---
title: Iterative Graph Normalization (IGN) Overview
url: https://www.emergentmind.com/topics/iterative-graph-normalization-ign
type: topic
---

# Iterative Graph Normalization (IGN) Overview

Searching arXiv for relevant papers on Iterative Graph Normalization and related terminology.
Iterative Graph Normalization (IGN) denotes the discrete-time iteration of a graph normalization map on node or assignment weights, introduced as a differentiable dynamical system for combinatorial optimization, especially Maximum Weight Independent Set (MWIS) and related matching formulations [2012.07764, 2605.05330]. In its basic form on a simple undirected graph with adjacency matrix \(A\), IGN repeatedly applies
\[
x^{k+1} \;=\; N_h(x^k), \qquad N(x) \;=\; x \oslash (A+I)x,
\]
optionally followed by a componentwise activation \(h\), thereby driving nonnegative weights toward binary indicators of independent sets [2012.07764]. A terminological caveat is essential: in the separate literature on higher-order graph representation learning, “IGN” also denotes invariant graph networks rather than iterative graph normalization; that usage is unrelated except for the acronym [2007.12035]. Within the optimization literature, IGN emphasizes the iterative dynamics induced by Graph Normalization (GN), while more recent work distinguishes a weighted, regularized variant, WRGN, as the practically useful form for weighted instances [2605.05330].

## 1. Terminology and formal setup

The original IGN formulation considers a simple undirected graph over \(V=\{1,\dots,n\}\), represented by a binary adjacency matrix \(A\in M_n(\{0,1\})\) with zero diagonal, together with a nonnegative node-weight vector \(x\in\mathbb{R}_+^n\); the weighted graph is \(G=(A,x)\) [2012.07764]. For node \(i\), the neighborhood is \(V(i)=\{j\in V\mid A_{ij}=1\}\), the degree is \(\deg(i)=|V(i)|\), and the support of a weight vector is \(\mathrm{supp}(x)=\{i\mid x_i>0\}\) [2012.07764].

The normalization map is defined by Hadamard division:
\[
N(x) \;=\; x \oslash (A+I)\,x,
\qquad
N_i(x) \;=\; \frac{x_i}{x_i + \sum_j A_{ij}x_j}.
\]
Activated normalization applies a componentwise nonlinearity \(h:[0,1]\to[0,1]\):
\[
N_h(x) \;=\; h\!\big(N(x)\big), \qquad N_{h,i}(x)=h\!\big(N_i(x)\big),
\]
and IGN is the iteration
\[
x^0=x,\qquad x^{k+1}=N_h(x^k)=N_h^{k+1}(x)
\]
[2012.07764].

Normalizability is defined by the condition \((A+I)x>0\) componentwise; equivalently, \(G=(A,x)\) is normalizable iff \(\mathrm{dens}_G(\mathrm{supp}(x))>0\), where
\[
\mathrm{dens}_G(S)
\;=\;
\min_{i\notin S}\Big|\;V(i)\cap S\;\Big|
\;=\;
\min_{i\notin S}\sum_{j\in S}A_{ij}.
\]
This condition ensures that every denominator in the normalization is positive [2012.07764].

The 2026 refinement rephrases the same iterative concept in terms of GN and IGN: GN is the normalization map, while IGN is explicitly the repeated application \(x^{t+1}=N_A(x^t)\) until convergence [2605.05330]. That work further introduces Weighted Regularized Graph Normalization (WRGN), which incorporates positive vertex weights \(w_i\), a bias \(v:=\sqrt{w}\), and a regularization parameter \(\gamma>1\) through
\[
A_{\gamma,v}
\;:=\;
\gamma\cdot \mathrm{diag}(v)^{-1}A\,\mathrm{diag}(v),
\]
with update
\[
x_i^{t+1}
\;=\;
\frac{x_i^t}{\,x_i^t+\gamma\sum_{j\in N(i)}\frac{v_j}{v_i}x_j^t\,}.
\]
This weighted regularized dynamics is presented as the weight-aware and provably binarizing version of IGN [2605.05330].

