---
title: Permutation-Preserving Ansatz
url: https://www.emergentmind.com/topics/permutation-preserving-problem-inspired-ansatz
type: topic
---

# Permutation-Preserving Ansatz

Searching arXiv for the primary paper and closely related work on permutation-preserving ansätze.
Permutation-preserving problem-inspired ansatz denotes a class of variational constructions in which the circuit architecture is tailored to permutation-structured feasible sets, so that the ansatz acts directly on valid permutations, tours, schedules, or closely related combinatorial objects rather than on a larger constrained space that must later be penalized. In the quantum-variational setting, the most explicit use of this phrase appears in QuPer, introduced for permutation-based combinatorial optimization such as quadratic assignment and graph isomorphism [2505.05981]. Closely related constructions arise in compactly encoded traveling-salesman solvers [2605.00739, 2508.21730], graph-controlled feasibility-preserving mixers for scheduling [2311.04100], and permutation-invariant circuit families exploiting symmetry reduction [2312.14909]. Across these works, the common design principle is that permutation structure is not treated as an afterthought: it determines the encoding, the elementary generators, the mixer or variational blocks, and often the post-processing map to the final discrete solution.

## 1. Concept and defining characteristics

In QuPer, the optimization target is written as
\[
\min_{P\in\Pi_n}\; f(P),
\]
where \(\Pi_n\) is the set of all \(n\times n\) permutation matrices and \(f:\Pi_n\to\mathbb R\) is an arbitrary cost [2505.05981]. The ansatz is problem-inspired because its circuit is derived from group-theoretical structure specific to permutations rather than from a generic hardware-efficient template. It is permutation-preserving in the sense that its elementary gates span a subgroup of permutation unitaries, and, with ancilla augmentation, generate convex combinations of such permutations represented as doubly-stochastic matrices before classical projection back to \(\Pi_n\) [2505.05981].

A related but stricter meaning of permutation-preserving appears in compact TSP formulations. In the Lehmer-factorial encoding of tours, every computational-basis state corresponds to exactly one valid tour, so “there is no need for additional penalty terms or one-hot constraints” [2508.21730]. In the resource-efficient TSP framework, the ansatz is built from controlled register-swaps acting on whole city registers; because these blocks “never produce invalid codes or duplicate-city configurations,” the evolution remains confined to the valid permutation subspace [2605.00739].

For scheduling-type problems, the same principle is implemented at the mixer level rather than in the encoding. Graph-controlled Permutation Mixers in CM-QAOA perform small permutations of assignment bits only when a Boolean feasibility predicate certifies that the move remains in the feasible set [2311.04100]. This preserves constraints by construction rather than by adding energetic penalties.

These variants suggest a useful taxonomy. One branch preserves permutations by **exact encoding**, another by **feasibility-preserving dynamics**, and another by **symmetry-preserving parameter tying**. A plausible implication is that the phrase encompasses a family of techniques rather than a single circuit template.

## 2. Group-theoretical and encoding foundations

QuPer starts from amplitude encoding with \(n=2^q\), where a \(q\)-qubit register in state \(\sum_{i=0}^{2^q-1}\alpha_i\ket i\) encodes an \(n\)-dimensional vector [2505.05981]. The relevant gate-generated group is
\[
LX_q=\langle X_q\cup CX_q\rangle,
\]
with
\[
X_q\cong\mathbb Z_2^q,\qquad CX_q\cong GL_q(\mathbb F_2),
\]
and every \(U\in LX_q\) uniquely factorized as
\[
U=U_XU_{CX},\quad U_X\in X_q,\;U_{CX}\in CX_q.
\]
The paper then invokes the Bruhat factorization
\[
GL_q(\F_2)=BWB,
\]
where \(B\) is the Borel subgroup of invertible upper-triangular matrices and \(W\cong S_q\) is the Weyl subgroup of permutation matrices [2505.05981]. This decomposition directly induces the three-block circuit structure of the ansatz.

In the TSP compact-encoding literature, the foundation is different but analogous. The Lehmer-factorial map sends a tour \(\pi\in S_n\) to an integer
\[
P(\pi)=\sum_{k=1}^n \ell_k\cdot (n-k)!,
\]
represented on
\[
m=\lceil\log_2(n!)\rceil = \sum_{k=1}^n \lceil\log_2 k\rceil = O(n\log n)
\]
qubits [2508.21730]. Since each basis state decodes to a unique tour, the encoding itself implements the combinatorial admissibility condition.

