---
title: Approximate Graph Propagation (AGP)
url: https://www.emergentmind.com/topics/approximate-graph-propagation-agp
type: topic
---

# Approximate Graph Propagation (AGP)

Approximate Graph Propagation (AGP) is a unified randomized framework for efficiently computing a broad class of graph propagation and node-proximity queries of the form
\[
\pi \;=\;\sum_{i=0}^\infty w_i\,(D^{-a}\,A\,D^{-b})^i\,x,
\]
where \(A\) is the adjacency matrix, \(D\) is the degree matrix, \(a,b\in[0,1]\) with \(a+b\ge 1\), \(\{w_i\}_{i=0}^\infty\) is a nonnegative weight sequence, and \(x\) is a nonnegative seed vector with \(\|x\|_1=1\). Introduced as a single algorithmic framework covering node-proximity computation and GNN feature propagation, AGP was later revisited to support dynamic parameterized queries and dynamic graph updates while preserving approximation guarantees [2106.03058] [2509.10036].

## 1. Origin, scope, and problem class

AGP was proposed to unify what earlier work had treated on a case-by-case basis: transition probabilities, Personalized PageRank, heat kernel PageRank, Katz, and the feature-propagation steps used in decoupled GNN architectures such as SGC, GDC, and APPNP [2106.03058]. The motivating observation is that many graph-mining and graph-learning tasks require repeated evaluation of the same algebraic template, yet prior methods either built bespoke algorithms for particular kernels or relied on expensive full-graph power iterations.

The framework treats graph propagation as a weighted sum of powers of the normalized operator \(D^{-a}AD^{-b}\). In the 2025 revisit, this formulation is presented as a formal query model with on-the-fly parameters \(a,b,\{w_i\},x\), together with an \(O(1)\)-time oracle \(O_w\) that returns \(w_i\) and the tail index \(L(\Delta)=\min\{k:\sum_{i\ge k} w_i\le \Delta\}\) [2509.10036]. This generality is explicitly stated to capture PageRank, feature propagation in Graph Neural Networks (GNNs), and graph-based Retrieval-Augmented Generation (RAG).

A central object in the framework is the \((\delta,c)\)-approximation notion: \(\hat y\) is a \((\delta,c)\)-approximation to \(\pi\) if, for every vertex \(v\) with \(\pi[v]\ge \delta\), one has
\[
|\pi[v]-\hat y[v]|\le c\cdot \pi[v].
\]
This places AGP in the class of relative-error algorithms for significant entries rather than uniform absolute-error approximators [2509.10036].

## 2. Unified formulation and truncation principle

The original AGP formulation emphasizes that the same propagation equation recovers a range of classical kernels and GNN propagators by varying \(a\), \(b\), the weight sequence, and the seed vector [2106.03058]. The 2025 treatment makes this explicit as a formal framework for parameterized queries. A truncation fact underlies both presentations: if \(Y_i:=\sum_{j\ge i} w_j\) and \(L:=L(c\delta)=O(\log(1/\delta))\), then the truncated sum
\[
\hat \pi\;=\;\sum_{i=0}^L w_i\,(D^{-a} A D^{-b})^i x
\]
already achieves a \((\delta,c)\)-approximation [2509.10036].

This truncation principle is operationally important because it converts an infinite propagation into a bounded-depth computation. In the 2021 presentation, deterministic propagation computes levelwise residues \(r^{(i)}=Y_i(D^{-a}AD^{-b})^ix\) and reserves \(q^{(i)}=(w_i/Y_i)r^{(i)}\), with \(\pi=\sum_{i=0}^\infty q^{(i)}\). Truncating at \(L=O(\log 1/\delta)\) ensures tail control, after which the computational problem becomes the efficient estimation of the first \(L\) levels [2106.03058].

The same formulation supports both query-centric and feature-centric regimes. When \(x\) is a one-hot seed, AGP instantiates personalized or local propagation tasks; when \(x\) is a feature column, it instantiates decoupled GNN propagation, including SGC, APPNP, and GDC [2106.03058]. This suggests that AGP is best understood not as a single kernel but as an algorithmic interface to truncated diffusion families.

## 3. Core randomized algorithm: push-and-sample propagation

The original AGP algorithm replaces full sparse matrix-vector multiplication by a randomized push-and-sample scheme [2106.03058]. At each level \(i\), it maintains an approximate residue \(\hat r^{(i)}\) and approximate reserve \(\hat q^{(i)}\). For a node \(u\) with \(\hat r^{(i)}(u)>0\), AGP computes
\[
\theta = (Y_{i+1}/Y_i)\,\hat r^{(i)}(u),
\]
then partitions neighbors into those whose propagated contribution exceeds a threshold \(\epsilon\) and those whose contribution is below it. “Heavy” neighbors are updated exactly, while “light” neighbors are sampled with probability proportional to the normalized propagation mass divided by \(\epsilon\). Accepted samples receive a fixed mass \(\epsilon\) [2106.03058].

