Research
Non-Convex Optimization via Quantum Tunneling Search on a Photonic Chip
Demonstrates quantum tunneling search for non-convex objectives on a femtosecond-written photonic waveguide chip, moving the algorithm from proposal to hardware.

- arXiv
- 2611.16791
- Published
- 2026-11-12
- Authors
- Xi-Jun Yuan, Shuo Zhou, Yizhou Liu, Xuan Chen, Hong-Jing Du, Xiao-Wen Shang, Wen-Hao Zhou, Yu-Xuan Fu, Bao-Jing Liu, Ze-Feng Lan, Hang Song, Fang-Song Nie, Yu-Lin Yu, Xiao-Yun Xu, Jian-Peng Dou, Tongyang Li, Hao Tang, Xian-Min Jin
AI summary
Overview
Research area: Quantum physics / photonic quantum computing — specifically analogue quantum simulation on integrated photonic chips applied to non-convex optimization for machine learning.
Technical level: Advanced. The paper assumes familiarity with Hamiltonian dynamics, waveguide arrays, quantum walks, and optimization theory, though its central claim is stated in accessible terms.
Scope (1 sentence): The paper proposes, implements, and benchmarks "quantum tunneling search" (QTS) — a quantum-tunneling-walk algorithm with local annealing — on femtosecond-laser-written photonic waveguide arrays encoding 1D and 2D non-convex landscapes.
What This Paper Is About
Machine-learning loss functions are generally non-convex, riddled with local minima and saddle points that trap gradient-descent methods before they reach a global optimum. The authors ask whether the quantum tunneling effect — in which a particle can pass through an energy barrier even without enough energy to climb it — can be harnessed as a practical optimizer. Their goal is to build such an optimizer physically, in a photonic chip, and test it against classical stochastic gradient descent (SGD).
Key Contributions
-
A named algorithm, quantum tunneling search (QTS): quantum tunneling walk (QTW) run from a specially prepared low-energy input state, followed by averaging the probability distribution over the evolution.
-
A physical encoding scheme: the objective function is written directly into the on-site potential term of the waveguide Hamiltonian by tuning each waveguide's refractive index with femtosecond direct laser writing, so that each grid point of the landscape corresponds to one waveguide.
-
An "incomplete local annealing" state-preparation step: rather than attempting full adiabatic quantum annealing to an exact ground state, the authors perform only a few staged annealing steps inside one known potential well, producing a state slightly above the ground energy but with faster subsequent search.
-
A varying-Hamiltonian extension: gradually increasing the potential strength relative to the fixed coupling along the propagation direction, which improves exploration of distant regions and, in one demonstration, isolates the global minimum at a specific waveguide.
Main Findings
-
Local minima are recovered experimentally. On a 2D double-well landscape built from Equation (6) of the paper (the landscape of Ref. [18]) implemented across a 21 × 9 waveguide array at 10 µm spacing, the average probability distribution over propagation lengths from 5 mm to 90 mm in 5 mm steps still highlights both wells, even though the injected state was not the exact ground state of the left well.
-
A "hitting" length of 55 mm is observed. At z = 55 mm the measured snapshot shows intensity in the right potential well; the authors define this as the minimum propagation length needed to reach the region of interest once, and note that assigning a first-arrival length would require an intensity threshold and is limited by the 5 mm sampling spacing.
-
1D multi-well landscapes behave as simulated. On 29 × 1 waveguide arrays hosting three- and five-well landscapes, the experimentally averaged distribution agrees well with simulation (simulation coupling C = 0.47, with perturbations applied to β). The annealed input state has energy E = 0.162, above the ground state E₀ = 0.034 but below the first excited state.
-
Convergence was not reached on the fixed-length chip. Running averages of well occupation show that the maximum available waveguide length is insufficient for the probability distribution to converge, which the authors give as the motivation for the varying-Hamiltonian approach.
-
The varying Hamiltonian spreads probability more evenly. On a five-well random landscape, both constant-H (averaged over z = 80 mm at 5 mm intervals) and varying-H forms highlight all local minima; varying-H enhances exploration of distant regions and yields a slightly more uniform distribution. The varying form used there is H(z) = C(−½Δ) + (z + z³)f(x).
-
The global minimum can be singled out. With H(z) = C(−½Δ) + (0.5z + 0.05z² + 0.2z³)f(x), the measured probability at the global minimum at x = 14 dominates other local minima at the final stage z = 90 mm.
-
The SGD comparison is explicitly qualified by the authors. SGD is run as a classical baseline (e.g. 1,600 iterations for the random-landscape comparison, 1,800 iterations for the global-minimum comparison, with positions scored as trajectories within |x − 14| < 0.5). The supplementary material states that the horizontal axes are fundamentally different (iteration count versus fabricated propagation length) and that the factor of 60 rescaling is "a display convention and not an independently established comparison of resources." It further states the data "do not allow such an end-to-end speedup estimate." The speedup over SGD is therefore argued theoretically (from the query-complexity separation of Ref. [18]) rather than measured end-to-end.
