Identity Sheaf Network

• Identity Sheaf Network is a simplification of sheaf neural networks that fixes every restriction map to the identity, removing learnable sheaf components.
• It retains a sheaf-based message passing architecture and achieves comparable results on standard heterophilic benchmarks such as Texas and Wisconsin.
• Empirical findings challenge the diffusion-based oversmoothing theory, suggesting that trivial restriction maps may suffice in certain graph learning tasks.

Identity Sheaf Network (ISN) is a deliberately simple ablation of Sheaf Neural Networks (SNNs) in which the sheaf-based architecture is retained but every restriction map is fixed to the identity, removing the key learned ingredient of the sheaf Laplacian (Caralt et al., 5 Mar 2026). It was introduced to test whether the empirical gains claimed for sheaf-learning models on heterophilous graphs actually derive from learning nontrivial restriction maps, or whether much of the benefit is already obtainable from a substantially simpler construction. In the formulation studied in "On the Necessity of Learnable Sheaf Laplacians" (Caralt et al., 5 Mar 2026), ISN is competitive with a range of SNN variants on standard heterophilic benchmarks, and the trained-model behavior does not support the diffusion-based oversmoothing narrative as strongly as prior sheaf-theoretic motivation would suggest.

1. Graph diffusion, oversmoothing, and the sheaf-theoretic premise

The starting point for ISN is the standard oversmoothing narrative for graph neural networks. For a graph $G=(V,E)$ with adjacency matrix $A$, degree matrix $D$, and node features $X\in\mathbb{R}^{n\times f}$, a GCN layer is written as

$$H^{(t+1)}=\sigma\!\left(D^{-1/2}AD^{1/2}H^{(t)}W_t\right).$$

Stacking layers is commonly interpreted as an Euler discretization of the graph heat equation

$\frac{dX(t)}{dt}=-\Delta_G X(t),$

where

$\Delta_G = I - D^{-1/2}AD^{1/2}.$

In the long-time limit,

$$\lim_{t\to\infty}X(t)\in \ker(\Delta_G)=\{x_u=x_v\mid (u,v)\in E\}.$$

This is the canonical oversmoothing phenomenon: features converge to constants on connected components (Caralt et al., 5 Mar 2026).

SNNs were introduced as an extension of Graph Convolutional Networks to address oversmoothing on heterophilous graphs by attaching a sheaf to the input graph and replacing the adjacency-based operator with a sheaf Laplacian defined by learnable restriction maps (Caralt et al., 5 Mar 2026). A cellular sheaf $F$ on $G$ assigns a stalk $A$0 to each node $A$1 and $A$2 to each edge $A$3, together with restriction maps

$A$4

for every incidence $A$5. These restriction maps are the crucial learnable objects in standard SNNs: instead of passing features directly along graph edges, node representations are first mapped into edge spaces and compared there.

The resulting sheaf Laplacian is defined through the coboundary operator $A$6, with

$A$7

and

$A$8

Equivalently,

$A$9

For sheaf diffusion,

$D$0

the limit lies in

$D$1

The theoretical justification is therefore that suitably learned non-identity restriction maps can shape the kernel of $D$2 so that adjacent nodes need not collapse to identical node-space representations, which is particularly relevant on heterophilous graphs (Caralt et al., 5 Mar 2026).

2. Formal definition of the Identity Sheaf Network

The central move of ISN is to test whether the additional complexity of learning restriction maps is necessary. In the notation of the paper, an Identity Sheaf is a sheaf $D$3 with

$D$4

and the collection of such sheaves is denoted $D$5. An Identity Sheaf Network is then

$D$6

Thus, the sheaf-based architecture is preserved while the learned sheaf component is removed (Caralt et al., 5 Mar 2026).

The SNN update used in the paper is

$D$7

where $D$8 denotes the learned sheaf at layer $D$9, and the restriction maps are computed via an MLP on concatenated node features: $X\in\mathbb{R}^{n\times f}$0 with additional hyperparameters that can fix one diagonal element of $X\in\mathbb{R}^{n\times f}$1 to $X\in\mathbb{R}^{n\times f}$2 or $X\in\mathbb{R}^{n\times f}$3. By contrast, ISN sets these maps to the identity and therefore removes this learned sheaf parameterization entirely.

The paper emphasizes that this makes the model essentially equivalent to a GIN-like architecture, except for the way the linear maps $X\in\mathbb{R}^{n\times f}$4 are applied; it is described as a sparser linear layer (Caralt et al., 5 Mar 2026). The purpose is explicitly ablative: if ISN performs similarly to full SNNs, then the empirical value of learned restriction maps is questionable. A plausible implication is that any observed advantage of sheaf-based message passing may not isolate the effect of learning a nontrivial sheaf Laplacian.

