---
title: 'GAS-SCF: Grover Adaptive SCF Optimization'
url: https://www.emergentmind.com/topics/grover-adaptive-search-self-consistent-field-gas-scf
type: topic
---

# GAS-SCF: Grover Adaptive SCF Optimization

Searching arXiv for the specified paper and closely related context.
Grover Adaptive Search Self-Consistent Field (GAS-SCF) is a hybrid quantum–classical algorithm for self-consistent field optimization in quantum chemistry that applies Grover’s amplitude amplification and the Dür–Høyer minimum-finding paradigm to the discrete inner optimization over single-determinant Fock states, while leaving the continuous outer optimization over molecular-orbital (MO) basis rotations to classical procedures. In this formulation, the SCF objective is treated as a nested minimization,
$$
E_{\mathrm{SCF}}=\min_{K}\min_{|x\rangle\in V}\langle x|H_{\mathrm{QUBO}}(K)|x\rangle,
$$
where $K$ denotes the current orthonormal MO basis and $V$ the target symmetry sector. GAS-SCF constructs a quantum oracle that marks occupation-number bitstrings whose energy lies below a threshold $\tau$ derived from a classical SCF solution, amplifies those marked states, and adaptively lowers $\tau$ whenever a better determinant is found. The method is presented as a rigorous baseline for discrete SCF optimization, with a theoretical quadratic speedup in the inner search relative to exhaustive search, and with numerical demonstrations ranging from small statevector simulations up to resource-oriented analyses for systems as large as 330 qubits [2606.20176].

## 1. Definition and SCF formulation

GAS-SCF is defined as a quantum-accelerated procedure for the inner discrete optimization step that appears inside Hartree–Fock or mean-field SCF calculations. The basic decomposition is: for a fixed MO basis, identify the best single determinant in the chosen symmetry sector; then update the MO basis classically and iterate. The quantum component targets the determinant search, whereas the orbital-rotation step remains classical [2606.20176].

The motivation is tied to the structure of SCF itself. For a given basis $K$, the algorithm evaluates the diagonal second-quantized energy of a determinant,
$$
f(x)=\langle x|H_{\mathrm{QUBO}}(K)|x\rangle,
$$
with $|x\rangle$ a Fock state encoded as a computational-basis bitstring. In the occupation basis,
$$
H_{\mathrm{QUBO}}(K)=\sum_w h_{ww}(K)x_w+\frac12\sum_{m\neq w}\big[g_{mmww}(K)-g_{mwwm}(K)\big]x_mx_w,
$$
where binary variables $x_w\in\{0,1\}$ encode spin-orbital occupancies. This makes the discrete SCF subproblem an explicit QUBO-style optimization over determinants in a constrained sector.

For restricted Hartree–Fock, the orbital-occupation energy can be written as
$$
E(n)=\sum_i n_i h_{ii}+\frac12\sum_{i,j} n_i n_j \big[2\langle ii|jj\rangle-\langle ij|ji\rangle\big].
$$
The outer mean-field update is linked to the Roothaan–Hall equations,
$$
\mathbb{F}\mathbb{C}=\mathbb{S}\mathbb{C}\boldsymbol{\epsilon},
$$
with $\mathbb{F}=\mathbb{H}_{\mathrm{core}}+\mathbb{G}(\mathbb{D})$ and, in RHF,
$$
D_{pq}=2\sum_{i\in \mathrm{occ}} C_{pi}C_{qi}.
$$
UHF uses spin-resolved densities and Fock matrices. GAS-SCF does not replace this classical outer-loop machinery; it modifies the way the determinant is selected inside that loop.

The method is motivated by the observation that the discrete SCF step is NP-complete in the worst case, and that standard classical heuristics such as Fock diagonalization or second-order orbital optimization can stall in local minima or miss lower-energy determinants, particularly near degeneracy, broken symmetry, and multiple-minima regimes. A plausible implication is that GAS-SCF is aimed less at routine weakly correlated cases than at regimes in which the quality of the single-determinant reference itself is unstable or ambiguous.

## 2. Oracle, encoding, and adaptive search mechanism

The quantum encoding assigns one qubit to each spin-orbital, so that a computational basis state $|x\rangle$ represents a determinant by its occupation pattern. Two initial-state strategies are used. The first prepares the full uniform superposition $|s\rangle$ over all bitstrings and enforces constraints with ancilla-assisted checks. The second prepares a Dicke state $|d\rangle$, i.e. an equal superposition over basis states with the required Hamming weights for fixed $N_\alpha$ and $N_\beta$, thereby restricting the search subspace ab initio [2606.20176].