Methodology in Plain English
-
Encode the problem in glass. A network of closely spaced optical waveguides is written into a chip with femtosecond laser pulses. Light hopping between neighbouring waveguides plays the role of a quantum particle moving across a landscape; each waveguide is one grid point of the objective function.
-
Make deep valleys optically deep. Waveguides corresponding to low objective values get a refractive index (propagation constant β) that acts as a deeper potential well. Light naturally tends to linger there, so bright spots in the output image mark candidate minima. The Hamiltonian is H = Σᵢ βᵢaᵢ†aᵢ + Σᵢ≠ⱼ Cᵢ,ⱼaᵢ†aⱼ, with only nearest-neighbour coupling.
-
Use propagation distance as time. Because ħ is set to 1, the evolution |Ψ(z)⟩ = e^{−iHz}|Ψ(0)⟩ is controlled simply by how long the waveguides are. Different chip lengths therefore give snapshots at different "times."
-
Prepare a good starting state cheaply. Instead of running a full quantum anneal to the exact ground state (impractical because of noise and the adiabatic condition), the authors anneal only inside one small, already-known region: they linearly interpolate H(t) = (1 − t/T)H_initial + (t/T)H_final while decreasing β, letting the state spread out a little. They do this in a few stages — for the 1D experiment, five stages, with β varying as 2^(6−10z) over each 0.1 mm stage.
-
Let tunneling do the search, then average. The prepared state is injected into the full landscape, and the probability distribution is averaged over many propagation lengths. This ensemble-averaging stands in for the ergodic average of a long run.
-
Strengthen the landscape over time. In the varying-Hamiltonian variant, the ratio β(z)/C is ramped up along the chip, deepening the wells so the state localizes with better position accuracy (Δx) near the lowest energy.
-
Read out optically. A 20× microscope objective couples light in at a nominal 1 mW, measurements are taken at 21 °C, and a CCD camera images the output facet. The reported quantities are normalized classical light-intensity fractions, not single-photon detection probabilities; processing involves background subtraction, anchoring on bright channels, rectangular regions of interest per waveguide, and thresholding.
Why This Matters
Impact on research. The work is one of relatively few experimental tests of a quantum-walk-based optimizer on real hardware for a machine-learning-relevant task. It shows that an objective function can be written into a chip's Hamiltonian at fabrication time and that tunneling dynamics plus a modest annealing step is enough to locate minima. It also contributes a carefully hedged account of how such experiments should — and should not — be compared with classical SGD.
Real-world applications (as motivated or implied by the paper):
- Training and loss-landscape exploration for deep neural networks, whose loss surfaces are high-dimensional and non-convex.
- General combinatorial and continuous optimization where gradient information is unreliable or gradients vanish.
- Analog/photonic accelerators for optimization workloads that are expensive on digital hardware.
- A testbed for Hamiltonian engineering as an optical computing primitive beyond simulation-oriented use.
Industry relevance. The author list includes an industry affiliation (TuringQ Company, Ltd.) alongside university and national-laboratory groups, and the paper's future-work section points at reconfigurable photonic platforms. That framing places the result in the photonic-computing hardware space rather than purely in academic quantum simulation.
Future Directions
-
Higher-dimensional landscapes. Because one grid point occupies one waveguide, adding a coordinate multiplies the waveguide count. The authors propose exploiting additive structure or low effective dimension, and hybrid optical–digital architectures, to keep chips realizable.
-
Reconfigurable chips. Fabrication currently fixes both the couplings and the refractive-index profile, so each objective function needs a new chip. Thermo-optic or electro-optic tuning, phase-change devices, and programmable photonic meshes are named as candidate routes to encode a family of objectives on one device.
-
Two-photon experiments. The present study uses coherent light and single-particle tunneling dynamics; measuring output coincidences for distinguishable and indistinguishable photon pairs would test predicted bunching (Figure S11) and how interference changes the chance of detecting both photons near the same minimum.
-
A stopping rule. The supplementary material states that determining when to stop for an unknown landscape "remains an open issue," since the measurement end length is currently set by available structures and modulation range rather than an online optimality certificate.
Target Audience
Researchers in photonic quantum computing and analogue quantum simulation; quantum-algorithms theorists interested in optimization speedups; machine-learning researchers curious about non-gradient-based and physical optimization substrates; and photonics engineers working on programmable or application-specific optical processors. The paper is best suited to readers comfortable with Hamiltonian formalism; the framing and conclusions are readable by a broader computational-science audience, particularly given the authors' explicit caveats about benchmark interpretation.
Authors’ abstract
Non-convex optimization plays an important role in modern machine learning since plenty of real-world problems have objective functions or constraints with non-convex properties. The non-convex nature of landscapes gives rise to numerous local minima and saddle points, dramatically hindering the ability of gradient descent-based methods to explore the landscape and obtain global optima. Here we present quantum tunneling search (QTS), an algorithm based on quantum tunneling…
The full abstract on arXiv