Continuous-Time Nonlinear Quantum Walk

• The topic introduces continuous-time nonlinear quantum walks characterized by a cubic on-site potential (Gross–Pitaevskii term) that leads to self-trapping phenomena.
• It details rigorous bounds showing how local graph degree and nonlinearity regulate the dynamics, converting ballistic spreading into tunable localization.
• The framework extends to applications such as nonlinear spatial search and quantum state transfer, offering constant-time search and robust quantum memory schemes.

A continuous-time nonlinear quantum walk is a quantum walk whose evolution is governed by a nonlinear Schrödinger equation rather than a linear one. In the one-dimensional setting studied for paths and cycles, the effective Hamiltonian combines nearest-neighbor hopping with an on-site cubic term of Gross–Pitaevskii type,

$$H \;=\; -J\,\sum_{\langle n,m\rangle}\Bigl(\lvert n\rangle\langle m\rvert+\lvert m\rangle\langle n\rvert\Bigr) \;+\; g\sum_{n=0}^{N-1}|\psi_n|^2\,\lvert n\rangle\langle n\rvert,$$

where $J>0$ is the hopping rate and $g\in\mathbb R$ is the nonlinearity coefficient (Shi et al., 19 May 2026). The resulting dynamics differs qualitatively from linear one-dimensional quantum walks: instead of rapid spreading, the walker can remain localized at the initially occupied site to arbitrary fidelity depending on the nonlinear coefficient, a phenomenon proved analytically for discrete paths and cycles with cubic nonlinearity (Shi et al., 19 May 2026). Closely related continuous-time models also underlie nonlinear spatial search on incomplete graphs, where on-site nonlinear self-potentials can accelerate search relative to the linear quantum-walk regime (Meyer et al., 31 May 2026).

1. Formal definition and governing equations

For a walker on a one-dimensional graph of $N$ sites $v_0,\dots,v_{N-1}$, with site basis $\{\lvert n\rangle:n=0,\dots,N-1\}$, the state is written as

$$\lvert\Psi(t)\rangle \;=\;\sum_{n=0}^{N-1}\psi_n(t)\,\lvert n\rangle, \qquad \sum_n|\psi_n|^2=1.$$

With $\hbar=1$ and often $J=1$, the Schrödinger equation becomes the discrete Gross–Pitaevskii equation

$$i\,\frac{d\psi_n}{dt} \;=\; -\sum_{m:\,m\sim n}\psi_m \;-\; g\,|\psi_n|^2\,\psi_n, \qquad n=0,\dots,N-1,$$

where $J>0$0 denotes neighbors of $J>0$1 (Shi et al., 19 May 2026).

The cubic term acts as an on-site potential proportional to the instantaneous probability. In the stated physical interpretation, such nonlinearities arise in Bose–Einstein condensates described by the Gross–Pitaevskii equation or by nonlinear optical waveguide arrays (Shi et al., 19 May 2026). In this formulation, the nonlinearity is local, state-dependent, and diagonal in the site basis.

A broader continuous-time nonlinear quantum-walk framework appears in spatial search. There the evolution is again nonlinear,

$J>0$2

with

$J>0$3

where $J>0$4 is the graph Laplacian, $J>0$5 is a hopping rate, $J>0$6 is the oracle marking the target vertex, and $J>0$7 specifies the nonlinearity (Meyer et al., 31 May 2026). Two forms are emphasized: the cubic nonlinearity $J>0$8, which reproduces the Gross–Pitaevskii-type self-potential, and the cubic–quintic nonlinearity $J>0$9 (Meyer et al., 31 May 2026).

2. Self-trapping on one-dimensional paths and cycles

The central analytical result for one-dimensional continuous-time nonlinear quantum walks is self-trapping. For a walk started at a single vertex $g\in\mathbb R$0, let $g\in\mathbb R$1 and $g\in\mathbb R$2, with the remaining amplitudes collected into a vector $g\in\mathbb R$3 satisfying $g\in\mathbb R$4. The proof of trapping proceeds through a conserved functional,

$g\in\mathbb R$5

which is constant in time. At $g\in\mathbb R$6, when all amplitude is at $g\in\mathbb R$7, one has $g\in\mathbb R$8, so

$g\in\mathbb R$9

for all $N$0 (Shi et al., 19 May 2026).

