---
title: 'MRGNN: Multiresolution Reservoir Graph Neural Network'
url: https://www.emergentmind.com/topics/multiresolution-reservoir-graph-neural-network-mrgnn
type: topic
---

# MRGNN: Multiresolution Reservoir Graph Neural Network

Multiresolution Reservoir Graph Neural Network (MRGNN), also described as a pyramidal Reservoir Graph Neural Network, is a graph-level learning architecture that interleaves reservoir-computing layers with topological graph-pooling layers so that node representations are repeatedly updated, coarsened, and finally aggregated into a graph embedding. In the 2021 formulation, each reservoir layer is defined as a contractive dynamical system whose node states converge to a unique fixed point, while each pooling layer reduces the support graph and the feature support, yielding a multiresolution or pyramidal computation that is explicitly analyzed in terms of convergence and complexity [2104.04710]. A later spectral reinterpretation revisits the reservoir definition through Laplacian filtering and introduces a Fairing-based variant designed to control smoothing more selectively, with an associated random-walk interpretation of redundancy [2507.12963].

## 1. Architectural organization

The core MRGNN architecture is organized into \(L\) blocks. For each block \(l=1,\dots,L-1\), the model alternates two operations: an RC layer that embeds every node via a contractive dynamical system, and a topological pooling layer that produces a coarsened graph with fewer nodes, fewer edges, and a smaller spectral radius. The final \(L\)-th block contains only an RC layer. The remaining node states are then aggregated by summation into a single \(H\)-dimensional graph embedding,
\[
\phi(G)=\sum_{n=1}^{N_L} x_n^{(L)*}.
\]
This graph embedding is the representation used for downstream graph-level prediction [2104.04710].

At layer \(l\), the hidden-state matrix is denoted \(X^{(l)}[t]\in\mathbb{R}^{N_l\times H}\), the input features are \(U^{(l)}\in\mathbb{R}^{N_l\times F_l}\), and the normalized adjacency is \(\tilde A^{(l)}\). The iterated update is
\[
X^{(l)}[t+1] = \tanh\!\bigl( \tilde A^{(l)} X^{(l)}[t] W^{(l)} + U^{(l)} W_{\mathrm{in}}^{(l)} \bigr).
\]
For \(l=1\), \(F_1\) is the original feature dimension; for \(l>1\), \(F_l=H\). The fixed-point state \(X^{(l)*}\) becomes the input \(U^{(l+1)}\) to the next block. This arrangement gives the model its multiresolution character: node dynamics are computed at one scale, then the graph is coarsened, and the process repeats on a smaller graph.

A later exposition of MRGNN reframes the reservoir as a fixed, non-learned graph filter of depth \(k\), using the symmetric normalized Laplacian
\[
L \equiv \Delta_{\mathrm{sym}} = I - D^{-1/2} A D^{-1/2},
\]
with \(\sigma(\cdot)=\tanh(\cdot)\). In that view, the reservoir step is written
\[
X^{(t+1)}=\sigma\bigl(H(L)X^{(t)}\bigr)=\tanh\bigl(LX^{(t)}\bigr),
\]
and the intermediate embeddings are concatenated into
\[
H^k \equiv [X^{(0)} \mid X^{(1)} \mid \dots \mid X^{(k)}] \in \mathbb{R}^{n\times (k+1)d}.
\]
The same source describes a simple two-layer MLP readout for graph-level classification or regression, which indicates that later work uses the MRGNN label both for the original multiresolution fixed-point construction and for a spectral reservoir interpretation built around the same general reservoir-on-graphs paradigm [2507.12963].

## 2. Reservoir dynamics and fixed-point theory

In the fixed-point formulation, a reservoir layer updates node states by
\[
X[t+1] = \tanh( \tilde A X[t] W + U W_{\mathrm{in}} + b ),
\]
where \(\tilde A\) is the symmetrically normalized adjacency matrix of the current graph, \(X[t]\in\mathbb{R}^{N\times H}\) is the node-state matrix at iteration \(t\), \(U\in\mathbb{R}^{N\times F}\) is the input feature matrix, \(W\in\mathbb{R}^{H\times H}\), \(W_{\mathrm{in}}\in\mathbb{R}^{F\times H}\), and \(b\in\mathbb{R}^{H}\) is broadcast to all nodes. The summary notes that, in practice, \(b\) is often omitted or absorbed; it also states that the paper uses \(W_{\mathrm{in}}\) for \(l=1\) of size \(F\times H\) and a hidden-scaling matrix for \(l>1\) of size \(H\times H\) [2104.04710].

