One-Hot QUBO (OHQ) Explained

• One-Hot QUBO (OHQ) is a discrete encoding technique that replaces a k-valued variable with k binary variables subject to an exactly-one constraint enforced by a quadratic penalty.
• It preserves the quadratic structure, allowing straightforward mapping to QUBO/Ising models for use in varied optimizers, including simulated annealing and quantum annealing.
• OHQ design involves trade-offs such as penalty tuning, clique density, and embedding challenges, prompting alternative approaches like sparse decompositions and domain-wall encodings.

Searching arXiv for recent and foundational papers on one-hot QUBO to ground the article. arXiv search query: "one-hot QUBO encoding penalty landscape domain-wall reduced alphabet cyclic peptide" One-Hot QUBO (OHQ) is a class of quadratic unconstrained binary optimization formulations in which each discrete or categorical decision is represented by a register of binary variables subject to an exactly-one condition. In its canonical form, a $k$-valued variable is replaced by $k$ binaries $x_{i,j}\in\{0,1\}$ with $\sum_{j=1}^k x_{i,j}=1$, and the constraint is enforced by a quadratic penalty of the form $\bigl(\sum_j x_{i,j}-1\bigr)^2$. The resulting model remains quadratic, admits direct QUBO and Ising representations, and has become a standard interface between constrained discrete models and classical or quantum-compatible binary optimizers (Zaborniak et al., 2023, Glover et al., 2018). Across the literature, OHQ appears not as a single application-specific method but as a reusable encoding pattern for assignments, permutations, integer variables, reduced alphabets, modular graph invariants, and structured design problems (Zhou et al., 22 Jun 2026, Nakano et al., 2023).

1. Core definition and algebraic structure

In the one-hot construction, every discrete variable $d_i$ that can assume one of $k$ distinct states is replaced by a one-hot register of $k$ binary variables $x_{i,j}$, with the exactly-one constraint

$\sum_{j=1}^k x_{i,j}=1.$

A valid solution is therefore a binary vector in which each register has exactly one active bit; any other assignment is invalid and must be raised in energy by a penalty term (Zaborniak et al., 2023).

The standard OHQ penalty for one register is

$k$0

or, for multiple groups,

$k$1

Using $k$2 for binary variables, the non-constant part expands into negative linear terms and positive intra-group quadratic couplings. Glover, Kochenberger and Du write the penalty for group $k$3 as

$k$4

up to an additive constant, so the QUBO matrix acquires $k$5 on diagonals and $k$6 on off-diagonals inside each one-hot clique (Glover et al., 2018). The reduced-alphabet cyclic-peptide formulation gives the same structure in block form,

$k$7

with a flattened bit-vector representation $k$8 and the standard Ising substitution $k$9 (Zhou et al., 22 Jun 2026).

A central property of OHQ is that the mapping from discrete variables to binary variables is already quadratic. In the DQM-to-QUBO treatment, the encoded cost takes the form

$x_{i,j}\in\{0,1\}$0

so no additional auxiliary variables are needed to quadratize higher-order terms introduced by the encoding itself (Zaborniak et al., 2023). This is one reason OHQ is routinely used as a generic reduction from discrete constrained models to annealer-compatible binary objectives.

2. Penalty semantics and energy-landscape effects

The one-hot penalty is not merely a syntactic device; it changes the topology of the optimization landscape. Valid solutions satisfy the exactly-one condition in every register and therefore incur zero one-hot penalty. Invalid solutions have strictly positive penalty, so their energy rises linearly with the penalty strength $x_{i,j}\in\{0,1\}$1 (Zaborniak et al., 2023). This separates the feasible and infeasible manifolds, but only if $x_{i,j}\in\{0,1\}$2 is chosen appropriately.

Several papers emphasize that penalty selection is non-trivial. If $x_{i,j}\in\{0,1\}$3 is too small, invalid assignments may lie below the best valid configurations, so the global QUBO optimum is not a valid solution (Zaborniak et al., 2023, Doucet et al., 7 Jan 2026). If $x_{i,j}\in\{0,1\}$4 is too large, the penalty can dominate the problem Hamiltonian, compress the effective dynamic range of the true objective, and degrade optimization. In the Job Shop Scheduling study, weak $x_{i,j}\in\{0,1\}$5 generated low-energy infeasible manifolds, while overly strong penalties suppressed the effective problem energy scale and increased irreversibility on quantum annealers (Doucet et al., 7 Jan 2026).