A second ingredient is a bound on the energy expectation,

$N$1

obtained by viewing the adjacency-matrix block coupling $N$2 and $N$3, then applying Cauchy–Schwarz together with the spectral norm of the remaining block (Shi et al., 19 May 2026). Combining these relations yields

$N$4

which is rearranged as

$N$5

The function $N$6 diverges as $N$7 and $N$8 and has a unique minimum $N$9. If $v_0,\dots,v_{N-1}$0, the trajectory $v_0,\dots,v_{N-1}$1 cannot cross the two roots $v_0,\dots,v_{N-1}$2 of $v_0,\dots,v_{N-1}$3. Since $v_0,\dots,v_{N-1}$4, one obtains

$v_0,\dots,v_{N-1}$5

so the probability at the initially occupied site is bounded below by $v_0,\dots,v_{N-1}$6, establishing self-trapping (Shi et al., 19 May 2026).

For the threshold values determined by degree alone, the rigorous minima are $v_0,\dots,v_{N-1}$7 when $v_0,\dots,v_{N-1}$8 is an endpoint of a path, so $v_0,\dots,v_{N-1}$9, and $\{\lvert n\rangle:n=0,\dots,N-1\}$0 when $\{\lvert n\rangle:n=0,\dots,N-1\}$1 is an interior site of a path or any site on a cycle, so $\{\lvert n\rangle:n=0,\dots,N-1\}$2 (Shi et al., 19 May 2026). The same source states that numerically the true trapping threshold is $\{\lvert n\rangle:n=0,\dots,N-1\}$3, while the proof yields non-tight bounds $\{\lvert n\rangle:n=0,\dots,N-1\}$4 or $\{\lvert n\rangle:n=0,\dots,N-1\}$5 (Shi et al., 19 May 2026). As $\{\lvert n\rangle:n=0,\dots,N-1\}$6, $\{\lvert n\rangle:n=0,\dots,N-1\}$7, and the walker becomes arbitrarily well-trapped; for moderate $\{\lvert n\rangle:n=0,\dots,N-1\}$8, $\{\lvert n\rangle:n=0,\dots,N-1\}$9 (Shi et al., 19 May 2026).

This establishes a precise contrast with linear one-dimensional quantum walks, which are known for spreading quickly in one dimension (Shi et al., 19 May 2026). A plausible implication is that the nonlinear term effectively converts one-dimensional transport from a ballistic-spreading regime into a tunably localized regime.

3. Dependence on graph structure and initial placement

On the finite path $$\lvert\Psi(t)\rangle \;=\;\sum_{n=0}^{N-1}\psi_n(t)\,\lvert n\rangle, \qquad \sum_n|\psi_n|^2=1.$$0, the dynamics depends on whether the walker starts at an endpoint or in the interior. If the walker starts at an endpoint, where $$\lvert\Psi(t)\rangle \;=\;\sum_{n=0}^{N-1}\psi_n(t)\,\lvert n\rangle, \qquad \sum_n|\psi_n|^2=1.$$1, trapping appears for $$\lvert\Psi(t)\rangle \;=\;\sum_{n=0}^{N-1}\psi_n(t)\,\lvert n\rangle, \qquad \sum_n|\psi_n|^2=1.$$2 numerically, while the proof requires $$\lvert\Psi(t)\rangle \;=\;\sum_{n=0}^{N-1}\psi_n(t)\,\lvert n\rangle, \qquad \sum_n|\psi_n|^2=1.$$3. If the walker starts in the interior, where $$\lvert\Psi(t)\rangle \;=\;\sum_{n=0}^{N-1}\psi_n(t)\,\lvert n\rangle, \qquad \sum_n|\psi_n|^2=1.$$4, trapping again appears for $$\lvert\Psi(t)\rangle \;=\;\sum_{n=0}^{N-1}\psi_n(t)\,\lvert n\rangle, \qquad \sum_n|\psi_n|^2=1.$$5 numerically, while the proof requires $$\lvert\Psi(t)\rangle \;=\;\sum_{n=0}^{N-1}\psi_n(t)\,\lvert n\rangle, \qquad \sum_n|\psi_n|^2=1.$$6 (Shi et al., 19 May 2026).

