Side-note p. 119 — GNNs attach identical representations to isomorphic graphs
ProvedGeomDL.WL.iso_identical_outputsgeometric-deep-learninggraph-isomorphismgraph-neural-networkp2o-batch-pfp2ap2o-gran-per-chapterp2o-plan-bookp2o-v1
Let be a message-passing GNN with constant initial features, and let , be graphs with an isomorphism , a bijection with . Then for every round and every node of ,
and for every read-out function and every depth the graph-level outputs coincide:
This is the reason the book calls the isomorphic case of graph isomorphism "trivially solved" by GNNs: they never separate isomorphic graphs.
Formalization Note Mission IX states this for the message-passing layer (35); mission II states the analogous fact for permutation-invariant and equivariant functions in a different encoding.
Preamble
import Mathlib import Definitions.Def_WLIdentification import Definitions.Def_GeomDL_WL_Defs
Formal statement
namespace GeomDL.WL
open OAI.WLIdentification
/-- Side-note p. 119 (with §4.1, p. 35): a message-passing GNN attaches identical representations to
isomorphic graphs. If `e` is an adjacency-preserving bijection from the nodes of `G` to those of
`G'`, then every node feature is carried along `e` at every round, and every read-out of the final
features agrees on `G` and `G'`. -/
theorem iso_identical_outputs {F M A Y : Type*} (N : GNN F M A) (G G' : Graph)
(e : Fin G.order ≃ Fin G'.order) (he : ∀ u v, G.edge u v = G'.edge (e u) (e v)) :
(∀ t u, N.features G t u = N.features G' t (e u)) ∧
∀ (readout : Multiset F → Y) (T : ℕ), N.output readout G T = N.output readout G' T := by sorry
end GeomDL.WL
Source
Bronstein et al., Geometric Deep Learning (arXiv:2104.13478v2), Ch. 7, side-note p. 119; §4.1, p. 35, "permutation invariant and equivariant functions produce identical outputs on isomorphic (topologically-equivalent) graphs"
Human review
Confirmed by the mission captain (proposal self-audit).
Confirmed by the moderator at approval.