In classical statistics and machine learning, optimal mass transport (OMT) provides a powerful way to compare probability distributions. By calculating the "Earth Mover's Distance" (or Wasserstein distance), OMT measures the minimum work required to reshape one probability distribution into another.
However, classical OMT assumes that the underlying quantities are scalar probabilities. This assumption breaks down in quantum mechanics and advanced signal processing, where states are represented by matrix-valued densities (such as quantum density operators). Because matrices do not commute ($AB \neq BA$), traditional geometric tools fail.
To bridge this gap, a growing body of research generalizes optimal transport and Riemannian geometry to the non-commutative setting. By treating the space of density matrices as a curved geometric manifold, these frameworks allow us to understand quantum evolution, optimization, and state tracking as natural geometric trajectories.
The Geometry of Quantum States
To understand why this framework works, we must first view the space of density matrices not just as a collection of arrays, but as a curved Riemannian manifold. In this space, the distance between two quantum states depends heavily on the chosen metric. As noted by Hao-Chung Cheng et al., the geometric insights we derive from quantum state spaces are fundamentally tied to the choice of metric and operator basis used to represent them.
Several distinct geometric metrics can be defined on these manifolds:
- Quantum Fisher Information Metric: Measures the sensitivity of a quantum state to perturbations of its parameters. Yoshio Uwano et al. demonstrated a fundamental connection between different geometric viewpoints by proving that a Riemannian metric derived from a fibered space structure is equivalent to the Symmetric Logarithmic Derivative (SLD) quantum Fisher metric.
- Fubini-Study Metric: Measures the physical overlap between quantum states. Julien Gacon et al. showed that the Quantum Fisher Information Matrix can be elegantly formulated as the Hessian of the Fubini-Study metric, enabling efficient second-order optimization without explicit matrix construction.
- Berry Curvature: Captures the geometric phase acquired by a quantum state under cyclic parameter variations. Vasilis Niarchos utilized Berry connections on vector bundles to analyze how superconformal primary states deform geometrically, while Guillaume Bal et al. connected this geometric phase directly to the robust dynamics of topological edge states moving along curved interfaces. Similarly, Xianglong Zeng et al. used the quantum geometric tensor—where the real part defines the quantum metric and the imaginary part represents the Berry curvature—to topologically characterize extended Su-Schrieffer-Heeger models.
By leveraging these metrics, researchers can treat quantum learning problems geometrically. For example, Bikram Khanal et al. utilized the Quantum Fisher Information Matrix to map the parameter space of a noisy quantum circuit as a Riemannian manifold, establishing data-dependent generalization bounds that determine whether a model has truly learned or merely memorized noise.
The Mechanism of Non-Commutative Transport
While metrics like the Quantum Fisher Information measure local sensitivity, they do not inherently capture the concept of "transporting" mass dynamically across a space. In classical physics, the Benamou-Brenier formulation solves this by framing the Wasserstein-2 distance as a fluid dynamics problem: finding a time-dependent density $\rho$ and velocity field $v$ that minimizes kinetic energy $\int \rho |v|^2 dt$, subject to the continuity equation:
$$\partial_t \rho + \nabla \cdot (\rho v) = 0$$
When extending this to matrix-valued densities, a major mathematical obstacle arises: the density matrix $\rho$ and the velocity operator $v$ do not commute. Therefore, the classical product $\rho v$ is ill-defined in a non-commutative setting.
To resolve this, Allen Tannenbaum et al. generalized the Benamou-Brenier action integral to density matrices by introducing non-commutative multiplication mechanisms. Instead of simple multiplication, they utilized the anti-commutator ($\rho v + v \rho$) and logarithmic means to define a matricial continuity equation. This formulation yields a novel matricial Wasserstein-2 metric, establishing a rigorous Riemannian structure on the space of density matrices.
Classical OMT: Scalar Density (ρ) ───► Standard Multiplication (ρv) ───► Standard Continuity
│
Non-Commutativity
▼
Matricial OMT: Density Matrix (ρ) ───► Anti-Commutator ({ρ, v}) ───► Matricial Continuity
For alternative transport costs, Allen Tannenbaum et al. extended the Wasserstein-1 distance to matrix densities by leveraging duality theory, creating a flexible framework capable of comparing complex multivariate structures like power spectra.
Applying these metrics to physical systems requires adapting them to specific quantum constraints. Giacomo De Palma et al. established a direct correspondence between quantum transport plans and physical quantum channels, revealing a deep geometric link between the quantum Wasserstein distance and the Wigner-Yanase metric. When moving to infinite systems, such as quantum spin lattices, Dario Trevisan et al. overcame the challenge of non-commuting operators across infinite dimensions by defining a per-site quantum Wasserstein-1 distance to bound entropy changes in high-temperature Gibbs states.
Quantum Dynamics as Gradient Flows
Why build this elaborate geometric machinery? The primary advantage is that it allows us to view complex dynamical processes as simple gradient flows—paths of steepest descent along a geometric landscape.
In open quantum systems, interaction with the environment causes dissipation. By equipping the state space with a Riemannian metric parameterized to align with a specific divergence measure, this dissipation can be modeled as an entropy-minimizing gradient flow.
For instance, Yu Cao et al. proved that the evolution of open quantum systems governed by primitive Lindblad equations follows a steepest descent path that minimizes the sandwiched Rényi divergence. This geometric framing provides a robust mathematical guarantee for the system's exponential decay toward its steady state.
▲ Entropy / Divergence Landscape
│
│ * (Initial State)
│ \
│ \ Steepest Descent Path
│ \ (Gradient Flow via Matricial Metric)
│ ▼
│ * (Steady State / Equilibrium)
└────────────────────────────────────────► State Manifold
This principle extends beyond physics into quantum game theory and optimization. Wayne Lin et al. modeled strategy evolution in cooperative quantum games as a continuous-time gradient flow with respect to the Quantum Shahshahani metric, demonstrating that decentralized learning dynamics naturally adapt toward Nash equilibria to solve the Best Separable State problem.
Projections and Computational Efficiency
Operating on these high-dimensional matrix manifolds is computationally demanding. To make these geometric insights practical for real-world engineering, researchers rely on projection and optimization algorithms.
Daoyi Dong et al. designed a quantum projection filter that mitigates computational costs by applying a Riemannian metric derived from Quantum Fisher Information to a state-space submanifold. This allows complex, high-dimensional quantum trajectories to be projected onto simpler, lower-dimensional representational spaces while strictly minimizing approximation errors.
For broader applications like diffusion tensor imaging and spectral analysis, Yongxin Chen et al. developed an efficient, scalable algorithm to solve matrix-valued optimal mass transport problems. Their approach utilizes sequential quadratic programming (SQP) combined with an inexact Hessian approximation and specialized preconditioning, bypassing the high computational complexity that historically hindered matrix-valued OMT adoption.
By reimagining matrix spaces through the lens of non-commutative geometry and fluid-like transport, this framework provides a unified language for tracking, optimizing, and understanding complex quantum and multivariate systems.
Key Foundational Work
Certain integrable system on a space associated with a quantum search algorithm (2007) — contributed 2 concept elements to this direction.
The Learnability of Unknown Quantum Measurements (2015) — contributed 1 concept elements to this direction.