Chronological feed of everything captured from Hartmut Neven.
paper / hneven / Jun 24
Quantum-Assisted Genetic Algorithms (QAGAs) integrate reverse quantum annealing as a mutation operator in classical genetic algorithms, leveraging quantum fluctuations for quasi-local/nonlocal search from classical states while using classical crossovers. Experiments on a D-Wave 2000Q processor demonstrate that QAGAs outperform standard forward quantum annealing on spin-glass benchmarks, rapidly finding good solutions and reliably reaching global optima where forward annealing stalls. This hybrid approach highlights a viable NISQ application for heuristic discrete optimization via complementary classical-quantum fluctuations.
quantum-annealinggenetic-algorithmsquantum-assisted-optimizationd-wave-processornisp-devicesspin-glass-problemsevolutionary-computing
“Reverse quantum annealing serves as a mutation operator in genetic algorithms by providing quasi-local or quasi-nonlocal search from classical states.”
paper / hneven / May 1
Researchers used qFlex, a tensor-network-based simulator, to model random quantum circuits (RQCs) on Summit, achieving 281 Pflop/s sustained performance in single precision. This classical simulation benchmarks NISQ quantum supremacy claims by replicating Google's experimental circuits. Results highlight NISQ devices' energy efficiency advantage over supercomputers by orders of magnitude and propose qFlex as a NISQ benchmark standard.
quantum-supremacynisc-computershpc-simulationrandom-quantum-circuitsqflex-simulatorsupercomputer-summit
“qFlex simulations on Summit sustained an average performance of 281 Pflop/s (true single precision) for hard random quantum circuits (RQC).”
paper / hneven / Mar 14
Proposes post-processing error decoders for quantum codes that mitigate errors on logical qubits using classical computation, without syndrome measurements or extra qubits. Employs stochastic sampling of subspace expansions to approximate projectors, enabling correction of logical errors, unencoded Hamiltonians, and approximate symmetries. Demonstrates pseudo-threshold of p≈0.50 for [[5,1,3]] code under depolarizing noise and improved fidelity for unencoded H2 molecule simulation.
quantum-error-correctionquantum-codessubspace-expansionquantum-error-decodingnear-term-quantumquantum-chemistryperfect-code
“Post-processing decoders mitigate errors using only logical qubits and classical post-processing, bypassing syndrome measurements and feed-forward.”
paper / hneven / Feb 28
Future quantum systems require integrated cryogenic circuits for qubit control. Researchers have developed a prototype 28nm bulk-CMOS cryogenic integrated circuit optimized for transmon qubit control. This circuit's performance has been validated in quantum control experiments, demonstrating a crucial step towards scalable quantum computing.
quantum-computingcryogenic-electronicsintegrated-circuitsqubit-controlcmos-technologyquantum-hardware
“Future quantum computing systems will require cryogenic integrated circuits to control and measure millions of qubits.”
paper / hneven / Feb 27
Optimized low-order Trotter-Suzuki product formulas with phase estimation achieve superior performance over linear combination of unitaries for fault-tolerant quantum simulation of correlated electron models like Hubbard and plane-wave basis, targeting relative precision metrics such as energies per unit cell. These methods enable simulations of up to N < 10^5 fermionic modes with O(1) T-complexity for Hubbard models and O(N^2) for plane-wave models. Fault-tolerant implementations using surface code gates at 10^{-3} error rates require only a few hundred thousand physical qubits for classically intractable instances.
quantum-simulationtrotterizationfault-tolerant-quantumcondensed-matterhubbard-modelelectronic-structurearxiv-paper
“Low-order Trotter methods with phase estimation compute relative precision quantities like energies per unit cell for condensed-phase systems.”
paper / hneven / Dec 19
The Majorana loop stabilizer code (MLSC) is a geometric locality-preserving fermion-to-qubit mapping where stabilizers are products of Majorana operators on closed paths of the fermionic hopping graph. On a 2D square lattice, MLSC corrects all single-qubit errors, surpassing prior geometric codes that only detect them, while mapping fermionic operators to lower-weight qubit operators despite higher code distance. MLSC integrates with quantum chemistry simulation algorithms, providing error correction without asymptotic overhead, ideal for near-term fermionic quantum simulations.
majorana-loop-stabilizer-codefermion-to-qubit-mappingquantum-error-correctiongeometric-localityfermionic-simulationsquantum-chemistrylattice-models
“MLSC preserves geometric locality in fermion-to-qubit mappings, avoiding overhead from non-local parity terms in Jordan-Wigner or Bravyi-Kitaev transformations.”
