Spectrum Projection Score (SPS)

• Spectrum Projection Score (SPS) is a projection-based metric with dual formulations: one defining the second eigenvalue of a graph Laplacian for clustering and the other computing residual norms for summary alignment in RAG.
• In clustering, SPS minimizes the second smallest eigenvalue to bisect data into weakly connected subsets and employs microcluster approximations to achieve significant speedups with minimal accuracy loss.
• In retrieval-augmented generation, SPS measures how well a candidate summary aligns with a reader model’s principal subspace, leading to improved performance in QA tasks.

Spectrum Projection Score (SPS) denotes two distinct projection-based quantities in the arXiv literature. In "Minimum Spectral Connectivity Projection Pursuit" (Hofmeyr et al., 2015), SPS is the second smallest eigenvalue of the graph Laplacian of projected data, used to find low-dimensional projections that maximize the separability of a binary partition of an unlabelled dataset. In "Spectrum Projection Score: Aligning Retrieved Summaries with Reader Models in Retrieval-Augmented Generation" (Hu et al., 8 Aug 2025), SPS is a supervision-free score defined from the residual of a summary representation outside a reader model’s principal subspace, used to rank retrieved summaries in retrieval-augmented generation. Both uses are projection-centric and spectrum-based, but they differ in mathematical object, optimization target, and application domain.

1. Dual usage of the term

The 2015 usage, developed by Hofmeyr et al., belongs to projection pursuit and spectral graph theory; the 2025 usage belongs to retrieval-augmented generation and inference-time control for LLMs. The identical name therefore does not identify a single standardized metric across fields, but two separate constructions with different semantics and different computational pipelines (Hofmeyr et al., 2015, Hu et al., 8 Aug 2025).

Setting	SPS definition	Operational role
Projection-pursuit clustering	$\mathrm{SPS}(V)\equiv \lambda_2(L(V))$	Minimize over projection subspaces to obtain a binary split into two weakly connected pieces
Retrieval-augmented generation	$\mathrm{SPS}(x)=\|(I-P)x\|_2$	Rank candidate summaries; lower SPS indicates stronger alignment with the reader’s principal subspace

This terminological overlap is a common source of confusion. In the first case, the score is attached to a projection matrix $V$ and a similarity graph on projected data. In the second, the score is attached to a candidate summary represented inside a reader LLM.

2. SPS in minimum spectral connectivity projection pursuit

In "Minimum Spectral Connectivity Projection Pursuit" (Hofmeyr et al., 2015), let $X=\{x_1,\dots,x_n\}\subset\mathbb{R}^d$ and let $V\in\mathbb{R}^{d\times r}$ be an orthonormal projection matrix with $r<d$. Writing $p_i=V^\top x_i$ and $P=[p_1\ \dots\ p_n]\in\mathbb{R}^{r\times n}$, the method forms a weighted similarity graph on the projected points with

$W_{ij}=k(\|p_i-p_j\|/\sigma),$

where $k:\mathbb{R}_+\to\mathbb{R}_+$ is a smooth, decreasing kernel and $\mathrm{SPS}(x)=\|(I-P)x\|_2$0 is the scale or bandwidth parameter. The degree matrix is $\mathrm{SPS}(x)=\|(I-P)x\|_2$1 with $\mathrm{SPS}(x)=\|(I-P)x\|_2$2, and the unnormalized graph Laplacian is

$\mathrm{SPS}(x)=\|(I-P)x\|_2$3

The Spectrum Projection Score of $\mathrm{SPS}(x)=\|(I-P)x\|_2$4 is the second smallest eigenvalue of $\mathrm{SPS}(x)=\|(I-P)x\|_2$5:

$\mathrm{SPS}(x)=\|(I-P)x\|_2$6

The intended interpretation is explicit: $\mathrm{SPS}(x)=\|(I-P)x\|_2$7 measures how well connected the graph is, and minimizing $\mathrm{SPS}(x)=\|(I-P)x\|_2$8 seeks a projection that bisects $\mathrm{SPS}(x)=\|(I-P)x\|_2$9 into two weakly connected pieces. The corresponding optimization problem is

$V$0

For $V$1, this reduces to finding a unit vector $V$2 with $V$3 that minimizes $V$4 of its induced Laplacian. An equivalent formulation adds a soft orthogonality penalty,

$V$5

with $V$6 large, to keep columns of $V$7 nearly orthonormal.

