Support-Projected Metric: Applications & Theory

• Support-projected metric is an intrinsic construction that restricts evaluations to specific support structures like support vectors, active sets, or graph supports.
• It underpins methods across machine learning, optimization, and quantum information by preserving properties such as positive semidefiniteness and connectivity.
• The metric is computed via projection techniques that ensure stability and improved empirical performance by focusing on the geometry relevant to the problem.

Support-projected metric denotes a class of constructions in which a metric, metric-like tensor, or metric evaluation rule is made intrinsic by restricting, parameterizing, or projecting it onto an application-specific support structure. In the most direct usage, this support can be the cone generated by positive semidefinite pair moments in relation alignment metric learning, the subspace generated by support vectors in large-margin methods, the observed support of a weighted graph, the active numerical range of a quantum Fisher information tensor, or the statistically significant support of a distribution estimated from samples. A plausible unifying interpretation is that the metric is not taken on the ambient representation alone, but on the effective directions, ranges, or sets that actually carry the geometry relevant to the problem (Gu, 2020, Do et al., 2013, Segarra et al., 2019, Kan et al., 2020, Pegoraro et al., 2021, Kim et al., 2023, Cho et al., 18 Sep 2025).

1. Conceptual pattern across the literature

Several of the cited works state that they do not explicitly define the phrase “support-projected metric,” but they nevertheless provide precise constructions that justify it as an operational description. In these constructions, a support object is first identified: support coefficients in support vector regression, support vectors in SVM dual representations, the edge support of a network, a support graph for graph-metric ball systems, a Krylov or active-set support in constrained optimization, a KDE superlevel support in generative-model evaluation, a constrained quantile cone in Wasserstein statistics, or the active range of an information tensor in quantum geometry (Gu, 2020, Do et al., 2013, Browne et al., 31 Mar 2026, Segarra et al., 2019, Kan et al., 2020, Goyens et al., 25 Mar 2026, Pegoraro et al., 2021, Cho et al., 18 Sep 2025).

A second recurring feature is that projection does not always mean Euclidean orthogonal projection. It may mean restriction to a cone, projection onto a positive semidefinite cone, metric projection onto a box under a variable metric, contraction of Voronoi cells to a support graph, or replacement of an unconstrained estimate by its closest element in a convex cone of admissible objects. What is preserved depends on the domain: positive semidefiniteness, connectedness of hit sets, triangle inequalities, bound feasibility, monotonicity of quantiles, or nondegenerate information directions. This suggests that “support-projected metric” is best understood as a structural principle rather than a single formalism.

2. Positive semidefinite support-projected metrics in relation alignment metric learning

In positive semidefinite support vector regression metric learning, the support-projected construction arises inside the relation alignment metric learning (RAML) framework, where relations between samples are modeled as continuous quantities rather than binary similar/dissimilar labels. The paper uses

$$g(\mathbf{y}_i,\mathbf{y}_j)=\|\mathbf{y}_i-\mathbf{y}_j\|_1$$

and aligns this decision-space relation with a Mahalanobis distance in feature space,

$$d_M^2(\mathbf{x}_i,\mathbf{x}_j)=(\mathbf{x}_i-\mathbf{x}_j)^\top M(\mathbf{x}_i-\mathbf{x}_j), \qquad M\succeq 0,$$

through

$$f(\mathbf{x}_i,\mathbf{x}_j;M,b)=\langle M,T_{ij}\rangle+b, \quad T_{ij}=(\mathbf{x}_i-\mathbf{x}_j)(\mathbf{x}_i-\mathbf{x}_j)^\top .$$

With SVR optimization, the recovered metric has the form

$M=\sum_{i=1}^n (a_i-a_i^*)\,T_i.$

If some coefficients are negative, the linear combination of PSD rank-1 matrices can become indefinite, so the original RAML-SVR may require post hoc spectral projection

$M=U\,\mathrm{diag}(\max(\lambda,0))\,U^\top,$

which the paper notes can hurt discriminative power (Gu, 2020).

The two proposed remedies enforce positive semidefiniteness during optimization. RAML-PCSVR imposes

$a_i\ge a_i^*,\quad \forall i,$

so that all coefficients $(a_i-a_i^*)\ge 0$, and therefore $M=\sum_i (a_i-a_i^*)T_i\succeq 0$. RAML-NCSVR instead parameterizes

$M=\sum_{i=1}^n \mu_i T_i,\quad \mu_i\ge 0,$

which guarantees $M\succeq 0$ by construction. The paper interprets both as PSD-enforcement strategies that amount to projecting or restricting the learned metric to the PSD cone via constraints or parameterization rather than by spectral truncation after the fact.