A separate terminological clarification arises from graph representation learning. In “The expressive power of kth-order invariant graph networks,” “IGN” denotes invariant graph networks, not iterative graph normalization. There, k-IGNs are higher-order, permutation-invariant neural architectures on \(k\)-tuple-indexed features with \(S_n\)-equivariant linear maps and invariant pooling [2007.12035]. This usage should not be conflated with graph normalization dynamics.

## 2. Core dynamics and structural properties

In the 2020 formulation, GN and IGN possess several basic structural properties. The map \(N_h\) sends weights to \([0,1]^n\), preserves support under \(N\), is scale invariant in the sense that \(N_h(\alpha x)=N_h(x)\) for \(\alpha>0\), and depends only on neighborwise ratios \(x_j/x_i\) [2012.07764]. GN also commutes with graph automorphisms: if \(M\) is a permutation satisfying \(MA=AM\), then \(N(Mx)=MN(x)\) [2012.07764]. Fixed points of the unactivated map satisfy a local condition: at node \(i\), either \(x_i=0\) or the \(i\)-th component of \((A+I)x\) equals \(1\) [2012.07764].

The activated map requires \(h(0)=0\) and \(h(1)=1\), together with either strict convexity on \([0,1]\), or the mixed convex–concave condition \(h(1/2)=1/2\), strict convexity on \([0,1/2]\), and strict concavity on \([1/2,1]\) [2012.07764]. These assumptions are used in the fixed-point and attractivity analysis.

The weighted regularized form recasts the same local closed-neighborhood normalization in a weight-tilted geometry. Letting \(y:=v\odot x\) and \(B:=I+\gamma A\), the WRGN update becomes
\[
y_i^{t+1}=\frac{v_i}{(By^t)_i}.
\]
The normalization remains local, preserves support, is projective or scale-invariant, and maps any normalizable state into \([0,1]^n\) [2605.05330].

The assignment specialization is structurally analogous. In the line graph \(L(K_{n,n})\), with variables \(x_{ij}\) representing matchings of row \(i\) to column \(j\), GN cross-normalizes each entry against all conflicting row and column competitors:
\[
x_{ij}^{t+1}
\;=\;
\frac{x_{ij}^t}{\,x_{ij}^t+\sum_{k\ne i}x_{kj}^t+\sum_{\ell\ne j}x_{i\ell}^t\, }.
\]
This provides a direct graph-normalization analogue of assignment normalization, but with hard combinatorial structure inherited from the line graph [2605.05330].

## 3. Independent sets, binary fixed points, and local stability

A central structural result of the 2020 paper is that the only binary fixed points are maximal independent sets (MIS) [2012.07764]. For a graph \(G\), an independent set \(S\subseteq V\) satisfies \(A_{ij}=0\) for all \(i,j\in S\), and its indicator vector \(\mathrm{ind}(S)\in\{0,1\}^n\) has entries equal to \(1\) exactly on \(S\) [2012.07764]. The fixed-point proposition states:
\[
\text{For any graph } G\text{ and any activation }h,\text{ a normalizable binary vector }x\text{ is a fixed point of }N\text{ iff }x\in\mathrm{MIS}(G).
\]
The reasoning is explicit in the paper: if \(x\) is the indicator of an MIS, then on support nodes the closed-neighborhood denominator equals the node’s own mass, yielding \(N_i(x)=h(1)=1\), while outside the support at least one neighboring \(1\) forces \(N_i(x)=h(0)=0\); conversely, if a binary vector is not independent, some adjacent pair with \(x_i=x_j=1\) yields \(N_i(x)=h(\le 1/2)<1\), so the point is not fixed [2012.07764].

