Sign Constraints in Mathematical Models

• Sign constraints are conditions that restrict or assign the sign (positive, negative, or unconstrained) of mathematical quantities across diverse fields.
• They are applied in statistical learning, geometric optimization, and cosmological models to enforce directional choices and reduce admissible structure.
• Their implementation enhances regularity in elliptic PDEs, governs phase diagrams in many-body theory, and guides discrete sign assignments in topological and genetic contexts.

Sign constraints are conditions that fix, restrict, compare, or exploit the sign of mathematical quantities. In the literature surveyed here, they appear as coefficient restrictions in statistical learning, symmetry-induced bounds in fermionic many-body theory, one-sided regularity hypotheses in elliptic PDE, discrete sign assignments in orientifolds, sign-sensitive interaction terms in cosmology, and sign-dependent mutation effects in fitness landscapes. Despite this diversity, the recurring technical role of a sign constraint is to reduce admissible structure: it can shrink a feasible set, force a correlation hierarchy, forbid a class of phases, or impose topological compatibility conditions (Zhang et al., 10 Mar 2025, Kato et al., 2017, Schiffer, 17 Jan 2026, Hekmati et al., 2019).

1. Formal types of sign constraints

The term has no single universal definition. Instead, distinct fields use related but non-equivalent formalizations.

In convex learning problems, sign constraints are usually coordinatewise inequalities on a parameter vector. One standard formulation partitions indices into $\mathcal{I}_+$, $\mathcal{I}_-$, and $\mathcal{I}_0$, or equivalently uses a sign vector $c \in \{0,\pm1\}^d$ and the feasible set

$$\mathcal{S} := \{w \in \mathbb{R}^d \mid c \odot w \geq 0_d\},$$

so that selected coefficients are required to be non-negative, non-positive, or unconstrained (Kato et al., 2017). In sign-constrained linear SVMs, the same idea is written as $$\mathcal{S} := \{ w \in \mathbb{R}^d \mid \sigma \odot w \geq 0\}$$ after absorbing non-positive constraints by negating features in advance (Tajima et al., 2021).

In geometric optimization, sign restrictions can be combined with nonlinear manifold constraints. The sign-constrained Stiefel manifold is

$$S_{n,r}^{P,N} := S_{n,r} \cap \mathbb{R}^{n \times r}_{P,N},$$

where $$S_{n,r} = \{X \in \mathbb{R}^{n \times r}: X^T X = I_r\}$$ and selected columns are required to be entrywise non-negative or non-positive (Chen et al., 2022).

In topological and string-theoretic settings, a sign constraint may be a discrete assignment rather than an inequality. For orientifolds $(M,\tau)$, a sign choice assigns a $\pm 1$ sign to each O-plane, i.e. to each connected component of the fixed set $\mathcal{I}_-$0, and is represented by an element of $\mathcal{I}_-$1 (Hekmati et al., 2019).

In cosmology, the relevant object may be the sign of an interaction term $\mathcal{I}_-$2 between dark matter and dark energy. Here the question is not merely whether $\mathcal{I}_-$3 is positive or negative, but whether its sign can reverse dynamically during cosmic evolution (Wei, 2010, Pan et al., 2019).

Domain	Representative form	Primary effect
Learning	$\mathcal{I}_-$4	Restricts coefficient signs
Many-body theory	$\mathcal{I}_-$5, $\mathcal{I}_-$6	Bounds correlation functions
Stiefel geometry	$\mathcal{I}_-$7 plus sign pattern	Couples orthogonality and sign
Orientifolds	$\mathcal{I}_-$8	Assigns O-plane signs
Cosmology	sign-changing $\mathcal{I}_-$9 or $\mathcal{I}_0$0	Reverses energy-transfer direction

This suggests that “sign constraint” is best understood as a family of one-sided or discrete admissibility conditions rather than a single formal apparatus.

