Flatiron's tensor network matches D-Wave's annealer on the same spin-glass dynamics

Joseph Tindall and colleagues at the Flatiron Institute in New York have simulated the time evolution of Ising spin glasses with a new tensor network scheme that, in many cases, is more accurate than the latest quantum annealers running the same problem ^[1][2]. That is consequential because a team of quantum computing researchers had recently run the same spin glass dynamics on D-Wave's Advantage2 annealer and claimed classical computers could not match their results ^[4].

The benchmark is a sensible one. An Ising spin glass is a lattice of spins pointing in random directions, a disordered state produced by different nearest-neighbour interactions for different pairs of spins ^[6]. Its difficulty grows with system size, which is exactly why it gets used to compare classical and quantum machines ^[7]. The Flatiron result is not a single lucky lattice: according to Physics World it scales in both two and three dimensions ^[3], and Tindall's group reports the classical scheme doing as well as, and sometimes better than, the annealer ^[5].

The mechanics matter for anyone judging how durable this is. Tensor networks encode a system as tensors whose legs, or bond indices, are joined by contraction; widening those legs raises the bond dimension and lets the network capture correlations more accurately ^[8]. The team wrote tensor network representations of both the spin glass Hamiltonian and the wave function ^[9]. The standing obstacle to time evolution is that entanglement accumulates as the state evolves, so bond indices must widen, driving up cost and sometimes making the calculation impossible ^[10].

Their workaround is belief propagation message passing for the contraction: instead of accounting exactly for every contribution, each tensor receives a compact summary of what the rest of the network looks like from its perspective ^[11]. Tindall describes this as a mean field approximation on each tensor's environment, because the messages arriving from each neighbour are assumed to be independent ^[12]. That approximation is the whole trade, and it bought enough headroom to evolve the system far longer than conventional tensor network schemes, far enough to reach the regime the annealer operates in ^[13].

Two cautions. First, this is one group's method against one machine, and Physics World has previously reported that some physicists were unconvinced by D-Wave's quantum advantage claim in the first place ^[14]. Second, the account gives no runtimes, node counts or hardware costs for the classical side ^[15], so "a classical computer can do this" is not yet the same statement as "a classical computer can do this cheaply."

What is settled is narrower and still useful: the premise that no classical method could reach this regime no longer holds for this problem ^[16]. The practical reading for anyone buying annealer hours is that the classical baseline is a moving target and has to be re-measured, not cited from a vendor deck.

Worth watching: whether an independent group reproduces the two- and three-dimensional results, whether the independence assumption behind belief propagation ^[12] degrades at larger sizes or longer times, and whether cost figures appear that let the comparison be made per dollar rather than per accuracy point.