The paper characterizes this objective as non-convex and non-smooth. Its significance lies in replacing variance-maximizing or reconstruction-based projections with a projection chosen directly for unsupervised separability.

3. Theoretical link to maximum-margin structure and computational method

A central theoretical result is the connection between minimum spectral connectivity and maximum-margin separation. In the univariate case, if $V$8 and the maximum-margin hyperplane normal $V$9 is unique, then

$X=\{x_1,\dots,x_n\}\subset\mathbb{R}^d$0

The paper states that, as $X=\{x_1,\dots,x_n\}\subset\mathbb{R}^d$1, the kernel becomes sensitive only to very small inter-point gaps, and proves that the optimal univariate projection based on spectral connectivity converges to the normal vector of the maximum-margin hyperplane through the data. Lemmas giving upper and lower spectral bounds in terms of the largest gap underpin this result (Hofmeyr et al., 2015).

The computational method is based on generalized gradient-based descent. To evaluate SPS at $X=\{x_1,\dots,x_n\}\subset\mathbb{R}^d$2, the procedure projects points in $X=\{x_1,\dots,x_n\}\subset\mathbb{R}^d$3, builds $X=\{x_1,\dots,x_n\}\subset\mathbb{R}^d$4 in $X=\{x_1,\dots,x_n\}\subset\mathbb{R}^d$5, forms $X=\{x_1,\dots,x_n\}\subset\mathbb{R}^d$6 and $X=\{x_1,\dots,x_n\}\subset\mathbb{R}^d$7, and computes $X=\{x_1,\dots,x_n\}\subset\mathbb{R}^d$8 and the associated eigenvector $X=\{x_1,\dots,x_n\}\subset\mathbb{R}^d$9 in $V\in\mathbb{R}^{d\times r}$0. When $V\in\mathbb{R}^{d\times r}$1 is a simple eigenvalue, standard matrix-perturbation gives

$V\in\mathbb{R}^{d\times r}$2

after which the chain rule is used to backpropagate to $V\in\mathbb{R}^{d\times r}$3. The reported per-step complexity is $V\in\mathbb{R}^{d\times r}$4.

To reduce the quadratic cost, the paper proposes a microcluster approximation. A fast $V\in\mathbb{R}^{d\times r}$5-means clustering such as BIRCH or K-means produces microclusters $V\in\mathbb{R}^{d\times r}$6 with centers $V\in\mathbb{R}^{d\times r}$7 and cardinalities $V\in\mathbb{R}^{d\times r}$8, where $V\in\mathbb{R}^{d\times r}$9. The full graph Laplacian is then approximated by an $r<d$0 weighted Laplacian $r<d$1 with

$r<d$2

and with degree $r<d$3 on the diagonal. The approximation error is bounded by

$r<d$4

where $r<d$5 is the radius of microcluster $r<d$6 and $r<d$7. The paper states that for small $r<d$8 relative to $r<d$9 this bound is tiny. Computing $p_i=V^\top x_i$0 of the $p_i=V^\top x_i$1 Laplacian costs $p_i=V^\top x_i$2, approximate gradients cost $p_i=V^\top x_i$3, and in practice $p_i=V^\top x_i$4 yields $p_i=V^\top x_i$5–$p_i=V^\top x_i$6 speedups with negligible loss in clustering quality.

4. Hierarchical divisive clustering and empirical behavior of SCPP

The SPS objective is embedded in a divisive hierarchical model to obtain more than two clusters. To obtain $p_i=V^\top x_i$7 clusters, the binary projection-pursuit split is applied recursively in a top-down fashion. At each node corresponding to a subset $p_i=V^\top x_i$8, the method finds the best projection $p_i=V^\top x_i$9 and then bipartitions $P=[p_1\ \dots\ p_n]\in\mathbb{R}^{r\times n}$0 by the sign or 2-means on the second eigenvector of its Laplacian. Splitting continues on the largest, or most connected, leaf until $P=[p_1\ \dots\ p_n]\in\mathbb{R}^{r\times n}$1 leaves are obtained. The resulting divisive tree permits different nodes to use different projection subspaces $P=[p_1\ \dots\ p_n]\in\mathbb{R}^{r\times n}$2 (Hofmeyr et al., 2015).