3. Heterophily regime and benchmark characterization

To explain why ISN can still perform strongly, the paper invokes the heterophily characterization of Wang et al. Using the gain statistic, for classes $X\in\mathbb{R}^{n\times f}$5,

$X\in\mathbb{R}^{n\times f}$6

where $X\in\mathbb{R}^{n\times f}$7 records the class composition of the neighborhood of class $X\in\mathbb{R}^{n\times f}$8, and $X\in\mathbb{R}^{n\times f}$9 is the average degree of class $$H^{(t+1)}=\sigma\!\left(D^{-1/2}AD^{1/2}H^{(t)}W_t\right).$$0. With threshold $$H^{(t+1)}=\sigma\!\left(D^{-1/2}AD^{1/2}H^{(t)}W_t\right).$$1, a dataset is called “good” heterophily if $$H^{(t+1)}=\sigma\!\left(D^{-1/2}AD^{1/2}H^{(t)}W_t\right).$$2, “bad” if $$H^{(t+1)}=\sigma\!\left(D^{-1/2}AD^{1/2}H^{(t)}W_t\right).$$3, and otherwise mixed (Caralt et al., 5 Mar 2026).

All five benchmarks used in the study are reported to be in the good heterophily regime: Texas, Wisconsin, Squirrel, Chameleon, and Cornell. Their reported minimum gains are $$H^{(t+1)}=\sigma\!\left(D^{-1/2}AD^{1/2}H^{(t)}W_t\right).$$4, and maximum gains are $$H^{(t+1)}=\sigma\!\left(D^{-1/2}AD^{1/2}H^{(t)}W_t\right).$$5, respectively. This benchmark characterization is used to argue that a simple model such as ISN can already be strong on these datasets without requiring learned restriction maps.

This framing is important because the justification for SNNs is typically strongest when local neighborhoods are label-discordant and ordinary diffusion is presumed to be harmful. The empirical result reported for ISN suggests that, on these standard benchmarks, the good heterophily structure captured by the gain statistic may already be sufficiently exploitable by the sheaf-style architecture even when the restriction maps are trivial. This suggests that benchmark choice is inseparable from claims about the necessity of learnable sheaf Laplacians.

4. Empirical comparisons with sheaf-learning architectures

The empirical comparison is centered on the heterophilic benchmarks originally used by Bodnar et al. ISN is compared against Best-RiSNN, Best-jDSNN, Conn-NSD, Best-SNN, Best-NSP, Best-NLSD, Best-DSNN, and Best-CSNN, using the best reported results from the literature (Caralt et al., 5 Mar 2026). The headline finding is that ISN matches or nearly matches these models across the board.

Dataset	ISN accuracy
Texas	$$H^{(t+1)}=\sigma\!\left(D^{-1/2}AD^{1/2}H^{(t)}W_t\right).$$6
Wisconsin	$$H^{(t+1)}=\sigma\!\left(D^{-1/2}AD^{1/2}H^{(t)}W_t\right).$$7
Squirrel	$$H^{(t+1)}=\sigma\!\left(D^{-1/2}AD^{1/2}H^{(t)}W_t\right).$$8
Chameleon	$$H^{(t+1)}=\sigma\!\left(D^{-1/2}AD^{1/2}H^{(t)}W_t\right).$$9
Cornell	$\frac{dX(t)}{dt}=-\Delta_G X(t),$0

Most competing models are statistically comparable to ISN; only a few show modest improvements or degradations beyond one standard deviation. For example, Best-SNN is slightly better on Squirrel $\frac{dX(t)}{dt}=-\Delta_G X(t),$1 and Chameleon $\frac{dX(t)}{dt}=-\Delta_G X(t),$2, while several other sheaf variants are essentially tied with ISN, and some perform worse (Caralt et al., 5 Mar 2026).

There is an additional appendix note on Film: the reported reproducible results are $\frac{dX(t)}{dt}=-\Delta_G X(t),$3 for SNN and $\frac{dX(t)}{dt}=-\Delta_G X(t),$4 for ISN. The best model there had only one layer, which is mentioned as an exception in the Rayleigh-plot behavior but not as a contradiction to the broader pattern.

The overall empirical interpretation given in the paper is that the gains attributed to learned sheaf structures are not robustly larger than the trivial identity baseline on these benchmarks. This does not prove that learnable restriction maps are never useful, but it directly challenges any stronger claim that they are necessary to match reported performance on standard heterophilic node-classification datasets.

5. Rayleigh quotient analysis and the oversmoothing controversy

To probe the oversmoothing story, the paper introduces a normalized measure based on the Rayleigh quotient. For a positive semidefinite matrix $\frac{dX(t)}{dt}=-\Delta_G X(t),$5,

$\frac{dX(t)}{dt}=-\Delta_G X(t),$6

This is used as a normalized Dirichlet energy so that comparisons across models and layers are meaningful (Caralt et al., 5 Mar 2026).

The hypothesis tested is presented as

$\frac{dX(t)}{dt}=-\Delta_G X(t),$7

and

$\frac{dX(t)}{dt}=-\Delta_G X(t),$8

where $\frac{dX(t)}{dt}=-\Delta_G X(t),$9 denotes the trained SNN representations and $\Delta_G = I - D^{-1/2}AD^{1/2}.$0 the ISN representations. If the diffusion theory were matching practice, the Rayleigh quotient computed with the “correct” sheaf Laplacian should be systematically smaller for the SNN than for the identity baseline, and the identity model should exhibit stronger oversmoothing under the identity Laplacian.