The resource-efficient TSP framework instead fixes one city and stores the remaining \(M=n-1\) tour positions in \(k=\lceil\log_2 M\rceil\)-qubit registers:
\[
|\pi\rangle = |a_0\rangle\otimes|a_1\rangle\otimes\cdots\otimes|a_{M-1}\rangle.
\]
Adjacent transpositions are realized by register-swap unitaries \(T_{i,i+1}\), which act as generators of \(S_M\) on valid codes [2605.00739]. Here, the group action is explicit at the circuit level.

For permutation-invariant quantum circuits, the symmetric group \(S_n\) is identified with SWAPs on \(n\) qubits, and the associated Lie-algebra generators \(P_{ij}\) yield continuous families
\[
U_{ij}(\theta)=e^{\,i\,\theta\,P_{ij}}.
\]
The parameter count of the fully permutation-invariant subalgebra scales as
\[
\dim\bigl(\pi\mathfrak{su}(2^n)\bigr)=\binom{n+3}{3}-1=\mathcal O(n^3),
\]
which quantifies symmetry-induced reduction in variational degrees of freedom [2312.14909].

## 3. Canonical circuit constructions

The clearest canonical form is the QuPer system ansatz
\[
\mathscr P_{\rm sys}(\theta)
\;=\;
\mathscr C_X(\theta_X)\;\Bigl[B(\theta_B)\;W(\theta_W)\;B(\theta'_B)\Bigr],
\]
where the \(X\)-block is built from parallel single-qubit \(\mathtt{Rx}\) gates, the \(B\)-blocks from parametrized \(\mathtt{CX}_{j,k}(\phi)\) gates corresponding to upper-triangular transvections, and the \(W\)-block from parametrized adjacent \(\mathtt{SWAP}(\phi)\) gates implementing simple reflections [2505.05981]. The resulting circuit has depth \(O(q)\) and parameter count
\[
\ell=q+3\binom q2=O(q^2),
\]
and spans exactly
\[
|LX_q|
=2^q\prod_{k=0}^{q-1}(2^q-2^k)
=2^{O(q^2)}\ll n!
\]
permutation unitaries without ancillas [2505.05981].

The same work augments this base circuit with \(m\) ancilla qubits. After Hadamards on the ancillas, ancilla-controlled applications of \(\mathscr P_{\rm sys}\), and uncomputation, the output matrix becomes
\[
\hat P
=
\sum_{r=1}^R \lambda_r^2\,[P_r],\qquad
R\le\min\{2^{2m},\,|LX_q|\},
\]
so the ansatz represents a convex combination of base permutations rather than a single one [2505.05981]. This is the “quantum boost” mechanism that enlarges the reachable region of the permutation polytope.

In the resource-efficient TSP formulation, a single adjacent-link variational block is
\[
U_{\rm swap}^{(i,i+1)}(\theta)
=
R_y^{(\rm aux)}(2\theta)\;\cdot\;
S_{i,i+1}^{(\rm aux)}\;\cdot\;
R_y^{(\rm aux)}(-2\theta),
\]
with \(S_{i,i+1}^{(\rm aux)}\) the ancilla-controlled swap of all \(k\) qubits in two neighboring registers [2605.00739]. A layer applies these blocks in ascending order:
\[
U_{\rm layer}\bigl(\boldsymbol\theta^{(\ell)}\bigr)
=
\prod_{i=0}^{M-2}
U_{\rm swap}^{(i,i+1)}\bigl(\theta_{\ell,i}\bigr),
\]
and the full ansatz of depth \(L\) is
\[
|\Psi(\boldsymbol\theta)\rangle
=
\Bigl[\prod_{\ell=1}^L U_{\rm layer}\bigl(\boldsymbol\theta^{(\ell)}\bigr)\Bigr]
\bigl(|0\rangle_{\rm aux}\otimes|\pi_0\rangle\bigr),
\]
with \(N_{\rm param}=L(M-1)=L(n-2)\) parameters [2605.00739].

