Ring of Disagrees in QAOA

• Ring of disagrees is defined as the Max-Cut problem on a cyclic graph that highlights locality and symmetry in shallow quantum algorithms.
• The study proves that for depth p, QAOA achieves an optimal cut fraction of (2p+1)/(2p+2) on cycles with n ≥ 2p+2, resolving a long-standing conjecture.
• Methodologies using free-fermion reduction and Laurent-polynomial optimization recast the QAOA performance analysis, supporting both theoretical and hardware benchmarking.

In the QAOA literature, the ring of disagrees is the Max-Cut problem on a cycle graph $C_n$: a connected $2$-regular graph with $n$ vertices and $n$ edges. It is one of the cleanest analytically tractable nontrivial QAOA instances, and its importance comes less from classical computational hardness than from what it reveals about locality, symmetry, and the limits of shallow algorithms on periodic structures. In recent work, the instance has moved from a canonical benchmark with known low-depth formulas to an exactly solved family: for depth $p$, the optimal QAOA cut fraction on cycles with $n \ge 2p+2$ is $(2p+1)/(2p+2)$, thereby resolving a long-standing conjecture of Farhi, Goldstone, and Gutmann (Marwaha, 28 Jun 2026).

1. Definition and graph-theoretic formulation

Let $C_n$ denote the cycle on $n$ vertices. For a bit string $x\in\{0,1\}^n$, an edge $2$0 is cut when $2$1, so the objective can be written as

$2$2

Equivalently, in the Pauli-$2$3 representation,

$2$4

The combinatorics of the optimum are elementary. If $2$5 is even, the alternating assignment cuts every edge, so the optimum cut size is $2$6. If $2$7 is odd, perfect alternation is impossible on the cycle, and the optimum is $2$8. This is why the instance is not used to study NP-hardness per se. Its role is instead to isolate how a local variational algorithm behaves when the globally optimal alternating pattern is obvious but cannot necessarily be coordinated from shallow local information (Marwaha, 28 Jun 2026).

A closely related formulation in the noisy-QAOA literature describes the same object as Max-Cut on a 2-regular connected graph, again emphasizing that the ring of disagrees is simply the unweighted cycle graph $2$9 with $n$0 edges and unit weights (Rabelo et al., 2024).

2. Standard QAOA formulation

For Max-Cut on the ring of disagrees, the standard QAOA ingredients are the cost Hamiltonian

$n$1

and the mixer

$n$2

The initial state is the uniform superposition

$n$3

and the depth-$n$4 ansatz is

$n$5

The objective function is

$n$6

with approximation ratio

$n$7

Because the cycle is translationally symmetric, one can evaluate a representative edge expectation and multiply by $n$8. For $n$9, the edge-level expectation is

$n$0

and the full objective is $n$1. For the $n$2, $n$3 example, the reported optimum is

$n$4

with approximation ratio

$n$5

This low-depth formula is one reason the ring of disagrees became a benchmark instance: it admits exact reference values while still exhibiting nontrivial depth dependence (Rabelo et al., 2024).

3. Exact depth-$n$6 performance and the locality threshold

The central modern theorem states that when the QAOA light cone does not see the whole graph,

$n$7

the optimal depth-$n$8 QAOA value on the cycle satisfies

$n$9

Thus the optimal expected cut fraction is exactly

$p$0

The same work also gives the complementary whole-graph-visible regime. If $p$1 and $p$2 is even, QAOA reaches cut fraction $p$3; if $p$4 and $p$5 is odd, it reaches $p$6. The result is therefore exact for every finite $p$7, with the transition at $p$8 (Marwaha, 28 Jun 2026).

Regime	Optimal QAOA cut fraction
$p$9	$n \ge 2p+2$0
$n \ge 2p+2$1, $n \ge 2p+2$2 even	$n \ge 2p+2$3
$n \ge 2p+2$4, $n \ge 2p+2$5 odd	$n \ge 2p+2$6

This theorem resolves a long-standing conjecture of Farhi, Goldstone, and Gutmann. It is also sharp from the local-algorithm viewpoint: prior locality arguments had shown that any symmetric local algorithm on the cycle that cannot see the whole graph at depth $n \ge 2p+2$7 cuts at most a $n \ge 2p+2$8 fraction of edges in expectation. QAOA exactly attains that ceiling. In this sense, the ring of disagrees is a rare family where QAOA is provably optimal within the class of symmetric depth-$n \ge 2p+2$9 local algorithms (Marwaha, 28 Jun 2026).

Small-depth special cases follow immediately. At $(2p+1)/(2p+2)$0, the exact optimum is $(2p+1)/(2p+2)$1. At $(2p+1)/(2p+2)$2, it is $(2p+1)/(2p+2)$3. More generally, the formula tends to $(2p+1)/(2p+2)$4 as $(2p+1)/(2p+2)$5 increases, but only reaches exact optimality before that limit once the whole cycle is visible.

4. Free-fermion reduction, Laurent polynomials, and quantum signal processing

The exact solution does not proceed by finding a closed-form optimal angle sequence. Instead, it reduces the cycle problem to one-qubit optimization and then recasts that one-qubit problem in the language of quantum signal processing. For even $(2p+1)/(2p+2)$6, the analysis uses the effective one-qubit Hamiltonian

$(2p+1)/(2p+2)$7

with momenta

$(2p+1)/(2p+2)$8

and one-qubit states

$(2p+1)/(2p+2)$9

where

$C_n$0

The cycle objective becomes

$C_n$1

The key technical step is an expressibility theorem: every depth-$C_n$2 one-qubit QAOA circuit corresponds to an admissible pair $C_n$3 of real Laurent polynomials, and conversely every such pair can be realized by some choice of QAOA angles. Defining

$C_n$4

the objective reduces to an extremal problem over bounded-degree Laurent polynomials. The optimal value is attained by

$C_n$5

which can also be written as

$C_n$6

where $C_n$7 is the Chebyshev polynomial of the second kind.

