⚛️ quantum physics

Interpreting Quantum Learning Models via Stochastic Processes

This paper proposes a probabilistic framework that interprets quantum learning models as stochastic processes by establishing a trade-off between representing quantum dynamics as Markovian maps with negative probabilities or as positive stochastic processes with higher-order memory dependencies, thereby bridging quantum mechanics with classical learning models like Projective Simulation.

Original authors: Johannes Fankhauser, Lukas J. Fiderer, Hans J. Briegel

Published 2026-07-21

📖 6 min read🧠 Deep dive

CC BY 4.0

Original authors: Johannes Fankhauser, Lukas J. Fiderer, Hans J. Briegel

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you are trying to understand how a complex machine makes a decision. In the world of classical computers, this is like watching a marble roll down a maze. You can see exactly where the marble is at every second, and you know that if it's at a specific junction, the path it takes next depends only on where it is right now. This is a "Markovian" process: the future is determined by the present, and the history doesn't matter. Scientists call this a "stochastic process," which is just a fancy way of saying a random walk with clear rules.

But now, imagine that machine is a quantum computer. It doesn't roll marbles; it waves like a ripple in a pond. In the quantum world, things can be in many places at once, and they can interfere with each other like overlapping waves. This makes it incredibly powerful, but it also makes it a nightmare to explain. If you try to trace the path of a quantum decision like a marble, the rules break. The future doesn't just depend on where the quantum "marble" is right now; it seems to depend on the entire history of where it could have been. This paper asks a big question: Can we force this weird, wave-like quantum behavior into a simple, marble-rolling story? And if we can, what do we have to give up to make it fit?

The authors of this paper, Johannes Fankhauser, Lukas J. Fiderer, and Hans J. Briegel, tackle this puzzle by trying to translate quantum machine learning models into the language of random walks. They discover that you can't have it all. You have to choose between two options, like a trade-off at a cosmic vending machine.

The First Option: The "Ghost" Walk
The first way to translate quantum dynamics is to use a map that is perfectly complete. Imagine you have a map of a city that shows every single street, alley, and rooftop. If you use this map, you can predict the next step of the journey just by looking at where you are right now. The rules are simple and immediate (this is called "Markovian"). However, there's a catch: to make the math work, the map has to include "ghost streets." These are paths that have negative probabilities. In our everyday world, a probability can't be negative—you can't have a -50% chance of rain. But in this quantum map, these negative numbers are necessary to cancel out the weird wave effects. So, you get a simple, step-by-step story, but it's a story that includes impossible, "ghostly" steps.

The Second Option: The "Memory" Walk
The second way is to use a map that only shows the main streets (like a standard street map). This map is "real"—every path has a positive, normal probability. There are no ghosts here. But because the map is incomplete, you can't predict the next step just by looking at where you are right now. To know where the traveler goes next, you have to remember where they were ten steps ago, or maybe twenty. The traveler needs a long memory. This is called a "non-Markovian" process. The rules aren't simple; they depend on the entire history of the walk.

The Big Trade-Off
The paper shows that these are the only two ways to tell the story of a quantum machine. You can either have a simple story with "ghost" steps (negative numbers), or a real story with a long memory (history dependence). You cannot have a story that is both simple and real. If you try to force a quantum process into a simple, real story, it breaks the rules of probability. If you try to force it into a simple, real story without memory, the math fails.

Why This Matters for Learning
The authors then take this idea and apply it to a model called "Projective Simulation." Think of this as a digital agent that learns by wandering through a network of memories (called "clips"). In the classical version, the agent walks from one memory to another, and you can see exactly which path it took. In the quantum version, the agent is still wandering, but the rules of the walk are different.

If you use the "Ghost" map, the agent's path is a straight line, but it's walking on a path that includes impossible steps. If you use the "Memory" map, the agent is walking on real paths, but its next step depends on a long chain of past memories. The authors show that even though the quantum agent is doing something weird, we can still understand it as a random walk if we are willing to accept either negative numbers or a long memory.

What the Paper Doesn't Say
It's important to note what this paper doesn't do. It doesn't say that quantum computers are actually walking through a maze of memories in the real world. The authors are careful to say this is just a way of interpreting the math, a way to make sense of the numbers. They don't claim to have solved the mystery of how the universe works, nor do they say that quantum computers are just classical computers with a longer memory. They simply show that if you want to describe quantum learning as a random walk, you have to choose your poison: either accept negative probabilities or accept that the system remembers everything.

The Bottom Line
In the end, this paper suggests that the "magic" of quantum learning isn't a single, mysterious force. It's just a different way of balancing the books. If you want the story to be simple, you have to write in a language with negative numbers. If you want the story to be real, you have to write a very long, detailed history. The authors provide a new dictionary to translate between these two languages, helping us see that even the strangest quantum decisions can be understood as a walk through a space of possibilities, provided we are flexible enough to change how we look at the path.