The landscape analysis in “Discrete quadratic model QUBO solution landscapes” formalizes this trade-off via thresholds such as $x_{i,j}\in\{0,1\}$6, above which the true optimal valid configuration becomes the global minimum, and additional thresholds governing whether invalid solutions remain local minima and whether valid solutions retain local-minimum status (Zaborniak et al., 2023). The paper further notes that classical heuristics may face larger barriers as $x_{i,j}\in\{0,1\}$7 grows, while quantum-annealing methods may experience reduced effective sampling temperature together with greater control-error sensitivity.

Practical guidance across the literature is consistent but context-dependent. A common rule of thumb in the cyclic-peptide formulation is to set the hard penalty larger than the sum of the magnitudes of the other weights, written as $x_{i,j}\in\{0,1\}$8, to guarantee feasibility (Zhou et al., 22 Jun 2026). For qubit allocation, the penalty was set equal to the largest gate-cost coefficient and increased only if sampled solutions violated the one-hot constraints; no circuit required more than three times the original maximum coefficient (Dury et al., 2020). In robot design, the one-hot penalties were chosen as twice the number of choices in each group, with explicit values $x_{i,j}\in\{0,1\}$9, $\sum_{j=1}^k x_{i,j}=1$0, $\sum_{j=1}^k x_{i,j}=1$1, $\sum_{j=1}^k x_{i,j}=1$2, and $\sum_{j=1}^k x_{i,j}=1$3, because all design rewards were normalized into $\sum_{j=1}^k x_{i,j}=1$4 and the authors wanted the penalty to exceed the maximum reward obtainable from violating the group constraint (Kang et al., 15 May 2026).

A common misconception is that larger penalties are always safer. The available results do not support that view. The thermodynamic study on quantum annealers identifies a “practical sweet spot” immediately above the feasibility transition, where one-hot is enforced without overwhelming the objective Hamiltonian (Doucet et al., 7 Jan 2026). This suggests that OHQ penalty tuning is part of model design rather than an afterthought.

3. Representational cost, clique structure, and alternative encodings

The principal structural cost of OHQ is density. A one-hot group of size $\sum_{j=1}^k x_{i,j}=1$5 induces a $\sum_{j=1}^k x_{i,j}=1$6-clique of pairwise penalties, which is benign for small groups but costly for large categorical domains or permutation kernels. In the conventional $\sum_{j=1}^k x_{i,j}=1$7 permutation encoding, enforcing exactly one $\sum_{j=1}^k x_{i,j}=1$8 in each row and column yields $\sum_{j=1}^k x_{i,j}=1$9 binary variables and $\bigl(\sum_j x_{i,j}-1\bigr)^2$0 nonzero quadratic terms, with maximum absolute Ising bias $\bigl(\sum_j x_{i,j}-1\bigr)^2$1 (Nakano et al., 2023). The same paper presents this scaling as a primary motivation for alternative encodings such as dual-matrix domain-wall.

Reduced alphabets are one response to this cost. In cyclic-peptide sequence design, starting from 20 natural amino acids at each of $\bigl(\sum_j x_{i,j}-1\bigr)^2$2 positions would require $\bigl(\sum_j x_{i,j}-1\bigr)^2$3 binary variables in naïve one-hot; replacing them by a reduced alphabet of size $\bigl(\sum_j x_{i,j}-1\bigr)^2$4 cuts the count to $\bigl(\sum_j x_{i,j}-1\bigr)^2$5 (Zhou et al., 22 Jun 2026). The default representation there uses a resource-aware eight-class alphabet motivated by MJ interaction-profile clustering, intended to balance coarse interaction-pattern preservation with encoding cost (Zhou et al., 22 Jun 2026). The trade-off is explicit: smaller $\bigl(\sum_j x_{i,j}-1\bigr)^2$6 reduces variables and can accelerate QUBO solves, but increases information loss and downstream decoding burden.

Compact encodings offer a more aggressive reduction. The cyclic-peptide paper notes that a compact binary code would use only $\bigl(\sum_j x_{i,j}-1\bigr)^2$7 bits per position, but arbitrary class-dependent scores then introduce higher-order terms requiring quadratization or auxiliary bits (Zhou et al., 22 Jun 2026). This is a standard OHQ trade-off: one-hot is variable-expensive but algebraically clean; compact encodings are variable-efficient but often structurally more complex.

A second response is sparsification by auxiliary variables. “Sparse QUBO Formulation for Efficient Embedding via Network-Based Decomposition of Equality and Inequality Constraints” replaces the dense one-hot clique on $\bigl(\sum_j x_{i,j}-1\bigr)^2$8 variables by a chain construction with $\bigl(\sum_j x_{i,j}-1\bigr)^2$9 auxiliary binaries and only $d_i$0 quadratic couplings, reducing edge count from $d_i$1 to $d_i$2 for the one-hot constraint (Suda et al., 26 Jan 2026). The paper reports that this sparse formulation leads to lower physical-qubit usage, shorter average chain lengths, lower chain break rates, and higher feasible solution rates on D-Wave hardware (Suda et al., 26 Jan 2026).