In “Freeze and Conquer,” the ansatz is not derived from explicit group decompositions but from a discrete search over 675 topologies of the form
\[
(P_1,E_2,P_3,E_4,P_5),
\]
with rotation axes \(P_i\in\{x,y,z\}\) and entanglement patterns \(E_j\in\{\text{linear, rev-lin, full, circ, SCA}\}\), under the alternating \(R\)–\(E\)–\(R\)–\(E\)–\(R\) constraint [2508.21730]. The fixed five-block circuit
\[
\mathcal A(\Theta)
=
U_R^{(5)}(\Theta^{(5)})\cdot U_E^{(4)}\cdot U_R^{(3)}(\Theta^{(3)})\cdot U_E^{(2)}\cdot U_R^{(1)}(\Theta^{(1)})
\]
is applied directly to the compact permutation encoding, so the gates “preserve the permutation encoding automatically” [2508.21730].

## 4. Feasibility preservation, invariance, and expressivity

The central technical question is whether the ansatz preserves the feasible combinatorial subspace and whether it can adequately traverse that subspace.

For graph-controlled permutation mixers, feasibility is formalized with a constraint graph \(G=(V,E)\) whose edges encode assignment-clash, precedence-clash, and machine-clash relations in the flexible job-shop problem [2311.04100]. A permutation \(\pi\in S_N\) is allowed on a bit string \(x\) only when the Boolean predicate
\[
\chi_\pi(x)=\bigwedge_{j=1}^N \chi_\pi^j(x)
\]
holds, with
\[
\chi_\pi^j(x)=\neg x_j\;\vee\;\bigwedge_{\ell\in nbhd(\pi(j))}\neg x_{\pi^{-1}(\ell)}.
\]
The controlled mixer
\[
B_\pi(\beta)\coloneqq \Lambda_{\chi_\pi}\bigl[|0\rangle\langle0|_c\otimes I+|1\rangle\langle1|_c\otimes U_\pi(\beta)\bigr]
\]
moves amplitude from \(|x\rangle\) to \(|\pi(x)\rangle\) only when \(\pi(x)\in F\), and Theorem 4.1 shows that the full CM-QAOA state remains inside the feasible subspace \(\mathcal F\) at every step [2311.04100]. Theorem 4.2 then establishes connectivity by showing that any two feasible schedules can be linked by a sequence of allowed transpositions.

In the resource-efficient TSP ansatz, feasibility is preserved more simply. Because the building blocks are controlled swaps of whole registers containing valid city codes, they “never produce invalid codes or duplicate-city configurations” and therefore keep the state within the valid permutation subspace [2605.00739]. The state can be expanded as
\[
|\Psi(\boldsymbol\theta)\rangle
=
\sum_{\pi\in S_M}\sum_{\alpha=0}^1 c_{\pi,\alpha}(\boldsymbol\theta)\,|\alpha\rangle_{\rm aux}\otimes|\pi\rangle,
\]
with marginal tour probabilities
\[
P(\pi;\boldsymbol\theta)=\sum_{\alpha\in\{0,1\}}
\left|
\langle \alpha|_{\rm aux}\langle\pi|\Psi(\boldsymbol\theta)\rangle
\right|^2.
\]
This is exact feasible-subspace evolution rather than approximate penalty suppression [2605.00739].

Permutation-invariant circuits address a different property: invariance under relabeling. The invariance condition
\[
S_{kl}\,\mathcal U(\vec\theta)\,S_{kl}=\mathcal U(\vec\theta)\qquad \forall\,(k,l)
\]
is enforced by symmetrizing parameters over orbits of \(S_n\), for example through blocks such as
\[
U^{(1)}(\alpha)=e^{\,i\alpha\sum_i Z_i},\qquad
U^{(2)}(\beta)=\exp\!\bigl[i\beta\sum_{i<j}P_{ij}\bigr].
\]
A two-layer example interleaves global \(Z\)-rotations and all-to-all SWAP-family generators [2312.14909]. This does not encode permutations as solutions; instead it restricts the variational search to the \(S_n\)-invariant sector.

A common misconception is that any ansatz acting on permutation-labeled data is automatically permutation-preserving. The literature distinguishes at least three non-equivalent properties: preserving the validity of encoded permutations [2605.00739, 2508.21730], preserving feasibility under constrained moves [2311.04100], and being invariant under qubit or label permutations [2312.14909].

## 5. Optimization workflows and cost evaluation

The optimization loop depends strongly on how the ansatz represents solutions.

