Pfaffian Activation Functions

• Pfaffian activation functions are neural network activations expressed as Pfaffian functions, where outputs are polynomials in auxiliary smooth functions solving triangular differential equations.
• They unify common smooth nonlinearities—such as sigmoid, tanh, arctan, and exponential—thereby providing a cohesive framework for analyzing VC-dimension in graph neural networks and classifier robustness.
• Their explicit analytic structure supports rigorous proofs that link architecture parameters with capacity and robustness, offering actionable insights for designing deep neural networks.

Pfaffian activation functions are neural-network activation functions that are Pfaffian functions, namely functions represented by a polynomial in auxiliary smooth functions that satisfy a triangular system of first-order differential equations with polynomial right-hand sides. In the cited literature, this class is used as a unifying analytic framework for smooth nonlinearities such as the logistic sigmoid, hyperbolic tangent, arctangent, exponential, and polynomials, and it supports two distinct lines of analysis: VC-dimension bounds for message-passing graph neural networks and tube-volume and condition-number bounds for classifier robustness near decision boundaries (D'Inverno et al., 2024, Lezeau et al., 9 Jul 2026).

1. Formal definition and Pfaffian format

Let $U \subset \mathbb{R}^n$ be an open domain. A Pfaffian chain of length $\ell \ge 0$ and degree $\alpha \ge 1$ on $U$ is a sequence of $C^\infty$ functions

$f_1(x), \ldots, f_\ell(x) : U \to \mathbb{R}$

such that, for each $1 \le j \le \ell$,

$$d f_j(x) = \sum_{i=1}^n g_{ij}(x,f_1(x),\ldots,f_j(x))\,dx_i,$$

where each $g_{ij}(x,y_1,\ldots,y_j)$ is a polynomial in $(x,y_1,\ldots,y_j)$ of total degree $\ell \ge 0$0. Given such a chain, a function $\ell \ge 0$1 is Pfaffian of format $\ell \ge 0$2 if

$\ell \ge 0$3

where $\ell \ge 0$4 is a polynomial of degree $\ell \ge 0$5 (D'Inverno et al., 2024).

A notationally equivalent formulation uses a chain of order $\ell \ge 0$6 and chain-degree $\ell \ge 0$7, with functions $\ell \ge 0$8 satisfying

$\ell \ge 0$9

where each $\alpha \ge 1$0 is a real polynomial of total degree $\alpha \ge 1$1; then $\alpha \ge 1$2 is Pfaffian of format $\alpha \ge 1$3 if $\alpha \ge 1$4 has degree $\alpha \ge 1$5 (Lezeau et al., 9 Jul 2026).

The central structural parameters are the chain length, written as $\alpha \ge 1$6 or $\alpha \ge 1$7; the maximal degree $\alpha \ge 1$8 of the polynomial coefficients in the differential system; and the degree $\alpha \ge 1$9 of the terminal polynomial representation. In the neural-network applications considered in the cited work, these parameters become explicit complexity descriptors. This suggests that “Pfaffian activation function” refers less to a specific nonlinearity family than to an analytic representation class whose format can be propagated into statistical-capacity and robustness bounds.

2. Canonical activation functions in the Pfaffian class

Several standard activations admit low-complexity Pfaffian representations. The logistic sigmoid and hyperbolic tangent are singled out because each satisfies a first-order ODE whose right-hand side is polynomial in the function itself. The arctangent also belongs to the class, although its representation uses a chain of length $U$0 rather than $U$1 (D'Inverno et al., 2024, Lezeau et al., 9 Jul 2026).

