Nonlinear Quantum Search

• Nonlinear quantum search is a framework using nonlinear Schrödinger equations, where state-dependent self-potentials (e.g., cubic, cubic–quintic, logarithmic) modify conventional quantum search dynamics.
• It leverages physical models like Bose–Einstein condensates to implement nonlinearities that can rescale evolution, potentially enabling constant-time search under specific scaling conditions.
• Analytical studies on complete graphs and lattices reveal that while nonlinear dynamics can surpass Grover’s speedups, the improvements come with tradeoffs in particle number, timing precision, and control resources.

Nonlinear quantum search denotes search protocols in which the effective evolution is governed not by the linear Schrödinger equation alone, but by a nonlinear Schrödinger equation or an equivalent state-dependent Hamiltonian. In the arXiv literature, the canonical setting is continuous-time quantum walk search on a graph, with a marked-vertex oracle and a self-potential of the form $-g\,f(|\psi|^2)$, where $f$ may be cubic, cubic–quintic, or logarithmic and often has a mean-field origin in Bose–Einstein condensates or nonlinear media. The subject spans several distinct but connected lines of work: effective Gross–Pitaevskii and cubic–quintic search models, nonlinear state-discrimination approaches to unstructured search, graph-dependent spatial search beyond the complete graph, and resource-theoretic analyses showing that apparent super-Grover speedups are paid for by particle number, timing precision, or feedback control (Meyer et al., 2013, Meyer et al., 2013, DalFavero et al., 2024).

1. Historical emergence and conceptual scope

The baseline problem is unstructured search over $N$ basis states, usually initialized in the uniform superposition

$$|s\rangle=\frac{1}{\sqrt{N}}\sum_{i=1}^N |i\rangle,$$

with a marked state or marked set encoded by an oracle term in a continuous-time Hamiltonian. In the linear setting, the complete-graph search Hamiltonian

$H_0=-\gamma L-|a\rangle\langle a|$

or equivalently $$H_0=-\gamma N|s\rangle\langle s|-|w\rangle\langle w|$$ yields Grover scaling $O(\sqrt{N})$ when $\gamma$ is tuned to its critical value (Meyer et al., 9 Mar 2025, Meyer et al., 2013).

Nonlinear quantum search arises when the linear Hamiltonian is supplemented by a density-dependent self-potential. The most common physically motivated cases are the Gross–Pitaevskii or cubic nonlinearity $f(p)=p$, the cubic–quintic nonlinearity $f(p)=p-p^2$ or $f$0, and the logarithmic nonlinearity $f$1. In the literature surveyed here, the cubic term is associated with two-body contact interactions in Bose–Einstein condensates, the quintic term with three-body interactions or effective higher-order corrections, and the logarithmic term with some effective descriptions of Bose liquids (Meyer et al., 2013, DalFavero et al., 2024).

The field developed along two partially divergent lines. One line studies effective nonlinearities as approximations to many-body linear quantum systems, especially BEC mean-field dynamics; this line emphasizes physical resource accounting and often preserves or only modestly improves Grover scaling unless the nonlinear coupling scales with system size. Another line studies nonlinear quantum mechanics as a computational model in its own right, where the central primitive is rapid discrimination of exponentially close states; in that setting, unstructured search can be reduced to nonlinear state separation and can become exponentially faster in $f$2 (Childs et al., 2015).

2. Canonical mathematical framework

The standard nonlinear search equation on a graph is

$f$3

Here $f$4 is the linear search Hamiltonian, $f$5 sets the nonlinear strength, and the nonlinear term is diagonal in the vertex basis. On the complete graph with $f$6 marked vertices, symmetry reduces the dynamics to a two-dimensional subspace spanned by the uniform marked state $f$7 and uniform unmarked state $f$8, with

$f$9

as the success probability. In this reduced description the nonlinear terms enter only through

$N$0

and the critical jumping rate becomes state dependent,

$N$1

With this choice, the nonlinear dynamics can be made to follow the same effective two-level path as linear Farhi–Gutmann search, but with a rescaled time variable (Meyer et al., 2013, DalFavero et al., 2024).