QuPer does not encode the objective as a Hamiltonian. Instead, it samples a doubly-stochastic matrix \(\hat P(\boldsymbol\theta)\), optionally projects it to a permutation \(\tilde P\in\Pi_n\), and evaluates the cost classically [2505.05981]. For quadratic assignment,
\[
f(P)=\mathrm{Tr}\bigl(W\,P\,D^\top\,P^\top\bigr),
\]
and for graph isomorphism,
\[
f(P)=\|A-PBP^\top\|_F^2
=\mathrm{Tr}(A^\top A)-2\,\mathrm{Tr}(A\,P\,B\,P^\top)+\mathrm{Tr}(B^\top B).
\]
The cost is computed after projection in \(O(n^3)\) time, while gradients are estimated by the parameter-shift rule plus classical postprocessing [2505.05981].

The TSP resource-efficient framework uses a diagonal distance Hamiltonian on the feasible subspace:
\[
E(\boldsymbol\theta)
=
\langle\Psi(\boldsymbol\theta)|
(I_{\rm aux}\otimes H_{\rm dist})
|\Psi(\boldsymbol\theta)\rangle.
\]
In the divide-and-conquer formulation, the full Hamiltonian is expanded as
\[
H=\sum_t c_t\bigotimes_i \tilde O_{t,i},\qquad \tilde O_{t,i}\in\{I,Z\}^{\otimes q_i},
\]
and, under a product-state approximation,
\[
E(\boldsymbol\theta)=\sum_t c_t\prod_{i=1}^N
\langle\psi_i(\theta_i)|\tilde O_{t,i}|\psi_i(\theta_i)\rangle.
\]
A single basis measurement yields one feasible tour sample \(\pi\), and after \(S\) shots the empirical estimator is
\[
\hat E=\frac1S\sum_{s=1}^S D\bigl(\pi^{(s)}\bigr)
\]
with \(D(\pi)\) the tour length [2605.00739].

In “Freeze and Conquer,” the optimize–freeze–reuse workflow separates structural search from parameter search. Training uses simulated annealing over the discrete topology space, with fitness
\[
F(\tau,\Theta)=P_{\rm opt},
\]
the probability that a measurement yields the optimal tour of the training instance [2508.21730]. The simulated-annealing loop uses \(T_0=1.0\), cooling rate \(\alpha=0.999\), \(T_{\min}=10^{-3}\), and iteration cap \(500\); each candidate topology is evaluated by a local VQE in which 100 random parameter vectors are sampled, the best 10 are refined by Powell, and final \(P_{\rm opt}\) is estimated from \(N=1024\) shots [2508.21730]. After training, the best topology is frozen and reused on new instances, with only Powell re-optimization of parameters from the trained initialization.

These workflows illustrate two contrasting strategies. QuPer and the divide-and-conquer TSP formulation rely on classically reconstructed costs from sampled feasible outputs; Freeze and Conquer amortizes architectural search across instances. This suggests that problem-inspired permutation ansätze can support both instance-specific and reusable regimes.

## 6. Resources, empirical behavior, and limitations

The principal resource advantage emphasized across the literature is qubit compression.

### Selected constructions and reported scaling

| Construction | Qubit / variable scaling | Key circuit or parameter scaling |
|---|---:|---:|
| QuPer system register | \(q=\log_2 n\) plus \(m\) ancillas | \(\ell=O((\log n)^2)\), depth \(O(\log n)\) [2505.05981] |
| Lehmer-encoded TSP | \(m=\lceil\log_2(n!)\rceil=O(n\log n)\) | 5-block ansatz, 675 candidate topologies [2508.21730] |
| Register-swap TSP ansatz | \(M\,k=O(n\log n)\) data qubits plus one ancilla | \(N_{\rm param}=L(n-2)\) [2605.00739] |
| GCPM for FJSP | \(N=|O|\cdot|M|\cdot|T|\) main qubits plus ancilla registers | Worst-case Toffoli count \(O(N^5)\), depth roughly \(O(N^4)\) [2311.04100] |
| Permutation-invariant circuits | symmetry-reduced parameter space | \(\mathcal O(n^3)\) parameters [2312.14909] |

QuPer is explicitly intended for near-term use because the number of system qubits scales logarithmically with permutation dimension [2505.05981]. The paper reports simulation up to \(n=256\), requiring \(20\) qubits, and states that QuPer is competitive with classical heuristics, with typical optimality gaps of a few percent on QAPlib instances [2505.05981].

