Möbius Convolutions: Structures & Applications

• Möbius convolutions are algebraic and geometric operators that aggregate over admissible intermediates in structured spaces like posets, Boolean lattices, catoids, and spheres.
• They enable the derivation of convolution identities and practical inversion formulas in incidence algebras, hyperplane arrangements, and matroid invariants.
• Applications span algorithm acceleration using fast subset convolution and designing equivariant spherical CNNs in geometric deep learning.

In the cited literature, Möbius-related convolutions arise in several technically distinct forms: convolution in incidence algebras of locally finite posets, convolution semirings on generalized Möbius categories, subset convolution accelerated through zeta and Möbius transforms on the Boolean lattice, and a Möbius-equivariant spherical convolution operator on $S^2$ (Wang, 2012, Cranch et al., 29 Aug 2025, Stoian, 2024, Mitchel et al., 2022). The common thread is the organization of convolution by an underlying Möbius-theoretic or Möbius-geometric structure, but the ambient objects, invariance principles, and algebraic goals differ substantially.

1. Scope of the term

The phrase “Möbius convolution” is not attached to a single canonical construction in current arXiv usage. In combinatorics and incidence algebra, it refers to convolution formulae derived from the Möbius function of a locally finite poset and to the Möbius conjugation map $\mu^\ast$ (Wang, 2012). In categorical algebra, it appears in the extension of convolution semirings and Kleene algebras to generalized Möbius categories, formulated as Möbius catoids (Cranch et al., 29 Aug 2025). In algorithm design, Möbius transforms are the inverse transforms underlying fast subset convolution on the Boolean algebra of subsets (Stoian, 2024). In geometric deep learning, Möbius convolution denotes an $SL(2,\mathbb C)$-equivariant spherical convolution operator on the Riemann sphere (Mitchel et al., 2022).

Setting	Core structure	Convolutional object
Incidence algebra	Locally finite poset $P$ and $\Int(P)$	$(a\ast b)(x,y)=\sum_{x\le z\le y} a(x,z)b(z,y)$
Generalised Möbius categories	Möbius catoid $(C,\odot,s,t)$	$(f\ast g)(x)=\sum_{x\in y\odot z} f(y)\cdot g(z)$
Boolean lattice algorithms	Set-functions on $2^{[n]}$	$$(f\star g)(S)=\sum_{T\subseteq S} f(T)g(S\setminus T)$$
Spherical CNNs	$\mu^\ast$0 with $\mu^\ast$1 action	$\mu^\ast$2

This suggests that the term is best understood as a family resemblance rather than a unique definition. The constructions share an intermediate-summation or intermediate-factorization pattern, but they are not interchangeable.

2. Incidence-algebra Möbius conjugation on posets

Let $\mu^\ast$3 be a locally finite poset and $\mu^\ast$4 a commutative ring with $\mu^\ast$5. The interval space is

$\mu^\ast$6

and the incidence algebra $\mu^\ast$7 consists of functions $\mu^\ast$8 with convolution

$\mu^\ast$9

The zeta function is $SL(2,\mathbb C)$0 for $SL(2,\mathbb C)$1, and its inverse in $SL(2,\mathbb C)$2 is the Möbius function $SL(2,\mathbb C)$3, characterized by

$SL(2,\mathbb C)$4

The ring $SL(2,\mathbb C)$5 of pointwise functions $SL(2,\mathbb C)$6 embeds into $SL(2,\mathbb C)$7 by the diagonal map $SL(2,\mathbb C)$8, and Wang defines the Möbius conjugation

$SL(2,\mathbb C)$9

Its fundamental property is Theorem 1.1:

$P$0

and $P$1 is injective, hence a ring-monomorphism $P$2 (Wang, 2012).

This formulation packages Möbius inversion into an algebra homomorphism. The cited paper presents the usual zeta–Möbius inversion as a corollary obtained by taking $P$3, so the classical inversion formula is recovered from the multiplicativity of $P$4 rather than introduced independently (Wang, 2012). A common misconception is to identify this construction with geometric Möbius transforms; here “Möbius” refers to the incidence-theoretic Möbius function of a poset.

