Source-linked AI summary
QH-GEM: Quantum-Hydrodynamic Generative Modeling
Harbir Antil, Alex Kaltenbach, Sarswati Shah
TL;DR
The paper addresses deterministic target-distribution generation without freely parameterizing a transport map or time-dependent velocity field. It controls the initial phase of a free-particle Madelung system, proves transport and reachability results, and demonstrates numerical phase identification, subject to additional analytical assumptions.
Problem
The paper studies how to generate target distributions while fixing a physically motivated evolution law instead of freely learning a transport map or time-dependent velocity field.
Method
It uses the initial phase of the free-particle Schrödinger equation as a PDE-constrained control variable, with Madelung dynamics inducing the density-coupled characteristic transport.
Results
The framework transports the reference density to the evolving Born density, characterizes exact isotropic Gaussian reachability, establishes local Wasserstein transport estimates, and demonstrates numerical phase identification.
Takeaways & Limitations
Once the initial phase is selected, generation is deterministic, with randomness entering only through the initial sample and subsequent motion governed by free Madelung dynamics.
Takeaways & Limitations
The guarantees require nonvanishing Born density and additional global velocity regularity conditions not ensured by standard global Schrödinger well-posedness.
Abstract
from arXiv · showhide
In this paper, we develop a deterministic, physically constrained generative framework based on the Madelung formulation of the free-particle Schrödinger equation. A reference Born probability density and a controllable initial phase function serve as initial data for the free Madelung system, which couples the Born probability density and phase function through the Bohm quantum potential, while the phase function determines the hydrodynamic velocity field. Provided that the Born probability density remains positive and the hydrodynamic velocity field generates a unique global characteristic flow, samples drawn from the reference density and transported along the characteristic flow are distributed according to the evolving Born probability density at every time. As a consequence, randomness enters only through the initial sampling; the subsequent generation is deterministic and involves neither stochastic dynamics nor an independently parameterized time-dependent velocity field. We formulate terminal-time distribution matching as a PDE-constrained phase-identification problem and derive the underlying Hamiltonian and Fisher-information structure. For isotropic Gaussian wave packets, we obtain explicit dynamics and a necessary and sufficient condition for exact reachability of isotropic Gaussian targets by quadratic initial phase functions, together with the corresponding sampling map. For a smooth prescribed potential initial velocity field, we further establish that the characteristic flow approximates the associated first-order transport map with an O(T^2) error, both uniformly and in the 1- and 2-Wasserstein distances. A numerical Gaussian benchmark validates the fully discrete forward solver, while full-grid PDE-constrained phase identification is demonstrated for asymmetric bimodal targets.
1. Introduction
The paper fixes a physically motivated free-particle Schrödinger evolution and controls only its initial phase to generate target distributions through deterministic Madelung transport. Under positivity and characteristic-flow assumptions, the framework provides transport guarantees, structural analysis, exact Gaussian reachability, local transport estimates, and numerical demonstrations.
- Model formulation: The framework uses the free-particle Schrödinger equation and Madelung transform to couple Born density, phase, and hydrodynamic velocity.The density satisfies probability-conserving transport, while the phase evolves through the quantum Hamilton–Jacobi equation and Bohm quantum potential.
- Deterministic transport: Under global Lipschitz velocity and positivity assumptions, the characteristic flow transports the reference density to the evolving Born density at every time.The transport representation is [X(t; ·)]#(ρ0 dx) = ρ(·, t) dx.
- Deterministic generation: Randomness enters only through the initial sample; after the initial phase is fixed, the characteristic transport and terminal generation are deterministic.The model simulates no stochastic process during generation and does not independently parameterize a time-dependent velocity field.
- Physical constraints and novelty: Unlike normalizing flows and continuous flow models, the method controls only the initial phase while density-coupled Hamiltonian dynamics determine subsequent transport.The Bohm quantum potential contributes dispersive, time-reversible dynamics rather than dissipative parabolic regularization.