The implementation relies on degree-aware adjacency organization. In the 2021 version, adjacency lists are sorted by neighbor degree so that suitable neighbors can be skipped in \(O(1)\) per heavy block, and the remaining neighbors are grouped into \(O(\log n)\) degree bands for binomial sampling [2106.03058]. In the 2025 revisit, AGP-Static is described more explicitly through bucket partitions
\[
B[u]_j=\{v:2^j\le d_v<2^{j+1}\},
\]
and a two-regime propagation rule: small-degree neighbors are fully scanned, while large-degree neighbors are sampled via a bucket-based binomial procedure, each sampled vertex receiving a fixed mass \(\epsilon=\Theta(\delta/L^2)\) in the original AGP-Static analysis [2509.10036].

The 2025 paper identifies two limitations in that state-of-the-art static solution. First, the claimed query complexity assumed an optimal subset-sampling algorithm that “did not exist” at the time, so an extra \(O(\log^2 n)\) factor appears in practice. Second, AGP-Static processes graph updates poorly, taking \(O(n\log n)\) time per update in the original discussion and \(O(n)\) if naively maintained in the narrative comparison [2509.10036]. The proposed AGP-Static++ replaces the two-stage binomial-plus-rejection mechanism with repeated draws from a bounded Geometric\((p^*)\), yielding \(O(\log n+\mu)\) overall per neighborhood rather than per bucket, and it also removes the need for sorted scans [2509.10036].

## 4. Approximation guarantees and complexity bounds

The theoretical guarantees of AGP are framed in terms of unbiasedness, variance, and relative error. For the 2021 algorithm, for all \(i,v\),
\[
E[\hat r^{(i)}(v)] = r^{(i)}(v), \qquad E[\hat q^{(i)}(v)] = q^{(i)}(v),
\]
and the variance bound
\[
\mathrm{Var}[\widehat{\mathrm{PPR}}(v)] \le (L(L+1)/2)\cdot \epsilon \cdot \mathrm{PPR}(v)
\]
implies, via Chebyshev’s inequality, that choosing \(\epsilon=\Theta(\delta/L^2)\) and \(L=\Theta(\log 1/\delta)\) yields
\[
\Pr\!\left[ |\widehat{\mathrm{PPR}}(v)-\mathrm{PPR}(v)| \ge (1/10)\mathrm{PPR}(v)\right]\le 0.01
\]
for every \(v\) with \(\mathrm{PPR}(v)>\delta\). In that sense, AGP is a \((\delta,10\%)\)-relative-error algorithm with success probability at least \(99\%\) [2106.03058].

The expected running time of the original method is
\[
E[T] = O\!\left( \frac{L^2}{\delta}\sum_{i=1}^L \left\|Y_i(D^{-a}AD^{-b})^i x\right\|_1 \right),
\]
with the summary statement that, in most measures, \(\sum \|\cdot\|_1 = O(1)\), giving \(O(\log^2(1/\delta)/\delta)\) up to logarithmic factors [2106.03058].

The 2025 refinement improves both the parameter choice and the query-time dependence on \(L\). AGP-Static++ sets
\[
\epsilon \leftarrow \frac{c^2\delta}{2(L+1)},
\]
and a variance analysis summarized through Lemmas 4–7 yields
\[
\mathrm{Var}[\hat \pi(v)] \le \epsilon \cdot L \cdot \pi(v).
\]
The resulting expected query time is
\[
O\Bigl(T_s\;\tfrac1{\epsilon}\;\sum_{i=1}^L\bigl\|\,Y_i\,(D^{-a}AD^{-b})^i x\bigl\|_1\;+\;C_{\rm init}\Bigr),
\]
where \(T_s=O(\log^2 n)\) for AGP-Static and \(T_s=O(\log n)\) for AGP-Static++, and
\[
C_{\rm init}=O(n)
\]
to read \(x\), or \(O(L/\delta)\) when \(x\) admits a compact representation [2509.10036]. Substituting \(\epsilon=\Theta(\delta/L)\) and \(T_s=\Theta(\log n)\) gives
\[
O\Bigl(\log n\,\tfrac{L}{\delta}\sum_{i=1}^L\|Y_i(D^{-a}AD^{-b})^i x\|_1+n\Bigr).
\]

The fully dynamic extension, AGP-Dynamic, preserves the AGP-Static++ query complexity and approximation errors while achieving \(O(1)\) amortized time per edge insertion or deletion [2509.10036]. It does so through lazy degree updates and reference degrees \(\tilde d_u\), maintaining the invariant
\[
\frac12 \tilde d_u \le d_u \le 2\tilde d_u,
\]
so that stale sampling upper bounds remain within a constant factor of the true probabilities.