3. Hyperplane arrangements and characteristic-polynomial convolution

A principal application of Möbius conjugation in Wang’s treatment is the intersection poset of a hyperplane arrangement $P$5. For a finite arrangement in a real or complex vector space $P$6 of dimension $P$7, the intersection poset $P$8 consists of all nonempty intersections of members of $P$9, ordered by reverse inclusion, with minimal element $\Int(P)$0; adjoining $\Int(P)$1 gives the reduced lattice $\Int(P)$2. The characteristic polynomial is

$\Int(P)$3

For $\Int(P)$4 in $\Int(P)$5, the local arrangement inside $\Int(P)$6 is

$\Int(P)$7

Theorem 2.1 gives the convolution identity

$\Int(P)$8

and, in particular,

$\Int(P)$9

The proof is obtained by taking $(a\ast b)(x,y)=\sum_{x\le z\le y} a(x,z)b(z,y)$0 and the functions $(a\ast b)(x,y)=\sum_{x\le z\le y} a(x,z)b(z,y)$1 and $(a\ast b)(x,y)=\sum_{x\le z\le y} a(x,z)b(z,y)$2, then applying $(a\ast b)(x,y)=\sum_{x\le z\le y} a(x,z)b(z,y)$3 (Wang, 2012).

In the real case $(a\ast b)(x,y)=\sum_{x\le z\le y} a(x,z)b(z,y)$4, the number of regions

$(a\ast b)(x,y)=\sum_{x\le z\le y} a(x,z)b(z,y)$5

and the number of relatively bounded regions

$(a\ast b)(x,y)=\sum_{x\le z\le y} a(x,z)b(z,y)$6

satisfy Zaslavsky’s formulas

$(a\ast b)(x,y)=\sum_{x\le z\le y} a(x,z)b(z,y)$7

Substituting these into the characteristic-polynomial convolution yields

$(a\ast b)(x,y)=\sum_{x\le z\le y} a(x,z)b(z,y)$8

and

$(a\ast b)(x,y)=\sum_{x\le z\le y} a(x,z)b(z,y)$9

For integral arrangements, the same framework yields a reciprocity theorem for $(C,\odot,s,t)$0: if $(C,\odot,s,t)$1 is a large prime and $(C,\odot,s,t)$2 is the reduction mod $(C,\odot,s,t)$3, then Athanasiadis’s theorem gives

$(C,\odot,s,t)$4

and Wang derives

$(C,\odot,s,t)$5

The paper further states that $(C,\odot,s,t)$6 equals the total number of integer points lying in the closures of the regions of the periodic arrangement

$(C,\odot,s,t)$7

inside the central cube $(C,\odot,s,t)$8 (Wang, 2012).

4. Matroids, Boolean lattices, and subset convolution

The same poset-based mechanism specializes to the Boolean lattice $(C,\odot,s,t)$9 and produces convolution identities for matroids. For a matroid $(f\ast g)(x)=\sum_{x\in y\odot z} f(y)\cdot g(z)$0 on ground set $(f\ast g)(x)=\sum_{x\in y\odot z} f(y)\cdot g(z)$1 with rank function $(f\ast g)(x)=\sum_{x\in y\odot z} f(y)\cdot g(z)$2, the rank-generating function and Tutte polynomial are

$(f\ast g)(x)=\sum_{x\in y\odot z} f(y)\cdot g(z)$3

On the Boolean lattice under inclusion, the classical Möbius function is $(f\ast g)(x)=\sum_{x\in y\odot z} f(y)\cdot g(z)$4. Wang shows that, by choosing

$(f\ast g)(x)=\sum_{x\in y\odot z} f(y)\cdot g(z)$5

one obtains the Kook–Reiner–Stanton convolution formula

$(f\ast g)(x)=\sum_{x\in y\odot z} f(y)\cdot g(z)$6

equivalently,

$(f\ast g)(x)=\sum_{x\in y\odot z} f(y)\cdot g(z)$7

The same method also recovers Kung’s multiplicative convolution identities for the subset-corank polynomial $(f\ast g)(x)=\sum_{x\in y\odot z} f(y)\cdot g(z)$8 and mixed “xy” convolution identities (Wang, 2012).