The central theoretical property is contraction. If
\[
F(X)=\tanh(\tilde A X W+\cdots),
\]
then the map is \(K\)-Lipschitz with
\[
K = \rho(\tilde A)\cdot\|W\|_2,
\]
where \(\|W\|_2\) is the spectral norm of \(W\) and \(\rho(\tilde A)\leq 1\) is the largest-magnitude eigenvalue of \(\tilde A\). When \(K<1\), described as the “Graph Embedding Stability” condition, \(F\) is a contraction, so a unique fixed point \(X^*\) exists and the iteration converges geometrically from any initialization. In the implementation pipeline, the initialization is \(X[0]=0\) [2104.04710].

The paper also provides an explicit bound on the convergence time. If the layer is iterated until \(\|X[t+1]-X[t]\|<\epsilon\), then the number of iterations \(T\) satisfies
\[
T \le \left\lceil \frac{\ln \epsilon + \ln(1-K) - \ln H_1}{\ln K} \right\rceil,
\]
where \(H_1=\|\tanh(UW_{\mathrm{in}})\|\) is the norm after one step. The stated consequence is that \(T\) grows as \(K\to 1\), that is, as \(\rho(\tilde A)\) or \(\|W\|\) grow. This establishes a direct connection between graph topology, reservoir scaling, and the runtime needed to obtain the fixed-point representation [2104.04710].

The original graph-level readout after the last reservoir layer is the row-wise summation of \(X^{(L)*}\), followed by a linear ridge-regression classifier trained on \(\phi(G)\). A common misconception is that the reservoir is merely an initialization stage for a conventional trainable GNN. In the reported formulation, the reservoir itself is the principal feature extractor, while the trainable component is the final readout. The later spectral reinterpretation preserves this reservoir-centered perspective, although it reformulates the propagation operator and the readout construction [2507.12963].

## 3. Graph coarsening and the multiresolution trade-off

The multiresolution component is defined through an assignment or selection matrix \(S\in\mathbb{R}^{N_l\times N_{l+1}}\) with \(N_{l+1}<N_l\). The coarsened adjacency and feature matrices are
\[
A^{(l+1)} = S^T A^{(l)} S \in \mathbb{R}^{N_{l+1}\times N_{l+1}},
\qquad
U^{(l+1)} = S^T X^{(l)*} \in \mathbb{R}^{N_{l+1}\times H}.
\]
Applying the coarsening operator successively reduces \(N\), \(M\), and typically the spectral radius \(\rho(\tilde A)\). The paper’s theoretical argument is that this affects not only storage and per-iteration cost, but also the contraction speed of the reservoir updates [2104.04710].

Three pooling instantiations are listed.

| Pooling method | Mechanism | Reported note |
|---|---|---|
| Graclus | Unsupervised spectral clustering into pairs, then average-pool | May introduce “fake” nodes |
| NMF pooling | \(A\approx Q S\) with non-negative \(Q,S\), then aggregate by \(S^T\) | Can yield dense coarsened graphs |
| Node-Decimation (NDP) | Drop \(\sim\)half the nodes via max-cut partition, reconnect via Kron-reduction, and sparsify | Reduces \(N\), \(M\), and \(\rho(\tilde A)\) most effectively |

The computational analysis quantifies the benefit. Each iteration of a reservoir layer costs
\[
O(MH + NH^2),
\]
corresponding to sparse \(A\)-times-\(X\) and dense \(XW\). Summed over \(T\) iterations, one layer costs
\[
O\bigl(T\cdot(MH + NH^2)\bigr)
=
O\bigl(\mathcal{T}(\rho(\tilde A))\cdot(MH + NH^2)\bigr).
\]
Because pooling reduces \(N\), \(M\), and typically \(\rho(\tilde A)\), it reduces the number of matrix multiplications and accelerates contraction. The theoretical point is therefore stronger than ordinary downsampling: pooling is analyzed as a mechanism that alters the convergence properties of the reservoir dynamics themselves [2104.04710].

This also clarifies an important trade-off. Pooling is not presented as uniformly accuracy-improving. Rather, it provides what the paper describes as a principled trade-off between accuracy and complexity. A plausible implication is that MRGNN is best understood as a family of operating points on an accuracy–efficiency frontier, rather than as a single architecture with a fixed optimum.

