Axes of Correspondence: Data, Neural & Algebra

• Axes of correspondence are specialized directions or invariant supports that make latent relations explicit in methods like correspondence analysis, HOSVD, and ICA.
• They provide actionable insights into multiway data decomposition, dimensionality reduction, and practical applications such as object-centric control in robotics.
• The concept extends to algebra and geometry, framing invariant data in axial algebras and cobordism correspondences to classify complex structural relations.

to=arxiv_search.search 仲博ি 心博json {"query":"\"axes of correspondence\" correspondence analysis tensor HOSVD ICA word embeddings soft-matching neural representational comparison", "max_results": 10}】【”】【assistant to=arxiv_search.search մեկնաբանություն 大发快三是不是json {"query":"Lusztig Correspondence Howe Correspondence finite reductive dual pairs (Pan, 2019)", "max_results": 5}】【：】【“】【assistant to=arxiv_search.search մեկնաբանություն 大发娱乐json {"query":"Partial Soft-Matching Distance Neural Representational Comparison Partial Unit Correspondence (Kapoor et al., 22 Feb 2026)", "max_results": 5}】【。analysis to=arxiv_search.search 天天中彩票中了json {"query":"Axis Tour Word Tour ICA transformed embeddings (Yamagiwa et al., 2024)", "max_results": 5} “Axes of correspondence” is not a single universally fixed term. Across several technical literatures, it denotes the directions, units, generators, or invariant supports through which a correspondence becomes explicit: principal factor axes in Correspondence Analysis, mode-wise HOSVD axes in multiway contingency analysis, matched neurons or coordinate axes in representational comparison, reordered ICA components in word embeddings, object-centric task axes in robotic manipulation, double axes in Matsuo algebras, and rigid invariant loci for correspondences of curves (Murtagh, 2015, Coulaud et al., 2021, Kapoor et al., 22 Feb 2026, Yamagiwa et al., 2024, Sharma et al., 2021, Galt et al., 2020, Saha, 2012). A common theme is that correspondence is treated not merely as a relation between whole objects, but as a structured alignment resolved into factor directions, eigenspaces, semantically meaningful units, or invariant geometric data.

1. Correspondence Analysis and principal factor axes

In Correspondence Analysis (CA), the axes of correspondence are the principal factor directions obtained from the geometry of profiles under the $\chi^2$ metric. Starting from a nonnegative matrix $k_{IJ}$, CA passes to frequencies $f_{IJ}$, row masses $f_I$, and column masses $f_J$. For a row $i \in I$, the column profile is

$$f^i_J = \{ f^i_j = f_{ij}/f_i = (k_{ij}/k)/(k_i/k) ; f_i \neq 0 ; j \in J \},$$

and the analysis is carried out on these profiles rather than on raw counts (Murtagh, 2015).

The resulting factor space is orthonormal, and the axes are ordered by decreasing eigenvalue, or inertia. The paper states that “Correspondence factor analysis provides a latent semantic or principal axes mapping” (Murtagh, 2015). Its inertia decomposition is written as

$$M^2(N_J(I)) = \sum_{\alpha=1..\nu} \lambda_\alpha = \sum_{i \in I} f_i \rho^2(i),$$

with

$$\lambda_\alpha = \sum_{i \in I} f_i F^2_\alpha(i) \qquad\text{and}\qquad \rho^2(i) = \sum_{\alpha=1..\nu} F^2_\alpha(i).$$

Here $\lambda_\alpha$ is the inertia explained by axis $k_{IJ}$0, $k_{IJ}$1 is the coordinate of point $k_{IJ}$2 on that axis, and $k_{IJ}$3 is the squared distance of $k_{IJ}$4 from the origin in factor space (Murtagh, 2015).

Interpretation in CA is contribution-driven. The paper emphasizes that “Inertia is more important than projection (i.e., position) per se,” and defines absolute and relative contributions such as $k_{IJ}$5 and $k_{IJ}$6 for pointwise interpretation of an axis (Murtagh, 2015). This makes the axes semantic only in relation to the masses that generate them.

A further claim of the high-dimensional study is that CA is particularly suitable for “an orthonormal mapping, or scaling, of power law distributed data,” including dimensions up to around one million. As dimensionality increases, the paper reports concentration of the cloud, decreasing mean absolute contribution to inertia, and effective simplification of the geometry around a small number of dominant directions (Murtagh, 2015).

