Quantum Spatial Best-Arm Identification

• QSBAI is a quantum algorithmic framework that extends best-arm identification to spatially constrained graph bandits by integrating Szegedy quantum walks with Grover-like amplitude amplification.
• It models the multi-armed bandit problem on a finite undirected graph where arm accessibility is limited to adjacent vertices, using a reversible Markov chain for exploration.
• The approach generalizes traditional quantum search methods, achieving quadratic speedups in optimal arm selection on structured graphs such as complete and complete bipartite graphs.

Quantum Spatial Best-Arm Identification (QSBAI) is a quantum algorithmic framework for best-arm identification in a graph bandit setting, where each arm is embedded in a finite, undirected graph and the agent’s actions are spatially constrained by graph adjacency. The framework models the graph bandit as a reversible Markov chain, lifts that chain to a Szegedy quantum walk on $\mathbb{C}^{V} \otimes \mathbb{C}^{V}$, and combines the walk with a marking reflection so that repeated walk-and-oracle steps implement a Grover-like amplitude amplification subject to locality constraints. In this sense, QSBAI generalizes earlier quantum best-arm identification algorithms for unconstrained multi-armed bandits and establishes a link between Grover-type search, quantum walks, and reinforcement-learning tasks with structural restrictions (Yamagami et al., 7 Sep 2025).

1. Graph bandits and the spatial BAI objective

In the formulation studied for QSBAI, a finite undirected graph

$G=(V,E)$

indexes the arm set: each vertex $v\in V$ corresponds to an arm with reward distribution $\nu_v$ and mean reward

$\mu_v=\mathbb{E}[R_v].$

The best arm is

$v^* \in \arg\max_{v\in V}\mu_v.$

Unlike unconstrained multi-armed bandits, the action space is not freely accessible at every step. If the agent is at vertex $v_t$, admissible choices are restricted to

$$\mathcal{N}(v_t)=\{u\in V:(v_t,u)\in E\}\cup\{v_t\}.$$

The agent therefore explores by following a path on $G$, and exploration must respect connectivity and local adjacency (Yamagami et al., 7 Sep 2025).

The paper treats best-arm identification as a pure-exploration objective: within a given number of time steps, output the index of the best arm with maximal success probability under graph-constrained movement. The term “spatial” has two explicit meanings in this setting. First, the quantum state space is constrained by a graph, because basis states encode graph positions and the evolution is implemented by a quantum walk along graph edges. Second, action choice is constrained by spatial position, because the agent may only explore arms reachable by traversing the graph rather than querying arbitrary labels as in a fully connected bandit (Yamagami et al., 7 Sep 2025).

This formulation is motivated by a limitation of earlier quantum bandit work. Prior quantum best-arm identification procedures generally assume oracle access to arbitrary arm labels in superposition, which is effectively a complete-graph access model. The unstructured quantum BAI model of “Quantum Bandits” (Casalé et al., 2020) and the fixed-confidence quantum exploration algorithm of Wang, You, Li, and Childs (Wang et al., 2020) both operate in that unconstrained regime. QSBAI extends the setting by embedding the arm set into a graph and using a Szegedy walk to enforce locality (Yamagami et al., 7 Sep 2025).

2. Szegedy-walk construction and generalized amplitude amplification

QSBAI acts on the Hilbert space

$\mathcal{H}=\mathbb{C}^{V}\otimes \mathbb{C}^{V},$

with computational basis $G=(V,E)$0. These basis states are interpreted as directed edges or origin–terminus pairs: the first register encodes the origin vertex and the second the terminus vertex (Yamagami et al., 7 Sep 2025).

The graph structure is encoded by choosing a reversible Markov chain with transition matrix

$G=(V,E)$1

where $G=(V,E)$2, $G=(V,E)$3, and $G=(V,E)$4 only if $G=(V,E)$5 or $G=(V,E)$6. From $G=(V,E)$7, QSBAI constructs the Szegedy walk via two reflection operators,

$G=(V,E)$8

and defines the Szegedy quantum walk operator as

$G=(V,E)$9

This unitary is the quantum analogue of two steps of the classical Markov chain, and its behavior is governed by the spectrum of the discriminant matrix associated with $v\in V$0 (Yamagami et al., 7 Sep 2025).

The best arm is encoded through a marked reflection $v\in V$1 that flips the phase of states associated with the best arm, or more generally with a “good” subspace. QSBAI then forms a Grover-like composite operator

$v\in V$2

or a closely related composition, and iterates it from an initial state $v\in V$3, often a coherent version of the stationary distribution: $v\in V$4 After a suitable number of time steps, measurement of the position register yields the best arm with high probability (Yamagami et al., 7 Sep 2025).

