Simulated Bifurcation Algorithm

• Simulated bifurcation algorithm is a heuristic that represents discrete optimization problems using continuous dynamics and bifurcation of state variables.
• It leverages nonlinear dynamics to efficiently solve Ising, QUBO, and higher-order pseudo-Boolean objectives across various applications.
• The method is adaptable to hardware-friendly implementations such as FPGA and ASIC, with enhancements from replica coupling and adaptive control.

Simulated bifurcation (SB) is a quantum-inspired, classical-mechanics-based heuristic for combinatorial optimization in which discrete spins or binary variables are represented by continuous dynamical variables whose signs encode the final solution. In the literature, SB is primarily used for Ising and QUBO optimization, but the same dynamical template has been extended to higher-order pseudo-Boolean objectives, replica-coupled quantum-inspired variants, adaptive population-based controllers, and several domain-specific reformulations, including MIMO detection, community detection, traffic control, finance, and configuration interaction in quantum chemistry (Zhang et al., 2022, Kanao et al., 2022, Pawłowski et al., 1 Apr 2026, Aiga et al., 6 Mar 2026).

1. Optimization setting and conceptual basis

The standard optimization target of SB is the Ising model or an equivalent QUBO. A common form used in the recent literature is

$$H(\mathbf{s}) = -\frac{1}{2}\sum_{i=1}^N \sum_{j=1}^N J_{ij}s_is_j - B\sum_{i=1}^N h_i s_i,$$

with $s_i\in\{-1,+1\}$, symmetric couplings $J$, and, when present, a bias vector $h$ (Orlando et al., 14 Oct 2025). Equivalent binary formulations use $s_i = 2q_i - 1$, so graph partitioning, Max-Cut, community detection, knapsack, MIMO maximum-likelihood detection, and several resource-allocation problems can be cast into the same algebraic form (Li et al., 2024, Orlando et al., 14 Oct 2025).

The defining idea of SB is to avoid direct combinatorial search over $\{-1,+1\}^N$. Instead, each spin is associated with a continuous position-like variable and a momentum-like variable, and the optimizer simulates a nonlinear Hamiltonian or Hamiltonian-inspired dynamical system. As a control parameter is increased, the symmetric state loses stability and the continuous variables bifurcate toward two sign-separated branches. The final discrete solution is then read out from the signs of the continuous coordinates (Kanao et al., 2022, Zhang et al., 2022).

This physical interpretation is consistent across the literature, but the exact algebraic form varies. Some papers emphasize Kerr-nonlinear parametric oscillators and quartic potentials, whereas others present a more implementation-oriented update rule that already incorporates clipping, discretization, or sign substitution (Bouquet et al., 2021, Kashimata et al., 2023, Qian et al., 19 Dec 2025). A practical implication is that “simulated bifurcation” denotes a family of closely related dynamical solvers rather than one immutable recurrence.

2. Canonical dynamics and numerical realization

A widely used formulation, particularly for ballistic SB and discrete SB, evolves continuous variables $x_i$ and $y_i$ according to

$\dot{x}_i=a_0y_i,$

$\dot{y}_i=-\left[a_0-a(t)\right]x_i+c_0f_i,$

where $s_i\in\{-1,+1\}$0 is a time-dependent control parameter, $s_i\in\{-1,+1\}$1 and $s_i\in\{-1,+1\}$2 are constants, and the interaction force $s_i\in\{-1,+1\}$3 distinguishes the algorithmic variant. In ballistic SB,

$s_i\in\{-1,+1\}$4

whereas in discrete SB,

$s_i\in\{-1,+1\}$5

With increasing $s_i\in\{-1,+1\}$6 from zero to $s_i\in\{-1,+1\}$7, bifurcations to $s_i\in\{-1,+1\}$8 occur, and the signs $s_i\in\{-1,+1\}$9 yield a solution to the Ising problem (Kanao et al., 2022).

A standard implementation detail is confinement by perfectly inelastic walls at $J$0. After each time step, if $J$1, the state is reset as

$J$2

This saturation rule is central in digital realizations because it enforces binary-like amplitudes and eliminates momentum after wall collisions (Kanao et al., 2022). Numerical integration is commonly performed by the symplectic Euler method, and several papers adopt a linearly increasing schedule for $J$3 together with $J$4 and a coupling normalization proportional to $J$5 or to an empirical coupling scale derived from $J$6 (Kanao et al., 2022, Zhang et al., 2022).

Other implementations keep more of the original nonlinear oscillator structure. In one discrete-time form used for hardware-aware SB, the state update is

$J$7

with $J$8 continuously increased from $J$9 to $h$0 (Qian et al., 19 Dec 2025). Multi-chip adiabatic implementations similarly retain cubic nonlinearities and a gradually ramped control parameter, but reorganize matrix-vector multiplication and time evolution for streaming hardware execution (Kashimata et al., 2023). These differences are implementation-level rather than conceptual: in each case, binary decisions emerge from continuous nonlinear dynamics driven through a bifurcation regime.