The plotted results, however, show the opposite trend in many cases: the expected separation between the red curve $\Delta_G = I - D^{-1/2}AD^{1/2}.$1 and the blue curve $\Delta_G = I - D^{-1/2}AD^{1/2}.$2 is not observed in the way the diffusion argument predicts (Caralt et al., 5 Mar 2026). The paper therefore concludes that, in trained networks, the sheaf-space energy is generally not lower in the manner suggested by the theory, and ISNs do not appear to oversmooth more than their SNN counterparts.

This is the principal controversy surrounding ISN. The original SNN motivation relies on a diffusion-based argument in which learned restriction maps reshape the kernel of $\Delta_G = I - D^{-1/2}AD^{1/2}.$3 and thereby mitigate destructive collapse on heterophilous graphs. The ISN results do not deny the formal theory of sheaf diffusion; rather, they indicate that the behavior predicted by that theory is not reflected empirically in the trained networks studied. A plausible implication is that optimization, residual structure, normalization, or architecture-level effects may dominate the practical behavior attributed to learned sheaf Laplacians.

6. Relation to other sheaf-network constructions

The term “identity sheaf” can also arise in a distinct sheaf-theoretic context unrelated to node-classification SNNs. In "Modeling wireless network routing using sheaves" (Robinson, 2016), wireless routing is modeled on an abstract simplicial complex built from a link complex $\Delta_G = I - D^{-1/2}AD^{1/2}.$4, where

$\Delta_G = I - D^{-1/2}AD^{1/2}.$5

The activation sheaf $\Delta_G = I - D^{-1/2}AD^{1/2}.$6 assigns to each cell $\Delta_G = I - D^{-1/2}AD^{1/2}.$7

$\Delta_G = I - D^{-1/2}AD^{1/2}.$8

with restriction maps

$\Delta_G = I - D^{-1/2}AD^{1/2}.$9

A global section is a complete, interference-free assignment of which node is active on each simplex (Robinson, 2016).

This construction is not the same as ISN, but it clarifies a broader point about sheaf-based network models: restriction maps encode local compatibility constraints, and global sections represent globally feasible configurations. In the wireless-routing setting, the sheaf formalism captures collision-free activation patterns and can be extended to a data payload sheaf $$\lim_{t\to\infty}X(t)\in \ker(\Delta_G)=\{x_u=x_v\mid (u,v)\in E\}.$$0 on a time-dependent link graph, with state updates such as

$$\lim_{t\to\infty}X(t)\in \ker(\Delta_G)=\{x_u=x_v\mid (u,v)\in E\}.$$1

The vector activation sheaf $$\lim_{t\to\infty}X(t)\in \ker(\Delta_G)=\{x_u=x_v\mid (u,v)\in E\}.$$2 satisfies

$$\lim_{t\to\infty}X(t)\in \ker(\Delta_G)=\{x_u=x_v\mid (u,v)\in E\}.$$3

The relevance of this comparison is conceptual rather than genealogical. In both settings, the sheaf formalism replaces undifferentiated adjacency with local-to-global consistency constraints. The difference is that ISN is an ablation inside a learned sheaf-Laplacian architecture for heterophilic node classification, whereas the wireless model uses sheaf restrictions to represent feasible simultaneous transmissions under CSMA/CD-style assumptions. Conflating these constructions would be a category error: they share sheaf language, but they solve different problems and instantiate different objects.

7. Significance, limitations, and interpretation

ISN is significant because it converts the necessity of learnable sheaf Laplacians from a theoretical presumption into an empirical question. By taking an SNN and fixing every restriction map to the identity $$\lim_{t\to\infty}X(t)\in \ker(\Delta_G)=\{x_u=x_v\mid (u,v)\in E\}.$$4, it provides a strong baseline that preserves the overall SNN-style message passing and training pipeline while removing the learned sheaf parameterization (Caralt et al., 5 Mar 2026). On the standard heterophilic benchmarks studied, the reported answer is largely that the learnable sheaf machinery is not necessary to match the reported performance.

The principal limitation is scope. The conclusion is benchmark-specific: Texas, Wisconsin, Squirrel, Chameleon, Cornell, and the appendix result on Film. The paper does not establish that learnable restriction maps are never useful; rather, it shows that the empirical improvements reported by sheaf-learning architectures are largely reproducible by the trivial identity baseline on the standard benchmarks examined. This suggests that claims about the indispensability of learned sheaf structure should be made cautiously and supported by stronger ablations.

A common misconception is that ISN refutes sheaf neural networks in general. The paper supports a narrower conclusion. It revisits a trivial sheaf construction to ask whether the additional complexity of learning restriction maps is necessary, and on the studied benchmarks the answer is largely no (Caralt et al., 5 Mar 2026). Another misconception is that the formal diffusion theory is mathematically invalid; the argument made is instead that trained-model behavior does not reflect the oversmoothing pattern predicted by that theory. Accordingly, ISN is best understood as a methodological baseline and an empirical critique of causal attribution, rather than as a rejection of sheaf-theoretic modeling itself.