- Contributions: The paper derives exact isotropic Gaussian reachability and local nonlinear transport results, and validates the approach numerically on Gaussian and asymmetric bimodal targets.Its contributions include explicit quadratic-phase dynamics, Wasserstein short-time bounds, solver validation, and full-grid phase identification.
2. The Free Madelung System as a Generative Transport Model
The free Madelung system couples a positive Born density and phase through the Bohm quantum potential, while the phase induces a hydrodynamic velocity for deterministic transport. Under classical-solution and flow assumptions, its characteristic representation supports sampling and phase-based terminal distribution matching.
- Model equations and structure: The Madelung system consists of a continuity equation for the Born density and a quantum Hamilton–Jacobi equation coupling phase and density through the Bohm potential.The velocity is induced by the phase gradient, and the system inherits Hamiltonian, time-reversible, and dispersive structure from free Schrödinger evolution.
- Analytical setting: The analysis assumes a unique classical solution, strict positivity of the density, and, when needed, regularity and growth conditions ensuring a unique global characteristic flow.Strict positivity makes the pointwise quantum potential well-defined, while global Lipschitz and linear-growth bounds support the characteristic-flow construction.
- Scope and limitations: The framework’s classical transport conclusions do not follow from Schrödinger well-posedness alone, because nonvanishing density and global velocity bounds remain additional hypotheses.Weak formulations may allow vacuum but do not furnish the classical characteristic flow required here.
- Deterministic transport representation: The characteristic flow is a unique global C1-diffeomorphic transport map generated by the hydrodynamic velocity field.It connects the Eulerian continuity equation with particle transport and yields the evolving density representation.
- Deterministic transport representation: Once the initial phase is fixed, trajectories and terminal samples are deterministic; randomness enters only through the initial sample.The flow is not independently prescribed: it is generated by the density-coupled Madelung dynamics.
PHASE IDENTIFICATION GENERATION
The framework controls the initial phase of the free Madelung system and uses its coupled Hamiltonian evolution to transport an initial distribution deterministically. Terminal distribution matching is posed as PDE-constrained phase identification, with explicit Gaussian analysis available but no general optimization existence guarantee asserted.
- Deterministic generation: An initial sample is transported deterministically by the characteristic flow once a phase function satisfying terminal-time matching has been found.The resulting terminal sample follows the sampling identity induced by the flow.
- Hamiltonian structure: The hydrodynamic velocity is a potential field coupled to the Born density through the Bohm quantum potential.The coupling arises from the phase equation and the Fisher-information contribution to the Hamiltonian.
- Hamiltonian structure: The free Madelung system has canonical Hamiltonian structure and conserves its Hamiltonian over time.The Hamiltonian contains hydrodynamic and quantum kinetic contributions, with the latter equal to 1/(8m) times the Fisher information.
- Phase control: Adding a constant to the initial phase leaves the velocity field, characteristic flow, and transported probability measure unchanged.Phase controls are therefore naturally defined modulo additive constants and can be normalized.
- PDE-constrained phase identification: Terminal distribution matching is formulated by minimizing a discrepancy between the terminal transported measure and a target over admissible normalized initial phases.The terminal state is represented as the pushforward of the reference measure by the characteristic map, and zero cost is equivalent to exact matching.
- Scope boundary: No general existence or differentiability result is asserted for the phase-identification optimization or its terminal control-to-state map.The Gaussian reference and quadratic phase family are treated separately because the corresponding map is explicit.
3. Gaussian Wave-Packet Dynamics and Phase-Controlled Reachability
The Gaussian specialization makes free Madelung dynamics explicit and reduces phase control to algebraic reachability conditions. Quadratic initial phases yield exact affine sampling maps for reachable isotropic Gaussian targets, while quantum spreading persists even from zero initial phase.
- Gaussian dynamics: The isotropic Gaussian specialization reduces the PDE dynamics to explicit evolution equations for the mean, width, and phase coefficients.The terminal-time control-to-state map is therefore explicitly evaluable within this finite-dimensional family.
- Gaussian dynamics: Zero initial phase and velocity still produce nonstationary Gaussian dynamics because the Bohm quantum potential induces radial dilation.The width satisfies σ(t) > σ0 for every t > 0, expressing intrinsic quantum spreading.