The marking rule is threshold-based:
$$
E(x)<\tau \iff f(x)-\tau<0.
$$
The oracle computes $f(x)$ into a cost register, adds $-\tau$, and uses the sign bit of the two’s-complement result to identify whether the determinant improves upon the current threshold. In the target symmetry sector $V$, the phase oracle acts as
$$
S_f|x\rangle=
\begin{cases}
-|x\rangle,& f(x)-\tau<0 \text{ and } |x\rangle\in V,\\
|x\rangle,& \text{otherwise}.
\end{cases}
$$

Amplitude amplification is then applied with reflection about the chosen initial state,
$$
U_s=2|\psi_0\rangle\langle\psi_0|-\mathbb{I},
$$
yielding the Grover iterate
$$
G=U_sS_f,
$$
or equivalently
$$
Q=-AS_0A^{-1}S_f.
$$
If the search space has $N$ states and $T$ marked states, then with $\sin\theta=\sqrt{T/N}$ the success probability after $L$ iterations is
$$
P_{\mathrm{success}}=\sin^2[(2L+1)\theta],
$$
and the expected optimal number of iterations is approximately
$$
k\approx \left\lfloor \frac{\pi}{4}\sqrt{\frac{N}{T}} \right\rfloor.
$$
When $T$ is known, the paper uses the exact arcsin-based expression labeled Eq. (6); when $T$ is unknown, randomized-$L$ or fixed-point strategies are proposed.

The “adaptive” aspect follows the Dür–Høyer minimum-finding paradigm. After measurement, the resulting bitstring $x^\ast$ is classically evaluated. If $E(x^\ast)<\tau$, the threshold is updated to $\tau\leftarrow E(x^\ast)$ and the search is repeated. As $\tau$ decreases, the marked set shrinks, so the optimal amplification depth changes during execution. This suggests that GAS-SCF is not a one-shot Grover search but an iterative threshold-tightening procedure embedded in the SCF loop.

## 3. Hybrid workflow and circuit components

The algorithm takes as inputs the one- and two-electron integrals $h_{pq}(K)$ and $g_{pqrs}(K)$ in the current MO basis, a target symmetry sector with fixed $N_\alpha$ and $N_\beta$ and possibly spin or point-group constraints, and an initial threshold $\tau$ obtained from a classical SCF calculation. The system register contains $n$ qubits for spin-orbital occupations. The cost register uses $m+1$ qubits, where $m$ is chosen from
$$
m=\left\lfloor \log_2 \max_x |f(x)| \right\rfloor \le \left\lfloor \log_2 \sum_{ij}|Q_{ij}|+\sum_i|b_i| \right\rfloor,
$$
with one extra sign qubit. In the uniform-$|s\rangle$ variant, optional number registers of sizes
$$
\mu=\nu=\left\lfloor \log_2(n/2)\right\rfloor
$$
encode electron counts for symmetry enforcement [2606.20176].

Energy evaluation is implemented with quantum arithmetic. The QFT-based construction prepares the cost register in the Fourier basis using Hadamards and phase gates, applies controlled additions for the linear and quadratic terms in $f(x)$, and then performs an inverse QFT to return to the computational basis. Alternatively, a fault-tolerant design without QFT uses singly and doubly controlled adders while retaining two’s-complement encoding. Noninteger integrals are handled by scaling by $2^E$ and rounding to integers, which preserves eigenstates and rescales eigenvalues; the corresponding register size becomes
$$
m_E=E+\left\lfloor \log_2 \max_x |f(x)| \right\rfloor.
$$
The paper states that, practically, $E\approx 50$ suffices to match typical 15-digit precision of integrals.

A single GAS round consists of preparing $|\psi_0\rangle=|s\rangle$ or $|d\rangle$, executing $L$ Grover iterations, measuring the system register, classically evaluating the sampled determinant, and possibly updating $\tau$. If a GAS round improves the threshold, the MO basis may then be updated classically using orbital-rotation heuristics such as Roothaan–Hall diagonalization or second-order methods, after which the integrals are rebuilt and the process repeated until SCF convergence or a stopping condition based on energy change, gradient norms, or stability tests.

The gate structure is dominated by energy accumulation and constraint handling. For each repetition, cost accumulation scales as $O(nm)$ singly controlled phase gates for linear terms and $O(n^2m)$ for quadratic terms in the QFT-based version; the fault-tolerant adder version has total T-count scaling $O(nm+n^2m)$ with $O(m)$ clean ancillas. In the uniform-$|s\rangle$ variant, symmetry enforcement adds $O(\log n)$ Hadamards for the count registers, $O(n)$ singly controlled phase gates for count accumulation, $O(1)$ multi-controlled X gates for exact Hamming weights, inverse QFTs on the count registers, and a Toffoli to combine $\alpha/\beta$ flags. Reflection about $|s\rangle$ is implemented through $H^{\otimes n}$, whereas Dicke-state reflection can cost up to $O(n^2)$ depending on $k$.

