Classical Computers Just Beat A Quantum Annealer At Its Own Benchmark

[Rachel_72]

Researchers at the Flatiron Institute in New York have used a new classical tensor network technique to simulate Ising spin glass dynamics, and in many cases their method turned out to be more accurate than D-Wave's Advantage2 quantum annealer running the exact same problem. The work, led by Joseph Tindall and published in Science, directly challenges an earlier claim that classical computers could not match the annealer's results on this specific benchmark.

The Ising spin glass model describes spins on a lattice pointing in essentially random directions due to competing interactions between neighboring pairs, and its difficulty scales up sharply with system size, which makes it a genuinely useful benchmark for comparing classical and quantum performance head to head. Tensor networks represent the state of a many particle system using interconnected mathematical objects that can be thought of a bit like LEGO pieces snapping together, where widening the connections between them lets the network capture more correlation detail at the cost of more computation.

The real trick behind this specific result was a technique called belief propagation, where each tensor receives a compact summary of what the rest of the network effectively looks like from its own local perspective rather than accounting for every single contribution exactly. That approach let the team's classical simulation push forward in time far enough to actually reach the same regime the quantum annealer operates in, something conventional tensor network methods have historically struggled to do without the computational cost exploding.

Across cylindrical, diamond, and cubic lattice geometries, the classical method's error in measuring how spins at different points on the lattice relate to each other came in lower than the annealer's on two of the three geometries, and roughly matched it on the third. That is a meaningfully different outcome than simply matching performance across the board, since it suggests the classical technique is not just catching up but genuinely outperforming the quantum hardware on some specific versions of this problem.

The team plans to extend this same belief propagation approach to interacting electronic systems and finite temperature problems next, continuing what has become a fairly consistent pattern of classical algorithms closing gaps that earlier quantum advantage claims assumed were permanent

No quantum advantage (yet) in the world of tensor networks - Physics Worldphysicsworld.com

[Molly32]

This is exactly the pattern that keeps playing out with quantum advantage claims and it never stops being a little bit funny to watch happen in real time. Someone announces a quantum computer beat classical methods, then a year or two later some grad student or research group finds a smarter classical algorithm that closes the gap right back up again.

[Coastal Otter]

The plan to extend this same belief propagation approach toward the Hubbard model next is honestly the part I am most excited about here. Interacting electronic systems are directly relevant to real materials science and superconductivity research, so any genuine progress there would matter a whole lot more practically than another spin glass benchmark eventually would

[Phil]

The specific detail that the classical method won on two geometries and only tied on the third is honestly the most interesting and most credible part of this whole story to me. A clean sweep across every single geometry tested would have actually made me more suspicious that something in the methodology or comparison itself was subtly off

[Jamie]

D-Wave's original claim was already getting side eye from a lot of physicists when it first came out, so this result does not feel like a huge shock to me personally.
Annealers in particular have a long and pretty rocky history of advantage claims that eventually got walked back once someone found a cleverer classical shortcut to the exact same answer.

[ShawnMichaels]

Curious how this direct classical result actually lands with the D-Wave side of things given how much of their overall commercial pitch has leaned specifically on quantum advantage claims tied to problems just like this one. A high profile paper in Science undercutting one of their headline results is not exactly a great look for their broader investor facing narrative

[Dark Elizabeth]

Open sourcing their actual tensor network simulator library is a genuinely good move for reproducibility in this field specifically. So much of the quantum versus classical debate over the past several years has hinged on results that were genuinely difficult for outside groups to independently verify or properly reproduce

[Stag]

Belief propagation as a mean field style approximation on each tensor's local environment is a genuinely clever way to sidestep the usual explosion in computational cost that comes with tracking every single correlation exactly. Trading a small amount of theoretical precision for a massive gain in how far forward in time you can actually simulate seems like an extremely worthwhile tradeoff here specifically.