Our paper: arXiv:2607.24014
Authors: Iordanis Kerenidis (Quantum Signals)
Quantum machine learning faces three problems at scale: models stop training (barren plateaus), there is often no reason to believe a classical computer could not do the same job, and computing gradients takes too many circuit runs. This paper proposes one architecture that handles all three.
Why we wrote it
Most quantum models solve one of these problems at the expense of another. A model that trains well is often easy to simulate classically, which removes any reason to run it on a quantum computer. We wanted a design where trainability and classical hardness are both guaranteed, and where a single parameter sets the trade-off between them.
How it fits our work
This is the theoretical foundation of our quantum program. It sets design rules for any quantum version of Pythia: train the circuit through statistics that can be computed classically, and take the value from samples that cannot be. Our research on quantum samplers, where a trained circuit runs on a quantum computer and a classical transformer reads its output, is built on the architecture in this paper.
What it adds to our approach
- Trainable. The unitary brick-wall reaches every unitary, so it can start from a random point, and two-body readouts keep gradient variance polynomial in the number of qubits, so there are no exponential barren plateaus.
- Expressive. A particle number k controls classical hardness: the best-known classical sampling algorithms run in time exponential in k.
- Efficient. A multi-layer parallel parameter-shift rule computes all O(n²) gradients from 4kn circuit evaluations instead of 4n², a factor n/k that grows with circuit size.
- Two hardware targets. A brick-wall layout for nearest-neighbour machines and a butterfly layout for all-to-all machines, with the same guarantees.
Key takeaway
Gradient cost drops by a factor n/k: at the paper's operating point of k = 100 on a 1,000-qubit circuit, that is 10x fewer circuit evaluations per training step, and the saving grows linearly with circuit size. At that same operating point, a single classical query is estimated at about 1015 operations, and at k = 120 even an exascale machine would manage less than one query per second.
Read the full paper on arXiv. For questions about our research, contact info[at]quantumsignals.ai.