A closely related algorithmic setting is subset convolution of set-functions on $(f\ast g)(x)=\sum_{x\in y\odot z} f(y)\cdot g(z)$9:

$2^{[n]}$0

The zeta transform is

$2^{[n]}$1

and the Möbius transform, inverse to the zeta transform, is

$2^{[n]}$2

Björklund, Husfeldt, Kaski and Koivisto gave an $2^{[n]}$3-time evaluation by ranking functions by cardinality, applying $2^{[n]}$4 zeta transforms, performing pointwise products, and applying $2^{[n]}$5 Möbius inversions. Stoian’s FFT-based algorithm “completely eliminates the need for set function transforms and maintains the running time of the original algorithm,” again obtaining overall $2^{[n]}$6 time (Stoian, 2024). The paper attributes the practical limitations of the classical approach to “very large intermediate partial sums” and “severe floating-point cancellation errors,” since every Möbius inversion involves alternating additions and subtractions over a height-$2^{[n]}$7 lattice (Stoian, 2024).

The Boolean-lattice case therefore links structural combinatorics and fast algorithms. In one direction, Möbius-theoretic identities yield convolution formulae for Tutte-type invariants; in the other, Möbius inversion becomes an implementation bottleneck that can be replaced by FFT without changing asymptotic complexity.

5. Generalised Möbius categories and convolution Kleene algebras

The categorical generalization replaces intervals in a poset by arrows in a catoid. A catoid is a quadruple $2^{[n]}$8 where $2^{[n]}$9 is a set, $$(f\star g)(S)=\sum_{T\subseteq S} f(T)g(S\setminus T)$$0 is an associative multi-operation, and $$(f\star g)(S)=\sum_{T\subseteq S} f(T)g(S\setminus T)$$1 satisfy

$$(f\star g)(S)=\sum_{T\subseteq S} f(T)g(S\setminus T)$$2

An $$(f\star g)(S)=\sum_{T\subseteq S} f(T)g(S\setminus T)$$3-decomposition of $$(f\star g)(S)=\sum_{T\subseteq S} f(T)g(S\setminus T)$$4 is a list $$(f\star g)(S)=\sum_{T\subseteq S} f(T)g(S\setminus T)$$5 of non-identity arrows with

$$(f\star g)(S)=\sum_{T\subseteq S} f(T)g(S\setminus T)$$6

and the length is

$$(f\star g)(S)=\sum_{T\subseteq S} f(T)g(S\setminus T)$$7

A catoid is Möbius iff every arrow admits only finitely many full decompositions; equivalently, it is finitely $$(f\star g)(S)=\sum_{T\subseteq S} f(T)g(S\setminus T)$$8-decomposable, each identity arrow is indecomposable, and the two incidence conditions

$$(f\star g)(S)=\sum_{T\subseteq S} f(T)g(S\setminus T)$$9

hold (Cranch et al., 29 Aug 2025).

Given a finitely $\mu^\ast$00-decomposable catoid and a semiring $\mu^\ast$01, convolution on $\mu^\ast$02 is defined by

$\mu^\ast$03

Theorem 4.1 states that $\mu^\ast$04 is again a semiring, where

$\mu^\ast$05

is the indicator of the identities. If $\mu^\ast$06 is a Kleene algebra and $\mu^\ast$07 is Möbius, a recursive star is defined on $\mu^\ast$08 by

$\mu^\ast$09

for identities $\mu^\ast$10, and for non-identities $\mu^\ast$11,

$\mu^\ast$12

Theorem 5.3 asserts that

$\mu^\ast$13

is itself a Kleene algebra (Cranch et al., 29 Aug 2025).

The examples include the free monoid $\mu^\ast$14, the shuffle catoid on $\mu^\ast$15, the interval category $\mu^\ast$16 of a locally finite poset $\mu^\ast$17, path categories of finite directed graphs, guarded-string categories, and higher categories (Cranch et al., 29 Aug 2025). The paper also makes an explicit conceptual distinction: classical Möbius inversion in incidence algebras is convolution invertibility of the zeta-function, whereas the Kleene star is “a genuine ‘closure’ operator, not an inverse.” This addresses a frequent confusion when incidence algebras, semirings, and program-logical semantics are discussed under the same convolutional vocabulary (Cranch et al., 29 Aug 2025).