3. Variants and algorithmic extensions

The core variants are adiabatic SB, ballistic SB, and discrete SB. Ballistic SB keeps the continuous amplitudes in the interaction term, whereas discrete SB replaces them by signs. This substitution reduces multiplier cost and is especially attractive for FPGAs and compute-in-memory hardware, because products with $h$1 reduce to sign selection or add-subtract logic (Orlando et al., 14 Oct 2025). In some problem classes, dSB trades a small loss in dynamical smoothness for higher hardware efficiency and strong empirical accuracy (Kanao et al., 2022, Orlando et al., 14 Oct 2025).

Heated simulated bifurcation augments the momentum dynamics with a deterministic heating term,

$h$2

motivated by a simplification of the Nosé–Hoover method. The purpose is to sustain momentum, counteract freezing caused by inelastic walls, and improve escape from local minima. On dense Sherrington–Kirkpatrick instances with up to 2000 spin variables, the heated ballistic and heated discrete variants improve step-to-solution over their unheated counterparts, with the strongest gain reported for heated ballistic SB (Kanao et al., 2022).

A separate extension targets genuinely higher-order cost functions. For third-order objectives,

$h$3

the higher-order SB literature replaces the pairwise force by the gradient of the higher-order polynomial and operates directly on the original variable set rather than quadratizing with ancillary spins (Kanao et al., 2022). On 3-regular 3-XORSAT, direct third-order ballistic SB outperforms second-order SB applied after quadratization and also outperforms direct simulated annealing on the third-order cost function (Kanao et al., 2022). Traffic-signal optimization on real Beijing road networks extends the same theme to multiple-spin interactions via high-dimensional expansion, with the reported empirical time complexity reduced to $h$4 (Zhao et al., 18 Feb 2025).

Recent quantum-inspired extensions introduce replica interactions reminiscent of Suzuki–Trotter quantum annealing. Simulated Bifurcation Quantum Annealing (SBQA) adds a nearest-neighbor inter-replica coupling

$h$5

and reports systematic improvement over standard SBM in sparse and rugged regimes while remaining competitive elsewhere (Pawłowski et al., 1 Apr 2026). Another recent direction, Adaptive Enhanced Quantum-inspired Simulated Bifurcation, monitors population diversity, freeze rate, flip rate, and improvement rate, then adapts step size, guidance, restart, and stopping policies through a closed-loop perception–decision–execution scheme (Liu et al., 23 Jun 2026).

4. Reformulations and applications

The most detailed communications application is MIMO detection. For an uplink model

$h$6

the maximum-likelihood detector solves

$h$7

For QPSK and 16-QAM, the complex system is embedded into a real-valued form, encoded as spins, and rewritten as an Ising Hamiltonian. In QPSK, SB substantially outperforms MMSE over a range of SNRs, especially around $h$8–$h$9 dB, and the performance gap becomes larger as the MIMO dimension grows; however, plain SB exhibits a high-SNR error floor, motivating MMSE-regularized SB and hybrid schemes (Zhang et al., 2022). A later deep-unfolded formulation makes the SB update differentiable, trains $s_i = 2q_i - 1$0 parameters, and reports a gain of about $s_i = 2q_i - 1$1 dB at BER $s_i = 2q_i - 1$2 for $s_i = 2q_i - 1$3 MIMO and about $s_i = 2q_i - 1$4 dB at BER $s_i = 2q_i - 1$5 for $s_i = 2q_i - 1$6 MIMO, relative to LM-SB and TPG (Takabe, 2023).

Community detection by modularity maximization offers a different QUBO route. The graph partition is encoded by one-hot binary assignment variables, the modularity objective is converted into a penalized QUBO with slack variables for nonempty-community constraints, and dSB is applied after a QUBO-to-Ising transformation. On Zachary’s Karate Club, SB achieves modularity $s_i = 2q_i - 1$7, matching Fujitsu’s Digital Annealer and exceeding the values reported for IBM and D-Wave; on the IEEE 33-bus system, SB reaches $s_i = 2q_i - 1$8, again matching Fujitsu and exceeding Gurobi and D-Wave in the reported comparison (Li et al., 2024).

Finance and trading trajectory optimization were among the early applied SB domains. Integer Markowitz portfolio optimization and multi-period trading problems are converted into Ising form by binary encoding of holdings and, in the trajectory setting, by adding time-coupling terms for rebalancing costs. One one-bit selection study on 441 S&P 500 assets reports a 5-second run time and a selected subset of about 120 assets, while another trajectory study reports preliminary results for portfolios of up to 1000 assets and solutions in the order of seconds on a single CPU (Bouquet et al., 2021, Steinhauer et al., 2020).

Further applications move beyond standard pairwise Ising usage. Real-world traffic signal control on four-way, eight-phase intersections uses higher-order SB with high-dimensional expansion and reports superiority over simulated annealing in both efficiency and effectiveness (Zhao et al., 18 Feb 2025). Quantum chemistry adapts the underlying SB philosophy more radically: Simulated Bifurcation-based Configuration Interaction replaces binary spins by continuous CI coefficients, uses the Rayleigh quotient as the potential term, and updates one or two target states through symplectic Euler dynamics with variationally selected parameters rather than adiabatic scheduling (Aiga et al., 6 Mar 2026).

