The Signal
Arclight Authors Compile Imaginary-Time Filters for Reuse
A paper by Hengyue Li and colleagues describes a compiled approach to probabilistic imaginary-time evolution whose circuit size per attempt stays fixed as evolution time changes. Benchmarks show that its cost per accepted output depends on
Quantum.world Team · Written with AI · fact-checked ·
“Numerical tests show reduced online cost relative to termwise Pauli-based methods at a common accuracy threshold.”
Hengyue Li and co-authors, the paper’s authors
A compiled basis replaces repeated time steps
Hengyue Li and colleagues posted the paper, “Probabilistic imaginary-time evolution via compiled diagonal representations,” to arXiv on October 29. The method, which they call fast imaginary-time evolution (FITE), targets quantum simulations that use imaginary-time evolution to suppress higher-energy components and prepare states closer to a system’s ground state. Because the operation is nonunitary, the proposed circuit uses an ancilla and postselection: failed attempts are discarded.
FITE first compiles an approximate basis transformation for a Hamiltonian, then expresses it using commuting Pauli-Z strings. Each filter can cover the requested total imaginary time, rather than repeating short filters across many product-formula steps. For a fixed compiled representation, changing the evolution time changes filter angles, not the circuit’s structure or gate count per attempt; the representation can also be reused for different input states.
The authors describe basis compilation and coefficient recovery as offline work. Their method projects the physical Hamiltonian onto the retained diagonal terms to recover coefficients. But the main numerical comparisons use a phase-conversion shortcut for those coefficients, a distinction that matters when interpreting the benchmark results.
Gate counts vary across three small models
The tests cover a two-site Hubbard model encoded in two qubits, a five-qubit open XY chain and a three-qubit XYZ chain. The researchers compare FITE with first- and second-order termwise Pauli filters and a grouped two-site filter. Their main circuit benchmarks are noiseless, use complex-Haar random input states, and count logical CX gates; they exclude input preparation, offline compilation, coefficient estimation and hardware routing.
At a common one-percent mean-infidelity threshold, FITE uses 12, 58 and 50 CX gates per full attempt for the Hubbard, XY and XYZ examples, respectively, independent of imaginary time. At an evolution time of β = 1.1, the reported mean postselection probabilities are about 21, 263 and 123 times those of the first-order Pauli baseline. The gains in relative success do not guarantee high absolute success: for XYZ, FITE succeeds on only about 0.000409 of attempts at that time.
To assess accuracy, the paper tests 16,384 input states per model and sets the mean-infidelity threshold using an estimate plus 1.96 standard errors. The threshold applies to the sampled input ensemble, not to every possible input. The authors also report that reducing the compilation interval can lower errors, while the main results’ remaining error is dominated by bias in the first-order Trotter target.
Retries and grouped filters change the comparison
A gate count per attempt is only part of the cost when postselection can fail. At β = 1.1 in the XYZ example, FITE’s expected cost per accepted output is about 236,000 CX gates, compared with about 3,300 for the grouped-filter baseline. That baseline uses slightly more gates per full attempt—56 against FITE’s 50—but its higher success probability more than offsets the difference.
At β = 5, the expected FITE cost per accepted XYZ output rises to 7.21 × 1014 CX gates, even though each full attempt still uses 50. The result illustrates why a time-independent circuit size does not by itself establish a cheaper simulation. The authors’ comparison finds lower accepted-output costs than the termwise first- and second-order baselines across sampled times, and lower costs than grouped filtering for XY, but not for every model and baseline.
The scaling test is still a warning
The paper does not establish efficient scaling to larger systems. In its size tests, four- and five-qubit XYZ configurations remain above one-percent mean infidelity despite larger basis circuits and optimization searches; the authors say those tests do not establish efficient scaling. The practical case for FITE therefore depends on finding a sufficiently accurate, shallow compiled basis, keeping postselection probabilities usable, and reusing the offline work enough times to repay its cost.
For the three-qubit XYZ benchmark at β = 5, the paper’s own accounting puts the expected FITE cost at 7.21 × 1014 CX gates per accepted output. That is the clearest limit in the reported results: fixed gates per attempt do not mean fixed, or low, cost for obtaining a successful state.