On the cycle $$\lvert\Psi(t)\rangle \;=\;\sum_{n=0}^{N-1}\psi_n(t)\,\lvert n\rangle, \qquad \sum_n|\psi_n|^2=1.$$7, all sites have $$\lvert\Psi(t)\rangle \;=\;\sum_{n=0}^{N-1}\psi_n(t)\,\lvert n\rangle, \qquad \sum_n|\psi_n|^2=1.$$8, so the same interior threshold applies (Shi et al., 19 May 2026). In the analytical framework of the proof, the degree of the initially occupied vertex is the sole graph-theoretic parameter entering the lower-bound criterion through $$\lvert\Psi(t)\rangle \;=\;\sum_{n=0}^{N-1}\psi_n(t)\,\lvert n\rangle, \qquad \sum_n|\psi_n|^2=1.$$9. This suggests that, within this one-dimensional class, the local coordination number controls the rigorous trapping bound more directly than the global graph size.

The spatial-search setting broadens the graph dependence in a different direction. There, nonlinear speedups are established for graphs that are “sufficiently complete,” meaning that under linear search the system effectively evolves in a two-dimensional subspace spanned by $\hbar=1$0, with asymptotic spectral gap $\hbar=1$1 for some $\hbar=1$2 (Meyer et al., 31 May 2026). The examples explicitly listed are Paley graphs and complete bipartite graphs with $\hbar=1$3, both having three symmetry classes of vertices $\hbar=1$4 and $\hbar=1$5 (Meyer et al., 31 May 2026). Hypercubes and arbitrary-dimensional cubic lattices are treated numerically rather than through the same closed-form proof (Meyer et al., 31 May 2026).

The two one-dimensional trapping problem and the spatial-search problem therefore emphasize different structural features. In the former, the decisive parameter in the proof is the local degree of the starting vertex; in the latter, the decisive condition is the approximate two-dimensional spectral structure of the graph under linear search (Shi et al., 19 May 2026, Meyer et al., 31 May 2026).

4. Quantum state transfer and quantum memory interpretation

The one-dimensional self-trapping effect is proposed as a timing mechanism for quantum state transfer. The stated protocol on a small spin chain such as $\hbar=1$6 is: initialize the walker, encoding the qubit, at the source node $\hbar=1$7; turn on a large nonlinearity $\hbar=1$8 to trap the amplitude at $\hbar=1$9 with fidelity at least $J=1$0; when ready to transfer, suddenly set $J=1$1; allow linear evolution to perform perfect or pretty-good state transfer to the target node in a known time $J=1$2; and upon arrival, restore $J=1$3 large to trap the qubit at the receiving node indefinitely until use (Shi et al., 19 May 2026).

The specific example given is the path $J=1$4, for which perfect state transfer from $J=1$5 occurs at

$J=1$6

The same summary states that one can hold with $J=1$7 for arbitrary waiting times before and after the linear transfer window (Shi et al., 19 May 2026).

Within this scheme, the same chain is interpreted as both storage and transmission line. The storage fidelity is the self-trapping bound $J=1$8, and the release fidelity is the fidelity of the underlying linear state transfer (Shi et al., 19 May 2026). The source explicitly describes this as a “quantum memory viewpoint,” with trapping and transfer corresponding respectively to storage and release of quantum information (Shi et al., 19 May 2026).