2. Anti-unitary symmetry and correlation bounds in sign-problem-free DQMC

In determinantal quantum Monte Carlo, a powerful route to sign-problem freedom is the presence of a Kramers anti-unitary symmetry acting only on fermionic degrees of freedom. For a quadratic Hamiltonian in a Majorana basis $\mathcal{I}_0$1, the symmetry acts as $\mathcal{I}_0$2, with $\mathcal{I}_0$3 and $\mathcal{I}_0$4 real antisymmetric. Under this structure, the fermion determinant is real and non-negative for every Hubbard–Stratonovich field configuration, which guarantees the absence of the sign problem; the paper explicitly cites the earlier criterion of Wu et al. in this context (Zhang et al., 10 Mar 2025).

The central result is that the same anti-unitary symmetry also constrains local correlations. For every such symmetry, there exists a unique local operator

$\mathcal{I}_0$5

whose Euclidean space-time correlation function upper bounds those of a large class of other local operators, including superconducting order parameters (Zhang et al., 10 Mar 2025). The symmetry implies the reality relation

$\mathcal{I}_0$6

and Wick-theoretic/Fierz decomposition expresses connected bilinear correlators as a direct part plus an exchange part. The exchange term is maximal when the bilinear matrix is proportional to $\mathcal{I}_0$7, so the anti-unitary-symmetry-matched channel dominates the accessible correlation hierarchy (Zhang et al., 10 Mar 2025).

The physical consequence is a restriction on phase diagrams of sign-problem-free models. For intra-unit-cell bilinears, the only possible tendency to long-range order is in the channel selected by the anti-unitary symmetry. For superconducting order parameters, the direct term vanishes by charge symmetry, so the maximal possible order is determined by the exchange term and therefore occurs in the superconducting channel $\mathcal{I}_0$8 (Zhang et al., 10 Mar 2025). The same analysis also explains why generic finite-momentum orders are difficult to realize in sign-problem-free models: the maximal correlation function is peaked at zero momentum unless the anti-unitary symmetry is projective with respect to translation. If several anti-unitary symmetries coexist, their associated operators must become simultaneously critical at any phase transition detected by intra-cell bilinear observables; otherwise, only order involving bosonic fields can become critical (Zhang et al., 10 Mar 2025).

A common misconception is that the absence of a sign problem is merely an algorithmic convenience. In these models it is also a structural restriction on what phases can dominate, because the symmetry that enforces non-negative Monte Carlo weights simultaneously imposes correlation inequalities.

3. Coefficient sign restrictions in learning, inverse problems, and hierarchical regression

In machine learning, sign constraints are a direct mechanism for incorporating domain knowledge about monotonic or directional feature effects. In sign-constrained regularized loss minimization, the feasible set is $\mathcal{I}_0$9, and two generic algorithms are obtained by inserting a sign-correction step into Pegasos and SDCA. The resulting methods, SC-Pega and SC-SDCA, preserve the convergence rate of the unconstrained algorithms; for SC-Pega,

$c \in \{0,\pm1\}^d$0

and for SC-SDCA the required iteration counts match the unconstrained case up to constants (Kato et al., 2017). Empirically, on an E. coli prediction task with 177 samples, 10 used for training and 10,000 random splits, SC-SVM achieved mean PRBEP $c \in \{0,\pm1\}^d$1 and ROC $c \in \{0,\pm1\}^d$2, versus PRBEP $c \in \{0,\pm1\}^d$3 and ROC $c \in \{0,\pm1\}^d$4 for a classic SVM (Kato et al., 2017).

For linear SVMs specifically, sign-constrained learning can be formulated in primal and dual form with the same sign-feasible region. A projected-gradient method and a Frank–Wolfe method both have $c \in \{0,\pm1\}^d$5 cost per iteration and sublinear convergence. The Frank–Wolfe analysis provides an explicit iteration bound: if $c \in \{0,\pm1\}^d$6 for all $c \in \{0,\pm1\}^d$7, then

$c \in \{0,\pm1\}^d$8

guarantees an $c \in \{0,\pm1\}^d$9-accurate dual solution, and the duality gap supplies a stopping criterion (Tajima et al., 2021). In sequence classification with SVM-pairwise features constructed from similarities to training examples, sign constraints improved average ROC scores on all 12 protein function prediction tasks reported in the paper (Tajima et al., 2021).