In this setting, the phrase “support-projected metric” is especially literal: RAML-PCSVR constrains support vector differences to be nonnegative, while RAML-NCSVR optimizes directly in the cone generated by the PSD atoms $$d_M^2(\mathbf{x}_i,\mathbf{x}_j)=(\mathbf{x}_i-\mathbf{x}_j)^\top M(\mathbf{x}_i-\mathbf{x}_j), \qquad M\succeq 0,$$0. Empirically, the PSD methods consistently match or outperform the original RAML-SVR on single-label, multi-label, and label-distribution tasks. Reported examples include MnistDat, where RAML-PCSVR is approximately $$d_M^2(\mathbf{x}_i,\mathbf{x}_j)=(\mathbf{x}_i-\mathbf{x}_j)^\top M(\mathbf{x}_i-\mathbf{x}_j), \qquad M\succeq 0,$$1 and RAML-NCSVR approximately $$d_M^2(\mathbf{x}_i,\mathbf{x}_j)=(\mathbf{x}_i-\mathbf{x}_j)^\top M(\mathbf{x}_i-\mathbf{x}_j), \qquad M\succeq 0,$$2 versus RAML-SVR approximately $$d_M^2(\mathbf{x}_i,\mathbf{x}_j)=(\mathbf{x}_i-\mathbf{x}_j)^\top M(\mathbf{x}_i-\mathbf{x}_j), \qquad M\succeq 0,$$3, and TDT2_20, where RAML-PCSVR is approximately $$d_M^2(\mathbf{x}_i,\mathbf{x}_j)=(\mathbf{x}_i-\mathbf{x}_j)^\top M(\mathbf{x}_i-\mathbf{x}_j), \qquad M\succeq 0,$$4 and RAML-NCSVR approximately $$d_M^2(\mathbf{x}_i,\mathbf{x}_j)=(\mathbf{x}_i-\mathbf{x}_j)^\top M(\mathbf{x}_i-\mathbf{x}_j), \qquad M\succeq 0,$$5 versus RAML-SVR approximately $$d_M^2(\mathbf{x}_i,\mathbf{x}_j)=(\mathbf{x}_i-\mathbf{x}_j)^\top M(\mathbf{x}_i-\mathbf{x}_j), \qquad M\succeq 0,$$6. The paper also reports improved metric validity and downstream KNN-style performance, while noting slower optimization than plain SVR, large pairwise Gram matrices, and the absence of formal convergence or generalization bounds (Gu, 2020).

3. Support-vector subspaces and margin-band geometry in SVM and MKL

In the metric-learning interpretation of SVM, the paper “A Metric-learning based framework for Support Vector Machines and Multiple Kernel Learning” does not explicitly use the term “Support-Projected Metric,” but it states that its formulations justify that operational interpretation. In the linear case, a diagonal transform $$d_M^2(\mathbf{x}_i,\mathbf{x}_j)=(\mathbf{x}_i-\mathbf{x}_j)^\top M(\mathbf{x}_i-\mathbf{x}_j), \qquad M\succeq 0,$$7 induces $$d_M^2(\mathbf{x}_i,\mathbf{x}_j)=(\mathbf{x}_i-\mathbf{x}_j)^\top M(\mathbf{x}_i-\mathbf{x}_j), \qquad M\succeq 0,$$8, and in kernel space the weight vector has the dual form

$$d_M^2(\mathbf{x}_i,\mathbf{x}_j)=(\mathbf{x}_i-\mathbf{x}_j)^\top M(\mathbf{x}_i-\mathbf{x}_j), \qquad M\succeq 0,$$9

For the band-constrained formulation, the dual becomes

$$f(\mathbf{x}_i,\mathbf{x}_j;M,b)=\langle M,T_{ij}\rangle+b, \quad T_{ij}=(\mathbf{x}_i-\mathbf{x}_j)(\mathbf{x}_i-\mathbf{x}_j)^\top .$$0