## 4. Complexity, baseline role, and relation to structured quantum optimization

At the level of query complexity, GAS-SCF inherits the standard amplitude-amplification scaling: the expected number of Grover iterations is $O(\sqrt{N/T})$, compared with classical $O(N/T)$ sampling over an unstructured domain. The total complexity is therefore
$$
O(\sqrt{N/T}\cdot C_{\mathrm{oracle}}),
$$
where $C_{\mathrm{oracle}}$ denotes the cost of energy evaluation and symmetry checks [2606.20176].

The central caveat is that the oracle may dominate the end-to-end runtime. For dense integral tensors, the two-electron term yields $O(n^2m)$ gate-level additions per repetition, so asymptotic quadratic speedup in the number of search queries does not automatically imply practical advantage in a fault-tolerant regime. The paper therefore positions GAS-SCF as a benchmark rather than as a claim of near-term superiority.

This baseline role is explicit. GAS-SCF marks all states that improve on the best known classical energy and then amplifies them without presuming additional exploitable structure. Structured approaches such as QAOA, DQI, and AQC are accordingly framed as methods that would need to beat this warm-started amplitude-amplification baseline by exploiting problem structure such as graph locality, low-rank integrals, tailored mixers or ansätze, or interference-based decoders. A plausible implication is that GAS-SCF serves as a reference point for distinguishing advantage arising from genuinely structure-aware design from advantage attributable merely to access to the same classical warm start.

The paper also distinguishes the two search-space variants in terms of robustness. In the full uniform-$|s\rangle$ formulation, unmarked states outside the symmetry sector remain in superposition, which keeps the marked fraction below $1/2$ for $n_s>1$ spatial orbitals and thereby avoids overbalancing. In the Dicke-state variant, by contrast, a poor threshold can mark more than half the valid states, causing the algorithm to amplify unmarked states instead. For such cases, overbalanced-domain methods such as Faro–Marino strategies are suggested.

## 5. Numerical demonstrations and classical simulations

The reported demonstrations are classically simulated proofs of concept, with the largest explicit statevector simulation using 26 qubits total including ancillas, and with additional analyses extending to larger systems up to 330 qubits [2606.20176].

For $\mathrm{H}_3^-$ in STO-3G, the uniform-$|s\rangle$ variant uses $N=2^6=64$ system states and, when $\tau$ is below the true ground energy, has $T=1$ and optimal $L\approx 6$. The observed Grover oscillations match
$$
\sin^2[(2L+1)\theta],\qquad \theta=\arcsin(1/\sqrt{64}).
$$
In the Dicke-$|d\rangle$ variant for the chosen $(N_\alpha,N_\beta)=(2,2)$ sector, the valid-state count is $N=9$, again with $T=1$ under a tight threshold and $L\approx 2$, with oscillations following the same formula and $\theta=\arcsin(1/3)$. For LiH in STO-3G, using only the Dicke variant, the valid-state count is $N=225$ with $T=1$, and the observed amplification is consistent with $\theta=\arcsin(1/15)$.

The overbalancing pathology is illustrated on $\mathrm{H}_3^-$ in the Dicke variant with $\tau=-4$, where more than half the valid states are marked and the algorithm amplifies unmarked states rather than improving ones. The uniform-$|s\rangle$ variant avoids this failure because off-sector states remain present and unmarked.

For $\mathrm{OH}^-$ in 6-31G with active space $(6o,(3e,3e))$, using 12 system qubits in the Dicke variant, the valid-state count is
$$
N=\binom{6}{3}^2=400.
$$
A classical RHF seed from PySCF gives $\tau\approx -780$ in rounded units and corresponds to a local minimum. With that threshold, $T=12$ states lie below $\tau$. Using $L\approx 4$ from Eq. (6), statevector sampling with 10,000 shots returned only marked bitstrings, all of which improve upon the RHF solution; the largest reported improvement is approximately $2.2\,\mathrm{eV}$. This is presented as a direct demonstration that amplitude amplification can boost the entire set of lower-energy determinants relative to a classical starting point.