This mechanism is explicitly a generalization of amplitude amplification. In Grover’s search one alternates a phase flip on marked items with a diffusion operator; in QSBAI the diffusion-like component is replaced by a Szegedy walk that respects graph adjacency. A plausible implication is that the usual two-dimensional rotation picture of amplitude amplification survives in a spatially constrained search space, but with the effective rotation angle determined by the walk spectrum rather than by uniform access to all arms.

3. Relation to Grover search and earlier quantum BAI

For complete graphs, QSBAI reduces to the non-spatial quantum BAI setting. If $v\in V$5 is complete with $v\in V$6 and the underlying Markov chain is uniform,

$v\in V$7

then the Szegedy operator acts essentially as a global diffusion operator over the entire vertex set. In that regime, QSBAI reproduces the known Grover-type behavior: sinusoidal success probability, optimal number of steps of order $v\in V$8, and maximum success probability approaching $v\in V$9 (Yamagami et al., 7 Sep 2025).

This reduction identifies Grover search as a special case of the Szegedy-walk framework. The marked reflection plays the role of the usual search oracle, while the walk operator replaces the standard global reflection about the initial state. Accordingly, QSBAI is presented not merely as a graph-specific construction but as a strict generalization of standard quantum best-arm identification (Yamagami et al., 7 Sep 2025).

The unstructured antecedents are explicit. “Quantum Bandits” models quantum best-arm identification through amplitude amplification on a joint action–environment Hilbert space $\nu_v$0, with a reward oracle

$\nu_v$1

and a Grover-like iteration operator $\nu_v$2 (Casalé et al., 2020). Wang, You, Li, and Childs later gave a fixed-confidence quantum BAI algorithm with quantum query complexity

$\nu_v$3

matching a quantum lower bound up to polylogarithmic factors in the unstructured index-oracle model (Wang et al., 2020). QSBAI differs by replacing unrestricted index access with graph-constrained exploration and replacing global amplitude amplification with a Szegedy-walk-based construction (Yamagami et al., 7 Sep 2025).

A common misconception is that QSBAI is simply Grover search on a graph. The paper’s formulation is narrower and more specific: the marked object is not an arbitrary marked vertex but the best arm in a graph bandit, and the walk operator is derived from a reversible Markov chain chosen to represent graph-constrained actions rather than from an unconstrained oracle model (Yamagami et al., 7 Sep 2025).

4. Mathematical behavior on complete and complete bipartite graphs

The paper analyzes two graph families in detail: complete graphs and complete bipartite graphs (Yamagami et al., 7 Sep 2025). In both cases the central observable is the success probability

$\nu_v$4

where $\nu_v$5 is the vertex obtained by measuring the relevant register after $\nu_v$6 applications of the walk-and-marking operator.

For a complete graph with a single marked arm, the dynamics reduces to a low-dimensional invariant subspace spanned by the marked component and its orthogonal complement. The state can be written as

$\nu_v$7

and consequently

$\nu_v$8

The optimal time is chosen so that

$\nu_v$9

For a single marked arm with uniform initialization, $\mu_v=\mathbb{E}[R_v].$0, yielding

$\mu_v=\mathbb{E}[R_v].$1

This is the same scaling as Grover search and is described as a quadratic speedup over classical exhaustive search on the complete-graph bandit (Yamagami et al., 7 Sep 2025).

For complete bipartite graphs $\mu_v=\mathbb{E}[R_v].$2, the walk spectrum is more complex. If the best arm lies in one partition, the relevant dynamics is described as evolving largely in a $\mu_v=\mathbb{E}[R_v].$3- or $\mu_v=\mathbb{E}[R_v].$4-dimensional subspace built from uniform superpositions over edges incident to each partition together with marked components. The success probability takes a more complicated form,

$\mu_v=\mathbb{E}[R_v].$5

where $\mu_v=\mathbb{E}[R_v].$6 and the eigenphases $\mu_v=\mathbb{E}[R_v].$7 depend on $\mu_v=\mathbb{E}[R_v].$8 and the location of the marked vertex. The corresponding maximizing time satisfies

$\mu_v=\mathbb{E}[R_v].$9

with $v^* \in \arg\max_{v\in V}\mu_v.$0 encoding effective degree or marked-set size (Yamagami et al., 7 Sep 2025).

The interpretation supplied by the paper is spectral. The Szegedy walk has eigenvalues $v^* \in \arg\max_{v\in V}\mu_v.$1, with $v^* \in \arg\max_{v\in V}\mu_v.$2 related to eigenvalues of the discriminant matrix built from $v^* \in \arg\max_{v\in V}\mu_v.$3. Marking perturbs this spectrum, and the smallest nonzero effective eigenphase controls the oscillation period of amplitude into the marked subspace. The resulting heuristic law is

$v^* \in \arg\max_{v\in V}\mu_v.$4

with $v^* \in \arg\max_{v\in V}\mu_v.$5 for complete graphs and an $v^* \in \arg\max_{v\in V}\mu_v.$6-dependent expression for complete bipartite graphs (Yamagami et al., 7 Sep 2025).