## 4. Training protocol, datasets, and empirical behavior

The reported training and inference workflow is explicit. First, pooling assignments \(S^{(l)}\) and coarsened \(A^{(l)}\) are precomputed for each block using Graclus, NMF, or NDP. For each input graph \(G\), the reservoir iteration is run from \(X[0]=0\) until \(\|\Delta X\|<\epsilon\) or \(t=\text{max\_iter}\); the result is then pooled by \(S^T\) to obtain the next layer’s input features, and the process is repeated. After the last RC layer, the node states are summed to form \(\phi(G)\). A ridge-regression readout is fitted on \(\{\phi(G),\text{label}(G)\}\). At inference time, the same embedding pipeline is used and only readout training is omitted [2104.04710].

The evaluation covers two synthetic datasets and a set of established benchmarks. The synthetic datasets are B-easy and B-hard, described as two new synthetic “structure-+-feature informative only jointly” datasets with 3 classes each and \(\sim 1{,}800\) graphs. The benchmark suite comprises ten TUD datasets spanning molecules, including Mutagenicity and Enzymes, and social networks, including IMDB-BIN and Reddit. The baseline is “No-pool,” defined as the same depth \(L\) but without pooling [2104.04710].

The key empirical findings are differentiated rather than uniform. No-pool achieves slightly higher accuracy, especially on small graphs, but pyramidal models run up to \(3\times\) faster on large graphs. Among the pooling methods, NDP gives the best accuracy/time trade-off, especially on Reddit-2K/5K, because it reduces \(N\), \(M\), and \(\rho(\tilde A)\) most effectively. The accuracy/time ratio, defined as accuracy divided by train-plus-test seconds, is reported as consistently higher for pyramidal models, and NDP-based MRGNN is best overall. Qualitative LDA visualizations of \(\phi(G)\) show only mild degradation in class separability under pooling [2104.04710].

These results address a frequent misconception that the pyramidal design is valuable only when it preserves accuracy exactly. The reported evidence instead supports a more specific claim: on large graphs, the principal benefit is improved efficiency with near-baseline accuracy, whereas on small graphs the unpooled baseline can still be slightly more accurate. The architecture is thus empirically positioned as an efficiency-oriented reservoir GNN rather than as a uniformly dominant alternative.

## 5. Spectral reinterpretation and the Fairing-based reservoir

A 2025 paper revisits the reservoir in MRGNN as a spectral reservoir model and proposes a variant based on a Fairing algorithm originally introduced in surface design in computer graphics [2507.12963]. The motivation is over-smoothing: repeated application of low-pass graph operators drives node features toward low-frequency Laplacian components, making node embeddings increasingly indistinguishable.

In the spectral formulation, with \(L=U\Lambda U^T\), any polynomial filter \(h(L)\) can be written as
\[
h(L)=U\,\mathrm{diag}\bigl(h(\lambda_1),\dots,h(\lambda_n)\bigr)\,U^T.
\]
Taubin’s Fairing algorithm alternates a shrinking low-pass step with an unshrinking high-pass component. For spectral frequency \(\omega\in\sigma(L)\), one full Fairing cycle with \(n\) shrink steps of parameter \(\lambda>0\) and \(m\) unshrink steps of parameter \(\mu<0\) gives
\[
H_{\mathrm{fair}}(\omega)=(1-\lambda\omega)^n(1-\mu\omega)^m.
\]
The source characterizes this as a pass-band filter that smooths intermediate frequencies while preserving the global norm and avoiding the monotonic decay of all but the zero-frequency component.

For a single Fairing cycle with parameter \(\alpha\), the operator is
\[
I-\alpha L = U\,\mathrm{diag}(1-\alpha\lambda_i)\,U^T,
\]
and iterating it \(k\) times yields
\[
X^{(k)} = (I-\alpha L)^k X^{(0)}
= \sum_{i=1}^n (1-\alpha\lambda_i)^k\,u_i u_i^T X^{(0)}.
\]
More generally, alternating shrink and unshrink gives
\[
X^{(k)}=(I-\lambda L)^n(I-\mu L)^m X^{(0)}
=
\sum_{i=1}^n (1-\lambda\lambda_i)^n(1-\mu\lambda_i)^m\,u_i u_i^T X^{(0)}.
\]
This makes the spectral selectivity of the reservoir explicit [2507.12963].

