Penalty-Encoded Hybrid Interface

• The paper demonstrates that converting the exact-cardinality constraint to a quadratic penalty enables the use of D-Wave’s quantum annealing in hybrid workflows.
• Penalty encoding introduces a dense, rank-one term to the QUBO matrix, which transforms sparse correlations into a fully connected interaction pattern affecting embedding and chain lengths.
• Empirical findings indicate that combining penalty encoding with native constraints or classical post-processing significantly reduces chain-break fractions and achieves near-optimal performance.

In recent quantum-annealing portfolio-optimization work, Penalty-Encoded Hybrid Interface can be understood as an Editor's term for a modeling and solver interface in which a hard combinatorial constraint—most prominently the exact-cardinality condition $\mathbf{1}^\top z=K$—is converted into a quadratic penalty inside a BQM/QUBO and then passed to a hybrid or direct quantum-annealing workflow. In the 2026 D-Wave portfolio literature, this interface is contrasted with two alternatives: a constraint-native hybrid interface based on CQM, and a penalty-free direct-QPU pipeline in which the QPU samples only the objective and exact feasibility is restored classically afterward (Lozano, 17 May 2026, Lozano, 17 May 2026).

1. Conceptual scope

In the mean-variance-turnover study, the term has a precise operational meaning: the penalty-encoded hybrid interface is the D-Wave LeapHybridBQMSampler path applied to a BQM/QUBO in which the cardinality equality is replaced by the penalty

$A\left(\sum_i z_i-K\right)^2,$

with $A=4.0$ in the main experiments and robustness checked at $A\in\{2,4,8\}$ (Lozano, 17 May 2026). In the direct-QPU portfolio study, the same interface appears in the canonical penalty-encoded Markowitz QUBO

$$\min_{x \in \{0,1\}^n} \; -\mu^\top x + \lambda\, x^\top \Sigma\, x + A\, (1^\top x - K)^2,$$

which is the baseline formulation submitted to D-Wave hardware (Lozano, 17 May 2026).

The underlying optimization classes are closely related but not identical. One paper studies the binary, equal-weight, cardinality-constrained Markowitz selection problem, while the other studies a cardinality-constrained mean-variance portfolio model with turnover costs,

$$\min_{z \in \{0,1\}^N} \; -\boldsymbol{\mu}^\top z + \lambda \, z^\top \Sigma \, z + \boldsymbol{\tau}^\top |z - z_{t-1}| \quad \text{s.t.} \quad \mathbf{1}^\top z = K.$$

Because $z$ and $z_{t-1}$ are binary, the turnover term linearizes exactly as

$$|z_i - z_{t-1,i}| = z_i(1 - z_{t-1,i}) + (1 - z_i)z_{t-1,i},$$

so the turnover contribution shifts only diagonal coefficients and does not change the penalty-density mechanism (Lozano, 17 May 2026).

2. Penalty encoding as a graph transformation

The central algebraic fact is that the exact-$K$ penalty contributes a dense rank-one term. Expanding

$A\left(\sum_i z_i-K\right)^2,$0

and using $A\left(\sum_i z_i-K\right)^2,$1, one obtains

$A\left(\sum_i z_i-K\right)^2,$2

so the full QUBO matrix becomes

$A\left(\sum_i z_i-K\right)^2,$3

with constant offset $A\left(\sum_i z_i-K\right)^2,$4 (Lozano, 17 May 2026). In the mean-variance-turnover formulation, the same mechanism yields

$A\left(\sum_i z_i-K\right)^2,$5

where $A\left(\sum_i z_i-K\right)^2,$6 (Lozano, 17 May 2026).

Because $A\left(\sum_i z_i-K\right)^2,$7 is dense, every pair $A\left(\sum_i z_i-K\right)^2,$8 receives a coupling. The direct-QPU paper makes the consequence explicit: $A\left(\sum_i z_i-K\right)^2,$9 Accordingly, even if the original covariance graph is sparse, the penalty injects a complete-graph interaction pattern (Lozano, 17 May 2026).

This effect is especially transparent in the betting case study, where the covariance is naturally block diagonal: $A=4.0$0 and

$A=4.0$1

The paper emphasizes that betting covariance consists of disconnected $A=4.0$2 blocks, one clique per match, yet the added term $A=4.0$3 makes the logical graph complete over all $A=4.0$4 outcomes (Lozano, 17 May 2026). The hybrid audit paper generalizes this observation as density-axis collapse: covariance density $A=4.0$5 becomes “irrelevant to the encoded logical graph density when using penalty encoding” (Lozano, 17 May 2026).