The proposed candidate implementations are Bose–Einstein condensates in optical lattices, where the Gross–Pitaevskii nonlinearity $J=1$9 is proportional to the scattering length and tunable via Feshbach resonance, and nonlinear optical waveguide arrays with Kerr nonlinearity, where $$i\,\frac{d\psi_n}{dt} \;=\; -\sum_{m:\,m\sim n}\psi_m \;-\; g\,|\psi_n|^2\,\psi_n, \qquad n=0,\dots,N-1,$$0 can be modulated by pump intensity or external fields (Shi et al., 19 May 2026). A plausible implication is that the timing protocol depends not only on the existence of a nonlinear trapping regime but also on the ability to switch or modulate the effective nonlinearity on operational timescales.

5. Nonlinear spatial search and runtime scaling

In continuous-time spatial search, the walker begins in the uniform superposition

$$i\,\frac{d\psi_n}{dt} \;=\; -\sum_{m:\,m\sim n}\psi_m \;-\; g\,|\psi_n|^2\,\psi_n, \qquad n=0,\dots,N-1,$$1

and the marked vertex $$i\,\frac{d\psi_n}{dt} \;=\; -\sum_{m:\,m\sim n}\psi_m \;-\; g\,|\psi_n|^2\,\psi_n, \qquad n=0,\dots,N-1,$$2 enters through the oracle term $$i\,\frac{d\psi_n}{dt} \;=\; -\sum_{m:\,m\sim n}\psi_m \;-\; g\,|\psi_n|^2\,\psi_n, \qquad n=0,\dots,N-1,$$3 in the linear part of the Hamiltonian (Meyer et al., 31 May 2026). In the linear case, one chooses a critical hopping rate $$i\,\frac{d\psi_n}{dt} \;=\; -\sum_{m:\,m\sim n}\psi_m \;-\; g\,|\psi_n|^2\,\psi_n, \qquad n=0,\dots,N-1,$$4 so that two eigenvalues of $$i\,\frac{d\psi_n}{dt} \;=\; -\sum_{m:\,m\sim n}\psi_m \;-\; g\,|\psi_n|^2\,\psi_n, \qquad n=0,\dots,N-1,$$5 become nearly degenerate with gap $$i\,\frac{d\psi_n}{dt} \;=\; -\sum_{m:\,m\sim n}\psi_m \;-\; g\,|\psi_n|^2\,\psi_n, \qquad n=0,\dots,N-1,$$6. In the nonlinear regime, this is generalized to a time-dependent rate

$$i\,\frac{d\psi_n}{dt} \;=\; -\sum_{m:\,m\sim n}\psi_m \;-\; g\,|\psi_n|^2\,\psi_n, \qquad n=0,\dots,N-1,$$7

where $$i\,\frac{d\psi_n}{dt} \;=\; -\sum_{m:\,m\sim n}\psi_m \;-\; g\,|\psi_n|^2\,\psi_n, \qquad n=0,\dots,N-1,$$8 are the nonlinear self-potentials on symmetry classes $$i\,\frac{d\psi_n}{dt} \;=\; -\sum_{m:\,m\sim n}\psi_m \;-\; g\,|\psi_n|^2\,\psi_n, \qquad n=0,\dots,N-1,$$9; this choice forces the nonlinear evolution to track the linear path in a rescaled time (Meyer et al., 31 May 2026).

For cubic nonlinearity $J>0$00, the theorem stated for sufficiently complete graphs assumes

$J>0$01

with $J>0$02. Under these conditions, up to the first success time the nonlinear evolution remains within $J>0$03 of the rescaled linear critical-search path, and the hitting time satisfies

$J>0$04

For fixed $J>0$05, this remains $J>0$06, while for $J>0$07 it becomes $J>0$08, giving constant-time search (Meyer et al., 31 May 2026).

Applied to Paley graphs or complete bipartite graphs with $J>0$09, the same summary states that for any $J>0$10 the conditions hold and

$J>0$11

A refined analysis for Paley graphs permits $J>0$12, in which case the nonlinear algorithm succeeds in constant time $J>0$13 with probability $J>0$14 (Meyer et al., 31 May 2026).