- Phase control: The initial linear phase coefficient u0 determines the terminal mean, whereas the quadratic coefficient a0 determines the terminal standard deviation.Thus, exact Gaussian matching reduces to algebraic conditions on the initial phase coefficients.
- Exact reachability: A quadratic initial phase reaches an isotropic Gaussian target if and only if the stated reachability inequality holds.When the inequality is strict, the admissible quadratic coefficients have two branches, while the additive constant does not affect the density, velocity, or flow.
- Exact reachability: For every reachable target, the characteristic flow has a closed-form affine sampling map that transports any reference sample to the target distribution.The pushforward identity gives [X⋆(X0)]#P = µ⋆.
- Scope: Terminal-time density matching does not uniquely determine the initial phase, and the isotropic result does not cover anisotropic targets or arbitrary initial phases.For d > 1, anisotropic Gaussian targets require matrix-valued quadratic phase coefficients; the reachability inequality is only asserted within the considered quadratic family.
4. Local-in-Time Realization of Nonlinear Transport Maps
For smooth potential initial velocity fields, the chosen initial phase reproduces the prescribed transport direction to first order. The resulting characteristic flow and transported measures differ from the first-order map by O(T^2), under the stated regularity, positivity, and flow assumptions.
- Construction: The initial phase is chosen as θ0 = mφ0 + c0 when the prescribed velocity field is potential, v0 = ∇φ0.This makes the initial hydrodynamic velocity equal to the prescribed transport direction.
- Construction: The first-order comparison map is the forward-Euler approximation of the characteristic flow at t = 0.The associated comparison measure is obtained by pushing the initial probability measure through this map.
- Error estimates: A uniform acceleration bound yields quadratic-in-time bounds for the characteristic-map error and the corresponding W1- and W2-distances.The measure estimates follow through the coupling induced by the initial measure.
- Error estimates: Under Assumption 2.1, the characteristic flow satisfies uniform estimates, and the transported density remains the pushforward of the initial measure.The theorem applies for p ∈ {1, 2} under the stated moment and regularity conditions.
- Scope: The result is local in time: the initial phase reproduces the prescribed direction to first order, while density-dependent material acceleration contributes the O(T^2) remainder.It does not assert exact or global reachability of an arbitrary target measure.
- Scope: The Gaussian result and the nonlinear transport result provide complementary benchmarks for exact invariant-family reachability and short-time approximation.The paper next uses the explicit Gaussian solution to assess numerical phase identification.
5. Numerical Validation
The numerical study validates the discrete forward solver on Gaussian dynamics and demonstrates PDE-constrained initial-phase identification for asymmetric bimodal targets in one and two dimensions.
- Discretization and optimization: The fully discrete method uses conservative finite volumes, Rusanov fluxes, Godunov–Lax–Friedrichs differences, and SSP–RK time integration under a CFL restriction.The discrete control is the normalized initial phase, and the solver induces a terminal-time control-to-state map.
- Gaussian forward-solver validation: At T = 0.3, the Gaussian benchmark compares computed and exact terminal densities and tracks discrete L2, L1, and Hellinger errors under mesh refinement.For Nx = 301, terminal errors are reported, while Figure 3 shows error decay as the mesh is refined.
- 1D phase identification: Full-grid phase identification deforms a unimodal reference density into an asymmetric bimodal target while generating subsequent transport through the discrete free Madelung evolution.The quadratic phase family cannot generate the bimodal target, motivating optimization over the full discrete phase space.
- Gaussian forward-solver validation: The Gaussian benchmark indicates linear convergence in the discrete L2, L1, and Hellinger metrics toward the exact dynamics.The validation uses explicit Gaussian dynamics and refines the time step alongside the spatial mesh.
- Initialization and control: The quantile transport map supplies only an initialization direction; optimization adjusts the initial phase while the Madelung system determines the later velocity field.The reported controls are quasi-Newton approximations obtained through reverse-mode automatic differentiation, without claims of stationarity or local optimality.