Activation $U$2	Pfaffian chain	Format $U$3
Polynomial $U$4	none	$U$5
$U$6	$U$7	$U$8
$U$9	$C^\infty$0	$C^\infty$1
$C^\infty$2	$C^\infty$3	$C^\infty$4
$C^\infty$5	$C^\infty$6	$C^\infty$7

For the logistic sigmoid,

$C^\infty$8

so in differential form

$C^\infty$9

and $f_1(x), \ldots, f_\ell(x) : U \to \mathbb{R}$0 has degree $f_1(x), \ldots, f_\ell(x) : U \to \mathbb{R}$1. Hence $f_1(x), \ldots, f_\ell(x) : U \to \mathbb{R}$2 is Pfaffian of format $f_1(x), \ldots, f_\ell(x) : U \to \mathbb{R}$3. For the hyperbolic tangent,

$f_1(x), \ldots, f_\ell(x) : U \to \mathbb{R}$4

so

$f_1(x), \ldots, f_\ell(x) : U \to \mathbb{R}$5

and $f_1(x), \ldots, f_\ell(x) : U \to \mathbb{R}$6 is also Pfaffian of format $f_1(x), \ldots, f_\ell(x) : U \to \mathbb{R}$7 (D'Inverno et al., 2024).

The class is therefore broad enough to include most of the “engineer-friendly” smooth activations discussed in the source material. A common misconception is that the term is essentially synonymous with sigmoid-type nonlinearities; the available examples show otherwise, since polynomial, exponential, and arctangent activations are also explicitly listed as Pfaffian (Lezeau et al., 9 Jul 2026).

3. VC dimension in graph neural networks with Pfaffian activations

In the graph-learning setting, the cited work studies a generic message-passing GNN with $f_1(x), \ldots, f_\ell(x) : U \to \mathbb{R}$8 layers, hidden dimension $f_1(x), \ldots, f_\ell(x) : U \to \mathbb{R}$9, input graphs of at most $1 \le j \le \ell$0 nodes, node attributes of dimension $1 \le j \le \ell$1, and total number of real parameters $1 \le j \le \ell$2. The analysis is motivated by prior results for piecewise polynomial activations and extends them to smooth activations such as sigmoid and hyperbolic tangent by using Pfaffian function theory. The bounds are also tied to the Weisfeiler-Lehman framework: the paper denotes by $1 \le j \le \ell$3 the number of distinct initial colors in the $1 \le j \le \ell$4-WL initialization, and by $1 \le j \le \ell$5 the total number of node colors produced over all $1 \le j \le \ell$6 iterations of $1 \le j \le \ell$7-WL color refinement, summed across graphs in the training set (D'Inverno et al., 2024).

For the most general case, if every combine, every aggregate, and the final readout are Pfaffian functions of formats

$1 \le j \le \ell$8

then Theorem 1 states that

$1 \le j \le \ell$9

where

$$d f_j(x) = \sum_{i=1}^n g_{ij}(x,f_1(x),\ldots,f_j(x))\,dx_i,$$0

and $$d f_j(x) = \sum_{i=1}^n g_{ij}(x,f_1(x),\ldots,f_j(x))\,dx_i,$$1 is a bound on the number of connected components of a certain Pfaffian variety. By a Gabrielov–Vorobjov bound,

$$d f_j(x) = \sum_{i=1}^n g_{ij}(x,f_1(x),\ldots,f_j(x))\,dx_i,$$2

with

$$d f_j(x) = \sum_{i=1}^n g_{ij}(x,f_1(x),\ldots,f_j(x))\,dx_i,$$3

$$d f_j(x) = \sum_{i=1}^n g_{ij}(x,f_1(x),\ldots,f_j(x))\,dx_i,$$4

$$d f_j(x) = \sum_{i=1}^n g_{ij}(x,f_1(x),\ldots,f_j(x))\,dx_i,$$5

A specialization in which all Pfaffian degrees are bounded by $$d f_j(x) = \sum_{i=1}^n g_{ij}(x,f_1(x),\ldots,f_j(x))\,dx_i,$$6 yields, up to lower-order terms,

$$d f_j(x) = \sum_{i=1}^n g_{ij}(x,f_1(x),\ldots,f_j(x))\,dx_i,$$7

For the widely used “linear-combine + sum-aggregate + $$d f_j(x) = \sum_{i=1}^n g_{ij}(x,f_1(x),\ldots,f_j(x))\,dx_i,$$8” model with a single scalar activation $$d f_j(x) = \sum_{i=1}^n g_{ij}(x,f_1(x),\ldots,f_j(x))\,dx_i,$$9 of format $g_{ij}(x,y_1,\ldots,y_j)$0, Theorem 2 gives

