H-Eigenscore: Eigen-Structured Scoring

• H-Eigenscore is an eigenvalue‐based scoring method that derives a scalar from the latent spectral structure of data.
• It is applied in diverse fields such as bibliometrics for coupled author–paper ranking and in uncertainty detection for confidence estimation.
• Method reliability depends on normalization, controlled spectral contamination, and convergence properties to ensure meaningful scores.

H-Eigenscore denotes, in current usage, an eigenvalue-based scoring idiom rather than a single universally standardized invariant. Its most explicit formulation appears in bibliometrics, where a coupled eigensystem simultaneously assigns an author score and a paper score from the author–paper bipartite network and the paper citation network (Ujum et al., 2015). Elsewhere, closely related usages treat an “H-Eigenscore”-like quantity as a spectral confidence statistic for hallucination detection, as a structured procedure for extracting a target eigenvalue from a large hierarchical matrix, or as a generalized-eigenvalue template whose interpretability depends on basis quality, parameter choice, and explicit control of spectral contamination (Shamsi, 8 Jun 2026, Apriansyah et al., 2023, 0808.1017). The range of usages indicates that the term functions primarily as a label for eigen-structured ranking, confidence, or extraction mechanisms.

1. Scope and family resemblance

The principal commonality across H-Eigenscore usages is that a scalar or ordered quantity is obtained from an eigensystem, and its reliability is tied to how faithfully the chosen spectral object captures the latent structure of interest. In bibliometrics, the latent structure is mutual reinforcement between authors and papers; in hallucination detection, it is the geometry of hidden-state response trajectories; in structured linear algebra, it is the position of the $k$-th eigenvalue within a compressed spectrum; and in generalized-eigenvalue workflows, it is the separation of physically relevant low-lying states from omitted or contaminating modes (Ujum et al., 2015, Shamsi, 8 Jun 2026, Apriansyah et al., 2023, 0808.1017).

Context	Scored object	Spectral mechanism
Bibliometrics	author score and paper score	coupled power iteration on $W$, $C$, and $L=WC$
Hallucination detection	selective-prediction confidence	comparison to EigenScore/H-Eigenscore-style spectral uncertainty statistics
Structured eigensolving	$k$-th eigenvalue	inertia-based bisection with generalized $LDL^T$
Variational spectroscopy / optimization	ordered states or descent directions	GEVP or Hessian eigensystem with explicit gap/error control

This breadth matters because it prevents a reductive definition. H-Eigenscore is not confined to citation ranking, nor is it identical to every construction involving an $H$-prefixed eigenvalue. Rather, the cited literature supports a narrower characterization: an H-Eigenscore is an eigen-structured score whose semantics depend on the ambient problem class and on the stability of the eigensystem used to define it.

2. Coupled author–paper eigenscores

The most concrete and fully specified H-Eigenscore construction is the dual author–paper framework studied as CITEX and then modified as CAPS, a HITS-like algorithm for simultaneous evaluation of researchers and their works (Ujum et al., 2015). The data model begins with a binary author–paper matrix $M\in\{0,1\}^{m\times n}$, where $M_{ij}=1$ if author $i$ has (co)authored paper $W$0, and a binary citation matrix $W$1, where $W$2 if paper $W$3 cites paper $W$4, with $W$5. A column-normalized authorship matrix $W$6 distributes paper credit equally among coauthors: $W$7

CITEX couples author and paper scores through

$W$8

The paper argues that this formulation embeds a problematic bias. In its own summary, CITEX effectively rewards authors who are “highly prolific with, or are highly cited by good authors,” and papers that “share the same authors with, or are cited by good papers” (Ujum et al., 2015). The inclusion of $W$9 acts as an artificial self-citation term, and the use of raw $C$0 amplifies team-structure effects.

CAPS removes that term and replaces $C$1 by $C$2, yielding the coupled eigensystem

$C$3

With

$C$4

the updates become

$C$5

The two governing criteria are stated explicitly: “a good author is cited by good authors” and “a good paper is cited by good authors” (Ujum et al., 2015). The recursive interpretation is equally explicit: “a good author has good papers that are cited by good authors who have good papers and so on.”

The bibliometric H-Eigenscore is therefore a mutually reinforcing spectral score over two coupled layers. It is not a scalar derived from a single adjacency matrix; it is a joint fixed point over author and paper spaces.

3. Normalization, convergence, and empirical behavior

Normalization is central to the CAPS interpretation. The replacement of $C$6 by $C$7 is intended to ensure conservation of citation count when switching from the paper citation network to the author citation network (Ujum et al., 2015). This prevents simple inflation by coauthorship multiplicity and makes the propagation rule closer to a fractional author-citation model. The paper defines $C$8 as an author–paper citation matrix, while $C$9 becomes the fractional author citation matrix.