The empirical evaluation covers Optical Digits, Pen Digits, Multiple-Feature Digits, Satellite, Image Segmentation, Breast-Cancer, Synthetic Chart, Isolet, Dermatology, Yeast, Smartphone Activity, Yale Faces, and Phoneme, with up to approximately $P=[p_1\ \dots\ p_n]\in\mathbb{R}^{r\times n}$3 samples, dimension up to approximately $P=[p_1\ \dots\ p_n]\in\mathbb{R}^{r\times n}$4, and clusters $P=[p_1\ \dots\ p_n]\in\mathbb{R}^{r\times n}$5 up to $P=[p_1\ \dots\ p_n]\in\mathbb{R}^{r\times n}$6. The competing methods are SC, SC_PC / SC_IC, DRSC, and iSVR_G. The evaluation metrics are Purity and Normalized Mutual Information.

The reported result is that SCPP, the SPS-based method, attains the highest Purity and NMI on over half the data sets and is never far behind the best on the rest. Across 14 benchmarks the average relative Purity and NMI of SCPP exceed all competitors. On large $P=[p_1\ \dots\ p_n]\in\mathbb{R}^{r\times n}$7, such as $P=[p_1\ \dots\ p_n]\in\mathbb{R}^{r\times n}$8, microclustered SCPP runs $P=[p_1\ \dots\ p_n]\in\mathbb{R}^{r\times n}$9–$W_{ij}=k(\|p_i-p_j\|/\sigma),$0 faster than the full $W_{ij}=k(\|p_i-p_j\|/\sigma),$1 method while preserving more than $W_{ij}=k(\|p_i-p_j\|/\sigma),$2 of clustering accuracy. The paper summarizes this as state-of-the-art clustering performance across a wide range of data sets.

5. SPS in retrieval-augmented generation and xCompress

In the 2025 RAG formulation, SPS is introduced as a lightweight, supervision-free metric for measuring the semantic alignment of a retrieved summary with a reader model’s hidden representation (Hu et al., 8 Aug 2025). The motivation is that token-level perplexity and single mean-pooled embeddings tend to favor “typical” or centrally clustered tokens rather than informative boundary tokens. SPS replaces this point-based view with an area-based view.

Let $W_{ij}=k(\|p_i-p_j\|/\sigma),$3 be the hidden dimension of the reader’s penultimate layer. From either the input-embedding matrix or a large bank of hidden representations, one forms

$W_{ij}=k(\|p_i-p_j\|/\sigma),$4

The singular value decomposition is

$W_{ij}=k(\|p_i-p_j\|/\sigma),$5

The method retains the top $W_{ij}=k(\|p_i-p_j\|/\sigma),$6 singular values whose cumulative energy is at least $W_{ij}=k(\|p_i-p_j\|/\sigma),$7 of $W_{ij}=k(\|p_i-p_j\|/\sigma),$8, with corresponding left singular vectors $W_{ij}=k(\|p_i-p_j\|/\sigma),$9. The projection matrix onto the principal subspace is

$k:\mathbb{R}_+\to\mathbb{R}_+$0

For a candidate summary of $k:\mathbb{R}_+\to\mathbb{R}_+$1 tokens, the reader produces penultimate hidden states $k:\mathbb{R}_+\to\mathbb{R}_+$2. The summary “area” vector is the element-wise max-pooled vector

$k:\mathbb{R}_+\to\mathbb{R}_+$3

SPS is then defined by decomposing $k:\mathbb{R}_+\to\mathbb{R}_+$4 into its component inside the principal subspace and its residual outside it:

$k:\mathbb{R}_+\to\mathbb{R}_+$5

A lower value means that the summary’s salient envelope lies mostly within the reader’s core directions.

This score is integrated into xCompress, which sits between the retriever or compressor LLM and the final reader LLM. For each query, the system retrieves top-$k:\mathbb{R}_+\to\mathbb{R}_+$6 documents, compresses them into an initial summary $k:\mathbb{R}_+\to\mathbb{R}_+$7, and computes two simple norms on $k:\mathbb{R}_+\to\mathbb{R}_+$8’s representation in the reader:

• $k:\mathbb{R}_+\to\mathbb{R}_+$9
• $\mathrm{SPS}(x)=\|(I-P)x\|_2$00

with ratio $\mathrm{SPS}(x)=\|(I-P)x\|_2$01. If $\mathrm{SPS}(x)=\|(I-P)x\|_2$02, the system skips sampling and feeds $\mathrm{SPS}(x)=\|(I-P)x\|_2$03 to the reader. Otherwise it samples $\mathrm{SPS}(x)=\|(I-P)x\|_2$04 diverse summaries. For text-to-text, the sampling uses stochastic decoding with temperature $\mathrm{SPS}(x)=\|(I-P)x\|_2$05 and repetition_penalty $\mathrm{SPS}(x)=\|(I-P)x\|_2$06; for text-to-embedding, it injects small Gaussian probes, keeps $\mathrm{SPS}(x)=\|(I-P)x\|_2$07 probes with the smallest semantic-deviation scores $\mathrm{SPS}(x)=\|(I-P)x\|_2$08, plus the original. SPS is computed for each candidate, and the summary with the lowest SPS is selected for final answer generation.