The same paper derives a random-walk interpretation. With transition matrix \(P=D^{-1}A\),
\[
L = D^{-1/2}(I-P)D^{1/2},
\]
so
\[
L^k = D^{-1/2}\sum_{t=0}^k \binom{k}{t}(-1)^t P^t D^{1/2}.
\]
Applying \(L^k\) therefore mixes \(t\)-step walk contributions with signed binomial weights, and the source states that these signed sums heavily over-count tottering walks. Fairing instead uses
\[
(\alpha I + (1-\alpha)P)^k
=
\sum_{t=0}^k \binom{k}{t}\alpha^{k-t}(1-\alpha)^t P^t.
\]
Under this filter, the probability that a random vacuous/non-vacuous walk has exactly \(t\) non-vacuous edges is
\[
\Pr[T=t]=\binom{k}{t}\alpha^{k-t}(1-\alpha)^t,
\qquad
\mathbb{E}[T]=k(1-\alpha).
\]
The stated interpretation is that \(\alpha\) near \(1\) suppresses moves, \(\alpha\) near \(0\) emphasizes \(k\)-hop walks, and intermediate \(\alpha\) spreads weight across walk lengths, thereby controlling both over-smoothing and expressive reach [2507.12963].

## 6. Configuration guidelines, comparative results, and open problems

The technical summary accompanying the original model provides concrete guidelines for adaptation to new graph tasks. For reservoir hyperparameters, hidden size \(H=50\)–\(200\) is reported to work well, with the caveat that larger \(H\) increases expressivity but also the per-update cost \(O(NH^2)\). The spectral radius \(\rho(W)\) is recommended in \([0.1,0.9]\) to satisfy contraction \(K<1\), and input or hidden scalings \((\omega_{\mathrm{in}},\omega_{\mathrm{hid}})\) are suggested in \([0.1,0.8]\). The convergence threshold \(\epsilon\approx 10^{-5}\) and \(\text{max\_iter}\approx 50\) are described as robust defaults. For pooling, NDP is recommended for large, sparse graphs; the sparsification parameter is tuned in \([0.05,0.2]\); Graclus is to be avoided on very large graphs because it may introduce fake nodes; NMF is to be avoided if \(N>500\) because it can yield dense coarsened graphs; and halving \(N\) per layer is recommended to balance abstraction against information loss. For readout, a linear or ridge-regression model is considered sufficient when \(\phi(G)\) is high-quality, while a small MLP is suggested for highly nonlinear tasks [2104.04710].

The 2025 Fairing study reports exploratory graph-classification results on four benchmarks, using the original MRGNN defaults and no further model selection. The setup uses reservoir depth \(k\in\{4,5,6\}\), hidden readout dimensions \(q\in\{15,50,100\}\), and learning rate in \(\{10^{-3},5\cdot 10^{-4}\}\), with dropout, weight-decay, and batch-size as in the original. Results are reported with 10-fold cross-validation \(\times\) 5 runs, as mean \(\pm\) std (%) [2507.12963].

| Dataset | MRGNN (\(\ddagger\)) | Fairing reservoir |
|---|---:|---:|
| PTC | \(57.60 \pm 10.01\) | \(59.58 \pm 1.73\) |
| NCI1 | \(80.58 \pm 1.88\) | \(79.02 \pm 0.27\) |
| PROTEINS | \(75.84 \pm 3.51\) | \(74.39 \pm 0.83\) |
| ENZYMES | \(68.20 \pm 6.86\) | \(68.37 \pm 1.89\) |

The paper states that, despite no hyperparameter-search for Fairing, accuracies are comparable to MRGNN, confirming that controlled spectral smoothing can match reservoir performance. It further identifies several open directions: learnable Fairing parameters \((\alpha,\lambda,\mu)\) per layer or per graph; adaptive choice of \(n,m\) based on graph spectra; extension to directed or weighted graphs through asymmetric Laplacians; combination with residual or jumping-knowledge readouts; and theoretical bounds on mixing times and representational capacity as functions of \((\alpha,k)\) [2507.12963].

Taken together, these sources define MRGNN as a reservoir-centric, multiresolution graph architecture in which coarsening is integral to both efficiency and dynamical stability. The original formulation emphasizes fixed-point reservoir embeddings under contraction and a mathematically grounded accuracy–complexity trade-off, while the later spectral work suggests that the same framework can be interpreted through graph signal processing, pass-band filtering, and random-walk redundancy control.

Source: https://www.emergentmind.com/topics/multiresolution-reservoir-graph-neural-network-mrgnn