paper / hneven / Dec 11
For fixed control parameters, QAOA's objective function value concentrates around a consistent mean for typical instances from reasonable distributions, rendering the optimization landscape instance-independent. This holds rigorously for low-depth circuits on MaxCut instances over large 3-regular graphs and extends numerically to higher depths via the Law of Large Numbers. Simulations further indicate that parameters optimized on small instances (e.g., 10 qubits) perform well on larger ones (e.g., 24 qubits), enabling parameter reuse and reducing quantum hardware calls.
qaoaquantum-approximate-optimizationmaxcutquantum-circuitsconcentration-phenomenonquantum-algorithmsarxiv-paper
“For fixed parameters and typical instances from reasonable distributions, QAOA objective function values concentrate such that typical instances yield nearly identical values.”
paper / hneven / Jul 25
New quantum algorithm simulates chemistry in plane wave basis with gate complexity \tilde{O}(N^{1/3} \eta^{8/3}), where N is orbitals and \eta electrons, beating prior \tilde{O}(N^{8/3} / \eta^{2/3}) scaling. Achieved in first quantization using rotating frame of kinetic operator and interaction picture. Excels when N \gg \eta, critical for low discretization error or non-Born-Oppenheimer simulations.
quantum-simulationquantum-chemistryplane-wavesfirst-quantizationquantum-algorithmssublinear-scaling
“New algorithm gate complexity is \tilde{O}(N^{1/3} \eta^{8/3})”
paper / hneven / Jul 12
Coherent multi-qubit quantum tunneling forms bands of non-ergodic extended (NEE) states, superpositions of numerous low-energy computational states, facilitating population transfer (PT) between them. In an n-spin transverse-field model encoding optimization problems, PT targets bitstrings within a narrow energy window around an initial low-energy state. Analytical results show PT runtime scales like multi-target Grover search, modulated by exp(n/(2B²)) for large driver field B, with numerical validation in quantum parallel tempering for binary optimization on dense graphs.
quantum-tunnelingnon-ergodic-statespopulation-transferquantum-searchquantum-optimizationspin-glassparallel-tempering
“NEE states from quantum tunneling enable population transfer between computational states with similar energies.”
paper / hneven / Jun 7
The paper introduces asymmetric qubitization for simulating SYK model dynamics with N Majorana modes over time t to precision ε using O(N^{7/2} t + N^{5/2} t polylog(N/ε)) gates. This scales sublinearly in Hamiltonian terms, exponentially better in 1/ε, and polynomially superior in N and t compared to prior O(N^{10} t^2 / ε) methods. The technique encodes the Hamiltonian via projections of a signal oracle onto states from Hadamard gates (for B) and random quantum circuits (for A).
quantum-simulationsachdev-ye-kitaevqubitizationmajorana-modesquantum-algorithmsarxiv-paper
“Asymmetric qubitization simulates SYK model with N Majorana modes for time t to precision ε using O(N^{7/2} t + N^{5/2} t polylog(N/ε)) gates”
paper / hneven / May 22
Quantum neural networks using universal near-term circuit topologies are trained via Adam optimization to implement generalized quantum measurements (POVMs) that optimally discriminate non-orthogonal quantum probability distributions. These circuits, parameterized through classical-quantum interactions, simulate unknown non-unitary operations and achieve performance comparable to theoretical optima in minimizing errors and inconclusives. Simulations demonstrate generalization to unseen mixed states, highlighting quantum machine learning's unique advantage over classical methods for this inherently quantum task.
quantum-neural-networksquantum-discriminationpovmquantum-machine-learningnear-term-quantum-circuitsadam-optimization
“Universal quantum circuit topologies guarantee the ability to learn arbitrary input-output mappings.”
paper / hneven / May 9
Quantum circuits exactly encode correlated electron spectra with T gate complexity O(N + log(1/ε)) for N orbitals, enabling qubitization-based phase estimation with optimal O(λ/ε) query complexity. For electronic structure Hamiltonians, this yields overall T complexity O(N³/ε + N² log(1/ε)/ε), asymptotically superior to prior methods in the classically intractable regime. Fault-tolerant compilation to surface code with 10^{-3} gate errors supports phase estimation on relevant instances using ~1 million superconducting qubits in hours.