For the cubic–quintic nonlinearity $J>0$15, the corresponding runtime is

$J>0$16

so $J>0$17 again yields constant-time search (Meyer et al., 31 May 2026). Numerical evidence extends this behavior to hypercubes of dimension $J>0$18, where simulations with $J>0$19 show nearly constant $J>0$20 and success probability near unity for $J>0$21, and to periodic $J>0$22-dimensional lattices, where cubic or cubic–quintic nonlinearity with $J>0$23 gives near-constant $J>0$24 and high success for $J>0$25, but fails to reach the same success probability as the linear case for $J>0$26 (Meyer et al., 31 May 2026). The threshold dimension is therefore reported as $J>0$27 (Meyer et al., 31 May 2026).

6. Relation to nonlinear Dirac limits, photonic models, and open issues

Although the one-dimensional self-trapping and spatial-search models are continuous-time systems from the outset, related nonlinear quantum-walk literature often studies discrete-time walks whose continuum limit is a nonlinear Dirac equation. In a general nonlinear quantum walk on the lattice $J>0$28, the one-step evolution operator is

$J>0$29

with a shift $J>0$30, a linear coin $J>0$31, and a nonlinear state-dependent coin $J>0$32. Using Shannon interpolation, the walker on $J>0$33 is shown to converge uniformly in Sobolev space $J>0$34 on fixed time intervals to the solution of a nonlinear Dirac equation as $J>0$35 (Maeda et al., 2019). The limiting equation is

$J>0$36

and the convergence theorem states

$J>0$37

under the stated regularity and Lipschitz assumptions (Maeda et al., 2019).

A different discrete nonlinear model, designed for photonic implementation using an optical nonlinear Kerr medium, yields in a space-time continuum limit a nonlinear Dirac equation with a nonlinear mass term,

$J>0$38

and supports approximate bright and dark solitons in the continuum analysis (Anglés-Castillo et al., 2023). Numerical iteration of the underlying map produces localized bright solitons, kick-activated motion, soliton–soliton collisions in which packets pass through one another largely unscathed, dark solitons, and modified behavior under synthetic electric fields (Anglés-Castillo et al., 2023).

These discrete-time continuum-limit results do not define the same model as the continuous-time Gross–Pitaevskii walk on a path or cycle. They nevertheless place continuous-time nonlinear quantum walks within a larger research program linking state-dependent walk dynamics, effective nonlinear field equations, and localized structures such as self-trapped states and solitons (Maeda et al., 2019, Anglés-Castillo et al., 2023). A common misconception is that “nonlinear quantum walk” denotes a single canonical construction; the cited literature instead includes at least continuous-time Gross–Pitaevskii-type walks on graphs, continuous-time search Hamiltonians with nonlinear self-potentials, and discrete-time state-dependent coin walks with nonlinear Dirac limits (Shi et al., 19 May 2026, Meyer et al., 31 May 2026, Maeda et al., 2019, Anglés-Castillo et al., 2023).

Several limitations and open questions are explicitly identified in the spatial-search setting. The effective nonlinear Schrödinger equation arises in the mean-field Gross–Pitaevskii limit of a Bose–Einstein condensate with two-body interactions, and the coupling $J>0$39 scales inversely with the number of particles $J>0$40, so achieving large $J>0$41 while maintaining the mean-field regime requires a trade-off between condensate size and interaction strength (Meyer et al., 31 May 2026). The same source notes that low-dimensional lattices and sparse graphs that fail to approximate the two-dimensional spectral structure of the complete graph do not benefit except at very large $J>0$42 that break the approximations, and it lists open questions concerning rigorous hypercube speedup proofs, universal lower bounds on $J>0$43 versus graph spectral properties, and particle-number versus connectivity trade-offs in a Bose–Einstein-condensate implementation (Meyer et al., 31 May 2026).