A complementary analytical perspective is furnished by conserved quantities. In the linear continuous-time walk, $N$2 is conserved. For a Gross–Pitaevskii-type nonlinearity with repulsive interactions and time-independent $N$3, the conserved quantity is instead

$N$4

whereas in the attractive case with a time-varying critical $N$5, the conserved object is

$N$6

These conservation laws formalize the fact that the effective nonlinear Hamiltonian itself is generally not the conserved energy functional (Meyer et al., 9 Mar 2025).

3. Complete-graph regimes: from constant-factor changes to constant time

The complete graph remains the main analytic laboratory for nonlinear quantum search. In one Gross–Pitaevskii formulation with fixed $N$7, repulsive interactions preserve the $N$8 runtime but worsen the multiplicative constant, while attractive interactions with a time-dependent critical hopping function $N$9 improve the constant factor by $$|s\rangle=\frac{1}{\sqrt{N}}\sum_{i=1}^N |i\rangle,$$0 yet still remain in $$|s\rangle=\frac{1}{\sqrt{N}}\sum_{i=1}^N |i\rangle,$$1 (Meyer et al., 9 Mar 2025). In that class of models, nonlinear search modifies the effective energy landscape without changing the asymptotic exponent.

A different regime appears when the nonlinear strength scales with problem size. For the cubic Gross–Pitaevskii model on the complete graph, the runtime

$$|s\rangle=\frac{1}{\sqrt{N}}\sum_{i=1}^N |i\rangle,$$2

becomes constant when $$|s\rangle=\frac{1}{\sqrt{N}}\sum_{i=1}^N |i\rangle,$$3, equivalently $$|s\rangle=\frac{1}{\sqrt{N}}\sum_{i=1}^N |i\rangle,$$4. The same model exhibits a sharp resource tradeoff: the success peak narrows as $$|s\rangle=\frac{1}{\sqrt{N}}\sum_{i=1}^N |i\rangle,$$5, so the constant-time algorithm requires time-measurement precision scaling as $$|s\rangle=\frac{1}{\sqrt{N}}\sum_{i=1}^N |i\rangle,$$6. When that clock resource is counted, the jointly optimized time–space product becomes $$|s\rangle=\frac{1}{\sqrt{N}}\sum_{i=1}^N |i\rangle,$$7 rather than $$|s\rangle=\frac{1}{\sqrt{N}}\sum_{i=1}^N |i\rangle,$$8 (Meyer et al., 2013).

General nonlinearities preserve this pattern while changing the detailed peak structure. For cubic search with $$|s\rangle=\frac{1}{\sqrt{N}}\sum_{i=1}^N |i\rangle,$$9 marked items, the runtime scales as

$H_0=-\gamma L-|a\rangle\langle a|$0

For cubic–quintic search, the leading runtime scaling is asymptotically the same, but the success-peak width behaves differently: when $H_0=-\gamma L-|a\rangle\langle a|$1, the peak width remains $H_0=-\gamma L-|a\rangle\langle a|$2 and does not shrink with increasing $H_0=-\gamma L-|a\rangle\langle a|$3; when $H_0=-\gamma L-|a\rangle\langle a|$4, the peak narrows in the same way as in the cubic case. This yields the counterintuitive conclusion that, once timing precision is included in the resource count, a single marked item can be easier than multiple marked items in the cubic–quintic model (Meyer et al., 2013).

The 2024 many-body analysis gives the most detailed account of cubic–quintic tuning. For a BEC mean-field model with

$H_0=-\gamma L-|a\rangle\langle a|$5

constant-time search is achieved in several regimes. If $H_0=-\gamma L-|a\rangle\langle a|$6, the runtime approaches $H_0=-\gamma L-|a\rangle\langle a|$7, but the success probability peak has width $H_0=-\gamma L-|a\rangle\langle a|$8. If $H_0=-\gamma L-|a\rangle\langle a|$9, the runtime approaches $$H_0=-\gamma N|s\rangle\langle s|-|w\rangle\langle w|$$0 and the peak widens as $$H_0=-\gamma N|s\rangle\langle s|-|w\rangle\langle w|$$1, eliminating the need for extreme clock precision. If $$H_0=-\gamma N|s\rangle\langle s|-|w\rangle\langle w|$$2, with