The O$_2$ triplet at bond length $2.55\ \text{\AA}$ is used to probe the quality of classical single-determinant references. In STO-3G, the overlap $(\mathrm{FCI}|\mathrm{ROHF})=2.17\times 10^{-7}$ indicates that ROHF is a poor single-determinant reference. Higher CI truncations give $(\mathrm{CI\text{-}12}|\mathrm{ROHF})=0.959$ and $(\mathrm{CI\text{-}1234}|\mathrm{ROHF})=0.0475$, showing that lower-order truncations can overstate ROHF’s importance. The study also finds single Fock states with substantially higher overlap than ROHF, such as $(\mathrm{FCI}|104804210)=-0.354$ in STO-3G and $(\mathrm{CI\text{-}1234}|6864771481610)=0.552$ in 6-31G, indicating the presence of better determinants outside the classical ROHF minimum.

For linear triplet O$_3$ with O–O bonds of $2\ \text{\AA}$, the paper considers cc-pVDZ, cc-pVTZ, and cc-pVQZ basis sets with 84, 180, and 330 data qubits respectively. The corresponding search-space sizes in the target sector are approximately $1.09\times 10^{20}$, $6.84\times 10^{28}$, and $2.92\times 10^{35}$. Across six initialization strategies—minao, 1e, atom, huckel, vsap, and sap—and two SCF flavors, ROHF and SO-ROHF/Newton, GAS-style brute-force sector searches with a five-minute limit found lower-energy single-determinant solutions than PySCF in 34 out of 36 cases. Reported ROHF improvements $\Delta E=E_{\mathrm{GAS}}-E_{\mathrm{PySCF}}$ include $-0.155568\ \mathrm{Ha}$ for sap in cc-pVTZ, $-0.133417\ \mathrm{Ha}$ for minao in cc-pVDZ, and $-0.123688\ \mathrm{Ha}$ for sap in cc-pVQZ. Reported SO-ROHF improvements include $-0.099283\ \mathrm{Ha}$ for 1e or sap in cc-pVDZ, $-0.078899\ \mathrm{Ha}$ for minao/vsap/sap in cc-pVTZ, and $-0.069503\ \mathrm{Ha}$ for sap in cc-pVQZ.

## 6. Resource requirements, limitations, and prospective developments

Chemically relevant applications of GAS-SCF are explicitly tied to large-scale, fault-tolerant quantum hardware. The resource picture includes system qubits for spin-orbitals, a cost register of size $m+1$, optional electron-count registers of size $\mu,\nu\approx \log_2(n/2)$ in the uniform-$|s\rangle$ formulation, and clean ancillas scaling as $O(m)$ for fault-tolerant adders or up to $O(n)$ when controlled adders are also used for number counts. In the O$_3$ examples, the data-qubit counts alone are 84, 180, and 330 for cc-pVDZ, cc-pVTZ, and cc-pVQZ respectively [2606.20176].

Fault-tolerant arithmetic dominates T-gate budgets. A singly controlled adder requires $4(m-1)$ T gates and $m-1$ clean ancillas; a doubly controlled adder costs $4m$ T gates and $m$ ancillas. Since each linear or quadratic term contributes one such adder, total T-count per repetition scales with $O(n)$ linear and $O(n^2)$ quadratic terms. By contrast, marking through controlled $R_z(-2\pi)$ and multi-controlled X/Z gates is described as Clifford-dominated and comparatively cheap in T-count.

The main limitations are threefold. First, overbalancing can break the Dicke-state variant when $\tau$ is too loose. Second, oracle dominance can erase the practical benefit of Grover scaling, especially with dense two-electron integrals and high precision. Third, the procedure is sensitive to arithmetic and phase errors, since the comparison depends on correct two’s-complement sign evaluation and near-degenerate states require enough precision to preserve ordering. The paper therefore frames fault tolerance as a precondition for chemically relevant deployment.

Several optimization directions are identified. These include exploiting integral sparsity and screening, low-rank factorization of two-electron integrals, use of point-group symmetry to reduce $N$, replacement of QFT-based arithmetic by optimized fault-tolerant adders, use of qubitization or more efficient arithmetic layouts, improved Dicke-state preparation, and the use of Weyl bounds on $T$ to choose $L$. Future work also includes systematic benchmarking against advanced classical SCF heuristics and post-HF pipelines, hybrid alternation between quantum GAS rounds and classical orbital optimization, integration of structure-aware optimizers such as QAOA and DQI on top of the GAS-SCF baseline, and extension of symmetry constraints to more refined spin sectors via Weyl-formula bounds.

Taken together, these features place GAS-SCF at the intersection of quantum search, arithmetic-heavy oracle design, and mean-field electronic-structure optimization. Its significance lies not only in the quadratic inner-loop speedup but also in its role as a concrete, verifiable baseline for identifying better single-determinant solutions in SCF regimes where conventional heuristics can become trapped or misleading.

Source: https://www.emergentmind.com/topics/grover-adaptive-search-self-consistent-field-gas-scf