2. Multiway correspondence: mode-wise axes and generalized barycenters

The tensor extension of CA replaces the row/column duality of a matrix by a family of mode-wise point clouds. For a $k_{IJ}$7-way contingency tensor

$k_{IJ}$8

the unfolding $k_{IJ}$9 along mode $f_{IJ}$0 defines a point cloud $f_{IJ}$1 whose rows correspond to the categories of that mode and whose columns correspond to all combinations of the remaining modes (Coulaud et al., 2021). The axes of correspondence are therefore no longer only row and column axes; each tensor mode has its own axes.

The tensor analogue of the SVD is given by the Tucker model and HOSVD: $f_{IJ}$2 where $f_{IJ}$3 is an orthonormal factor matrix for mode $f_{IJ}$4, $f_{IJ}$5 is the core tensor, and $f_{IJ}$6 is the multilinear rank. The mode-$f_{IJ}$7 principal component matrix is

$f_{IJ}$8

Accordingly, the columns of $f_{IJ}$9 are the axes for mode $f_I$0, and the rows of $f_I$1 are the coordinates of the points of mode $f_I$2 projected on those axes (Coulaud et al., 2021).

A central structural result is that the axes are coupled across modes. For general order $f_I$3, the paper gives

$f_I$4

with $f_I$5 (Coulaud et al., 2021). This means that one mode’s principal coordinates are linear combinations of Kronecker products of the others, so the mode-wise axes are not isolated decompositions.

To recover genuine CA geometry, the paper introduces CA-style marginals

$f_I$6

and the isometry

$f_I$7

Under this metric, the barycentric relation survives in tensor form. The authors define

$f_I$8

and prove generalized barycentric formulas in which each scaled coordinate of one mode is a weighted barycenter of coordinates from the other modes (Coulaud et al., 2021). The reported advantages are that MWCA keeps each mode separate, reveals global relationships among all modes at once, and preserves the CA-type barycentric interpretation; the reported limitations are dependence on the chosen tensor approximation, possible sign flips of bases, and the fact that classical matricized CA can expose relations specific to a chosen unfolding more sharply (Coulaud et al., 2021).

3. Axis-level correspondence in learned representations

In neural representational comparison, the question of axes of correspondence is posed explicitly as a question about which individual units or coordinate axes line up across systems, and which do not. The partial soft-matching framework argues that rotation-invariant similarity measures such as CKA, RSA, and CCA-style methods discard axis identity and are therefore blind to this question. Its central contribution is a partial optimal transport formulation of soft-matching that allows only a fraction $f_I$9 of the total mass to be matched, leaving the rest unmatched. The transport plan $f_J$0 then induces a ranking of units: row sums and column sums indicate how much each unit participates in cross-system alignment, and near-zero sums identify effectively unmatched units (Kapoor et al., 22 Feb 2026).

This changes the interpretation of representational correspondence from a forced global matching to a partition into aligned and unmatched subpopulations. The paper is explicit that partial OT does not satisfy the triangle inequality, so it is not a proper metric in the strict sense, but it is proposed as a comparative tool. Its empirical results are organized around robustness to outliers, ranking of neurons and voxels by alignment quality, improved precision of homologous alignment in fMRI, and the observation that matched deep-network units tend to have similar maximally exciting images whereas unmatched units show divergent patterns (Kapoor et al., 22 Feb 2026). Most directly for the axis question, the authors rotate one representation by an orthogonal matrix $f_J$1 and show that rotation consistently reduces alignment at all thresholds, including among the most highly matched units; they interpret this as evidence for “privileged representational axes” (Kapoor et al., 22 Feb 2026).

A related but distinct use of the term appears in ICA-transformed word embeddings. There, ICA yields semantically interpretable components, but “ICA does not determine any meaningful order among the axes”: any permutation of the components is equally valid (Yamagiwa et al., 2024). Axis Tour addresses this by defining an axis embedding from the top $f_J$2 words on each ICA axis and solving a TSP-like ordering problem so that neighboring axes have high cosine similarity. The objective is semantic continuity rather than new decomposition. The paper reports that Axis Tour shifts the histogram of adjacent-axis cosine similarities toward more positive values, gives a Spearman correlation of $f_J$3 between skewness and local semantic continuity compared with $f_J$4 for Skewness Sort, and at $f_J$5 yields average downstream performance of $f_J$6 on Analogy, $f_J$7 on Similarity, and $f_J$8 on Categorization, compared with PCA’s $f_J$9, $i \in I$0, and $i \in I$1 (Yamagiwa et al., 2024). At full dimension $i \in I$2, all methods coincide because Axis Tour, Random Order, and Skewness Sort are orthogonal reparameterizations of the same ICA space (Yamagiwa et al., 2024).