quantum-circuitsquantum-phase-estimationelectronic-structurehubbard-modelt-complexityqubitizationquantum-chemistry
“Quantum circuits encode Hubbard model and electronic structure spectra exactly up to rotation synthesis errors”
paper / hneven / Mar 29
Parameterized quantum circuits trained via hybrid quantum-classical optimization suffer from barren plateaus when initialized with random circuits. The probability of non-zero gradients to fixed precision vanishes exponentially with qubit count due to Hilbert space dimensionality and 2-design properties. This limits scalability beyond few qubits, necessitating new initialization strategies.
quantum-neural-networksbarren-plateausquantum-traininghybrid-quantum-classicalrandom-circuitsquantum-machine-learning
“For wide class of reasonable parameterized quantum circuits, probability of gradient along any reasonable direction being non-zero to fixed precision is exponentially small in number of qubits”
paper / hneven / Mar 8
High-fidelity qubit control demands precise, device-specific parameters found via iterative bootstrapped calibrations, which drift over time. The proposed framework models calibration dependencies as a directed acyclic graph (DAG), reducing the process to graph traversal. This enables automation and extensibility of calibration strategies for quantum devices.
quantum-physicsphysical-qubitsqubit-calibrationdirected-acyclic-graphquantum-computingcalibration-strategyarxiv-paper
“High-fidelity qubit control requires precisely tuned control parameters.”
paper / hneven / Mar 5
Deep reinforcement learning with trust-region-policy-optimization optimizes two-qubit unitary gates for quantum simulation, enhancing speed and fidelity under leakage and stochastic errors. Training incorporates control noise to boost robustness, yielding solutions two orders of magnitude better in average gate error than stochastic gradient descent baselines. Gate times reduce by up to one order of magnitude compared to optimal synthesis methods.
quantum-physicsquantum-controlreinforcement-learningdeep-rlquantum-gatesquantum-computingarxiv-paper
“Deep RL reduces average gate error by two orders of magnitude over stochastic-gradient-descent baselines for two-qubit unitary gates.”
paper / hneven / Feb 26
Coherent tunneling under a transverse field creates bands of delocalized states for efficient population transfer (PT) between low-energy marked states in a narrow energy window. In a model with M marked bit-strings amid 2^n states, PT from one marked state yields a superposition over ~Ω marked states with runtime scaling as in multi-target Grover's algorithm, multiplied by exp(n / (2 B_⊥²)) for n ≫ B_⊥² ≫ 1. The system's non-integrable Hamiltonian produces mini-bands of non-ergodic delocalized states with heavy-tailed width distribution, analytically derived from nonlinear cavity equations for random matrix ensembles.
quantum-physicsdelocalized-statespopulation-transferquantum-optimizationnon-ergodic-statestransverse-fieldgrover-algorithm
“PT starts at a marked state and ends in a superposition of ~Ω marked states inside the PT window.”
paper / hneven / Feb 16
The paper introduces a quantum neural network (QNN) using parameter-dependent unitary transformations on input quantum states, with binary classification via measurement of a Pauli operator on a readout qubit. It demonstrates exact representation of any Boolean function on n-bit classical inputs using two-qubit unitaries, though some require exponential circuit depth, and shows trainable parameters successfully distinguish downsampled handwritten digit images via classical simulation. QNNs extend to quantum superposition inputs and general quantum state labeling, designed for execution on near-term gate-model quantum hardware beyond simulation limits.
quantum-neural-networksqnnquantum-classificationsupervised-learningnear-term-quantumquantum-machine-learningarxiv-paper
“A QNN circuit made from two-qubit unitaries can exactly represent the label of any Boolean function on n-bit inputs.”
paper / hneven / Dec 14
A classical algorithm simulates sampling from universal random quantum circuits on a single workstation, achieving larger qubit counts and depths than previously reported, such as 5x9 qubits at depth 37, 7x8 at depth 27, and 10x(κ>10) at depth 19. The method, akin to the Feynman path approach, scales exponentially with depth times the smaller lateral dimension (treewidth), confirming Boixo et al. bounds; e.g., 7x7 qubits at depth 40 remains infeasible. Sampling and observable estimation are trivially parallelizable up to certain depths, with supercomputer scaling expected to extend limits.
quantum-circuitsclassical-simulationquantum-supremacylow-depth-circuitsgraphical-modelsquantum-computing
“Circuits with 5×9 qubits and depth 37 are easy to sample on a single workstation”
paper / hneven / Aug 31
Analysis of quantum annealing machines using real-world fault diagnosis in multiplier circuits shows these instances are harder than comparable random spin-glass benchmarks. State-of-the-art classical optimization algorithms outperform transverse-field quantum annealing on these problems. These small, challenging instances from real data are ideal for testing near-term quantum annealers and alternative quantum optimization strategies.