The same paper analyzes stability through the Jacobian of \(N_h\). Writing \(B=A+I\), the Jacobian entries are
\[
J_{ii}(x)
=
h'\!\left(\frac{x_i}{\sum_j B_{ij}x_j}\right)
\frac{\sum_j A_{ij}x_j}{\big(\sum_j B_{ij}x_j\big)^2},
\]
\[
J_{ij}(x)
=
h'\!\left(\frac{x_i}{\sum_j B_{ij}x_j}\right)
\frac{-\,x_i\,A_{ij}}{\big(\sum_j B_{ij}x_j\big)^2},
\qquad i\ne j,
\]
or in matrix form,
\[
J(x)=H(x)\,\big(W(x)-\mathrm{diag}(x)A\big),
\]
with diagonal \(H\) and \(W\) as specified in the paper [2012.07764].

At an MIS indicator \(x\), the spectrum simplifies to
\[
\Lambda(J(x))
=
\{0\}\cup
\left\{
\frac{h'(0)}{\sum_j A_{ij}x_j}\,;\; i\notin \mathrm{supp}(x)
\right\},
\]
and the spectral radius is
\[
\rho(J(x))=\frac{h'(0)}{\mathrm{dens}_G(x)}.
\]
Hence an MIS is attractive whenever \(h'(0)<\mathrm{dens}_G(x)\); if \(h'(0)<1\), then every MIS in any graph is attractive; and if \(h'(0)=0\), convergence to an MIS is quadratic in a neighborhood [2012.07764].

Non-maximal independent sets are treated differently. Under the condition \(h'(y)\ne 0\) for all \(y\in[0,1]\), any non-maximal independent set is a repulsive point for \(N_h\) [2012.07764]. The paper notes a practical trade-off: this repulsivity condition conflicts with quadratic convergence to MIS, since the latter requires \(h'(0)=0\) [2012.07764]. A plausible implication is that activation design controls not only local rate but also which combinatorial structures can become attractors.

The 2026 WRGN analysis sharpens the picture. For any connected simple graph \(A\), any \(\gamma>0\), any positive bias \(v\), and any normalizable initialization \(x^0\), the WRGN sequence converges to a unique normalizable fixed point \(x^*\) [2605.05330]. For \(\gamma>1\), every non-binary fixed point is strictly repulsive, and a binary fixed point with support \(M\) is asymptotically stable iff
\[
\mathrm{stab}_{\gamma,w}(M)
:=
\min_{i\in M}
\gamma\sum_{j\in N(i)\cap M^c}\frac{\sqrt{w_j}}{\sqrt{w_i}}
>1.
\]
Every MWIS is a stable attractor, and any WRGN sequence with \(\gamma>1\) converges to a \(y\)-stable MIS while strictly increasing the relaxed MWIS objective \(\sum_i w_i x_i^t\) until convergence [2605.05330].

## 4. Relation to MWIS and optimization theory

IGN is motivated by the Maximum Weight Independent Set problem:
\[
\max_{x\in\{0,1\}^n}\ \sum_{i=1}^n w_i x_i
\quad\text{s.t.}\quad
x_i+x_j\le 1
\quad \forall (i,j)\in E,
\]
or equivalently by the binary quadratic constraint \(x^\top A x=0\) [2012.07764, 2605.05330]. In the 2020 formulation, IGN’s binary fixed points are MIS rather than necessarily maximum-weight solutions, so the method is characterized as a differentiable heuristic aiming to find a high-weight MIS by reweighting dynamics [2012.07764].

The 2026 work strengthens this interpretation by introducing a relaxed primal objective on fuzzy memberships \(x\in[0,1]^n\),
\[
\sum_{i=1}^n w_i x_i,
\]
and proving that WRGN increases this quantity strictly at each iteration unless at a fixed point [2605.05330]. This moves the method from heuristic dynamics toward a monotone optimization scheme on a relaxation of MWIS.