The experimental validation covers HotpotQA, 2WikiMultiHop, Natural Questions, TriviaQA, and Musique, using LLaMA-3.1-8B-Instruct, Gemma3-12B-Instruct, Qwen3-8B, and Mistral-7B. The paper reports that PPL and LongPPL have near-zero or negative Pearson correlations with downstream Exact Match and F1, whereas SPS shows strong positive correlations and higher AUROC. For example, on HotpotQA, SPS achieves $\mathrm{SPS}(x)=\|(I-P)x\|_2$09, $\mathrm{SPS}(x)=\|(I-P)x\|_2$10, and AUROC of approximately $\mathrm{SPS}(x)=\|(I-P)x\|_2$11. In end-to-end QA on HotpotQA with LLaMA-3.1, naïve retrieval-to-reader gives EM/F1 of $\mathrm{SPS}(x)=\|(I-P)x\|_2$12, CompAct summary gives $\mathrm{SPS}(x)=\|(I-P)x\|_2$13, and xCompress + SPS gives $\mathrm{SPS}(x)=\|(I-P)x\|_2$14, corresponding to a $\mathrm{SPS}(x)=\|(I-P)x\|_2$15 EM and $\mathrm{SPS}(x)=\|(I-P)x\|_2$16 F1 improvement over CompAct. In text-to-embedding with Mistral-7B + xRAG, xCompress + SPS improves EM/F1 on HotpotQA from $\mathrm{SPS}(x)=\|(I-P)x\|_2$17 to $\mathrm{SPS}(x)=\|(I-P)x\|_2$18.

The ablations report that max-pooling outperforms mean-pooling and last-token pooling by 2–4 points EM/F1, penultimate-layer embeddings give the best downstream performance, $\mathrm{SPS}(x)=\|(I-P)x\|_2$19 variance retention is optimal, and performance saturates at $\mathrm{SPS}(x)=\|(I-P)x\|_2$20. Practical details include precomputing the SVD or PCA once per reader model, an $\mathrm{SPS}(x)=\|(I-P)x\|_2$21 projection cost per candidate at inference, adaptive norm-guided filtering that skips sampling for approximately $\mathrm{SPS}(x)=\|(I-P)x\|_2$22 of queries and cuts $\mathrm{SPS}(x)=\|(I-P)x\|_2$23 of sampling cost with minimal impact on accuracy, greedy decoding for the final answer, and an overhead of approximately $\mathrm{SPS}(x)=\|(I-P)x\|_2$24–$\mathrm{SPS}(x)=\|(I-P)x\|_2$25 over a single-pass RAG pipeline on standard GPU hardware such as NVIDIA A100.

6. Comparative interpretation and disambiguation

The two SPS definitions are mathematically unrelated beyond a broad reliance on projection and spectrum. In the clustering formulation, SPS is a second-smallest-eigenvalue objective on a graph Laplacian constructed from pairwise similarities in a projected space. In the RAG formulation, SPS is an $\mathrm{SPS}(x)=\|(I-P)x\|_2$26 residual norm after projecting a max-pooled summary representation onto a reader-specific PCA subspace (Hofmeyr et al., 2015, Hu et al., 8 Aug 2025).

Their optimization logic also differs. The clustering method minimizes SPS over projection matrices $\mathrm{SPS}(x)=\|(I-P)x\|_2$27 to discover unsupervised binary partitions, and then composes these splits into a divisive hierarchy. The RAG method ordinarily holds the reader subspace fixed, computes SPS for each candidate summary, and selects the summary with the lowest score inside xCompress. In both cases, lower values are preferred, but “lower” means different things: weaker graph connectivity in the first case, and smaller residual outside the reader’s principal directions in the second.

A common misconception is that SPS denotes a single canonical score transferable across application areas. The published record summarized here does not support that interpretation. A more precise usage is to specify either the spectral-connectivity SPS of projection pursuit clustering or the reader-alignment SPS used in retrieval-augmented generation. This suggests that explicit citation is necessary whenever the acronym appears without context.