Generalized T-Permutant Measures

• Generalized T-permutant measures are finite signed measures on X^Y that remain invariant under the group action induced by a homomorphism T.
• They provide an algebraic framework to construct equivariant non-expansive operators (GENEOs), transporting source symmetries to target domains.
• This framework extends classical permutant measures to heterogeneous data spaces, enabling applications in graph processing and geometric machine learning.

Searching arXiv for the cited papers to ground the article in the relevant literature. arXiv search query: (Ahmad et al., 2022) Generalized T-permutant measures are finite signed measures on the function set $X^Y$ that are invariant under the action induced by a homomorphism $T:G\to K$, where $G\subseteq \mathrm{Aut}(X)$ and $K\subseteq \mathrm{Aut}(Y)$. They were introduced in order to construct Group Equivariant Non-Expansive Operators (GENEOs) between graphs and, more generally, between heterogeneous perception pairs $(\mathbb R^X,G)$ and $(\mathbb R^Y,K)$. In this setting, they provide the algebraic device that transports source symmetries to target symmetries and yields equivariant linear operators, with non-expansivity enforced by a total-variation bound (Ahmad et al., 2022, Conti et al., 7 Jan 2026).

1. Definition and invariance

For finite nonempty sets $X$ and $Y$, a subgroup $G$ of $\mathrm{Aut}(X)$, a subgroup $T:G\to K$0 of $T:G\to K$1, and a homomorphism $T:G\to K$2, a finite signed measure $T:G\to K$3 on $T:G\to K$4 is called a generalized permutant measure with respect to $T:G\to K$5 if each subset $T:G\to K$6 of $T:G\to K$7 is measurable and

$T:G\to K$8

Equivalently, in the formulation used later for the representation theorem, $T:G\to K$9 is a finite signed measure on the power set $G\subseteq \mathrm{Aut}(X)$0 such that

$G\subseteq \mathrm{Aut}(X)$1

or, at the level of singletons,

$G\subseteq \mathrm{Aut}(X)$2

The underlying action is

$G\subseteq \mathrm{Aut}(X)$3

that is,

$G\subseteq \mathrm{Aut}(X)$4

Accordingly, a generalized T-permutant measure is constant on the $G\subseteq \mathrm{Aut}(X)$5-orbits of this action (Ahmad et al., 2022).

2. The parameter $G\subseteq \mathrm{Aut}(X)$6 and heterogeneous symmetry transport

In the generalized setting one works with two possibly different domains of measurement,

$G\subseteq \mathrm{Aut}(X)$7

together with a group homomorphism

$G\subseteq \mathrm{Aut}(X)$8

The role of $G\subseteq \mathrm{Aut}(X)$9 is to encode how symmetries of the source domain $K\subseteq \mathrm{Aut}(Y)$0 are transported to symmetries of the target domain $K\subseteq \mathrm{Aut}(Y)$1. If $K\subseteq \mathrm{Aut}(Y)$2 is a symmetry of $K\subseteq \mathrm{Aut}(Y)$3, then $K\subseteq \mathrm{Aut}(Y)$4 is the corresponding symmetry of $K\subseteq \mathrm{Aut}(Y)$5. The invariance condition

$K\subseteq \mathrm{Aut}(Y)$6

states that, for each measurable subset $K\subseteq \mathrm{Aut}(Y)$7, pushing its elements $K\subseteq \mathrm{Aut}(Y)$8 forward by

$K\subseteq \mathrm{Aut}(Y)$9

leaves the total signed measure unchanged. In this sense, $(\mathbb R^X,G)$0 determines the precise right action needed to compare orbits in $(\mathbb R^X,G)$1. The later algebraic theory adds a structural assumption: if the action of $(\mathbb R^X,G)$2 on $(\mathbb R^X,G)$3 is transitive, then linear GENEOs admit a complete representation by generalized T-permutant measures; if it is not transitive, one works orbit by orbit (Conti et al., 7 Jan 2026).

3. From generalized T-permutant measures to linear GENEOs

The basic construction associates to a generalized permutant measure $(\mathbb R^X,G)$4 the operator

$(\mathbb R^X,G)$5

This map is a linear GEO from $(\mathbb R^X,G)$6 to $(\mathbb R^X,G)$7 with respect to $(\mathbb R^X,G)$8. If, in addition,

$(\mathbb R^X,G)$9

then $(\mathbb R^Y,K)$0 is non-expansive and hence a GENEO. The proof uses linearity, the change of variable $(\mathbb R^Y,K)$1, and the estimate

$(\mathbb R^Y,K)$2

(Ahmad et al., 2022).

A later representation theorem turns this construction into a characterization. Assume that $(\mathbb R^Y,K)$3 is a homomorphism and that $(\mathbb R^Y,K)$4 acts transitively on the finite set $(\mathbb R^Y,K)$5. Then $(\mathbb R^Y,K)$6 is a linear GENEO

$(\mathbb R^Y,K)$7

if and only if there exists a generalized $(\mathbb R^Y,K)$8-permutant measure $(\mathbb R^Y,K)$9 on $X$0 with

$X$1

such that

$X$2

The proof outline is explicitly algebraic. One writes $X$3 as a matrix $X$4, derives from $X$5-equivariance that

$X$6

decomposes $X$7, applies the Birkhoff–von Neumann theorem to the positive and negative parts, and averages the resulting coefficients over the $X$8-orbits in $X$9. The norm constraint is exact: $Y$0 Thus non-expansivity is equivalent to the total variation bound on $Y$1 (Conti et al., 7 Jan 2026).

