The Signal
Adiabatic VQA Study Tests Entanglement Detection on Three Qubits
A University of Granada-led study tests adiabatic-inspired training for variational quantum algorithms that classify entangled and separable three-qubit states. Its 25-qubit reference describes earlier methods, not a result demonstrated by
Quantum.world Team · Written with AI · fact-checked ·
A paper posted on October 29 describes a test of adiabatic-inspired training for a variational quantum algorithm, or VQA, that classifies entangled and separable states. The test covers three-qubit systems—not 25-qubit detection. That larger number appears in the paper’s discussion of earlier methods, which the authors say can handle systems up to 25 qubits but require training separately for each state.
What the 25-qubit figure refers to
Ernesto Acosta, Carlos Cano, Daniel Manzano and Guillermo Botella propose encoding VQA training as a quadratic unconstrained binary optimization problem, or QUBO. Their method discretizes the circuit’s parameter space, divides it into regions, and refines the search around promising candidates in successive levels. The QUBO is designed for an adiabatic quantum annealer, which would search for a low-energy solution to the optimization problem.
The 25-qubit figure is background, not the new method’s demonstrated reach. The paper cites previous VQA techniques using a geometric measure to distinguish entangled from separable states at that scale, then identifies repeated, state-by-state training as a computational burden. The new study instead examines whether adiabatic-style optimization can train a classifier on a smaller, defined test set.
A 96-record test set, built from three-qubit states
The researchers generated 96 records across four categories: 48 separable states and 16 examples each of bipartite-entangled, GHZ and W states. They compared the adiabatic-trained VQA with a conventionally trained VQA and a classical artificial neural network. The reported experiments used a server with 32 CPU cores and 33.6 gigabytes of RAM; the paper does not report a demonstration on quantum-annealing hardware.
The adiabatic model averaged 67% accuracy on training data and 61% on validation data. The artificial neural network did better, averaging 75% on training and 65% on validation. Within the adiabatic model’s entangled-state results, accuracy was 65% for bipartite states, 60% for GHZ states and 55% for W states.
Accuracy is preliminary; runtime remains unmeasured
The authors report that performance improved at successive levels of the hierarchical search, but they also flag a trade-off: discretizing the parameter space sacrifices resolution and can miss a global minimum. They say execution-time comparisons still need to be evaluated, so the paper does not establish that the approach trains faster than its alternatives. The evolutionary-training baseline also remains incomplete because it exceeded the capacity of the team’s experimental infrastructure.
“While execution time comparisons remain to be fully evaluated,” write Ernesto Acosta, Carlos Cano, Daniel Manzano and Guillermo Botella, “the translation of the VQA training into an adiabatic framework introduces several advantages, including a fixed and potentially reduced number of computational steps independent of data volume, as the adiabatic process explores the entire state space in a single run.” Those possible advantages are not yet a measured speedup. The reported validation result for the adiabatic model is 61% on the three-qubit test.
The next test is scale and hardware
The authors argue that a VQA can work with quantum data without requiring a complete classical description of each state, a potential route to entanglement classification without full tomography. Their current results are a preliminary test of that idea, not a 25-qubit demonstration. The paper calls for more extensive experiments, including a completed evolutionary baseline; it also leaves execution-time comparisons open.