For CITEX, convergence is formulated through the recursion

$L=WC$0

with initial vectors $L=WC$1 and $L=WC$2, and termination when

$L=WC$3

The paper explicitly invokes Perron–Frobenius for the nonnegative matrix $L=WC$4, so the iterates converge in direction to a stationary vector $L=WC$5 satisfying $L=WC$6 (Ujum et al., 2015). CAPS is presented in the same fixed-point spirit, although without a separate derivation at the same level of detail.

The empirical study uses Thomson ISI / Journal Citation Reports data in the “Information Science & Library Science” category from 1980–2012, comprising 213,530 papers, 471,191 total inter-paper citations, and 73,597 author keywords, with no author or bibliographic reference disambiguation (Ujum et al., 2015). On this corpus, CAPS is reported to correlate more strongly with the $L=WC$7-index than CITEX does: Spearman $L=WC$8 for CAPS versus $L=WC$9-index, compared with $k$0 for CITEX versus $k$1-index, both significant at $k$2. For papers, the reported correlations are $k$3 between citation count and CAPS, $k$4 between citation count and CITEX, and $k$5 between CAPS and CITEX (Ujum et al., 2015).

The score distributions are highly unequal under both schemes, but the paper treats CITEX as markedly more pathological. It reports that Rogers alone accounts for 83.6% of the CITEX author score distribution, Rogers plus Cassada together account for 96%, and the CITEX author-score Gini coefficient is 0.9999. Under CAPS, the top 20% of authors account for about 99.96% of total scores and the author-score Gini coefficient is 0.9891 (Ujum et al., 2015). On the paper side, CAPS has a Gini coefficient of 0.9912, CITEX 0.9785; about 81.2% of the lowest-scoring population has exactly zero CAPS score, whereas CITEX has no zero-scoring population because of the artificial paper self-citations. The paper also notes that the first 3819 CITEX paper ranks are occupied by papers authored by Rogers, and the next 2610 by papers authored by Cassada.

These results define the main limitation of bibliometric H-Eigenscores. The method can simultaneously rank authors and papers, but the resulting distributions remain extremely concentrated, and the authors explicitly recommend extreme caution in substantive interpretation (Ujum et al., 2015).

4. Conditions for reliable eigenscoring

A recurring theme in eigenvalue-based scoring is that the eigenvalue is meaningful only if the underlying eigensystem has controlled contamination, stable geometry, or analytically tractable projection. In lattice gauge theory, the generalized eigenvalue problem is used to extract energies and matrix elements from a correlator matrix

$k$6

with $k$7 (0808.1017). The paper’s central methodological conclusion is that eigenvalues become efficient, systematically controlled estimators only when the operator basis spans the lowest $k$8 states reasonably well and $k$9, $LDL^T$0, and basis construction are chosen to suppress excited-state contamination. In the regime $LDL^T$1, the dominant corrections to the $LDL^T$2-th level are suppressed by the gap to the first omitted state, $LDL^T$3, and choosing $LDL^T$4 sufficiently large improves convergence (0808.1017). Stable eigenvectors and plateaus in the effective energies are treated as diagnostics of a reliable basis.

An analogous message appears in mesh distortion optimization. There, projected Newton replaces an indefinite Hessian $LDL^T$5 by its SPD projection,

$LDL^T$6

so that the local search direction is descending (Zhu, 2021). The bottleneck is eigendecomposition of each element Hessian. For principal-stretch distortion energies on a $LDL^T$7D $LDL^T$8 tetrahedron, the paper derives an analytic Hessian eigensystem: the full $LDL^T$9 Hessian has a $H$0D null space from rigid translations, reduces to a $H$1 matrix, and six eigenpairs of that reduced Hessian are always available in closed form from three $H$2 pairwise stretch blocks; only an energy-specific $H$3 block may still require numerical handling (Zhu, 2021). The practical consequence is that SPD projection can be reduced from dense numerical eigendecomposition to mostly analytic clamping.

Taken together, these works imply a general principle for H-Eigenscore-like procedures. A spectral score is robust when the eigensystem is not treated as a black box: basis completeness, omitted-state gaps, null-space structure, and blockwise analytic diagonalization determine whether the score is interpretable and computationally efficient (0808.1017, Zhu, 2021).

5. H-Eigenscore-style uncertainty scores in selective prediction

In hallucination detection for LLMs and VLMs, H-Eigenscore appears as part of a family of unsupervised sampling detectors that includes Semantic Entropy and EigenScore (Shamsi, 8 Jun 2026). That literature frames hallucination detection as selective prediction: a detector assigns a scalar confidence score, and the system abstains when confidence is low. The 2026 density-ridge method explicitly contrasts itself with EigenScore/H-Eigenscore-style methods, which it characterizes as spectral or geometry-aware scalar summaries of hidden representations or sampled outputs (Shamsi, 8 Jun 2026).