Related ideas appear in inverse problems and hierarchical regression. In multiple penalized least squares for EEG source imaging, the modified Newton–Raphson algorithm does not straightforwardly handle sign constraints, but the active-set AMNR extension enforces them during optimization by admitting variables only with allowed sign and deactivating any coefficient that tries to cross zero. This supports nonnegative versions of Smooth Garrote and Smooth LASSO and yielded sparser solutions with improved localization and reduced ghost sources in EEG inverse problems (Vega-Hernández et al., 2019). In hierarchical marketing mix models, sign restrictions are imposed directly on selected regional and global coefficients, such as $$\mathcal{S} := \{w \in \mathbb{R}^d \mid c \odot w \geq 0_d\},$$0 and $$\mathcal{S} := \{w \in \mathbb{R}^d \mid c \odot w \geq 0_d\},$$1, and estimation is carried out by constrained maximum likelihood or Hamiltonian Monte Carlo with truncated-normal structure and constraint-aware dynamics (Chen et al., 2020).

The practical significance is not that sign constraints are universally beneficial, but that they can be statistically effective when prior sign information is credible. The reported gains are therefore conditional on the validity of the imposed directional knowledge.

4. Discrete, geometric, and topological sign structure

Some of the most technical uses of sign constraints concern discrete combinatorics, manifold geometry, and topological classification rather than coefficient shrinkage.

In evolutionary genetics, sign epistasis occurs when the sign of a mutation’s effect depends on genetic background. For two loci, one circuit is

$$\mathcal{S} := \{w \in \mathbb{R}^d \mid c \odot w \geq 0_d\},$$2

and sign epistasis can also be detected from the sign of

$$\mathcal{S} := \{w \in \mathbb{R}^d \mid c \odot w \geq 0_d\},$$3

Crona generalizes this viewpoint using circuits and “rectangular perturbations,” where blocks of $$\mathcal{S} := \{w \in \mathbb{R}^d \mid c \odot w \geq 0_d\},$$4 loci are flipped at once. The number of size-$$\mathcal{S} := \{w \in \mathbb{R}^d \mid c \odot w \geq 0_d\},$$5 rectangular perturbations is

$$\mathcal{S} := \{w \in \mathbb{R}^d \mid c \odot w \geq 0_d\},$$6

and the total number is

$$\mathcal{S} := \{w \in \mathbb{R}^d \mid c \odot w \geq 0_d\},$$7

For three loci, only 96 of the $$\mathcal{S} := \{w \in \mathbb{R}^d \mid c \odot w \geq 0_d\},$$8 possible genotype rank orders are compatible with additive fitness (Crona, 2018). In the four-locus Plasmodium vivax example, the paper reports 54 and 55 size-1 perturbations in the drug and drug-free landscapes, respectively, together with size-2 counts of 39 and 21 and size-3 counts of 9 and 5 (Crona, 2018).

In manifold optimization, the sign-constrained Stiefel manifold couples orthogonality and sign admissibility. The global error bound

$$\mathcal{S} := \{w \in \mathbb{R}^d \mid c \odot w \geq 0_d\},$$9

and the corresponding local bound with coefficient $$\mathcal{S} := \{ w \in \mathbb{R}^d \mid \sigma \odot w \geq 0\}$$0 quantify the distance to feasibility by separate residuals for sign and orthogonality violations (Chen et al., 2022). For $$\mathcal{S} := \{ w \in \mathbb{R}^d \mid \sigma \odot w \geq 0\}$$1, the two square-root terms are necessary and the exponents cannot be improved except in multiplicative constants. When $$\mathcal{S} := \{ w \in \mathbb{R}^d \mid \sigma \odot w \geq 0\}$$2 or $$\mathcal{S} := \{ w \in \mathbb{R}^d \mid \sigma \odot w \geq 0\}$$3, linear bounds using $$\mathcal{S} := \{ w \in \mathbb{R}^d \mid \sigma \odot w \geq 0\}$$4-type quantities become possible (Chen et al., 2022).