3. Structural pathologies on current D-Wave workflows

On sparse hardware topologies such as Pegasus and Zephyr, dense logical graphs require minor embedding, longer chains, and larger qubit overhead. The direct-QPU study reports that this is the dominant failure mode. On live D-Wave Advantage and Advantage2 hardware, dense penalty-encoded equity QUBOs at $A=4.0$6 ($A=4.0$7) required mean chain lengths around $A=4.0$8–$A=4.0$9 and produced chain-break fractions near $A\in\{2,4,8\}$0. At $A\in\{2,4,8\}$1 ($A\in\{2,4,8\}$2), mean chain lengths rose to $A\in\{2,4,8\}$3 on Pegasus and $A\in\{2,4,8\}$4 on Zephyr, with chain-break fractions of $A\in\{2,4,8\}$5 and $A\in\{2,4,8\}$6, respectively. Betting QUBOs rise from about $A\in\{2,4,8\}$7 chain breaks at $A\in\{2,4,8\}$8 to $A\in\{2,4,8\}$9 at $$\min_{x \in \{0,1\}^n} \; -\mu^\top x + \lambda\, x^\top \Sigma\, x + A\, (1^\top x - K)^2,$$0, and across all tested scales in the standard penalized formulation the feasible sample rate is $$\min_{x \in \{0,1\}^n} \; -\mu^\top x + \lambda\, x^\top \Sigma\, x + A\, (1^\top x - K)^2,$$1 (Lozano, 17 May 2026).

The hybrid audit identifies the same structural bottleneck on the direct-QPU path: for fully connected penalty-encoded graphs, chain-break fractions rise monotonically with size, reaching

$$\min_{x \in \{0,1\}^n} \; -\mu^\top x + \lambda\, x^\top \Sigma\, x + A\, (1^\top x - K)^2,$$2

while embedding overhead reaches

$$\min_{x \in \{0,1\}^n} \; -\mu^\top x + \lambda\, x^\top \Sigma\, x + A\, (1^\top x - K)^2,$$3

with statistics “nearly identical across density families” (Lozano, 17 May 2026).

A natural attempted remedy is topology-aware sparsification, but the direct-QPU paper shows that this introduces constraint dilution. Since the penalty itself lives in the off-diagonals, removing edges also removes penalty mass: $$\min_{x \in \{0,1\}^n} \; -\mu^\top x + \lambda\, x^\top \Sigma\, x + A\, (1^\top x - K)^2,$$4 Sparsified QUBOs can reduce chain lengths to nearly unit and drive chain-break fractions below the resolvable floor $$\min_{x \in \{0,1\}^n} \; -\mu^\top x + \lambda\, x^\top \Sigma\, x + A\, (1^\top x - K)^2,$$5, yet raw samples remain infeasible and tend to become near-all-ones vectors with $$\min_{x \in \{0,1\}^n} \; -\mu^\top x + \lambda\, x^\top \Sigma\, x + A\, (1^\top x - K)^2,$$6 (Lozano, 17 May 2026).

4. Interface variants

Three solver-facing designs are central in this literature.

Interface	Solver/model	Constraint treatment
Penalty-encoded hybrid	LeapHybridBQMSampler / BQM	$$\min_{x \in \{0,1\}^n} \; -\mu^\top x + \lambda\, x^\top \Sigma\, x + A\, (1^\top x - K)^2,$$7 encoded as $$\min_{x \in \{0,1\}^n} \; -\mu^\top x + \lambda\, x^\top \Sigma\, x + A\, (1^\top x - K)^2,$$8
Constraint-native hybrid	LeapHybridCQMSampler / CQM	$$\min_{x \in \{0,1\}^n} \; -\mu^\top x + \lambda\, x^\top \Sigma\, x + A\, (1^\top x - K)^2,$$9 kept as an explicit native constraint
Penalty-free direct QPU	Objective-only QUBO	No penalty term; exact-$$\min_{z \in \{0,1\}^N} \; -\boldsymbol{\mu}^\top z + \lambda \, z^\top \Sigma \, z + \boldsymbol{\tau}^\top |z - z_{t-1}| \quad \text{s.t.} \quad \mathbf{1}^\top z = K.$$0 restored classically

For the constraint-native model, the paper writes

$$\min_{z \in \{0,1\}^N} \; -\boldsymbol{\mu}^\top z + \lambda \, z^\top \Sigma \, z + \boldsymbol{\tau}^\top |z - z_{t-1}| \quad \text{s.t.} \quad \mathbf{1}^\top z = K.$$1