quantum-annealingquantum-optimizationbenchmarkingfault-diagnosisdigital-circuitsquantum-computingindustrial-applications
“Random spin-glass problems are not well suited to detect quantum speedup in annealing machines.”
paper / hneven / Aug 23
Path-integral QMC with open-boundary conditions (OBC) simulates tunneling decay in fully connected quantum spin models, but the conjectured quadratic speedup over periodic-boundary QMC (PBC) due to half-instanton action does not hold. Exact instanton endpoints require specific positions and non-zero momenta, unlike the conjecture's zero-momentum assumption, and OBC instantons reside in the symmetric subspace of maximum total angular momentum across all temperatures, while PBC instantons occupy lower-angular-momentum subspaces at finite temperatures. This results in less than quadratic speedup at finite temperatures. The method generalizes to non-permutation-symmetric many-qubit systems via spin-coherent-state path integrals.
quantum-monte-carlopath-integralquantum-tunnelingopen-boundary-conditionsinstantonquantum-spin-modelarxiv-paper
“OBC QMC escape rate is quadratically larger than PBC QMC according to prior conjecture”
paper / hneven / Aug 6
Noisy chaotic quantum circuits converge exponentially to uniform bit-string distributions under general noise models, making random sampling sufficient for approximation. Classical supercomputers cannot simulate high-fidelity versions with ~50 qubits and depth 40. A proposed Fourier-based approximation algorithm for noisy distributions fails to outperform random guessing in polynomial time, reinforcing quantum advantage claims.
quantum-circuitsquantum-supremacynoisy-quantumfourier-analysisquantum-samplingchaos-quantumcross-entropy-benchmarking
“Under general noise models, chaotic quantum circuit output distributions converge exponentially to uniform as gate count increases.”
paper / hneven / May 31
The paper introduces a dual plane wave basis that diagonalizes the potential operator, reducing the second-quantized Hamiltonian terms for electronic structure from O(N⁴) in Gaussian orbitals to O(N²). This enables single Trotter steps with linear gate depth on planar lattices and overall circuit depths of O(N^{7/2}) for Trotter and Õ(N^{8/3}) for Taylor series simulations at fixed charge densities, surpassing prior methods. Variational algorithms benefit from fewer measurements, with a proposed low-depth ansatz for simulating low-density jellium on near-term devices to demonstrate quantum supremacy.
quantum-simulationelectronic-structurequantum-algorithmstrotterizationvariational-quantumplane-wave-basisjellium-model
“Dual plane wave basis reduces second-quantized Hamiltonian terms to O(N²)”
paper / hneven / Jul 31
Random quantum circuit sampling demonstrates quantum supremacy as it requires exponential classical simulation time due to chaotic dynamics. Supercomputer simulations up to 42 qubits confirm convergence to this regime, with supremacy feasible in near-term devices using about 50 superconducting qubits despite error sensitivity. Cross-entropy benchmarking approximates circuit fidelity, enabling practical supremacy tests via extrapolation beyond tractable simulation sizes.
quantum-supremacyquantum-computingrandom-circuitsquantum-benchmarkingnear-term-devicescross-entropysuperconducting-qubits
“Sampling from (pseudo-)random quantum circuit output distributions requires exponential time on classical computers.”
paper / hneven / Jul 21
Pontryagin's minimum principle optimizes variational quantum algorithms by deriving bang-bang (square pulse) evolution as optimal for fixed computation time in both closed and open quantum systems with Markovian decoherence. This supports the ansatz in the Quantum Approximate Optimization Algorithm (QAOA) and yields a system-size independent pulse duration distribution for the Sherrington-Kirkpatrick spin-glass, scaled by inverse Hamiltonian coupling constants. Numerical results show these nonadiabatic bang-bang protocols outperform quantum annealing under weak noise and Redfield thermal bath dynamics.
variational-quantum-algorithmspontryagins-minimum-principlebang-bang-controlquantum-approximate-optimizationquantum-annealingspin-glassquantum-optimization
“Optimal evolution in variational quantum algorithms for fixed time is bang-bang (square pulse) form”
paper / hneven / Jun 27
The paper derives a general, non-perturbative interaction Hamiltonian for superconducting qubits coupled inductively via a nonlinear Josephson coupler under the Born-Oppenheimer approximation. The interaction splits into a classical component from circuit equations and a quantum component from the coupler's zero-point energy, enabling non-stoquastic and many-body terms beyond linear theories. Explicit series expansions are provided for any Hamiltonian term, applicable to any qubit type, with numerical validation on flux qubits assessing the approximation's regime.