In orientifold theory, sign choices on O-planes are not arbitrary combinatorial labels. They arise from invariant $$\mathcal{S} := \{ w \in \mathbb{R}^d \mid \sigma \odot w \geq 0\}$$5-gerbes classified by equivariant sheaf cohomology groups $$\mathcal{S} := \{ w \in \mathbb{R}^d \mid \sigma \odot w \geq 0\}$$6 and are related by coboundary maps satisfying

$$\mathcal{S} := \{ w \in \mathbb{R}^d \mid \sigma \odot w \geq 0\}$$7

The allowed geometric sign choices form a filtration

$$\mathcal{S} := \{ w \in \mathbb{R}^d \mid \sigma \odot w \geq 0\}$$8

which need not equal the full sign-assignment group at low $$\mathcal{S} := \{ w \in \mathbb{R}^d \mid \sigma \odot w \geq 0\}$$9 (Hekmati et al., 2019). Explicitly, for $$S_{n,r}^{P,N} := S_{n,r} \cap \mathbb{R}^{n \times r}_{P,N},$$0 with $$S_{n,r}^{P,N} := S_{n,r} \cap \mathbb{R}^{n \times r}_{P,N},$$1, only the diagonal subgroup of $$S_{n,r}^{P,N} := S_{n,r} \cap \mathbb{R}^{n \times r}_{P,N},$$2 is geometric; for $$S_{n,r}^{P,N} := S_{n,r} \cap \mathbb{R}^{n \times r}_{P,N},$$3, only sign choices with an even number of $$S_{n,r}^{P,N} := S_{n,r} \cap \mathbb{R}^{n \times r}_{P,N},$$4’s are allowed; and in $$S_{n,r}^{P,N} := S_{n,r} \cap \mathbb{R}^{n \times r}_{P,N},$$5 there are further “face” constraints (Hekmati et al., 2019). The same paper identifies the Gao–Hori construction as a special case of twisted $$S_{n,r}^{P,N} := S_{n,r} \cap \mathbb{R}^{n \times r}_{P,N},$$6-theory (Hekmati et al., 2019).

Across these examples, sign data serve not merely as labels but as compatibility conditions encoded by circuits, residual inequalities, or equivariant cohomology.

5. One-sided sign information and higher integrability in elliptic PDE

In elliptic regularity, sign constraints appear as one-sided assumptions on determinants or derivative components. These conditions sharpen higher-integrability conclusions that are not available under sign-indefinite hypotheses.

For quasiconcave $$S_{n,r}^{P,N} := S_{n,r} \cap \mathbb{R}^{n \times r}_{P,N},$$7-homogeneous functions $$S_{n,r}^{P,N} := S_{n,r} \cap \mathbb{R}^{n \times r}_{P,N},$$8, the paper proves that if $$S_{n,r}^{P,N} := S_{n,r} \cap \mathbb{R}^{n \times r}_{P,N},$$9 satisfies

$$S_{n,r} = \{X \in \mathbb{R}^{n \times r}: X^T X = I_r\}$$0

then

$$S_{n,r} = \{X \in \mathbb{R}^{n \times r}: X^T X = I_r\}$$1

for both $$S_{n,r} = \{X \in \mathbb{R}^{n \times r}: X^T X = I_r\}$$2 and $$S_{n,r} = \{X \in \mathbb{R}^{n \times r}: X^T X = I_r\}$$3, with different technical arguments in the two regimes (Schiffer, 17 Jan 2026). In the determinant case, if $$S_{n,r} = \{X \in \mathbb{R}^{n \times r}: X^T X = I_r\}$$4 almost everywhere, then at the critical regularity level one obtains

$$S_{n,r} = \{X \in \mathbb{R}^{n \times r}: X^T X = I_r\}$$5

The paper explicitly presents this as a version of Müller’s higher-integrability result under non-negativity of the determinant (Schiffer, 17 Jan 2026).