An exact Majorization–Minimization interpretation is given in the weighted state space \(y\). The energy is
\[
E_{\gamma,v}(y)
:=
\frac{1}{2}\,y^\top B y - v^\top y,
\qquad B:=I+\gamma A.
\]
At iterate \(y^t\), define the separable quadratic majorant
\[
G(y\mid y^t)
:=
\frac{1}{2}\sum_i (B y^t)_i\, y_i^2 - \sum_i v_i y_i.
\]
This majorant upper bounds \(E\), is tangent at \(y^t\), and is strictly convex; minimizing it yields
\[
y_i^{t+1}=\frac{v_i}{(By^t)_i},
\]
which is exactly the WRGN update. The resulting MM inequalities are
\[
E_{\gamma,v}(y^{t+1})
\le
G(y^{t+1}\mid y^t)
\le
G(y^t\mid y^t)
=
E_{\gamma,v}(y^t),
\]
so the energy decreases monotonically [2605.05330].

The same update also admits a quasi-Newton description:
\[
y^{t+1}-y^t
=
-
\mathrm{diag}\big(y^{t+1}\odot v\big)\,\nabla E_{\gamma,v}(y^t),
\qquad
\nabla E_{\gamma,v}(y)=By-v.
\]
The paper characterizes this as a preconditioned gradient step in which the diagonal preconditioner approximates inverse curvature and acts like an adaptive step size with implicit log-barriers [2605.05330].

The optimization viewpoint is complemented by a weighted extension of the Motzkin–Straus theorem. On the weight-tilted simplex
\[
\Delta_{w^{-1}}
:=
\Big\{
x\in\mathbb{R}_+^n \ \Big| \ \sum_{i=1}^n w_i x_i = 1
\Big\},
\]
consider
\[
Q(x)=x^\top (I+yA)x
\]
for \(y>1\). The local minima of \(Q\) on \(\Delta_{w^{-1}}\) are in one-to-one correspondence with the \(y\)-stable MISes of \(G\); if \(M\) is such an MIS, then
\[
Q(x^*)=\frac{1}{\sum_{i\in M}w_i}.
\]
In particular, any MWIS is \(y\)-stable and corresponds to a global minimum of \(Q\) over the tilted simplex [2605.05330]. This establishes a precise bridge between the dynamical system, quadratic optimization, and combinatorial structure.

## 5. Replicator dynamics, Fisher-type monotonicity, and geometric interpretation

WRGN is also equivalent to a discrete-time replicator dynamics of a nonlinear evolutionary game [2605.05330]. Defining the simplex state
\[
p^t
:=
\frac{v\odot y^t}{M_v(y^t)},
\qquad
M_v(y^t):=\sum_i v_i y_i^t,
\]
the dynamics becomes
\[
p_i^{t+1}
=
p_i^t\,\frac{f_i(p^t)}{\bar f(p^t)},
\]
with fitness
\[
f_i(p)
=
\frac{v_i}{(B(p\odot v))_i},
\qquad B:=I+\gamma A,
\]
and average fitness \(\bar f(p):=\sum_i p_i f_i(p)\) [2605.05330].

A key identity links this evolutionary interpretation to the optimization objective:
\[
M_v(y^{t+1})=\bar f(p^t).
\]
The paper interprets this as a Fisher-type statement: the population average fitness equals the next iteration’s weighted mass, and this quantity strictly increases off equilibrium [2605.05330]. Although the induced game is nonlinear and non-potential, the dynamics still admits monotone functionals, namely decreasing MM energy and increasing weighted mass.

The earlier 2020 paper also offered geometric evidence for structured dynamics. For any graph \(A\) of size \(n\), the image
\[
I_A=\{N(x);x\in N_A\}
\]
is a hypersurface intersecting each line at most once, and on trees \(N(x)=N(y)\Rightarrow y=kx\) for some \(k>0\) when \(x,y\) are positive [2012.07764]. For the path graph \(P_3\), the image of GN in \((0,1]^3\) is the “Taco” hypersurface
\[
uvw+uw-u-v-w+1=0,
\]
with two binary fixed points, \((1,0,1)\) and \((0,1,0)\), acting as attractors under suitable \(h\) [2012.07764]. This suggests that the normalization map has a nontrivial projective geometry already on very small graphs.

