vqe-circuits

name

vqe-circuits

description

How the eigenstate-preparation pipeline is built from circuits: gate-based statevector simulation, the Pauli-rotation primitive, gate-level UCCSD and the hardware-efficient ansatz, the prolongation operator, adiabatic resolution refinement, and the rodeo algorithm. Use this skill WHENEVER the task involves VQE, UCCSD/HEA circuits, prolongation, resolution refinement, the adiabatic path, OR the rodeo algorithm (energy scans, eigenstate preparation, the cos^2 success probability, fidelity-vs-M, acceptance). Read `references/rodeo.md` before writing or modifying any rodeo code so the formulas and conventions match the published algorithm.

VQE circuits, resolution refinement, and the rodeo algorithm

The four pipeline stages, all implemented in pairinglib and reused (never re-implemented) by the notebook and analysis scripts.

1. Coarse-state preparation (gate-level UCCSD-VQE)

• Trotterised product ansatz prod_mu exp(theta_mu (tau_mu - tau_mu^dag)) |HF>.
• Each excitation compiles to commuting Pauli rotations via Jordan-Wigner: a single -> 2 Pauli strings, a double -> 8. Use gate_uccsd_setup/_vqe/_state.
• Optimise with the EXACT analytic product gradient. The naive two-term pi/4 parameter shift is WRONG beyond the first factor (later factors gain frequency-1 cross terms; the spectrum is {1,2}). This is a known trap — always use the analytic gradient already implemented in the library.
• Singles vanish for the pairing force (seniority): expect a doubles theory.

2. Prolongation

• prolong_matrix(k_low, k_high, N) embeds every low-res basis state into the same occupation in the larger space, leaving new levels empty: gate-free for basis refinement, O(N_sites) two-qubit gates per dimension for lattices.
• The prolonged state satisfies <Phi0|H_high|Phi0> = E_low exactly.

3. Resolution refinement (adiabatic phase)

• Path H(s) = cos^2(pi s/2) P (H_low - mu) P^dag + sin^2(pi s/2)(H_high - mu), s = t/T, with shift mu = E_low + 0.6 to keep tracked states below the large null space of P^dag. Use refine_state / refine_NK.
• The path stays gapped, so T ~ 1/Delta E (not 1/Delta E_min^2). Overlap rises toward 1; energy falls to FCI from above.

4. Rodeo algorithm (final filter) — see references/rodeo.md

• One cycle: ancilla Hadamard, controlled e^{-iHt_k}, phase P(E t_k), Hadamard, measure; post-select all-zero. The post-selected map is (I + e^{iEt} e^{-iHt})/2 (rodeo_post0), equal to the explicit ancilla circuit (rodeo_cycle_ancilla).
• Feed the rodeo the RESOLUTION-REFINED state (high overlap p): that is the whole point — cost ~ 1/p and a few cycles when p ~ 1.

Validation anchors (assert these when you touch the code)

• rodeo_cycle_ancilla == rodeo_post0 to ~1e-16.
• Eigenstate input reproduces prod cos^2(t_k (E_obj-E)/2) (Eq. 1); Gaussian average reproduces [(1+e^{-(E_obj-E)^2 sigma^2/2})/2]^M (Eq. 2).
• Refined input (N=4, k=4, p~0.998) -> fidelity ~ 0.999995 within ~2 cycles, acceptance -> p; off-target background ~ 2^-M.