Technical Summary: Interpreting Quantum Learning Models via Stochastic Processes

Problem Statement
Quantum machine learning (QML) models define probabilistic input-output maps through parametrized quantum evolution and measurement. While these models offer potential computational advantages, their internal functioning resists interpretation as stochastic trajectories through intermediate configurations. In classical Markovian stochastic models, processes satisfy the Chapman-Kolmogorov divisibility condition, allowing probabilities to be decomposed into sequences of transitions between well-defined states. Quantum dynamics generically violate this condition; the probability of an outcome cannot generally be expressed as a sum over probabilities of intermediate configurations. Consequently, QML models lack a classical interpretation in terms of stochastic deliberation or paths. The central question addressed is whether quantum processes can be represented as stochastic dynamics on a suitably chosen configuration space, and if so, which features of such representations encode the non-classical structure of the underlying model.

Methodology
The authors develop a probabilistic framework representing quantum learning models as stochastic processes over configuration spaces, where dynamics are modeled as linear maps on probability distributions. The approach relies on fixing a set of positive operators (a POVM) to define the configuration space.

General Framework: A quantum state $\rho$ is mapped to a probability distribution $p_i = \text{Tr}(\rho E_i)$ via a fixed POVM $\{E_i\}$ . Quantum channels induce linear transition maps on these distributions.
Representation Analysis: The authors systematically compare two distinct regimes of configuration spaces:
- Informationally Complete (IC) POVMs: Representations where the POVM spans the space of operators (e.g., SIC-POVMs).
- Projective Representations: Representations based on orthogonal projectors (e.g., PVMs), which are not informationally complete.
Projective Simulation (PS) Integration: The framework is applied to Projective Simulation, a reinforcement learning model where deliberation is a random walk on a memory network. The authors generalize PS to accommodate the specific stochastic properties (quasi-stochasticity or non-Markovianity) derived from quantum representations.
Approximation Analysis: The paper investigates whether finite-order stochastic kernels can approximate quantum dynamics, particularly by exploiting periodicity in unitary evolutions or specific structures like stabilizer mechanics.

Key Contributions and Results

The Positivity-Divisibility Trade-off: The primary result is the establishment of a fundamental trade-off between the positivity of transition probabilities and the divisibility (Markovianity) of the process:
- IC-POVM Representations: These yield Markovian (first-order) dynamics that satisfy the Chapman-Kolmogorov condition. However, the transition kernels are generally quasi-stochastic, containing negative entries. The non-classicality of the quantum channel is encoded in this negativity. Crucially, while the kernel contains negative values, the resulting probability distributions over configurations remain positive and normalized because the kernel acts only on the subset of the probability simplex corresponding to valid quantum states.
- Projective Representations: These yield positive (strictly stochastic) transition kernels, preserving a direct interpretation in terms of stochastic trajectories. However, they generally violate the Chapman-Kolmogorov condition, requiring non-Markovian (higher-order) dynamics. The transition probabilities depend on a history of past configurations rather than just the current state.
Specific Examples:
- Qubit Dynamics: For a qubit undergoing unitary oscillation, the projective representation (basis states) requires a non-Markovian process where the transition depends on the initial state. In contrast, a SIC-POVM representation yields a first-order Markovian process with a quasi-stochastic kernel (negative entries).
- Stabilizer Dynamics: For single-qudit stabilizer mechanics, the projective representation results in a positive but non-Markovian process with a finite Markov order of $L=2$ , derived from the symplectic structure of the phase space.
Generalized Projective Simulation: The authors extend the PS framework to interpret quantum deliberation.
- In IC representations, the "deliberation" is a trajectory of admissible probability vectors (distributions) rather than a random walk over definite memory clips (clips). The "clips" act as coordinate labels rather than sharply occupied states.
- In Projective representations, the deliberation retains the interpretation of a random walk over clips, but the transition rules depend on the history of the walk (non-Markovian), recovering a notion of trajectories at the level of extended deliberation histories.
Finite-Horizon Approximations: The paper outlines that for certain regimes, such as unitary dynamics with periodic eigenphases or stabilizer mechanics, the history dependence required for positive stochastic representations can be bounded. This suggests that quantum dynamics can sometimes be approximated by finite-order stochastic processes with limited memory.

Significance and Claims
The paper claims to provide a unified framework for interpreting quantum learning models by distributing the non-classical features of quantum theory between negativity (in the transition kernel) and temporal dependence (memory/history).

It challenges the notion that quantum dynamics must be either non-classical in a way that forbids stochastic interpretation or classical in a way that preserves Markovianity. Instead, it shows that quantum dynamics can be represented stochastically by relaxing one of these two classical constraints.
The framework offers a route to interpret quantum deliberation processes (as in quantum PS) not as "spooky" non-stochastic events, but as either quasi-stochastic Markovian walks or positive non-Markovian walks.
The authors position their work as a representation of quantum probability dynamics rather than a commitment to a specific underlying ontology (e.g., hidden variables). They note that while the projective representation aligns with the "stochastic-quantum correspondence" proposed by Barandes, their broader framework demonstrates that such indivisible stochastic descriptions are just one realization within a larger class of probabilistic models.
Finally, the paper suggests that the complexity of quantum resources might be characterized by the Markov order required for their positive stochastic representations, though this remains an open question.

The work concludes that there is no unique stochastic description of a quantum process; rather, different representations distribute the non-classical features (negativity vs. memory) in different ways, providing a common language to compare classical stochastic models, probabilistic representations of quantum theory, and QML models.

Technical Summary: Interpreting Quantum Learning Models via Stochastic Processes

More like this