6. Möbius-equivariant spherical convolution in geometric deep learning

In geometric deep learning, Möbius convolution has a different meaning. The relevant Möbius transformations are the conformal automorphisms of the Riemann sphere. Identifying the unit sphere $\mu^\ast$18 with $\mu^\ast$19 by stereographic projection $\mu^\ast$20, the group $\mu^\ast$21 acts by

$\mu^\ast$22

with $\mu^\ast$23. Under this action, the round metric and area form transform conformally with scale factor

$\mu^\ast$24

For a multi-channel signal $\mu^\ast$25 and a multi-channel filter $\mu^\ast$26, the paper defines a Möbius-equivariant spherical convolution by

$\mu^\ast$27

where $\mu^\ast$28 is a lower-triangular frame operator, $\mu^\ast$29 is a density, and equivariance follows if

$\mu^\ast$30

A key observation is that although $\mu^\ast$31 is $\mu^\ast$32-dimensional, the relative transformation

$\mu^\ast$33

always lies in the $\mu^\ast$34-dimensional lower-triangular subgroup $\mu^\ast$35, so filter transforms need only be evaluated under $\mu^\ast$36 rather than the full group (Mitchel et al., 2022).

To compute these convolutions at scale, the paper develops a spectral-domain approximation of transformed filters. A family of log-polar basis functions on $\mu^\ast$37 is chosen:

$\mu^\ast$38

and a band-limited filter is expanded as

$\mu^\ast$39

After truncation and quadrature, the implementation uses the fast Spherical Harmonic Transform. The stated costs are $\mu^\ast$40 for FSHT and inverse FSHT, “practically $\mu^\ast$41 on GPU,” a one-time precompute of the $\mu^\ast$42 tensors costing $\mu^\ast$43, and each identity convolution reduction costing $\mu^\ast$44; with $\mu^\ast$45, $\mu^\ast$46, and $\mu^\ast$47 up to $\mu^\ast$48, this is reported as practical (Mitchel et al., 2022).

Within a spherical CNN, a typical block is

$\mu^\ast$49

with conformal filter-response normalization

$\mu^\ast$50

and pointwise activation

$\mu^\ast$51

Each filter $\mu^\ast$52 has $\mu^\ast$53 real coefficients $\mu^\ast$54; with $\mu^\ast$55, that is $\mu^\ast$56 parameters per channel pair (Mitchel et al., 2022).

The reported empirical results are specific. On genus-zero shape classification using SHREC ’11 with $\mu^\ast$57 classes and conformal augmentation by random Möbius transforms, a single MCResNet block with $\mu^\ast$58 channels and $\mu^\ast$59 achieves $\mu^\ast$60 accuracy on original data and $\mu^\ast$61 on conformally deformed data. The cited baselines are DiffusionNet at $\mu^\ast$62, Field Convolutions at $\mu^\ast$63, and Rotation-equiv. CubeNet at $\mu^\ast$64. On omni-directional image segmentation using Stanford 2D3DS, a U-Net style model with MCResNet encoder achieves average per-class accuracy / IoU of $\mu^\ast$65, compared with SWSCNN at $\mu^\ast$66, SphCNN at $\mu^\ast$67, and spatial rotational nets such as CubeNet at $\mu^\ast$68 (Mitchel et al., 2022).

7. Conceptual distinctions and recurrent confusions

The cited work supports several distinctions that are easy to blur if the term is treated as uniform. First, Möbius inversion in incidence algebras and Möbius-equivariance under $\mu^\ast$69 are unrelated notions of “Möbius”: the former is combinatorial and order-theoretic, the latter conformal and geometric (Wang, 2012, Mitchel et al., 2022). Second, in generalized Möbius categories, the Kleene star is explicitly not the inverse of the zeta function; it is a closure operation defined recursively from finite decomposability (Cranch et al., 29 Aug 2025). Third, fast subset convolution uses zeta and Möbius transforms on the Boolean lattice, but Stoian’s contribution is precisely to avoid explicit set-function transforms while preserving the $\mu^\ast$70 bound (Stoian, 2024).

At the same time, these constructions exhibit a structural parallel. In each case, convolution aggregates over admissible intermediates: $\mu^\ast$71 in intervals of a poset, $\mu^\ast$72 in a catoid, $\mu^\ast$73 in subset convolution, and integration over $\mu^\ast$74 with local frame transport in the spherical setting. This suggests that Möbius-related convolutions are unified less by a single formula than by a recurring principle: algebraic or geometric structure restricts the admissible decompositions over which convolution is formed.