When the Whole Story Arrives at Once
26 Apr 2026, Yanjiang
Imagine trying to watch a film one frame at a time. You see the opening shot, then pause. You advance to the next frame, pause again. Frame by frame, you reconstruct the plot — but something essential is lost. The momentum of a chase scene, the crescendo of a musical moment, the way a character’s expression shifts over seconds rather than stills. You can piece together what happened, but you cannot feel how it unfolded.
This is the predicament that has long faced physicists simulating quantum dynamics. The standard approach — time-stepping — advances a quantum state incrementally, one tiny slice of time at a time. It works, but it is expensive, and for long-time simulations it accumulates errors like a game of telephone whispered across generations. A preprint (arXiv:2412.11778) from a team at the Institute of Physics, École Polytechnique, and CNRS proposes something different: what if you could compute the entire trajectory at once?
The Global View
The idea, developed by Alessandro Sinibaldi, Douglas Hendry, Filippo Vicentini, and Giuseppe Carleo, is deceptively elegant. Instead of marching forward in tiny steps, they define a loss function that measures how well a proposed trajectory satisfies the Schrödinger equation across a finite time window. Then they minimize that loss — all at once, for the whole path.
Think of it like planning a road trip. The conventional method would be to drive one mile, check the map, drive another mile, check again. You will eventually reach your destination, but you might take wrong turns, and the longer the trip, the more those small errors compound. Sinibaldi and colleagues instead propose: look at the entire route from start to finish, and adjust it so that every segment simultaneously satisfies the laws of motion. Unlike a road trip, where the whole-route approach might be impractical, for quantum systems this global-in-time perspective unlocks something the step-by-step method cannot easily provide: a rigorous bound on the error.
“Because the loss function measures deviation from exact evolution over the entire window,” the authors explain in their preprint, “we can bound the error of the final state by the value of the loss itself.” This is not merely a technical convenience — it is a conceptual shift. In time-stepping, errors accumulate silently. In the global approach, the loss function acts as a diagnostic: if the loss is small, the trajectory is trustworthy. If it is large, you know immediately that something has gone wrong.
Neural Quantum States as Building Blocks
To represent the quantum state trajectory, the team turns to a technology that has reshaped computational physics in recent years: neural quantum states. These are artificial neural networks trained to represent quantum wavefunctions — the mathematical objects that encode everything knowable about a quantum system. Unlike traditional basis sets (plane waves, atomic orbitals), neural networks can efficiently represent highly entangled states that would otherwise require exponentially many parameters.
The innovation in this work is how the neural network is deployed. The team uses a Galerkin-inspired ansatz — a term borrowed from engineering mathematics that essentially means: express the time-dependent state as a linear combination of time-independent building blocks, each weighted by a time-dependent coefficient. The building blocks are neural quantum states, each a neural network trained to represent a particular configuration. The coefficients are simple functions of time.
This structure is powerful because it separates two kinds of complexity. The neural networks handle the spatial complexity — capturing how particles are correlated across the system. The time-dependent coefficients handle the temporal dynamics — how the state evolves. The result is a method that can simulate long-time dynamics without the computational cost exploding.
Testing the Method: Quantum Quenches
The team tested their method on a classic problem: global quantum quenches in the Transverse-Field Ising model, a paradigmatic system for studying phase transitions and quantum magnetism. A quench is what happens when you suddenly change a parameter in the Hamiltonian — like abruptly turning up the magnetic field. The system, initially in equilibrium, is thrown into a non-equilibrium state and must evolve toward a new equilibrium.
Simulating such dynamics is notoriously difficult. The system explores a vast region of Hilbert space, and the entanglement grows rapidly, making conventional methods struggle. The team’s approach, which they call the Time-Dependent Neural Galerkin method (tDNG), proved competitive with state-of-the-art time-dependent variational Monte Carlo — the current workhorse of neural quantum dynamics.
But the real surprise came in two dimensions. In the 2D Ising model, the team observed signatures of ergodicity breaking — the system failed to thermalize, remaining trapped in a subset of its possible states. This is not what one would naively expect. In large systems, one typically assumes that dynamics will explore all available phase space, eventually reaching thermal equilibrium. The 2D Ising model, under certain quench conditions, appears to resist this expectation.
Why this happens remains an open question. The team’s method provides a window into this behavior, but the mechanism — whether it is related to integrability, many-body localization, or something else entirely — will require further investigation. What matters is that tDNG can access regimes that were previously inaccessible, offering a new tool for exploring the rich landscape of non-equilibrium quantum matter.
What This Means
At its core, this work is about a fundamental tension in how we understand quantum dynamics. The Schrödinger equation is deterministic: given an initial state, the future is uniquely determined. But computing that future is exponentially hard as the system grows. Every method we have is a compromise — trading accuracy for tractability, or vice versa.
The Time-Dependent Neural Galerkin method does not solve this problem entirely, but it shifts the trade-off. By computing entire trajectories rather than stepping through time, it offers a different kind of access to quantum dynamics — one that is particularly suited to the long-time behavior that matters for understanding thermalization, ergodicity, and the foundations of statistical mechanics.
For a field that has long treated time as something to be marched through one step at a time, this framework opens a door. Whether that door leads to new understanding of non-equilibrium phases, or to practical algorithms for quantum simulation, depends on what comes next. But the path is now visible.
Perhaps, in the future, when physicists want to understand how a quantum system evolves, they will no longer ask “What happens at each moment?” but instead “What is the whole story?” — and the answer will come all at once.
Yanjiang is an online editor of Loom Science
References
- Alessandro Sinibaldi et al., Time-dependent Neural Galerkin Method for Quantum Dynamics, arXiv:2412.11778