Only points with nonzero $$f(\mathbf{x}_i,\mathbf{x}_j;M,b)=\langle M,T_{ij}\rangle+b, \quad T_{ij}=(\mathbf{x}_i-\mathbf{x}_j)(\mathbf{x}_i-\mathbf{x}_j)^\top .$$1 or $$f(\mathbf{x}_i,\mathbf{x}_j;M,b)=\langle M,T_{ij}\rangle+b, \quad T_{ij}=(\mathbf{x}_i-\mathbf{x}_j)(\mathbf{x}_i-\mathbf{x}_j)^\top .$$2 contribute, so the effective geometry is governed by support vectors. The paper therefore defines, as an inferred notion,

$$f(\mathbf{x}_i,\mathbf{x}_j;M,b)=\langle M,T_{ij}\rangle+b, \quad T_{ij}=(\mathbf{x}_i-\mathbf{x}_j)(\mathbf{x}_i-\mathbf{x}_j)^\top .$$3

and states that the learned transform acts primarily within $$f(\mathbf{x}_i,\mathbf{x}_j;M,b)=\langle M,T_{ij}\rangle+b, \quad T_{ij}=(\mathbf{x}_i-\mathbf{x}_j)(\mathbf{x}_i-\mathbf{x}_j)^\top .$$4 (Do et al., 2013).

This interpretation is strengthened by the paper’s within-class distance terms. Besides the standard margin, it introduces a bandwidth-based within-class measure

$$f(\mathbf{x}_i,\mathbf{x}_j;M,b)=\langle M,T_{ij}\rangle+b, \quad T_{ij}=(\mathbf{x}_i-\mathbf{x}_j)(\mathbf{x}_i-\mathbf{x}_j)^\top .$$5

and the convex band-constrained SVM

$$f(\mathbf{x}_i,\mathbf{x}_j;M,b)=\langle M,T_{ij}\rangle+b, \quad T_{ij}=(\mathbf{x}_i-\mathbf{x}_j)(\mathbf{x}_i-\mathbf{x}_j)^\top .$$6

The second constraint confines outputs to a narrow band around the margin and introduces $$f(\mathbf{x}_i,\mathbf{x}_j;M,b)=\langle M,T_{ij}\rangle+b, \quad T_{ij}=(\mathbf{x}_i-\mathbf{x}_j)(\mathbf{x}_i-\mathbf{x}_j)^\top .$$7 multipliers, so the geometry is shaped by support information from both sides of the margin band. The same logic extends to MKL, where

$$f(\mathbf{x}_i,\mathbf{x}_j;M,b)=\langle M,T_{ij}\rangle+b, \quad T_{ij}=(\mathbf{x}_i-\mathbf{x}_j)(\mathbf{x}_i-\mathbf{x}_j)^\top .$$8

and the induced transform is block diagonal,

$$f(\mathbf{x}_i,\mathbf{x}_j;M,b)=\langle M,T_{ij}\rangle+b, \quad T_{ij}=(\mathbf{x}_i-\mathbf{x}_j)(\mathbf{x}_i-\mathbf{x}_j)^\top .$$9

The significance of this work is not the invention of a new named metric object, but the clarification that SVM and MKL can be interpreted as Mahalanobis metric learning problems whose effective directions are selected by active constraints. In that sense, the support projection is onto the support-vector subspace and the margin-aligned band. The empirical results reported in the paper show that the bandwidth-controlled SVM frequently achieves the lowest error and that the MKL extension yields modest improvements over SimpleMKL, especially when the chosen kernel is poorly matched to the data (Do et al., 2013).

4. Graph, network, and metric-cone projections

In graph and network settings, support projection takes three distinct but related forms. First, in “Metric Representations of Networks: A Uniqueness Result,” the support of an observed network is the edge set $M=\sum_{i=1}^n (a_i-a_i^*)\,T_i.$0 together with the weights $M=\sum_{i=1}^n (a_i-a_i^*)\,T_i.$1 defined only on $M=\sum_{i=1}^n (a_i-a_i^*)\,T_i.$2. The canonical metric projection completes the partial dissimilarity to a full metric by shortest paths,

$M=\sum_{i=1}^n (a_i-a_i^*)\,T_i.$3

and the paper proves that this is the unique admissible projection satisfying the Axiom of Projection and the Axiom of Transformation. In that paper’s terminology, the support-projected metric is the shortest-path metric or metric closure on the support (Segarra et al., 2019).