The proposed alternative maps each hidden-state trajectory $H$4 into a six-dimensional kinematic feature vector

$H$5

fits a Gaussian KDE on correct trajectories, extracts a $H$6-D density ridge $H$7 by SCMS, and scores a test query by negated mean off-ridge distance: $H$8 The paper’s critique is that unsupervised detectors such as Semantic Entropy and EigenScore plateau because they reduce uncertainty to coarse dispersion statistics and may miss the manifold-level organization of generation dynamics (Shamsi, 8 Jun 2026).

The evaluation uses a deliberately label-scarce protocol with $H$9 calibration queries and $M\in\{0,1\}^{m\times n}$0 generations, on seven QA benchmarks and nine text and vision LLMs. Against Semantic Entropy, SAR, EigenScore, SAPLMA, and log-probability, the density-ridge score is reported to improve AUROC by 5–20 absolute points while degrading more gracefully than supervised probes when calibration labels are scarce (Shamsi, 8 Jun 2026). The paper’s own summary of the comparison is concise: EigenScore measures spectral uncertainty, whereas the ridge-based detector measures manifold proximity to a learned correct-response ridge.

This comparison clarifies an important semantic shift. In bibliometrics, H-Eigenscore ranks entities. In selective prediction, an H-Eigenscore-like quantity is a confidence score. The common element is not what is being scored, but the use of spectral structure as a proxy for latent regularity.

6. Neighboring notions and terminological boundaries

Several nearby constructions are distinct from H-Eigenscore proper. The term should not be conflated with $M\in\{0,1\}^{m\times n}$1-eigenvalues of Hermitian tensors. That work introduces $M\in\{0,1\}^{m\times n}$2-eigenvalues for $M\in\{0,1\}^{m\times n}$3-th order complex tensors, derives inclusion sets for them, and proves positivity criteria for Hermitian and CPS tensors; it also states explicitly that it does not define an object called “H-Eigenscore” (Chen et al., 17 Aug 2025). Likewise, the tensor-spectral literature on E-eigenvalues studies the E-characteristic polynomial and obtains a closed formula for the product of E-eigenvalues in terms of the gradient resultant and the $M\in\{0,1\}^{m\times n}$4-discriminant, generalizing the matrix identity $M\in\{0,1\}^{m\times n}$5 (Sodomaco, 2018). These are spectral invariants, not scoring algorithms.

A different usage appears in structured numerical linear algebra. For symmetric $M\in\{0,1\}^{m\times n}$6-matrices, “H-Eigenscore” is used to denote a fast, structured way to extract the $M\in\{0,1\}^{m\times n}$7-th smallest eigenvalue by combining Sylvester inertia-based bisection with a generalized $M\in\{0,1\}^{m\times n}$8 factorization of the compressed matrix (Apriansyah et al., 2023). The method has $M\in\{0,1\}^{m\times n}$9 storage and $M_{ij}=1$0 arithmetic complexity under the stated rank assumptions, and it can reuse interval information when several eigenvalues are required. Here the “score” is an ordered spectral location rather than a prestige or confidence value.

Quantum algorithms produce yet another variant. The Quantum Heaviside Eigen Solver defines a “quantum judge” that approximates a Heaviside step on the spectrum,

$M_{ij}=1$1

and uses dichotomy to recover the lowest eigenvalue with error smaller than $M_{ij}=1$2 in $M_{ij}=1$3 iterations of Hamiltonian shifting and threshold testing (Sun et al., 2021). A companion “quantum selector” then amplifies the corresponding eigenspace. This is an eigenspectrum-filtering mechanism, not a bibliometric or statistical score.

Finally, there is unrelated $M_{ij}=1$4-terminology in algebra. H-depth is the $M_{ij}=1$5-$M_{ij}=1$6-bimodule analogue of ordinary subring depth, with H-depth $M_{ij}=1$7 equivalent to H-separability (Kadison, 2011). Its appearance in the same lexical neighborhood is purely terminological.

The main encyclopedic point is therefore negative as much as positive: H-Eigenscore is not synonymous with every object bearing an $M_{ij}=1$8-prefix and an eigenvalue interpretation. It is best reserved for eigen-structured scoring procedures, most concretely the coupled author–paper eigenscore of CAPS/CITEX, and more loosely for related spectral confidence or extraction mechanisms whose semantics are determined by the application domain (Ujum et al., 2015, Shamsi, 8 Jun 2026, Apriansyah et al., 2023).