The same paper conjectured global convergence of IGN and convergence to binary solutions for suitable nonlinear activations [2012.07764]. The later WRGN theory supplies stronger guarantees, but for the regularized weighted dynamics rather than all original activated variants. This suggests a historical progression from an experimentally motivated discrete dynamical system to a more fully characterized optimization method.

## 6. Assignment problems, Sinkhorn-related normalization, and ML integration

IGN has a direct assignment-problem specialization. The maximum weight bipartite matching problem can be represented as MWIS in the line graph of the bipartite graph, with optimal solution a permutation matrix maximizing \(\sum_{ij}W_{ij}P_{ij}\) [2012.07764]. In this context, the 2020 paper defines cross normalization
\[
X(A)
=
A \oslash \big(A\mathbf{1}\mathbf{1}^\top + \mathbf{1}\mathbf{1}^\top A - A\big),
\]
and in vectorized form
\[
\mathrm{vec}(X(A))
=
\mathrm{vec}(A)\oslash
\Big(
\big(
\mathbf{1}\mathbf{1}^\top\otimes I_n + I_n\otimes \mathbf{1}\mathbf{1}^\top - I_{n^2}
\big)\,\mathrm{vec}(A)
\Big)
=
N_C(\mathrm{vec}(A)),
\]
where
\[
C
=
\mathbf{1}\mathbf{1}^{\top}\otimes I_n
+
I_n\otimes \mathbf{1}\mathbf{1}^{\top}
-
2I_{n^2}.
\]
Iterating activated cross normalization \(X_h\) yields Iterative Cross Normalization (ICN), intended to obtain a crisp assignment [2012.07764].

This leads to a comparison with Sinkhorn–Knopp (SK). SK alternates row and column normalization,
\[
R(A)=A \oslash (A\mathbf{1}\mathbf{1}^\top),
\qquad
C(A)=A \oslash (\mathbf{1}\mathbf{1}^\top A),
\]
and converges toward the Birkhoff polytope of doubly stochastic matrices, whereas ICN projects toward permutation matrices by normalizing simultaneously against row-and-column cross sums [2012.07764]. The distinction is summarized explicitly in the source: SK projects to \(B_n\), a soft doubly stochastic solution set, whereas ICN projects to \(P_n\), a crisp permutation set [2012.07764].

The relation to Softassign is also direct:
\[
SA_\tau(X)=SK(\exp(X/\tau)),
\]
with entropy-regularized interpretation
\[
SA_\tau(X)
=
\arg\max_{D\in B_n}
\mathrm{Tr}(D^\top X)-\frac{h(D)}{\tau},
\qquad
h(D)=-\sum_{ij}D_{ij}\log D_{ij}.
\]
As \(\tau\to 0\), Softassign tends almost surely to the optimal assignment under mild randomness assumptions, but remains soft in practice and often needs a cleanup step; the paper reports that combining SA with ICN yields crisp permutations efficiently [2012.07764].

Both papers emphasize differentiability. GN, IGN, and ICN are differentiable on the normalizable interior because they are compositions of sparse neighbor sums or row/column sums and elementwise divisions, so they can be unrolled for a fixed number of iterations and trained by backpropagation [2012.07764, 2605.05330]. The cited application classes include graph and hypergraph matching, sequence alignment, clustering, ranking, structured sparse attention, dynamic network pruning, Mixture-of-Experts routing, and broader constrained resource allocation tasks [2012.07764, 2605.05330]. A plausible implication is that IGN is most naturally viewed not merely as a standalone solver but as a trainable combinatorial decision layer.