Second, in planar graph metrics, “Single-Criteria Metric $M=\sum_{i=1}^n (a_i-a_i^*)\,T_i.$4-Dominating Set Problem via Minor-Preserving Support” constructs a support graph $M=\sum_{i=1}^n (a_i-a_i^*)\,T_i.$5 for a family of balls $M=\sum_{i=1}^n (a_i-a_i^*)\,T_i.$6 by equalizing radii, forming graph-Voronoi cells

$M=\sum_{i=1}^n (a_i-a_i^*)\,T_i.$7

and then contracting each cell to its center. The main structural theorem states that the resulting dual support $M=\sum_{i=1}^n (a_i-a_i^*)\,T_i.$8 is a contraction-minor of $M=\sum_{i=1}^n (a_i-a_i^*)\,T_i.$9, and for planar $M=U\,\mathrm{diag}(\max(\lambda,0))\,U^\top,$0, $M=U\,\mathrm{diag}(\max(\lambda,0))\,U^\top,$1 is planar. The paper explicitly notes that distances are not preserved in $M=U\,\mathrm{diag}(\max(\lambda,0))\,U^\top,$2; what is preserved is the connectivity of the hit sets $M=U\,\mathrm{diag}(\max(\lambda,0))\,U^\top,$3. This support-projected representation yields the shallow-cell complexity bound $M=U\,\mathrm{diag}(\max(\lambda,0))\,U^\top,$4 and underpins the first single-criteria polynomial-time $M=U\,\mathrm{diag}(\max(\lambda,0))\,U^\top,$5-approximation for the vertex-weighted metric $M=U\,\mathrm{diag}(\max(\lambda,0))\,U^\top,$6-dominating set on planar graphs (Browne et al., 31 Mar 2026).

Third, “A Projection Method for Metric-Constrained Optimization” studies projection of an observed dissimilarity matrix $M=U\,\mathrm{diag}(\max(\lambda,0))\,U^\top,$7 onto the metric cone $M=U\,\mathrm{diag}(\max(\lambda,0))\,U^\top,$8 intersected with support constraints $M=U\,\mathrm{diag}(\max(\lambda,0))\,U^\top,$9. Its weighted Euclidean form is

$a_i\ge a_i^*,\quad \forall i,$0

while the $a_i\ge a_i^*,\quad \forall i,$1 form is

$a_i\ge a_i^*,\quad \forall i,$2

Support constraints may fix entries or impose bounds. The paper develops a Dykstra/Hildreth projection algorithm with per-constraint $a_i\ge a_i^*,\quad \forall i,$3 updates and reports instances with approximately $a_i\ge a_i^*,\quad \forall i,$4 variables and up to $a_i\ge a_i^*,\quad \forall i,$5 constraints (Veldt et al., 2018).

Taken together, these works show that in graph settings the support-projected metric may mean metric completion from observed support, projection onto semimetric feasibility under structural support constraints, or compression of a ball system onto a support graph that preserves the incidence geometry needed for approximation algorithms.

5. Support-aware metrics in numerical optimization

In large-scale bound-constrained optimization, “PNKH-B: A Projected Newton-Krylov Method for Large-Scale Bound-Constrained Optimization” defines the metric used in the projection subproblem so that it is consistent with a low-rank Hessian approximation on a Krylov support. With

$a_i\ge a_i^*,\quad \forall i,$6

the projection metric is

$a_i\ge a_i^*,\quad \forall i,$7

where $a_i\ge a_i^*,\quad \forall i,$8. For $a_i\ge a_i^*,\quad \forall i,$9, $(a_i-a_i^*)\ge 0$0, while for $(a_i-a_i^*)\ge 0$1, $(a_i-a_i^*)\ge 0$2. The projected point is the metric projection

$(a_i-a_i^*)\ge 0$3

The KKT system identifies active and free variables, and variants of the method explicitly use active-set estimates to restrict the Hessian approximation to the free subspace. The paper proves global convergence to a stationary point under standard assumptions and emphasizes that the consistent Hessian metric leads to fast early convergence (Kan et al., 2020).

A closely related but more geometric development appears in “The Riemannian Landing Method: From projected gradient flows to SQP.” There, support-projected metrics are built from a smooth field of oblique projectors and tangent and normal metric restrictions. With a projector $(a_i-a_i^*)\ge 0$4 onto $(a_i-a_i^*)\ge 0$5 and complement $(a_i-a_i^*)\ge 0$6, the ambient metric is

$(a_i-a_i^*)\ge 0$7

This yields explicit tangent and normal steps,

$(a_i-a_i^*)\ge 0$8

and the paper shows that, under suitable metric choices, landing methods recover projected and null-space gradient flows, SQP, and a form of the augmented Lagrangian method. It further states that quadratic convergence arises only when the normal support and tangent restriction reproduce the SQP geometry (Goyens et al., 25 Mar 2026).

In both works, support projection is a way of making the metric intrinsic to the active computational degrees of freedom. The support is not a data set but a dynamically selected subspace, and the projection determines both stability and convergence behavior.