A notable consequence is methodological. The exact optimum value is proved without closed-form optimal QAOA parameters in the nontrivial regime $C_n$8. The proof establishes existence and optimality through the Laurent-polynomial and QSP correspondence instead. This is one of the paper’s most distinctive contributions: it reframes variational optimization on the ring of disagrees as an extremal polynomial problem (Marwaha, 28 Jun 2026).

5. Noisy QAOA and hardware-aware benchmarking

Before the exact theorem was proved, the ring of disagrees had already become a practical benchmark for noisy QAOA studies because its ideal behavior is known well enough to isolate device and compilation effects. A case study using Qiskit fake backends examined Fake-Lagos (7 qubits), Fake-Kolkata (27 qubits), and Fake-Washington (127 qubits) with AerEstimator, method="density_matrix", backend-specific coupling_map, backend-specific noise_model, backend-specific basis_gates, and shots=50000 (Rabelo et al., 2024).

The study considered ring sizes

$C_n$9

and depths

$n$0

with COBYLA as classical optimizer. Two scenarios were compared: noisy simulations with standard transpilation, and simulations with error-mitigation-oriented compilation using $2$17

The central empirical finding was structural rather than purely algorithmic: when the ring graph can be mirrored in the hardware coupling graph, transpilation overhead drops markedly and performance improves. For a representative $n$1, $n$2 ring circuit, the virtual circuit had depth $n$3 and 60 operations. After transpilation to Fake-Kolkata, optimization_level=0 gave depth $n$4, 271 operations, and 126 non-local gates, whereas optimization_level=3 gave depth $n$5, 108 operations, and 24 non-local gates.

The approximation-ratio results reflect this. On Fake-Kolkata with mitigation, the $n$6 ring achieved $n$7 at $n$8, $n$9 at $x\in\{0,1\}^n$0, $x\in\{0,1\}^n$1 at $x\in\{0,1\}^n$2, and $x\in\{0,1\}^n$3 at $x\in\{0,1\}^n$4. On Fake-Washington with mitigation, the corresponding $x\in\{0,1\}^n$5 values were $x\in\{0,1\}^n$6, $x\in\{0,1\}^n$7, $x\in\{0,1\}^n$8, and $x\in\{0,1\}^n$9. These remain below the ideal even-ring benchmark

$2$00

which gives $2$01 at $2$02, but they track the theoretical trend much more closely when ring structure aligns with device topology. The same study also reported that mitigation restored clear high-probability regions near the analytic $2$03 optimum $2$04 in success-probability contour plots (Rabelo et al., 2024).

6. Related ring-structured disagreement models

A related use of ring-structured disagreement appears in nonlocal opinion dynamics. In a ring of agents with circular opinions and mixed attractive/repulsive interactions, one finds not only global consensus but also local consensus, in which adjacent agents agree while global agreement fails, and chimera consensus, in which one domain preserves local agreement and another breaks it. In this setting the interaction radius is

$2$05

opinions are circular variables $2$06, and the average adjacent disagreement

$2$07

obeys the empirical scaling law

$2$08

The notation $2$09 and $2$10 classifies states by how the opinion field winds around the opinion circle as one traverses the spatial ring, so persistent disagreement is associated with nonzero winding rather than isolated clusters (Gao et al., 2019).

Another related framework studies attraction, repulsion, and neglect on a general random dynamic network and notes that it can be specialized to a ring by allowing only nearest-neighbor meetings. There the central disagreement measure is the state diameter

$2$11

and under symmetric constant updates the balance between agreement and disagreement is governed by

$2$12

If $2$13, global agreement convergence is achieved almost surely; if $2$14, disagreement divergence is achieved in expectation for almost all initial values. The paper explicitly remarks that, under symmetric constant updates, this sign threshold is topology-independent, so a ring is a natural specialization but not a special exception (Shi et al., 2012).

These dynamical constructions are not the graph-theoretic ring of disagrees of QAOA. They nonetheless illuminate the broader intuition behind the phrase: on a cycle, disagreement can be globally organized, locally constrained, and strongly shaped by locality.

7. Conceptual significance

The ring of disagrees occupies a distinctive place in quantum optimization because it is simultaneously simple and exacting. Its classical optimum is trivial, but shallow local algorithms cannot in general coordinate the globally alternating solution. That tension makes the instance a precise probe of locality. The exact formula

$2$15

therefore does more than solve one benchmark: it quantifies, layer by layer, how much of the cycle’s global structure depth-$2$16 QAOA can exploit before the light cone spans the whole graph (Marwaha, 28 Jun 2026).

A plausible implication is that the enduring importance of the ring of disagrees lies less in its role as a hard optimization problem than in its function as a calibration instance for theory, compilation, and hardware. It links exact variational analysis, free-fermion reduction, quantum signal processing, and noise-aware transpilation within a single family of graphs whose symmetries are fully explicit. In that respect, it has become both a solved model of local quantum optimization and a controlled laboratory for studying how disagreement—here, the desire to anti-align neighboring vertices—propagates around a ring.