superconducting-qubitsinductive-couplingquantum-physicsborn-oppenheimer-approxflux-qubitsnonlinear-regime
“Tunable interactions in superconducting qubits are achieved via mutual inductive coupling to a coupler with a nonlinear Josephson element.”
paper / hneven / Mar 3
Researchers derive an analytical instanton-based expression for the thermally-assisted tunneling decay rate in a fully connected quantum spin model, mapping it to the Kramers escape problem under a classical random dynamical field simulated via path integral QMC. They prove analytically that the exponential scaling with spin number N is identical for both the quantum tunneling rate and QMC escape rate, attributed to a dominant instantonic path governed by nonlinear mean-field equations for single-site magnetization. Scaling relations are established for spiky barriers where tunneling and QMC rates scale polynomially in N, contrasting classical activation's exponential scaling.
quantum-tunnelinginstantonsquantum-monte-carlothermally-assisted-tunnelingmean-field-theoryscaling-analysisquantum-spin-models
“Thermally-assisted tunneling decay in fully connected quantum spin model maps to Kramers escape problem of a classical random dynamical field”
paper / hneven / Dec 7
D-Wave 2X quantum annealer exploits finite-range tunneling to solve crafted problems with tall, narrow energy barriers, achieving time-to-99% success probability ~10^8 times faster than single-core simulated annealing for 945-variable instances. It also outperforms optimized quantum Monte Carlo emulation by the same factor, revealing a substantial constant overhead in classical simulation of physical QA. Numerical studies on number partitioning indicate quantum tunneling simulations scale better than SA, with implications for future annealer designs outpacing Chimera-specific classical heuristics.
quantum-annealingd-wavequantum-tunnelingsimulated-annealingquantum-advantageoptimizationquantum-computing
“D-Wave 2X is ~10^8 times faster than single-core SA to reach 99% success probability on 945-variable crafted problems”
paper / hneven / Oct 27
Quantum Monte Carlo (QMC) simulations of Ising ferromagnet ground-state tunneling exhibit the same O(Δ²) scaling with system size as incoherent tunneling, where Δ is the tunneling splitting. This equivalence enables QMC to predict quantum annealer performance through barriers. Open boundary conditions in imaginary time yield a quadratic QMC speedup, achieving linear scaling in Δ, explained via instanton theory.
quantum-tunnelingquantum-monte-carloquantum-annealingising-ferromagnetinstantonqmc-simulations
“QMC tunneling rate scales as O(Δ²) with system size, matching incoherent tunneling.”
paper / hneven / Oct 15
The paper introduces a theoretical framework for engineering a quantum thermal bath that drives a target quantum system with Hamiltonian H to the exact thermal equilibrium state e^{-H/T}/Tr(e^{-H/T}) using a tunable temperature parameter T, decoupled from the ambient environment. This analog approach mirrors the digital quantum Metropolis algorithm. For superconducting qubits, it proposes a circuit-QED implementation via driven lossy resonators, enabling simulation of large many-body quantum systems beyond classical reach and acting as a temperature knob for hybrid quantum-thermal annealers.
quantum-thermal-bathartificial-temperaturequantum-engineeringsuperconducting-qubitscircuit-qedmany-body-systemsquantum-annealing
“An engineered quantum bath coupled to a quantum system with Hamiltonian H drives it to the equilibrium state e^{-H/T}/Tr(e^{-H/T}) with tunable T.”
paper / hneven / Apr 7
Proposes a totally corrective boosting algorithm with explicit cardinality regularization, solving NP-hard combinatorial problems intractable for classical methods but promising for quantum optimizers. Experiments using a distributed classical heuristic demonstrate superior generalization and sparsity on benchmark datasets compared to standard boosting. Early stopping in unregularized boosting implicitly provides suboptimal cardinality control; fully solving the combinatorial problem yields better results, with quantum hardware expected to reduce training costs significantly.
totally-corrective-boostingcardinality-penalizationquantum-optimizationmachine-learningsparsity-regularizationboosting-algorithmsquantum-machine-learning
“Cardinality-penalized totally corrective boosting achieves better generalization performance than standard boosting on public benchmark datasets.”