$$H_0=-\gamma N|s\rangle\langle s|-|w\rangle\langle w|$$3

the oscillatory peak turns into a plateau at height $$H_0=-\gamma N|s\rangle\langle s|-|w\rangle\langle w|$$4, reached in constant time near $$H_0=-\gamma N|s\rangle\langle s|-|w\rangle\langle w|$$5; this realizes constant-time search without fine time resolution, at the expense of plateau height below 1 when $$H_0=-\gamma N|s\rangle\langle s|-|w\rangle\langle w|$$6 (DalFavero et al., 2024).

4. Extensions beyond the complete graph

A major question has been whether nonlinear speedups depend on all-to-all connectivity. Earlier work on “sufficiently complete” graphs answered this negatively for several graph families. A graph is sufficiently complete when, at the critical $$H_0=-\gamma N|s\rangle\langle s|-|w\rangle\langle w|$$7, the two lowest eigenstates and eigenvalues of the linear search Hamiltonian asymptotically match those of the complete graph. Strongly regular graphs and the hypercube satisfy this condition, with error parameters $$H_0=-\gamma N|s\rangle\langle s|-|w\rangle\langle w|$$8 and $$H_0=-\gamma N|s\rangle\langle s|-|w\rangle\langle w|$$9, respectively. On these graphs, cubic and cubic–quintic nonlinearities with $O(\sqrt{N})$0 again yield constant-time search, while the loglinear nonlinearity can do so on strongly regular graphs but not on the hypercube, where the nonlinear corrections become too large and numerically unstable (Meyer et al., 2015).

The same program has recently been pushed further. For sufficiently complete graphs, nonlinear speedups have been analytically proved for Paley graphs and for complete bipartite graphs whose two partite sets both have size $O(\sqrt{N})$1, using suitable cubic and cubic–quintic nonlinearities. Numerical evidence extends the claim to stronger nonlinearities and to hypercubes. The 2026 work also studies arbitrary-dimensional cubic lattices and reports that certain nonlinearities speed up search on sufficiently high-dimensional lattices, indicating that incomplete connectivity does not by itself eliminate nonlinear advantage (Meyer et al., 31 May 2026).

Two-dimensional lattices are a particularly stringent test case because linear spatial search is weak there. Numerical work using the Childs–Ge free lattice Hamiltonian with linear dispersion, plus a nonlinear phase of Wong–Meyer type, found that on the finite 2D torus the marked vertex can be found in

$O(\sqrt{N})$2

steps with success probability $O(\sqrt{N})$3, for overall complexity

$O(\sqrt{N})$4

That study also reports an optimal choice of walker parameters that avoids any additional asymptotic cost from time-measurement precision (Herzog et al., 2020).

5. Many-body implementations and resource tradeoffs

The physically dominant implementation model is a Bose–Einstein condensate in an optical lattice. Each lattice site corresponds to a database vertex, a large number of bosons occupy the same motional mode, and pairwise and three-body interactions induce cubic and quintic terms in the mean-field equation. In this picture, each particle interacts with the oracle in parallel, so a condensate with $O(\sqrt{N})$5 atoms behaves as a search model with $O(\sqrt{N})$6 parallel oracles (DalFavero et al., 2024).

This immediately connects nonlinear quantum search to the parallel-query lower bound

$O(\sqrt{N})$7

In the many-body BEC interpretation, constant-time search $O(\sqrt{N})$8 therefore requires $O(\sqrt{N})$9. More generally,

$\gamma$0

The earlier Gross–Pitaevskii analysis sharpened this into lower bounds on the number of particles required for the effective nonlinear approximation itself to remain self-consistent: the $\gamma$1 regime implies $\gamma$2, while the constant-time cubic regime implies $\gamma$3 (DalFavero et al., 2024, Meyer et al., 2013). For general nonlinearities, cubic–quintic search yields the same $\gamma$4 particle requirement in its constant-time regime, while the optimized loglinear regime demands $\gamma$5 (Meyer et al., 2013).