For TSP, the resource-efficient framework reports best average success rates of \(100\%\), \(100\%\), and \(95.5\%\) for 4-, 5-, and 6-city instances, respectively, in numerical simulation [2605.00739]. The related Freeze and Conquer study reports average optimal-trip sampling probabilities of \(100\%\) for 4-city cases, \(90\%\) for 5-city cases, \(80\%\) for 6-city cases, and approximately \(20\%\) for 7-city cases after reuse, indicating a marked onset of scalability limitations at 7 cities [2508.21730].

GCPMs provide strong formal guarantees but with substantial hardware overhead. Implementing the sequential mixer requires the main register, an auxiliary copy register, three further \(N\)-qubit ancilla registers \(a,b,z\), two single-qubit ancillae including \(c\), and a global-AND qubit; for \(|P|=O(N^2)\), the total Toffoli count in one mixer layer is \(O(N^5)\), with gate depth roughly \(O(N^4)\) [2311.04100]. The work explicitly notes that this is demanding for current NISQ devices.

Permutation-invariant circuits reduce parameter complexity to \(\mathcal O(n^3)\), but the same paper presents this as “an indication that symmetry restricts the applicability of quantum computing” [2312.14909]. This caution is important: symmetry reduction improves tractability only when the target low-energy sector or feasible set genuinely respects the imposed symmetry.

## 7. Related directions and broader significance

The phrase “permutation-preserving problem-inspired ansatz” is most directly associated with QuPer [2505.05981] and with the resource-efficient TSP framework [2605.00739], but the surrounding literature shows that the underlying idea has broader methodological significance.

One extension concerns **reusable ansätze**. Freeze and Conquer demonstrates that a permutation-preserving compact encoding can be paired with an optimize–freeze–reuse pipeline in which all heavy structural search is performed once and only parameter re-optimization is repeated on new instances [2508.21730]. This suggests that structural priors derived from permutation geometry may transfer across instances of the same problem family.

A second extension concerns **constraint graphs and local feasibility checks**. GCPMs show that problems whose feasible solutions are independent sets or fixed-size vertex subsets of a graph can be treated by redefining the edge set \(E\) and the control function \(\chi\) [2311.04100]. A plausible implication is that permutation-preserving ansätze are part of a larger class of feasibility-preserving variational constructions for structured combinatorial spaces.

A third extension concerns **symmetry-adapted variational design** beyond combinatorial optimization. Permutation-invariant circuits provide a Lie-algebraic recipe for enforcing full \(S_n\) symmetry in quantum circuits and for symmetrizing an existing ansatz by orbit-averaging parameters [2312.14909]. This is conceptually adjacent to permutation-preserving combinatorial ansätze, though the operational goal is different.

Finally, there is an instructive contrast with **qubit permutation for layout optimization**. PermVQE adds an outer optimization loop that permutes qubits to minimize a mutual-information-based cost
\[
C(\pi)=\sum_{i<j} I_{ij}\,d\bigl(\pi(i),\pi(j)\bigr)^2,
\]
thereby reducing ansatz depth for chemistry problems [2009.04996]. This work is about permuting the representation of a problem on hardware rather than preserving combinatorial permutations as feasible solutions. The contrast clarifies that “permutation-preserving” in the problem-inspired-ansatz literature refers primarily to solution-space structure, not merely to relabeling qubits.

Taken together, these works define a technically coherent paradigm: exploit the algebra, encoding geometry, and feasibility structure of permutation-based problems to design ansätze whose state space is already adapted to the discrete objects being optimized. The resulting circuits may preserve valid tours exactly [2605.00739, 2508.21730], preserve constrained schedules under controlled local permutations [2311.04100], or generate rich subsets and mixtures of permutation matrices from group-theoretic building blocks [2505.05981]. The main benefits are reduced qubit overhead, elimination or reduction of penalty terms, and a variational search space better aligned with the combinatorial target; the main limitations are expressivity bottlenecks, classical post-processing overhead, and, in some constructions, substantial ancilla and gate costs.

Source: https://www.emergentmind.com/topics/permutation-preserving-problem-inspired-ansatz