For the penalty-free direct-QPU redesign, the annealer-facing matrix is

$$\min_{z \in \{0,1\}^N} \; -\boldsymbol{\mu}^\top z + \lambda \, z^\top \Sigma \, z + \boldsymbol{\tau}^\top |z - z_{t-1}| \quad \text{s.t.} \quad \mathbf{1}^\top z = K.$$2

and the QPU samples

$$\min_{z \in \{0,1\}^N} \; -\boldsymbol{\mu}^\top z + \lambda \, z^\top \Sigma \, z + \boldsymbol{\tau}^\top |z - z_{t-1}| \quad \text{s.t.} \quad \mathbf{1}^\top z = K.$$3

with no penalty term $$\min_{z \in \{0,1\}^N} \; -\boldsymbol{\mu}^\top z + \lambda \, z^\top \Sigma \, z + \boldsymbol{\tau}^\top |z - z_{t-1}| \quad \text{s.t.} \quad \mathbf{1}^\top z = K.$$4 included (Lozano, 17 May 2026, Lozano, 17 May 2026).

The conceptual distinction is therefore not merely solver choice. It is a distinction between three interface philosophies: encoding hard constraints into the quadratic model, preserving them natively in the constrained model, or removing them from the annealer-facing model and restoring them afterward.

5. Classical feasibility restoration and the hybrid split

The penalty-free direct-QPU paper makes the hybrid split explicit. Feasibility is enforced by a greedy post-processing rule on the exact-cardinality set

$$\min_{z \in \{0,1\}^N} \; -\boldsymbol{\mu}^\top z + \lambda \, z^\top \Sigma \, z + \boldsymbol{\tau}^\top |z - z_{t-1}| \quad \text{s.t.} \quad \mathbf{1}^\top z = K.$$5

Operationally, the rule is:

• if $$\min_{z \in \{0,1\}^N} \; -\boldsymbol{\mu}^\top z + \lambda \, z^\top \Sigma \, z + \boldsymbol{\tau}^\top |z - z_{t-1}| \quad \text{s.t.} \quad \mathbf{1}^\top z = K.$$6, iteratively flip a selected variable from $$\min_{z \in \{0,1\}^N} \; -\boldsymbol{\mu}^\top z + \lambda \, z^\top \Sigma \, z + \boldsymbol{\tau}^\top |z - z_{t-1}| \quad \text{s.t.} \quad \mathbf{1}^\top z = K.$$7 to $$\min_{z \in \{0,1\}^N} \; -\boldsymbol{\mu}^\top z + \lambda \, z^\top \Sigma \, z + \boldsymbol{\tau}^\top |z - z_{t-1}| \quad \text{s.t.} \quad \mathbf{1}^\top z = K.$$8 with the smallest marginal contribution to the objective;
• if $$\min_{z \in \{0,1\}^N} \; -\boldsymbol{\mu}^\top z + \lambda \, z^\top \Sigma \, z + \boldsymbol{\tau}^\top |z - z_{t-1}| \quad \text{s.t.} \quad \mathbf{1}^\top z = K.$$9, iteratively flip an unselected variable from $z$0 to $z$1 with the largest marginal contribution (Lozano, 17 May 2026).

This same paper argues that favorable sparsify-and-project results can be explained by the projector alone. In betting, after settlement-graph sparsification, post-processing yields zero regret relative to a greedy reference at $z$2, but an ablation shows that all-ones projection on the same settlement-graph QUBO gives the same result: zero regret and perfect Jaccard overlap with the greedy solution. In the authors’ interpretation, “the classical backward-elimination projector alone explains the gains” (Lozano, 17 May 2026).

The penalty-free interface changes that attribution. There, all-ones projection regrets are $z$3 at $z$4 and $z$5 at $z$6, so the final outcome is no longer trivially attributable to the projector alone (Lozano, 17 May 2026). A plausible implication is that the term hybrid becomes more substantive only after the penalty artifact is removed.