paper / hneven / Mar 24
Quantum chemical modeling applies inelastic electron tunneling spectroscopy (IETS) to explain GPCR activation in mammalian nervous system, focusing on serotonin receptor agonists. Non-endogenous agonists share a characteristic IET spectral peak whose intensity correlates with their known potencies, mirroring serotonin itself. The model proposes experimental validation using deuterated lysergic acid dimethylamide (DAM-57) isotopologues, potentially advancing in silico drug potency prediction.
g-protein-coupled-receptorsneuroreceptor-activationinelastic-electron-tunnelingquantum-chemical-modelingdrug-designserotonin-receptorbiological-physics
“Non-endogenous agonists of the serotonin receptor exhibit a shared IET spectral peak intensity that scales with their known agonist potencies.”
paper / hneven / Mar 4
pHEX generalizes the HEX model by incorporating probabilistic relations in label graphs to handle uncertainty in hierarchy and exclusion (e.g., "sort of" furry). The pHEX graph converts to an Ising model, enabling off-the-shelf inference unlike HEX's custom methods. Experiments demonstrate pHEX outperforms HEX on large-scale ImageNet hierarchy-based visual object classification tasks.
probabilistic-hexising-modelslabel-relationshierarchical-classificationcrf-dnnimage-classificationmachine-learning
“pHEX models soft probabilistic relations between labels, such as an antelope being 'sort of' furry unlike a grizzly bear.”
paper / hneven / Feb 20
Researchers demonstrate that multiqubit tunneling in a programmable quantum annealer facilitates escape from false minima, unlike classical thermal hopping. They develop a non-perturbative theory predicting many-body dissipative tunneling rates under realistic noise. A 16-qubit primitive shows quantum evolution reaching the global minimum where classical paths fail, validated experimentally up to 200 qubits.
quantum-tunnelingquantum-annealingmultiqubit-tunnelingquantum-optimizationdwavearxiv-paperquantum-computing
“Multiqubit tunneling plays a computational role in a programmable quantum annealer”
paper / hneven / Dec 22
Invariant subspace methods, computed systematically with the Lanczos algorithm, reduce the effective Hilbert space for continuous-time quantum walks without prior symmetry knowledge. This yields optimal spatial search on non-regular graphs like complete graphs with broken links, complete bipartite graphs, and stars, disproving needs for regularity or high connectivity. The approach simplifies quantum transport efficiency calculations, boosts efficiencies by link removal in symmetric graphs, and bounds qubit transfer fidelity in XY spin networks.
quantum-walksdimensionality-reductionspatial-searchquantum-transportlanczos-algorithmnon-regular-graphsquantum-algorithms
“Lanczos algorithm systematically computes invariant subspaces for quantum walks without needing specific symmetry knowledge”
paper / hneven / Nov 14
First experimental demonstration on D-Wave Two annealer shows multiqubit quantum tunneling enables escape from false minima in non-convex optimization landscapes, where classical thermal hopping fails. NIBA Quantum Master Equation model captures dissipative tunneling under device noise, predicting suppression of collective quantum environment effects in critical phase and facilitation later. For 16-qubit primitive, 8-qubit tunneling inverts temperature dependence of success probability versus classical paths; scales to 200-qubit problems with embedded primitives.
quantum-tunnelingquantum-annealingd-wavequantum-optimizationmultiqubit-tunnelingquantum-master-equation
“Quantum tunneling enables finding global minimum while classical paths trap in false minimum for frustrated qubit cluster primitive.”
paper / hneven / Jun 17
Quantum annealing requires mapping problems to low-degree PUBO problems due to hardware constraints. The paper introduces non-convex polynomial loss functions based on the margin for training binary classifiers on quantum annealers. These losses are robust to label noise, outperform convex methods, and compile to regularized risk expressions evaluable in constant time relative to training examples.
quantum-annealingbinary-classificationnon-convex-losspolynomial-optimizationmachine-learningquantum-algorithms
“Quantum annealing objective functions are constrained to low-degree PUBO due to experimental hardware limitations.”
paper / hneven / May 5
Proposes q-loss, a non-convex training objective for binary classification tailored to adiabatic quantum optimization, formulated as quadratic unconstrained binary optimization (QUBO) with low bit-depth binary parameter expansions to match quantum hardware constraints. Validated using classical heuristic solvers on popular datasets, showing superior test error robustness compared to existing losses under increasing label noise. Non-convexity of q-loss drives the robustness gains.
quantum-physicsadiabatic-optimizationrobust-classificationbinary-classificationq-lossqubolabel-noise
“q-loss is formulated as quadratic unconstrained binary optimization (QUBO) for compatibility with adiabatic quantum optimization hardware.”