5. Algorithmic structure and access model

A conceptual QSBAI run has five components (Yamagami et al., 7 Sep 2025). First, initialization prepares a coherent state, typically

$v^* \in \arg\max_{v\in V}\mu_v.$7

where $v^* \in \arg\max_{v\in V}\mu_v.$8 is the stationary distribution of the chosen Markov chain. For regular graphs, this is often the uniform superposition over edges.

Second, the reward-encoding stage supplies a marking oracle. In the idealized best-arm identification setting of the analysis, this is modeled as a phase oracle such as

$v^* \in \arg\max_{v\in V}\mu_v.$9

The paper also notes a more general reward oracle of the form

$v_t$0

followed by thresholding or other processing to decide whether to apply a phase flip (Yamagami et al., 7 Sep 2025).

Third, the Szegedy walk step implements

$v_t$1

In circuit terms, the reflections $v_t$2 and $v_t$3 are realized by structured unitaries defined from the square-root transitions $v_t$4. Fourth, the algorithm iterates

$v_t$5

Fifth, after a number of steps selected from the spectral analysis, the algorithm measures the relevant vertex register and outputs the observed arm (Yamagami et al., 7 Sep 2025).

The access model is idealized. The graph structure and the Markov chain $v_t$6 are assumed known, the corresponding Szegedy walk is assumed implementable, and no noise or decoherence is modeled. The marking operator is also assumed available, at least as an abstract bandit oracle. The paper is explicit that mapping noisy reward samples to such a marking oracle is beyond scope (Yamagami et al., 7 Sep 2025).

This access model parallels earlier quantum BAI work. In the unstructured setting of “Quantum Bandits,” one assumes a unitary $v_t$7 preparing environment states conditioned on arms and a reward phase oracle $v_t$8 on the composite action–environment space (Casalé et al., 2020). In the fixed-confidence setting of Wang, You, Li, and Childs, one assumes coherent oracle access

$v_t$9

which supports superposition over arm indices and enables variable-time amplitude amplification and estimation (Wang et al., 2020). QSBAI modifies the access structure by enforcing locality through the walk rather than through unrestricted index access (Yamagami et al., 7 Sep 2025).

6. Position within quantum decision-making, limitations, and open questions

QSBAI lies at the intersection of three established lines of work. From quantum walks and spatial search, it inherits the use of Szegedy’s framework to search marked structure on graphs. From amplitude amplification and Grover search, it inherits the two-reflection mechanism that transfers amplitude into a marked subspace. From quantum reinforcement learning and quantum bandits, it inherits the decision-theoretic objective of identifying a best action under limited interaction (Yamagami et al., 7 Sep 2025).

Relative to earlier quantum bandit results, QSBAI contributes a structural extension rather than a stronger general complexity theorem. The unstructured model of “Quantum Bandits” shows that amplitude amplification can recommend the optimal arm with probability

$$\mathcal{N}(v_t)=\{u\in V:(v_t,u)\in E\}\cup\{v_t\}.$$0

at an iteration count

$$\mathcal{N}(v_t)=\{u\in V:(v_t,u)\in E\}\cup\{v_t\}.$$1

under its joint action–environment oracle model (Casalé et al., 2020). Wang, You, Li, and Childs establish fixed-confidence identification with quantum query complexity $$\mathcal{N}(v_t)=\{u\in V:(v_t,u)\in E\}\cup\{v_t\}.$$2, where $$\mathcal{N}(v_t)=\{u\in V:(v_t,u)\in E\}\cup\{v_t\}.$$3, and prove a matching lower bound up to polylogarithmic factors (Wang et al., 2020). QSBAI does not replace those results; rather, it shows how Grover-type amplification can be reformulated when actions are constrained by a graph (Yamagami et al., 7 Sep 2025).

Several limitations are explicit. The analysis is restricted to complete and complete bipartite graphs. The graph and the reversible Markov chain are assumed known in advance. The reward oracle is idealized, and the procedure is analyzed in the absence of noise and decoherence. The algorithmic structure is essentially non-adaptive or one-shot: it fixes a number of walk steps and measures, rather than interleaving quantum evolution with classical processing of reward observations (Yamagami et al., 7 Sep 2025).

The stated future directions are correspondingly structural. They include extending the analysis to arbitrary regular graphs, sparse graphs, or graphs with community structure; characterizing speedups in terms of spectral gaps or mixing times; designing adaptive QSBAI strategies; moving from pure best-arm identification to regret minimization; studying robustness to noise; and considering unknown graph structure or unknown transition dynamics (Yamagami et al., 7 Sep 2025). This suggests that the principal open problem is no longer whether quantum amplitude amplification can accelerate unstructured best-arm identification—that point is already established in earlier work (Casalé et al., 2020, Wang et al., 2020)—but how much of that advantage persists once locality, graph topology, and realistic oracle construction are treated as first-class constraints.