## 7. Empirical performance, limitations, and acronym ambiguity

The 2020 experimental study compared IGN with the greedy MWIS approximation algorithm of Kako–Ono–Hirata–Halldórsson, denoted WG. In that algorithm, the relative weighted degree is
\[
d_x(i)=\frac{\sum_{j\in V(i)}x_j}{x_i}
=
\frac{\sum_j A_{ij}x_j}{x_i},
\]
and the greedy rule repeatedly selects the node of minimum \(d_x(i)\), adds it to the solution, and removes it with its neighbors [2012.07764]. The connection to normalization is exact:
\[
N_i(x)^{-1}=1+d_x(i).
\]
Thus IGN is described as a “soft” reweighting counterpart of WG, pushing up small-\(d_x(i)\) nodes [2012.07764].

On binomial random graphs \(G_{n,p}\) with weights uniform in \((0,1]\), IGN’s solution weights were reported to be close to WG’s. For \(p=0.5\) and a power activation with \(a=2,t=0.01\), \(n=8\) instances yielded solutions “almost always identical,” and up to \(n=512\) the average relative gaps were reported as \(<5\%\), with probability that IGN beats WG exceeding \(0.5\) up to about \(n\approx 128\) for some settings [2012.07764]. For assignment problems, SA+ICN exactly solved AP with median gap \(0\) for \(n\le 16\) and \(\tau=0.01\), while for larger \(n\) up to \(512\), median and average gaps remained \(\lesssim 3\)–\(3.5\%\); SA was reported to reduce ICN iterations by about \(10\times\) at \(n=512\), with example averages \(36\) versus \(361\) iterations [2012.07764].

The 2026 work reports larger-scale MWIS results for warm-started WRGN. On real benchmarks up to \(882\text{K}\) nodes and \(344\text{M}\) edges, warm-start WRGN from a Bregman–Sinkhorn fractional solution typically reached solutions within about \(1\%\) of best-known MWIS in seconds on CPU; specific examples include AVR_024 and AVR_034 at \(0\%\) gap in \(\le 0.4\)s, AVR_023 and AVR_027 at \(\le 0.06\%\) gap in \(\le 1.4\)s and \(\le 0.2\)s, and MSCD instances with warm-start gaps often \(0.4\%\)–\(0.9\%\) and runtimes \(0.7\)–\(1.9\)s [2605.05330]. Time per iteration is \(O(|E|)\) for graph instances and \(O(n^2)\) for ICN on assignment matrices; memory can become the main bottleneck on very large sparse graphs [2012.07764, 2605.05330].

Several limitations remain explicit. In the 2020 framework, general convergence on arbitrary graphs was open, although complete graphs were solved exactly under a unique maximum-weight condition, and an \(L_1\)-norm monotonicity conjecture was proposed as a route to proving convergence [2012.07764]. The paper also noted sensitivity to the activation derivative at zero: \(h'(0)=0\) gives quadratic convergence near MIS but can permit convergence to non-maximal independent sets, whereas \(h'(0)>0\) ensures repulsivity of non-maximal IS but slows local rates [2012.07764]. The 2026 paper, despite proving convergence and binarization for WRGN, still identifies multiple binary MIS attractors and dependence on initialization; \(\gamma\)-pursuit and warm starts are proposed to mitigate convergence to suboptimal MIS [2605.05330].

Finally, the acronym ambiguity around “IGN” is itself a recurring source of confusion. In the higher-order GNN literature, k-IGN refers to invariant graph networks with linear \(S_n\)-equivariant layers on \(k\)-tuple-indexed features and invariant pooling, and the main result there is that k-IGNs are exactly as expressive as \(k\)-WL in graph distinction power [2007.12035]. That usage is unrelated to graph-normalization dynamics, despite the identical acronym. For work on iterative graph normalization in the optimization sense, the relevant objects are GN, IGN, ICN, and WRGN rather than invariant graph networks [2012.07764, 2605.05330].

Source: https://www.emergentmind.com/topics/iterative-graph-normalization-ign