and the target is the uniform distribution on $[0,1]$, then $T(x)=F_X(x)$ maps $X$ to a uniform variable, and the inverse $T^{-1}(u)=F_X^{-1}(u)$ maps uniform variables to $P_X$. The multidimensional, non-separable case requires a fundamentally different approach: mass transport via a potential function $T_{\#}P_X$0 such that $T_{\#}P_X$1 pushes $T_{\#}P_X$2 to a uniform density on a domain $T_{\#}P_X$3 by solving the Monge–Ampère equation:

$T_{\#}P_X$4

subject to mass conservation and specific boundary conditions (Kee, 2016).

2. Deterministic Distribution Matching via Shell Mapping

A major operational setting is fixed-to-fixed length distribution matching (DM) utilized in coding for communication systems. Here, the aim is to convert a uniform binary input $T_{\#}P_X$5 into sequences $T_{\#}P_X$6 whose empirical distribution matches a prescribed target distribution $T_{\#}P_X$7 on $T_{\#}P_X$8. The shell mapping distribution matcher (SMDM) is a deterministic, invertible mapping $T_{\#}P_X$9 that selects the $\mathcal{Y}$0 lowest-weight sequences according to a weight function $\mathcal{Y}$1 (or an integer approximation). This procedure achieves the minimum informational divergence $\mathcal{Y}$2 among all invertible fixed-length DMs, guaranteeing that the output is as close as possible (in KL-divergence) to the target i.i.d. distribution. Implementation is efficient for moderate $\mathcal{Y}$3 and small $\mathcal{Y}$4 via dynamic programming (Schulte et al., 2018).

DM Scheme	Map Type	Optimality Criterion
SMDM	Deterministic, invertible	KL-divergence to $\mathcal{Y}$5
CCDM	Deterministic, composition	Empirical type matching

In communication with probabilistic amplitude shaping (PAS), SMDM offers substantial shaping gains for short blocklengths (e.g., 0.59 dB gain over CCDM at 3 bpcu, rate-adaptive granularity $\mathcal{Y}$6 without changing shell parameters) (Schulte et al., 2018). Shell mapping also unifies divergence- and energy-optimal sequence design, as energy constraints are a special case under Maxwell–Boltzmann targets.

3. Quantum Analog-to-Digital Conversion (QADC)

In quantum information, deterministic distribution conversion is exemplified by the quantum analog-to-digital conversion (QADC) protocol. Given a quantum state $\mathcal{Y}$7 (amplitudes $\mathcal{Y}$8 encode analog information), QADC deterministically maps these amplitudes onto digital qubit registers, extracting m-bit approximations $\mathcal{Y}$9 to $P_X$0, $P_X$1, or $P_X$2 for each address $P_X$3. The process is fully unitary and requires no post-selection, achieving accuracy $P_X$4 up to phase estimation infidelity $P_X$5, using controlled swap-tests, phase estimation, and quantum arithmetic (Mitarai et al., 2018):

$P_X$6

QADC is a core utility for high-level quantum algorithms including nonlinear amplitude transformations and quantum neural networks, and contrasts with probabilistic or post-selected digital-to-analog conversions pervasive in quantum algorithms (Mitarai et al., 2018).

4. Deterministic Transformation to Uniformity and Mass Transport

For arbitrary, smooth, multivariate, and potentially non-separable densities $P_X$7 on a convex bounded domain $P_X$8, deterministic uniformization requires finding a scalar potential function $P_X$9 such that $F_X$0 is a measure-preserving map pushing $F_X$1 to a uniform density $F_X$2 (Kee, 2016). This approach generalizes the one-dimensional CDF transformation to $F_X$3 dimensions, replacing coordinate-wise marginalization (which fails for non-separable $F_X$4) with the solution to a nonlinear elliptic PDE, the Monge–Ampère equation:

$F_X$5

Subject to Neumann boundary conditions and appropriate normalization, this map is computed numerically via nonlinear conjugate gradients and multigrid, verifying that pushing a uniform mesh through $F_X$6 yields samples distributed as $F_X$7 within discretization error (Kee, 2016). In 1D, this recovers the inverse CDF transform; in higher dimensions, it supports deterministic transformation for highly structured non-separable targets.

5. Derandomizing Distribution Learning and Inference

In distribution learning, deterministic conversion methods replace randomized outputs with single deterministic hypotheses with provable guarantees. In multi-distribution or collaborative learning, converting a randomized predictor (e.g., a distribution over $F_X$8) into a deterministic function $F_X$9 achieving optimal worst-case error is, in general, computationally hard—NP-hardness is established via a reduction to discrepancy minimization (Larsen et al., 2024). However, if the underlying distributions are label-consistent (i.e., the label bias is consistent across distributions for each instance), a deterministic black-box reduction is proven possible: by identifying "heavily biased" points and deterministically setting predictions via majority vote, and otherwise rounding using the learned distribution, full derandomization is achieved in polynomial time and sample complexity (Larsen et al., 2024). This principle formalizes deterministic distribution conversion within learning-theoretic agnostic minimax frameworks.

6. Deterministic Inference: Bootstrapping and Simulation

Deterministic bootstrapping transforms the stochastic construction of the bootstrap distribution for linear functionals into a deterministic calculation. Linear bootstrap estimators $[0,1]$0 (weighted sums of independent discrete random variables $[0,1]$1) have exact distributions given by the $[0,1]$2-fold convolution of their component distributions. Using the discrete Fourier transform (characteristic functions), the convolution is computed deterministically: characteristic functions are multiplied on a frequency grid and inverted by FFT, producing the full distribution $[0,1]$3 (Pitschel, 2019). This approach eliminates the need for Monte Carlo sampling, enabling high-precision quantile estimation, hypothesis tests, and other applications where the statistic is a linear combination of independent inputs.

Scenario	Conversion Target	Deterministic Mechanism
Bootstrap mean	Empirical distribution	Fourier-based convolution
Block bootstrap	Block-averaged aggregates	Convolution via characteristic functions
Permutation test	Null distribution	Enumerative summation and FFT

Broader applicability extends to any statistics reducible to linear combinations of independent (possibly permuted or block-aggregated) random variables (Pitschel, 2019).

7. Connections, Limitations, and Open Problems

Deterministic distribution conversion encompasses exact matching under information-theoretic criteria (minimum KL-divergence), efficient mass transport (via Monge–Ampère PDEs), and derandomization of learning and simulation procedures. These approaches are united by deterministic, invertible mappings with provable sample complexity or information divergence optimality.

Limitations emerge in the general case—computational hardness under label-inconsistency in collaborative learning (Larsen et al., 2024), nontrivial numerical complexity for nonlinear PDE-based transport in high dimensions (Kee, 2016), and scalability concerns for deterministic bootstrapping in non-linear or high-dimensional contexts (Pitschel, 2019). Open problems include extending uniformization transforms beyond two dimensions, tightening sample and computational overheads in learning-theoretic derandomization, and the search for efficient deterministic learners in general agnostic settings (Larsen et al., 2024).

Deterministic distribution conversion remains foundational for reliable and reproducible probabilistic computation, efficient communication, robust inference, and quantum algorithm design.