5. Hardware realizations and scalability

Because SB updates all variables simultaneously and its dominant kernel is a dense or structured matrix-vector product, it maps naturally onto FPGA, ASIC, and compute-in-memory architectures. Recent work has emphasized co-design: simplification of the dynamical update, quantization of state variables, and hardware-native realization of the interaction term $s_i = 2q_i - 1$9 or $\{-1,+1\}^N$0 (Kashimata et al., 2023, Qian et al., 19 Dec 2025).

Work	Platform	Reported result
(Kashimata et al., 2023)	8-FPGA streaming cluster	32,768-spin problem with 97.9% pipeline efficiency
(Dee et al., 14 Apr 2025)	180 nm mixed-signal SRAM CIM	above 93% of the ground state solution in 0.6us with 10.8mW average power
(Qian et al., 19 Dec 2025)	32x256 FeFET CiM chip	up to 175.9x speedup over a GPU-based simulated bifurcation implementation
(Orlando et al., 14 Oct 2025)	AMD Kria KV260 FPGA	256-variable proof-of-concept with 254 ns per step

The multi-chip streaming architecture is particularly important for dense fully connected problems. A dual-ring design overlaps inter-chip communication with matrix-vector computation, hides communication time except for the unavoidable latency term, and demonstrates linear strong scaling up to the vicinity of the ideal communication limit. The measured 8-FPGA cluster reaches 77,775 GMAC/s on a 32,768-spin problem, and a validated performance model projects a 79-FPGA, 100,000-spin design comparable to a state-of-the-art 100,000-spin optical Ising machine (Kashimata et al., 2023).

Mixed-signal and compute-in-memory realizations take different algorithmic shortcuts. One SRAM-CIM SB solver implements the synchronous update

$\{-1,+1\}^N$1

and explicitly exploits analog-domain noise as part of the optimization process (Dee et al., 14 Apr 2025). FeFET compute-in-memory work removes the cubic term, ternarizes the state to $\{-1,+1\}^N$2, and compensates the resulting approximation loss with an attention-inspired initialization that reduces iterations by up to 80% (Qian et al., 19 Dec 2025). FPGA-oriented dSB work arrives at a similar conclusion from the digital side: replacing $\{-1,+1\}^N$3 by $\{-1,+1\}^N$4 removes multipliers and permits configurable parallelism across rows, columns, and replicated matrix-vector/time-evolution blocks (Orlando et al., 14 Oct 2025).

6. Limitations, misconceptions, and current directions

A recurring misconception is that SB is a quantum algorithm. The literature is explicit that SB is quantum-inspired, but implemented classically, typically on GPUs, FPGAs, or mixed-signal hardware (Li et al., 2024, Orlando et al., 14 Oct 2025). Another misconception is that SB denotes a single canonical update equation. In practice, the literature includes adiabatic, ballistic, discrete, heated, higher-order, replica-coupled, and hardware-simplified forms, often with different state variables, nonlinearities, and clipping rules (Kanao et al., 2022, Qian et al., 19 Dec 2025).

SB is also a heuristic rather than an exact optimizer. Several papers explicitly report local-minimum trapping, parameter sensitivity, or instance-dependent failure modes. In MIMO detection, plain SB develops a high-SNR error floor because residual optimization failures dominate once channel noise becomes very small (Zhang et al., 2022). In higher-order optimization, direct third-order SB outperforms quadratization on sparse cubic instances, but the literature remains benchmark-specific and does not yet provide a broad theory for dense higher-order tensors (Kanao et al., 2022). In sparse and rugged regimes, standard SBM can underperform, which is the main motivation behind SBQA’s replica coupling and its reported systematic gains (Pawłowski et al., 1 Apr 2026).

Open problems therefore remain at several levels. One is algorithmic tuning: step size, schedule shape, clipping policy, regularization strength, and heating or replica-coupling parameters still matter, even when papers report robust operating regions (Kanao et al., 2022, Pawłowski et al., 1 Apr 2026). Another is hybridization: MIMO studies show that MMSE-regularized SB and deep-unfolded SB can materially outperform plain SB, suggesting that model-based seeds and learned schedules are not incidental but structurally important (Zhang et al., 2022, Takabe, 2023). A third is adaptive control. Population-state methods that monitor diversity, saturation, sign flips, and recent progress indicate that fixed schedules are often too rigid for heterogeneous instance classes (Liu et al., 23 Jun 2026).

Taken together, the literature presents simulated bifurcation not as a single fixed solver but as a broad optimization paradigm: represent discrete decisions by continuous nonlinear dynamics, drive the system through bifurcation, and exploit parallel interaction structure in software or specialized hardware. Its strongest empirical profile is on problems that admit dense or structured quadratic formulations and benefit from massively parallel collective updates, while its main unresolved issues concern principled tuning, local-minimum avoidance, and generalization across sparse, rugged, or highly constrained problem families.