4. Object-centric task axes in robotic manipulation

In robotics, axes of correspondence arise as object-centric directions used to parameterize modular controllers. The task-axes-controller framework is designed for manipulation tasks with large variation in object shape, size, and geometry and without access to CAD models. Rather than learning a monolithic image-to-action policy, the method composes controllers for reaching, forcing, aligning, and rotating, each defined relative to semantically meaningful object properties inferred from vision (Sharma et al., 2021).

The paper uses three controller families: position controllers parameterized by $i \in I$3, force controllers parameterized by $i \in I$4, and rotation controllers parameterized by $i \in I$5. With current end-effector translation $i \in I$6, rotation $i \in I$7, and force $i \in I$8, the position error is

$i \in I$9

and the rotational correction is

$$f^i_J = \{ f^i_j = f_{ij}/f_i = (k_{ij}/k)/(k_i/k) ; f_i \neq 0 ; j \in J \},$$0

These formulas make the axis $$f^i_J = \{ f^i_j = f_{ij}/f_i = (k_{ij}/k)/(k_i/k) ; f_i \neq 0 ; j \in J \},$$1 a literal geometric constraint on translation or alignment (Sharma et al., 2021).

The visual side is bootstrapped by multi-view dense correspondence learning. A dense descriptor network $$f^i_J = \{ f^i_j = f_{ij}/f_i = (k_{ij}/k)/(k_i/k) ; f_i \neq 0 ; j \in J \},$$2 is used together with reference pixels $$f^i_J = \{ f^i_j = f_{ij}/f_i = (k_{ij}/k)/(k_i/k) ; f_i \neq 0 ; j \in J \},$$3 so that corresponding pixels in a new image are obtained by nearest-neighbor matching in descriptor space,

$$f^i_J = \{ f^i_j = f_{ij}/f_i = (k_{ij}/k)/(k_i/k) ; f_i \neq 0 ; j \in J \},$$4

after which depth yields the 3D keypoint $$f^i_J = \{ f^i_j = f_{ij}/f_i = (k_{ij}/k)/(k_i/k) ; f_i \neq 0 ; j \in J \},$$5 (Sharma et al., 2021). For axes, the method constructs a candidate axes set $$f^i_J = \{ f^i_j = f_{ij}/f_i = (k_{ij}/k)/(k_i/k) ; f_i \neq 0 ; j \in J \},$$6 containing object axes and world axes, and associates each position target with each candidate axis instead of relying on hand-crafted axis selection (Sharma et al., 2021).

The system is evaluated on Button Press, Block Tumble, and Door Open. Mean success over $$f^i_J = \{ f^i_j = f_{ij}/f_i = (k_{ij}/k)/(k_i/k) ; f_i \neq 0 ; j \in J \},$$7 environment variations is reported as follows: for Button Press, EE-Space $$f^i_J = \{ f^i_j = f_{ij}/f_i = (k_{ij}/k)/(k_i/k) ; f_i \neq 0 ; j \in J \},$$8, TAC (Manual) $$f^i_J = \{ f^i_j = f_{ij}/f_i = (k_{ij}/k)/(k_i/k) ; f_i \neq 0 ; j \in J \},$$9, TAC (Keypoints+Axes) $$M^2(N_J(I)) = \sum_{\alpha=1..\nu} \lambda_\alpha = \sum_{i \in I} f_i \rho^2(i),$$0; for Block Tumble, EE-Space $$M^2(N_J(I)) = \sum_{\alpha=1..\nu} \lambda_\alpha = \sum_{i \in I} f_i \rho^2(i),$$1, TAC (Manual) $$M^2(N_J(I)) = \sum_{\alpha=1..\nu} \lambda_\alpha = \sum_{i \in I} f_i \rho^2(i),$$2, TAC (Keypoints+Axes) $$M^2(N_J(I)) = \sum_{\alpha=1..\nu} \lambda_\alpha = \sum_{i \in I} f_i \rho^2(i),$$3; for Door Opening, EE-Space $$M^2(N_J(I)) = \sum_{\alpha=1..\nu} \lambda_\alpha = \sum_{i \in I} f_i \rho^2(i),$$4, TAC (Manual) $$M^2(N_J(I)) = \sum_{\alpha=1..\nu} \lambda_\alpha = \sum_{i \in I} f_i \rho^2(i),$$5, TAC (Keypoints+Axes) $$M^2(N_J(I)) = \sum_{\alpha=1..\nu} \lambda_\alpha = \sum_{i \in I} f_i \rho^2(i),$$6 (Sharma et al., 2021). The paper attributes this improvement to the object-centric inductive bias and to the modular separation of perception and control.