6. Empirical performance and audit results

The empirical contrast between interfaces is sharp. In the direct-QPU study, mean chain-break fractions on D-Wave Advantage and Advantage2 drop from roughly $z$7–$z$8 in the penalized pipeline to at most $z$9 in the penalty-free one, across equities up to $z_{t-1}$0 and betting up to $z_{t-1}$1. For equities, post-processed regret relative to the greedy classical reference is $z_{t-1}$2 at $z_{t-1}$3, $z_{t-1}$4 at $z_{t-1}$5, $z_{t-1}$6 at $z_{t-1}$7, and $z_{t-1}$8 at $z_{t-1}$9, i.e. at most about $$|z_i - z_{t-1,i}| = z_i(1 - z_{t-1,i}) + (1 - z_i)z_{t-1,i},$$0. For betting, the penalty-free pipeline matches greedy at $$|z_i - z_{t-1,i}| = z_i(1 - z_{t-1,i}) + (1 - z_i)z_{t-1,i},$$1, then returns lower-energy feasible portfolios than the greedy heuristic at $$|z_i - z_{t-1,i}| = z_i(1 - z_{t-1,i}) + (1 - z_i)z_{t-1,i},$$2 and $$|z_i - z_{t-1,i}| = z_i(1 - z_{t-1,i}) + (1 - z_i)z_{t-1,i},$$3, with reported regrets $$|z_i - z_{t-1,i}| = z_i(1 - z_{t-1,i}) + (1 - z_i)z_{t-1,i},$$4 and $$|z_i - z_{t-1,i}| = z_i(1 - z_{t-1,i}) + (1 - z_i)z_{t-1,i},$$5; the authors state explicitly that these are energy improvements over a heuristic baseline, not proofs of global optimality (Lozano, 17 May 2026).

The hybrid audit reaches a complementary conclusion on D-Wave’s cloud solvers. On cardinality-constrained mean-variance-turnover instances spanning $$|z_i - z_{t-1,i}| = z_i(1 - z_{t-1,i}) + (1 - z_i)z_{t-1,i},$$6 to $$|z_i - z_{t-1,i}| = z_i(1 - z_{t-1,i}) + (1 - z_i)z_{t-1,i},$$7, the constraint-native service matches Gurobi’s proven optimum on all 54 instances where Gurobi proves optimality, whereas the penalty-encoded BQM gap grows with $$|z_i - z_{t-1,i}| = z_i(1 - z_{t-1,i}) + (1 - z_i)z_{t-1,i},$$8. CQM returns identical solutions at every tested wall-clock budget from 5 to 300 seconds and across 10 repeated calls, and mean QPU access time is only $$|z_i - z_{t-1,i}| = z_i(1 - z_{t-1,i}) + (1 - z_i)z_{t-1,i},$$9 seconds out of a 5-second wall-clock budget, roughly $K$0 of the run. The paper therefore reads the successful hybrid path as “CQM’s classical pipeline, with a small QPU contribution,” not as a quantum-sampling win (Lozano, 17 May 2026).

Taken together, these results support a common diagnosis: the binding constraint at currently accessible scales is the penalty encoding of cardinality, not a generic “sparse hardware versus dense finance” mismatch.

7. Broader design principle and related usages

The direct-QPU paper states the broader lesson in architectural terms: if a hard constraint admits a cheap projection or repair operator, it may be better to omit the penalty from the annealer-facing problem entirely and let the QPU explore only the objective landscape (Lozano, 17 May 2026). The hybrid audit reaches an adjacent conclusion for cloud solvers: practitioners should prefer the constraint-native CQM interface first, benchmark against a strong MIQP baseline, and report run_time, charge_time, qpu_access_time, and the QPU fraction of wall-clock when claiming hybrid performance (Lozano, 17 May 2026).

Related but nonidentical constructions appear outside quantum finance. In optimization with multiple Hankel-rank constraints, a hybrid penalty method combines penalty subproblems with a post-processing pseudo-projection stage (Liu et al., 2019). In CPS falsification, input admissibility can be encoded as a higher-priority lexicographic penalty or as the implication $K$1, so that interface validity is enforced within the search objective (Zhang et al., 2020). In numerical PDEs, several interface methods encode transmission or boundary behavior through penalty terms rather than explicit interface unknowns, including unfitted interface penalty DG–FE formulations (Han et al., 2023), interior-penalty methods for fracture interface models without additional interface degrees of freedom (Liu et al., 2024), and sharp-interface active penalty methods that reconstruct boundary data inside a volumetric penalty operator (Shirokoff et al., 2013).

These analogies do not establish a single universal definition. They do, however, point to a consistent pattern: a penalty-encoded hybrid interface is a solver architecture in which a difficult interface condition is pushed into an auxiliary penalty representation, and performance depends critically on whether that encoding preserves—or destroys—the structure on which the underlying algorithm or hardware relies.