Clock precision is a second recurring resource. In the sharp-peak cubic regime, the peak width scales as $\gamma$6, so an entangled atomic clock requires $\gamma$7 ions to resolve the success peak (Meyer et al., 2013). The many-body cubic–quintic analysis makes the same point operationally: $\gamma$8 gives sharp peaks and high timing cost, $\gamma$9 gives broad peaks, and $f(p)=p$0 gives plateaus that all but remove timing as a bottleneck (DalFavero et al., 2024).

Not all physically motivated nonlinearities produce super-Grover scaling. For interacting BECs on the complete graph described directly by a discrete nonlinear Schrödinger equation, the runtime remains $f(p)=p$1, with complete search only when $f(p)=p$2 is tuned below a critical $f(p)=p$3; stronger interactions produce self-trapping rather than speedup. That study supports the view that BECs are natural search substrates while simultaneously showing that realistic interaction models need not yield asymptotic gains (Kahou et al., 2013).

6. Complexity-theoretic interpretation, misconceptions, and controversies

The central complexity-theoretic distinction is between effective and fundamental nonlinearity. Gross–Pitaevskii, cubic–quintic, and logarithmic equations in this literature are effective mean-field descriptions of underlying linear many-body systems. Their apparent super-Grover runtimes do not contradict standard lower bounds once particle number, clock precision, or other physical overheads are counted (Meyer et al., 2013, Meyer et al., 2013, DalFavero et al., 2024).

A different body of work studies nonlinear quantum mechanics more abstractly through state discrimination. In that setting, two states of overlap $f(p)=p$4 can be separated in time $f(p)=p$5 for Gross–Pitaevskii-type nonlinearities, and unstructured search can be solved in

$f(p)=p$6

which becomes $f(p)=p$7 for constant $f(p)=p$8. The same work interprets this as evidence that the Gross–Pitaevskii approximation cannot remain valid arbitrarily long if one insists on consistency with the underlying linear many-body theory (Childs et al., 2015).

A common misconception is that every nonlinear search speedup should be attributed directly to nonlinearity. The control-theoretic literature shows otherwise. In “Controlled Quantum Search,” a complete-graph search with time-dependent site-dependent nonlinear strengths attains runtime

$f(p)=p$9

when $f(p)=p-p^2$0, and $f(p)=p-p^2$1 when $f(p)=p-p^2$2; yet on the complete graph the same scaling can be achieved with $f(p)=p-p^2$3, that is, with linear dynamics plus time-dependent local potentials. In that model, the speedup is due to control and symmetry rather than irreducibly nonlinear quantum mechanics (Lacy et al., 2017).

Another misconception is that nonlinear search uniformly beats Grover once any physically realistic nonlinearity is present. The conserved-quantity analysis of Gross–Pitaevskii search shows that repulsive interactions can merely worsen constants, and attractive interactions can improve only constants while preserving $f(p)=p-p^2$4 scaling, unless additional $f(p)=p-p^2$5-dependent couplings or more elaborate nonlinear structures are introduced (Meyer et al., 9 Mar 2025). A plausible implication is that “nonlinear quantum search” names a family of models rather than a single asymptotic phenomenon.

Recent work on reinforcement-based search pushes this boundary further by using a state-dependent Hamiltonian term $f(p)=p-p^2$6, i.e. effectively nonlinear feedback, to reduce search time from $f(p)=p-p^2$7 to $f(p)=p-p^2$8 and to improve noise tolerance. Because this model requires access to the evolving state or multi-copy simulation of $f(p)=p-p^2$9, it lies outside the standard linear-oracle framework and is best viewed as an adjacent nonlinear-search paradigm rather than a direct refinement of Grover’s setting (Homayouni-Sangari et al., 5 Apr 2026).

Nonlinear quantum search is therefore best understood as a technically diverse research area at the intersection of continuous-time quantum walks, many-body mean-field theory, state discrimination, and resource-sensitive complexity analysis. Its strongest speedups occur in effective models where additional physical or control resources are explicit; its most robust conceptual contribution is to show how state-dependent dynamics can radically alter search trajectories, graph dependence, and the geometry of amplitude amplification without invalidating the lower bounds of linear quantum computation once all resources are included.