5. Representation-theoretic correspondences and algebraic axes

In finite reductive character theory, correspondence does not refer to geometric axes, but to compatible parametrizations indexed by semisimple and unipotent data. For a finite reductive dual pair $$M^2(N_J(I)) = \sum_{\alpha=1..\nu} \lambda_\alpha = \sum_{i \in I} f_i \rho^2(i),$$7 of symplectic and orthogonal type over $$M^2(N_J(I)) = \sum_{\alpha=1..\nu} \lambda_\alpha = \sum_{i \in I} f_i \rho^2(i),$$8 with $$M^2(N_J(I)) = \sum_{\alpha=1..\nu} \lambda_\alpha = \sum_{i \in I} f_i \rho^2(i),$$9 odd, the Howe correspondence is defined from the decomposition of the Weil representation restricted to $$\lambda_\alpha = \sum_{i \in I} f_i F^2_\alpha(i) \qquad\text{and}\qquad \rho^2(i) = \sum_{\alpha=1..\nu} F^2_\alpha(i).$$0, while the Lusztig correspondence identifies a Lusztig series $$\lambda_\alpha = \sum_{i \in I} f_i F^2_\alpha(i) \qquad\text{and}\qquad \rho^2(i) = \sum_{\alpha=1..\nu} F^2_\alpha(i).$$1 with unipotent characters of the centralizer $$\lambda_\alpha = \sum_{i \in I} f_i F^2_\alpha(i) \qquad\text{and}\qquad \rho^2(i) = \sum_{\alpha=1..\nu} F^2_\alpha(i).$$2. The main theorem proves that these correspondences commute up to twisting by the sign character, and as a consequence the Howe correspondence can be described explicitly in Lusztig’s parametrization for classical groups (Pan, 2019). A plausible implication is that, in this setting, correspondence is resolved into semisimple labels and unipotent parameters rather than into literal coordinate directions.

A genuinely axis-based algebraic usage appears in Matsuo algebras. There, an axial algebra is a commutative non-associative algebra generated by idempotents called axes, subject to a fusion law. In the Matsuo algebra $$\lambda_\alpha = \sum_{i \in I} f_i F^2_\alpha(i) \qquad\text{and}\qquad \rho^2(i) = \sum_{\alpha=1..\nu} F^2_\alpha(i).$$3 attached to a $$\lambda_\alpha = \sum_{i \in I} f_i F^2_\alpha(i) \qquad\text{and}\qquad \rho^2(i) = \sum_{\alpha=1..\nu} F^2_\alpha(i).$$4-transposition group, each basis element is a single axis satisfying the Jordan-type fusion law $$\lambda_\alpha = \sum_{i \in I} f_i F^2_\alpha(i) \qquad\text{and}\qquad \rho^2(i) = \sum_{\alpha=1..\nu} F^2_\alpha(i).$$5, and for orthogonal single axes $$\lambda_\alpha = \sum_{i \in I} f_i F^2_\alpha(i) \qquad\text{and}\qquad \rho^2(i) = \sum_{\alpha=1..\nu} F^2_\alpha(i).$$6 with $$\lambda_\alpha = \sum_{i \in I} f_i F^2_\alpha(i) \qquad\text{and}\qquad \rho^2(i) = \sum_{\alpha=1..\nu} F^2_\alpha(i).$$7, the sum

$$\lambda_\alpha = \sum_{i \in I} f_i F^2_\alpha(i) \qquad\text{and}\qquad \rho^2(i) = \sum_{\alpha=1..\nu} F^2_\alpha(i).$$8

is again an idempotent, called a double axis (Galt et al., 2020). The paper proves that if $$\lambda_\alpha = \sum_{i \in I} f_i F^2_\alpha(i) \qquad\text{and}\qquad \rho^2(i) = \sum_{\alpha=1..\nu} F^2_\alpha(i).$$9, then $\lambda_\alpha$0 satisfies the Monster-type fusion law $\lambda_\alpha$1, with eigenspaces