A third response is to abandon one-hot for other discrete encodings when the application permits it. The DQM landscape study compares one-hot and domain-wall encodings (Zaborniak et al., 2023), while the permutation study argues that dual-matrix domain-wall reduces the quadratic-term count and the maximum absolute coefficient values from $d_i$3 and $d_i$4 to $d_i$5 and $d_i$6, respectively (Nakano et al., 2023). A plausible implication is that OHQ is often preferred for conceptual simplicity and exact quadraticity, whereas alternatives become attractive when embedding density or coefficient range dominates runtime or hardware feasibility.

4. Solver mappings and hardware-specific considerations

Because OHQ produces quadratic objectives, the same logical model can be mapped to multiple solver families. The cyclic-peptide formulation states that classical simulated annealing, tabu search, and QA/QAOA samplers can operate on the same $d_i$7 or $d_i$8 representation without changing the model (Zhou et al., 22 Jun 2026). This solver-agnostic property is one of OHQ’s main practical advantages.

Solver-agnosticism, however, is not hardware-agnosticism. Classical annealers and heuristic solvers can often handle large sparse QUBOs and tolerate large penalty ratios (Zhou et al., 22 Jun 2026). Quantum annealers require minor embedding and are constrained by qubit connectivity and coupling-precision limits; penalty strengths must fit within the device’s coupling-strength range, and dense one-hot cliques increase chain lengths and embedding overhead (Zhou et al., 22 Jun 2026, Glover et al., 2018). The sparse one-hot decomposition paper addresses precisely this issue by reducing logical density before embedding (Suda et al., 26 Jan 2026).

Gate-model algorithms expose a different set of OHQ issues. In “Two-Step QAOA,” the one-hot constraint is first handled as a soft penalty

$d_i$9

using a standard Pauli-$k$0 mixer, and then converted into a hard Hamming-weight-preserving condition in a second stage via an $k$1-mixer

$k$2

which preserves Hamming weight $k$3 (Minato, 2024). This decomposition turns the one-hot condition from a soft energetic preference into a subspace restriction during the second optimization stage.

Thermodynamic analyses further complicate the picture. In the Job Shop study, reverse-annealing experiments on a D-Wave Advantage processor showed nearly vertical crossovers in mean energy change, entropy-production bounds, work and heat bounds, and efficiency at the same $k$4 boundary where feasibility was lost (Doucet et al., 7 Jan 2026). The paper’s conclusion is that QUBO penalties act as “thermodynamic control knobs.” This does not alter the formal correctness of OHQ, but it does mean that identical logical constraints can induce materially different physical annealing behavior depending on penalty scaling.

5. Domain-specific formulations and application patterns

OHQ has been instantiated across a wide variety of domains, usually by combining the one-hot block with application-specific quadratic costs and additional penalties.

In cyclic-peptide design, positions are encoded by binary residue-class assignment variables $k$5, and the total Hamiltonian is a weighted sum of one-hot validity, cyclization constraints, optional target-compatibility terms, motif rules, composition rules, and coarse developability proxies. By modifying the relevant constraint terms, the same framework can represent head-to-tail, disulfide-bridged, stapled, and bicyclic peptide designs (Zhou et al., 22 Jun 2026). The model is explicitly described as an early-stage search-space reduction and prioritization layer that outputs low-energy residue-class sequences rather than final molecular candidates (Zhou et al., 22 Jun 2026).

In general discrete quadratic models, OHQ serves as a direct reduction from multi-valued variables to binary optimization, preserving the original quadratic interaction structure in the encoded cost and adding only one-hot penalties (Zaborniak et al., 2023). In integer optimization on quantum annealers, the same approach underlies subproblem partitioning schemes. The 2019 partitioning paper distinguishes multivalued partitions, which retain one-hot penalties and therefore include infeasible states, from binary partitions, which eliminate the one-hot penalty inside subproblems so that all subproblem assignments are feasible (Okada et al., 2019).

In black-box optimization, FMQA applies one-hot encoding to integer or discretized continuous variables, then constructs a QUBO from a factorization-machine surrogate. The final objective is

$k$6

and the 2026 study argues that uniform random initial sampling can leave many one-hot bits never activated, so the corresponding FM parameters receive no direct gradient updates. To address this, the paper introduces LHS-FMQA and Sobol’-FMQA with complete marginal bit coverage and reports numerically higher mean final cruising speeds than the baseline FMQA on 17- and 32-variable human-powered-aircraft wing-shape benchmarks, with the advantage more pronounced on the 32-variable problem (Hayashi et al., 6 May 2026).

