Four ideas, in motion
QUBO decomposition, the HADOF loop, quantum genome assembly, and the optimisation results they produce — animated, in the order they actually run.
One big problem, many small ones
A QUBO over hundreds or thousands of variables is a dense matrix of pairwise couplings — far more than any current QPU can hold at once.
HADOF repeatedly carves out small, overlapping sub-blocks of that matrix, each one small enough to fit on a handful of physical qubits. Only the highlighted 3×3 block runs on the QPU at any moment — its couplings to the rest enter as expected values under the current marginals.
Encode, solve, estimate, update, aggregate
Each sub-Hamiltonian is solved on real hardware, its samples turned into marginal probabilities for every variable, and those marginals fed back in to steer the next round of sub-problems.
The loop repeats until the global candidate solutions stop improving — a federated search, not a single giant circuit.
Read the full method →Reads → overlap graph → one sequence
Sequencing reads become nodes; suffix–prefix overlaps become edges. Reconstructing the genome means finding the right path through that graph — structurally a travelling-salesman problem.
HADOF solved this directly on real sequencing data: a 7.3 Mbp bacterial genome, a graph of 524 nodes and 2,313 edges, requiring far more qubits than any current device offers end to end.
Decomposition holds up under noise
On IBM's 156-qubit ibm_aachen, topology-aware HADOF pushed sub-circuits from 5 qubits to 150, and SA-HADOF reached 98.5–98.8% of full Simulated Annealing's solution quality.