$\lambda_\alpha$2

and Miyamoto involution

$\lambda_\alpha$3

It also classifies primitive subalgebras generated by single and double axes in small ranks and introduces the flip construction, which yields infinite series $\lambda_\alpha$4, $\lambda_\alpha$5, and $\lambda_\alpha$6 of axial algebras of Monster type (Galt et al., 2020).

6. Correspondence in algebraic geometry: cobordism and invariant forms

A correspondence-based reformulation of bivariant ideas appears in algebraic cobordism. Instead of indexing a theory by a morphism $\lambda_\alpha$7, the bi-variant theory $\lambda_\alpha$8 is indexed by a pair $\lambda_\alpha$9 and built from isomorphism classes of correspondences

$k_{IJ}$00

where $k_{IJ}$01 is proper, $k_{IJ}$02 is smooth or quasi-smooth, and $k_{IJ}$03 is a vector bundle on $k_{IJ}$04 (Yokura, 2022). Product is defined by fiber product and direct sum of pulled-back bundles, and the theory carries product, pushforward, pullback, units, base-change compatibilities, and a projection formula. The paper proves that

$k_{IJ}$05

so the theory is a correspondence-based bi-variant extension of Lee–Pandharipande cobordism of bundles and Levine–Morel algebraic cobordism (Yokura, 2022).

For correspondences of curves, the relevant invariant object is a differential form rather than a bundle class. A correspondence is a tuple

$k_{IJ}$06

with morphisms $k_{IJ}$07. A nonzero form $k_{IJ}$08 is invariant if $k_{IJ}$09 and semi-invariant if $k_{IJ}$10 for some $k_{IJ}$11 (Saha, 2012). The paper studies the associated group $k_{IJ}$12 of semi-invariant classes and the conductor

$k_{IJ}$13

When $k_{IJ}$14 and both maps are tamely ramified, Theorem 1.1 gives a bound on the conductor determined by the genera and the degrees, and Proposition 2.1 shows that $k_{IJ}$15 has rank at most $k_{IJ}$16 (Saha, 2012).

On $k_{IJ}$17, under the hypotheses that the maps are tamely ramified, conjugate to maps completely ramified at a point, and satisfy $k_{IJ}$18, Theorem 1.2 shows that any primitive semi-invariant form is flat: either a weight-$k_{IJ}$19 form of type $k_{IJ}$20 or a weight-$k_{IJ}$21 form of type $k_{IJ}$22 (Saha, 2012). In the number-field setting, Theorem 1.3 states that if $k_{IJ}$23 is non-trivial and generated by primitives of weight $k_{IJ}$24 for infinitely many places $k_{IJ}$25, then the pair is conjugate to a pair obtained by composing a common endomorphism with multiplicative maps (Saha, 2012). The same discussion situates these results alongside multiplicative, Chebyshev, and Lattès classifications. In the paper’s geometric interpretation, the support of $k_{IJ}$26 acts as the invariant skeleton of the correspondence; this is the sense in which the support pattern functions as the correspondence’s “axis” (Saha, 2012).

Across these literatures, “axes of correspondence” therefore denotes a family of structurally similar but mathematically distinct devices. In CA and MWCA, the axes are orthonormal factor directions and mode-wise principal components; in neural and lexical models, they are matched or ordered coordinate directions; in robotics, they are semantically meaningful control directions inferred from correspondences; in axial algebras, they are idempotent generators governed by fusion laws; and in algebraic geometry, the corresponding invariant may be a cobordism correspondence or the support of a semi-invariant form. This suggests a shared technical pattern: a correspondence becomes most informative when it can be resolved into explicit directional, eigenspace, or support data rather than treated as a purely existential relation.