4. Extension of the classical permutant formalism

The generalized theory extends the classical notion of permutant measure in three explicit directions. In the classical theory one works only with

$Y$2

and $Y$3 is a measure on $Y$4. The generalized version instead allows $Y$5, allows a non-trivial homomorphism $Y$6, and defines $Y$7 on the full function set $Y$8 rather than on a subgroup of bijections (Ahmad et al., 2022).

This enlargement addresses a concrete limitation of earlier representation results for linear GENEOs, which characterized operators acting on data of the same type. The generalized theory is designed for operators between different perception pairs and therefore accommodates heterogeneous data spaces. The stated applications include passage from “vertex-weighted graphs” to “edge-weighted graphs,” and from “large” graphs to “small auxiliary” graphs, as in the $Y$9 construction. A common misconception is that permutant-based representations are intrinsically tied to bijections or to source and target spaces of the same kind; the generalized framework explicitly removes both restrictions (Conti et al., 7 Jan 2026).

5. Graph GENEOs and canonical examples

In the graph setting, one typically takes

$G$0

with $G$1, $G$2, and groups $G$3, $G$4. A graph-permutant measure $G$5 on $G$6 then yields the GENEO

$G$7

When $G$8 is supported on a small orbit, one obtains a very efficient equivariant operator (Ahmad et al., 2022).

A first explicit example is the $G$9-graph construction on $\mathrm{Aut}(X)$0 with $\mathrm{Aut}(X)$1. Let $\mathrm{Aut}(X)$2 be the permutant of all edge-swaps induced by vertex transpositions, and define the uniform measure

$\mathrm{Aut}(X)$3

Then

$\mathrm{Aut}(X)$4

is exactly the GENEO $\mathrm{Aut}(X)$5 of Section 5.1 whose action on $\mathrm{Aut}(X)$6-$\mathrm{Aut}(X)$7 edge-vectors of $\mathrm{Aut}(X)$8 produces the so-called “$\mathrm{Aut}(X)$9-codes.” In the subgraph encoding $T:G\to K$00, one has a permutant $T:G\to K$01 consisting of the six transpositions of the vertex set and

$T:G\to K$02

The paper reports that isomorphic subgraphs have codes equal up to permutation.

A second example concerns cycle graphs $T:G\to K$03. Here

$T:G\to K$04

For each base map $T:G\to K$05, one considers its orbit $T:G\to K$06 under the $T:G\to K$07-action

$T:G\to K$08

The uniform measure on $T:G\to K$09 yields the GENEO

$T:G\to K$10

The orbit sizes computed in the paper are $T:G\to K$11. These examples exhibit the intended role of generalized T-permutant measures: they are not only existence devices for equivariant operators, but also concrete combinatorial objects from which graph operators can be built directly.

6. Orbit structure, polytope geometry, and machine-learning use

The orbit decomposition of $T:G\to K$12 under $T:G\to K$13 controls the linear theory. If $T:G\to K$14 are the $T:G\to K$15 orbits of the action, the corresponding measures

$T:G\to K$16

form a basis of the real vector space of generalized $T:G\to K$17-permutant measures. Consequently, linear GENEOs can be written as

$T:G\to K$18

where

$T:G\to K$19

The space $T:G\to K$20 of all linear GENEOs is therefore a compact convex polytope in $T:G\to K$21, namely

$T:G\to K$22

This identifies the generalized T-permutant formalism with a finite-dimensional convex geometry of equivariant non-expansive linear maps (Conti et al., 7 Jan 2026).

The same paper gives an explicit application to autoencoder preprocessing. Let

$T:G\to K$23

with translation groups

$T:G\to K$24

For a unit-vector $T:G\to K$25 with $T:G\to K$26 modulo $T:G\to K$27, define

$T:G\to K$28

For each $T:G\to K$29, define

$T:G\to K$30

The set $T:G\to K$31 is invariant under $T:G\to K$32 and forms a single orbit $T:G\to K$33. The measure

$T:G\to K$34

is a generalized $T:G\to K$35-permutant measure, and

$T:G\to K$36

is a linear GENEO, equivariant to translations and non-expansive. In practice, one generates a family of such GENEOs, one for each “direction” $T:G\to K$37, and uses their outputs as preprocessed features before feeding data into a standard convolutional autoencoder. Experimental results on MNIST with salt-and-pepper noise show that the GENEO-enhanced autoencoder preserves classification accuracy and reduces reconstruction error much better than an ordinary autoencoder. This suggests that generalized T-permutant measures are not only a representation-theoretic tool but also a usable design principle for equivariant preprocessing in geometric machine learning.