Graph-theoretic constructions also use OHQ. Lotfi et al. encode nowhere-zero $k$7-flows by assigning one-hot variables $k$8 to nonzero flow residues on edges and auxiliary quotient bits $k$9 to modular conservation equations. Their Hamiltonian

$k$0

has ground-state energy zero if and only if the graph has a nowhere-zero $k$1-flow, and the zero-energy degeneracy equals the flow polynomial $k$2 (Lotfi et al., 27 Jun 2026). The paper proves that correctness is independent of orientation, root choice, and any positive $k$3 (Lotfi et al., 27 Jun 2026).

Other OHQ applications include qubit allocation, where $k$4 indicates that logical qubit $k$5 is assigned to physical qubit $k$6, with row and column one-hot penalties enforcing bijectivity (Dury et al., 2020); robot design optimization, where one-hot groups select a single design alternative per subsystem (Kang et al., 15 May 2026); and balanced $k$7-way hypergraph partitioning, where $k$8 indicates assignment of vertex $k$9 to part $x_{i,j}$0, combined with balance penalties and quadratic cut objectives (Li et al., 4 May 2026).

Quantum error correction provides a specialized but structurally revealing example. In SOME, the decoding task is formulated as an OHQ over syndrome-pair match variables $x_{i,j}$1, where the one-hot constraint enforces exactly one match per flipped syndrome (Guo et al., 31 Jul 2025). Valid solutions correspond to symmetric one-hot structures interpreted as self-inverse permutation matrices, and the paper reports up to a 99.9x reduction in variable count together with microsecond-scale decoding on a single-threaded commodity CPU (Guo et al., 31 Jul 2025).

6. Interpretation, limitations, and recurring misconceptions

OHQ is sometimes treated as synonymous with “the” QUBO encoding of a constrained discrete problem. The literature does not support that simplification. One-hot is a widely used encoding, but it coexists with domain-wall, log-encoding, half-hot reductions, sparse decompositions, and problem-specific structured parametrizations (Zaborniak et al., 2023, Nakano et al., 2023). This suggests that OHQ is best understood as a baseline or reference encoding rather than a universally dominant one.

A second misconception is that one-hot validity alone solves the modeling problem. In practice, the one-hot block only enforces discrete-choice semantics; the substantive problem still lies in the application-specific Hamiltonian. The cyclic-peptide framework, for example, combines one-hot validity with cyclization, target compatibility, motif, composition, and developability terms, and explicitly states that its outputs are not final molecular candidates (Zhou et al., 22 Jun 2026). Similarly, in hypergraph partitioning, one-hot assignment does not enforce balanced partitions unless a separate balance term is added (Li et al., 4 May 2026).

A third misconception is that exact quadraticity guarantees efficient execution on annealers. Dense one-hot cliques can be the dominant bottleneck in embedding, precision, and chain robustness (Glover et al., 2018, Suda et al., 26 Jan 2026). The sparse-chain one-hot construction and dual-matrix domain-wall permutation encoding were developed precisely because the conventional OHQ kernel becomes impractical at scale on present hardware (Suda et al., 26 Jan 2026, Nakano et al., 2023).

Finally, penalty robustness depends on context. In the nowhere-zero-flow construction, any positive $x_{i,j}$2 preserve correctness because every violated squared term is a nonnegative integer and zero energy is achieved exactly when all constraints hold (Lotfi et al., 27 Jun 2026). In contrast, optimization-focused OHQ models generally require nontrivial penalty calibration to prevent low-energy infeasible states or over-penalized landscapes (Zaborniak et al., 2023, Doucet et al., 7 Jan 2026). The difference is not contradictory: it reflects whether the objective is feasibility detection with exact zero-energy certificates or cost optimization under competing soft and hard terms.

Taken together, the literature presents OHQ as a modular quadratic encoding primitive with broad applicability, exact expressibility for exactly-one constraints, and direct compatibility with QUBO and Ising solvers. Its strengths are algebraic simplicity, transparent semantics, and solver portability; its liabilities are dense clique structure, penalty-sensitive landscapes, and hardware-dependent embedding costs. The current research trajectory therefore combines continued use of OHQ in domain formulations with increasingly explicit attention to reduced alphabets, sparse decompositions, alternative encodings, and thermodynamics-aware penalty design (Zhou et al., 22 Jun 2026, Suda et al., 26 Jan 2026, Doucet et al., 7 Jan 2026).