6. Distributional support estimation and Wasserstein projection

For evaluation of generative models, “TopP&R: Robust Support Estimation Approach for Evaluating Fidelity and Diversity in Generative Models” makes support estimation itself the central object. With real features $(a_i-a_i^*)\ge 0$9 and generated features $M=\sum_i (a_i-a_i^*)T_i\succeq 0$0, the KDE-based estimated supports are

$M=\sum_i (a_i-a_i^*)T_i\succeq 0$1

where $M=\sum_i (a_i-a_i^*)T_i\succeq 0$2 and $M=\sum_i (a_i-a_i^*)T_i\succeq 0$3 are bootstrap confidence bands. Topological Precision and Recall are then defined by membership tests against these support estimates: $M=\sum_i (a_i-a_i^*)T_i\succeq 0$4

$M=\sum_i (a_i-a_i^*)T_i\succeq 0$5

The paper proves statistical consistency under weak assumptions and adversarial noise conditions and reports robustness to outliers and non-IID perturbations (Kim et al., 2023).

A different distributional use appears in “Projected Statistical Methods for Distributional Data on the Real Line with the Wasserstein Metric.” For $M=\sum_i (a_i-a_i^*)T_i\succeq 0$6, the one-dimensional Wasserstein geometry is represented by quantiles,

$M=\sum_i (a_i-a_i^*)T_i\succeq 0$7

and unconstrained estimates are projected back to the cone of nondecreasing quantile functions

$M=\sum_i (a_i-a_i^*)T_i\succeq 0$8

through

$M=\sum_i (a_i-a_i^*)T_i\succeq 0$9

If supports are known to lie in $M=\sum_{i=1}^n \mu_i T_i,\quad \mu_i\ge 0,$0, the projection incorporates $M=\sum_{i=1}^n \mu_i T_i,\quad \mu_i\ge 0,$1. The paper implements this with quadratic B-splines, for which monotonicity reduces to monotonicity of the spline coefficients, and proves asymptotic consistency for projected PCA and projected regression (Pegoraro et al., 2021).

These works use support in two different senses: statistical support of a probability distribution and admissible support of a quantile-valued random object. In both cases, the projection step removes unreliable or invalid geometry before metric comparison or statistical inference is carried out.

7. Support projection in quantum information geometry

The paper “Support-Projected Petz Monotone Geometry of Two-Qubit Families: Three-Channel Identity and Non-Reduction of Curvatures” gives the most explicit differential-geometric use of the term. For a pure two-qubit variational family, it pulls back a Petz monotone metric to parameter space via the reduced one-qubit state $M=\sum_{i=1}^n \mu_i T_i,\quad \mu_i\ge 0,$2,

$M=\sum_{i=1}^n \mu_i T_i,\quad \mu_i\ge 0,$3

and then makes this tensor intrinsic by projecting onto the active numerical range of the QFI. If

$M=\sum_{i=1}^n \mu_i T_i,\quad \mu_i\ge 0,$4

the support-projected metric is

$M=\sum_{i=1}^n \mu_i T_i,\quad \mu_i\ge 0,$5

on $M=\sum_{i=1}^n \mu_i T_i,\quad \mu_i\ge 0,$6. This removes null directions and stabilizes curvature (Cho et al., 18 Sep 2025).

Its main structural theorem is a universal three-channel decomposition on any smooth two-parameter slice. Writing the reduced state in terms of $M=\sum_{i=1}^n \mu_i T_i,\quad \mu_i\ge 0,$7, $M=\sum_{i=1}^n \mu_i T_i,\quad \mu_i\ge 0,$8, and concurrence $M=\sum_{i=1}^n \mu_i T_i,\quad \mu_i\ge 0,$9, the metric satisfies

$M\succeq 0$0

The three channels are population, coherence, and concurrence. A second main result shows that neither the slice Gaussian curvature nor the ambient scalar curvature of the support-projected metric can, on any nonempty open set, be written solely as a function of concurrence or of the one-qubit entropy. The paper therefore rejects the expectation that these curvature invariants could serve as universal entanglement monotones. At the same time, it identifies a natural-gradient interpretation in which support projection removes inactive directions before inversion of the metric tensor (Cho et al., 18 Sep 2025).

In this quantum setting, the support-projected metric is not merely a regularization device. It is the intrinsic Riemannian geometry of the active information-bearing directions, and its curvature retains dependence on all three dynamical channels rather than collapsing to entanglement alone.