$g_{ij}(x,y_1,\ldots,y_j)$1

where $g_{ij}(x,y_1,\ldots,y_j)$2 is an absolute constant for $g_{ij}(x,y_1,\ldots,y_j)$3, with $g_{ij}(x,y_1,\ldots,y_j)$4, or $g_{ij}(x,y_1,\ldots,y_j)$5, with $g_{ij}(x,y_1,\ldots,y_j)$6. In big-$g_{ij}(x,y_1,\ldots,y_j)$7 form, this is again $g_{ij}(x,y_1,\ldots,y_j)$8.

The color-refined formulation sharpens dependence on graph structure. By merging nodes of the same color, Theorem 3 yields

$g_{ij}(x,y_1,\ldots,y_j)$9

This suggests that, in these bounds, combinatorial compression through $(x,y_1,\ldots,y_j)$0-WL colors can be more informative than raw node count.

The theoretical analysis is supported by two prototype experiments on the TUDatasets PROTEINS, NCI1, and PTC-MR using arctan or tanh activations. As the hidden dimension $(x,y_1,\ldots,y_j)$1 or the number of layers $(x,y_1,\ldots,y_j)$2 grows, the training-test accuracy gap widens in a manner consistent with an $(x,y_1,\ldots,y_j)$3, $(x,y_1,\ldots,y_j)$4 rise in complexity. When graphs are partitioned by increasing ratio $(x,y_1,\ldots,y_j)$5, the same gap increases, matching the $(x,y_1,\ldots,y_j)$6 dependence. The reported test-training gap curves track the polynomial VC-dimension growth predicted by the theory (D'Inverno et al., 2024).

4. Proof architecture: from Pfaffian equations to capacity bounds

The proof strategy for the GNN VC-dimension bounds proceeds in three stages. First, the forward pass of the network is rephrased as the solution to a system of

$(x,y_1,\ldots,y_j)$7

equations of the form

$(x,y_1,\ldots,y_j)$8

where $(x,y_1,\ldots,y_j)$9 collects hidden-unit variables together with graph-input variables, and $\ell \ge 0$00 denotes the network parameters. Each $\ell \ge 0$01 is a Pfaffian function of some shared Pfaffian chain of length $\ell \ge 0$02 and degrees $\ell \ge 0$03 (D'Inverno et al., 2024).

Second, a theorem of Gabrielov–Vorobjov bounds the number of connected components $\ell \ge 0$04 of the solution set

$\ell \ge 0$05

in terms of $\ell \ge 0$06. This is the geometric core of the argument: the topology of a Pfaffian variety becomes a quantitative input to a statistical-capacity estimate.

Third, a result of Karpinski–Macintyre, also attributed in the paper to Scarselli–Tsoi–Hagenbuchner in this context, yields

$\ell \ge 0$07

Combining the connected-component bound with this sign-pattern bound produces the stated polynomial or near-polynomial dependencies on $\ell \ge 0$08, and on the color counts $\ell \ge 0$09 and $\ell \ge 0$10.

The significance of this proof architecture is methodological. The derivation does not depend on piecewise-polynomial partitioning; instead it replaces combinatorial region counting by Pfaffian geometry. This suggests a route for analyzing smooth activations while still obtaining explicit, architecture-level capacity bounds.

5. Tubular neighbourhoods and robustness of Pfaffian classifiers

A second research direction studies Pfaffian activations through the geometry of decision boundaries. Let $\ell \ge 0$11 be a bounded smooth hypersurface defined by a Pfaffian function $\ell \ge 0$12 of format $\ell \ge 0$13, with $\ell \ge 0$14 on $\ell \ge 0$15. For a point $\ell \ge 0$16, radius $\ell \ge 0$17, and a random point $\ell \ge 0$18 uniform in the Euclidean ball $\ell \ge 0$19, writing $\ell \ge 0$20, Theorem 3.3 states

$\ell \ge 0$21

where

$\ell \ge 0$22

A weaker form drops the negative term:

$\ell \ge 0$23

(Lezeau et al., 9 Jul 2026).