For very weak solutions of $$S_{n,r} = \{X \in \mathbb{R}^{n \times r}: X^T X = I_r\}$$6-Laplace-type equations, the sign asymmetry may be placed directly on partial derivatives. If $$S_{n,r} = \{X \in \mathbb{R}^{n \times r}: X^T X = I_r\}$$7 for every coordinate index $$S_{n,r} = \{X \in \mathbb{R}^{n \times r}: X^T X = I_r\}$$8, with $$S_{n,r} = \{X \in \mathbb{R}^{n \times r}: X^T X = I_r\}$$9, while $(M,\tau)$0, then the full gradient satisfies

$(M,\tau)$1

A corollary states that if $(M,\tau)$2 and $(M,\tau)$3 with $(M,\tau)$4, then $(M,\tau)$5 (Schiffer, 17 Jan 2026).

Methodologically, the key tool is Lipschitz truncation. The paper further introduces asymmetric Lipschitz truncation, which treats positive and negative parts differently and is essential when the two signs have different integrability. For the $(M,\tau)$6-Laplacian result, convexity of the domain and vanishing boundary values are assumed (Schiffer, 17 Jan 2026).

These results show that a sign constraint can function as an “integrability propagator”: one-sided control on the negative part of a quantity can force improved regularity of the full object.

6. Sign-changing interactions and sign-switching energy densities in cosmology

Cosmological uses of sign constraints focus on dynamical sign reversals rather than static one-sided inequalities. The central quantity is usually an interaction term $(M,\tau)$7 in the continuity equations of dark matter and dark energy, where the sign determines the direction of energy transfer.

One class of models inserts the deceleration parameter $(M,\tau)$8 into the interaction,

$(M,\tau)$9

and the constrained analyses in the paper specialize to $\pm 1$0 and the forms $\pm 1$1, $\pm 1$2, and $\pm 1$3. Because $\pm 1$4 changes sign when the universe transitions from deceleration to acceleration, $\pm 1$5 does so as well (Wei, 2010). Using 557 Union2 SNIa, the WMAP7 shift parameter, and the BAO distance parameter, the paper reports best fits such as $\pm 1$6 for $\pm 1$7 and transition redshifts near $\pm 1$8–$\pm 1$9 (Wei, 2010).

A second class uses density differences directly: $\mathcal{I}_-$00 In these models the sign of $\mathcal{I}_-$01 reverses when the relative magnitudes of the dark densities change. The paper finds stable perturbations, slight preference for non-zero interaction, but also that zero interaction is allowed within $\mathcal{I}_-$02 confidence level. For IDE1 with the full dataset combination, it reports

$\mathcal{I}_-$03

which reduces the $\mathcal{I}_-$04 tension to about $\mathcal{I}_-$05, while the $\mathcal{I}_-$06 tension persists (Pan et al., 2019). The same analysis states that the generalized second law is always satisfied and that the second derivative of the total entropy becomes negative at late times, implying approach to thermodynamic equilibrium (Pan et al., 2019).

A more recent development replaces sign-changing interactions by a smooth sign-switch of the dark-energy density itself. In ECDM,

$\mathcal{I}_-$07

so $\mathcal{I}_-$08 evolves from negative values at high redshift to positive values at late times. The formal equation-of-state parameter diverges at the crossing, but the perturbation problem is regularized by evolving

$\mathcal{I}_-$09

instead of $\mathcal{I}_-$10 itself (Bouhmadi-López et al., 6 Jul 2026). With Planck 2018, ACT DR6, SPT-3G, DESI DR2, Pantheon+, and SH0ES, the paper finds $\mathcal{I}_-$11 for full data combinations and reports that abrupt transitions are strongly disfavored, while the model remains compatible with precision data and alleviates the Hubble tension (Bouhmadi-López et al., 6 Jul 2026).

A common misconception is that any sign-changing dark-sector model is automatically favored once late-time tensions are considered. The observational papers do not support that stronger claim: some sign-changing scenarios remain difficult to distinguish from non-interacting cosmologies at $\mathcal{I}_-$12, and Bayesian evidence can still favor $\mathcal{I}_-$13CDM (Pan et al., 2019).