These tube-volume bounds are then transferred to classification robustness. For a classifier $\ell \ge 0$24, define

$\ell \ge 0$25

so that the decision boundary $\ell \ge 0$26 satisfies

$\ell \ge 0$27

The local condition number at $\ell \ge 0$28 is

$\ell \ge 0$29

If each $\ell \ge 0$30 is a bounded smooth Pfaffian hypersurface of format $\ell \ge 0$31, then for $\ell \ge 0$32 uniform in $\ell \ge 0$33, Theorem 5.1 gives

$\ell \ge 0$34

Theorem 5.3 states the same bound for Gaussian input $\ell \ge 0$35:

$\ell \ge 0$36

The conclusions drawn in the source are explicit. The tube-volume bound shows that the probability that a random point lies within $\ell \ge 0$37 of a Pfaffian decision boundary decays polynomially in $\ell \ge 0$38 with coefficients depending on the format $\ell \ge 0$39, hence on network depth, activation complexity, and number of neurons. The uniform and Gaussian tail bounds quantify robustness by bounding the chance that a relative perturbation of size $\ell \ge 0$40 suffices to misclassify. For deep networks with Pfaffian activations, the worst-case prefactor $\ell \ge 0$41 grows exponentially in the chain length $\ell \ge 0$42, hence in the total hidden units $\ell \ge 0$43, reflecting the complexity of fewnomial intersection theory (Lezeau et al., 9 Jul 2026).

6. Single-hidden-layer sigmoid networks and scope of the theory

The robustness analysis has a sharper special case for single-hidden-layer sigmoid networks with rational weights. Consider

$\ell \ge 0$44

where $\ell \ge 0$45 is logistic, $\ell \ge 0$46 have common denominator $\ell \ge 0$47, and

$\ell \ge 0$48

If $\ell \ge 0$49 is smooth and contained in $\ell \ge 0$50, then Theorem 4.1 and Section 4.4 show that the section-degree of $\ell \ge 0$51, described as the maximal Gauss-map degree, satisfies

$\ell \ge 0$52

and that all tube-section degrees $\ell \ge 0$53 can be bounded by

$\ell \ge 0$54

Hence, for $\ell \ge 0$55 uniform in $\ell \ge 0$56, Theorem 4.4 yields

$\ell \ge 0$57

and Corollary 5.5 gives

$\ell \ge 0$58

Moreover, Lemma 4.9 states that the leading-order term is governed by line intersections,

$\ell \ge 0$59

which yields the sharp regime

$\ell \ge 0$60

(Lezeau et al., 9 Jul 2026).

Within the available sources, this special case marks an important distinction. In the general deep Pfaffian setting, the prefactor in the robustness bound can grow exponentially with chain length. In the single-hidden-layer rational-weights sigmoid setting, that fewnomial factor is replaced by a polynomial $\ell \ge 0$61, and the leading tail sharpens to $\ell \ge 0$62. A plausible implication is that Pfaffian analysis separates two phenomena that are often conflated: the analytic tractability of the activation itself and the geometric complexity induced by depth and hidden-unit proliferation.

The same source formulates these implications for network design directly. Limiting depth, choosing activations of low Pfaffian format, or restricting to single-layer rational-weight sigmoid networks yields better provable robustness guarantees under random or adversarial perturbations. Together with the GNN VC-dimension results, this places Pfaffian activation functions at an interface between o-minimal-style geometric control and neural-network learning theory: the activation is smooth, but its differential-algebraic format remains sufficiently explicit to support nonasymptotic bounds on both capacity and robustness (D'Inverno et al., 2024, Lezeau et al., 9 Jul 2026).