## 5. Instantiations across proximity queries and GNN propagation

AGP specializes to a range of graph kernels and propagation rules by choosing parameters in the unified equation [2106.03058].

| Task | Parameters | Seed/input |
|---|---|---|
| L-step transition | \(a=0, b=1, w_i=1\) if \(i=L\) else \(0\) | \(x=e_s\) |
| PageRank / PPR | \(a=0, b=1, w_i=\alpha(1-\alpha)^i\) | \(x=e_s\) or uniform teleport |
| Heat-kernel PR | \(a=0, b=1, w_i=e^{-t}t^i/i!\) | \(x=e_s\) |
| Katz | \(a=0, b=0, w_i=\beta^i\) | \(x=e_s\) |
| SGC | \(a=b=\tfrac12, w_i=1[i=L]\) | feature column |
| APPNP | \(a=b=\tfrac12, w_i=\alpha(1-\alpha)^i\) | feature column |
| GDC | \(a=b=\tfrac12, w_i=e^{-t}t^i/i!\) | feature column |

The original paper emphasizes two concrete application domains. In local clustering via heat-kernel PageRank, AGP was evaluated on YouTube, Orkut, Friendster, and Twitter, using MaxError and sweep-cut conductance as metrics. The reported result is that AGP attains the lowest MaxError at all running points and finds clusters of strictly lower conductance/time trade-off, with \(10\)–\(100\times\) speedups at \(\delta\approx 10^{-4}\) [2106.03058].

In node classification with GNNs, AGP was used as the feature-propagation engine for SGC, APPNP, and GDC on Reddit, Yelp, Amazon, and Papers100M. Across all four graphs, the AGP-based versions match exact-propagation accuracy within \(0.1\)–\(0.2\%\) while cutting propagation time by \(5\)–\(20\times\), and on Papers100M the exact SGC/APPNP pipelines require “tens of thousands of seconds” whereas AGP finishes in “a few thousand seconds.” The memory overhead is reported as \(O(n)\) extra memory, contrasted with \(O(n\log n)\) or \(O(nk)\) for PPRGo and GBP [2106.03058].

The 2025 revisit broadens the stated application range to dynamic graphs and dynamic parameterized queries, explicitly naming social networks, knowledge-graph RAG for LLMs, real-time recommender systems, and dynamic GNNs as scenarios where evolving graphs and on-the-fly choices of \(w_i,a,b\) matter [2509.10036]. This suggests an expansion of AGP from static precomputation toward online graph-serving settings.

## 6. Later developments, comparisons, and disambiguation

A significant later development is the distinction between AGP and newer acceleration methods based on Chebyshev approximation. The paper "Scaling Up Graph Propagation Computation on Large Graphs: A Local Chebyshev Approximation Approach" introduces ChebyPower and ChebyPush for graph propagation, with ChebyPush reported as \(3\)–\(8\times\) faster than state-of-the-art Push and randomized Push methods including AGP, PWPush, and PWPushSOR for SSPPR, and \(4\)–\(10\times\) faster than AGP and TEA+ for HKPR under comparable error [2412.10789]. That comparison places AGP as a strong prior baseline for local diffusion, but not as the final word on graph propagation computation.

The conceptual contrast is also precise. AGP is built around randomized push-and-sample propagation on a truncated power expansion, whereas the Chebyshev work replaces the Taylor basis \(\{x^k\}\) with the orthogonal Chebyshev basis \(\{T_k\}\), obtaining a \(\sqrt{N}\) acceleration predicted by approximation theory and deriving both a global ChebyPower routine and a local ChebyPush routine with provable error guarantees [2412.10789]. A plausible implication is that AGP and Chebyshev-based propagation occupy adjacent points in the design space: both target approximate graph diffusion, but they exploit different approximation bases and locality mechanisms.

The term “AGP” also appears in unrelated literatures. In probabilistic graphical models, "Join-Graph Propagation Algorithms" uses AGP to denote a family of parameterized message-passing algorithms, including Mini-Clustering and Iterative Join-Graph Propagation, for approximate inference in graphical models [1401.3489]. In systems work, GraphGuess is described as an “Approximate Graph Propagation (AGP)” framework with adaptive correction for iterative graph algorithms such as PageRank, SSSP, BP, and WCC [2104.10039]. Distance-estimation work based on seed propagation, such as PROPAGATE, also uses propagation-oriented approximation language but addresses distance-based metrics rather than the weighted diffusion framework of AGP proper [2301.06499]. These usages are terminological collisions rather than extensions of the 2021 Approximate Graph Propagation framework.

Within graph learning, propagation itself is also studied outside AGP. For example, propagation-embracing MLP distillation explicitly models a propagation operator \(\Pi\), often instantiated by PPR, to transfer structural information from a teacher GNN to a student MLP [2311.11759]. This is related at the level of propagation semantics, but it is not an AGP algorithm in the formal sense of the randomized \((\delta,c)\)-approximation framework.

## 7. Significance and open directions

The 2021 formulation characterizes AGP as the first single algorithm that handles popular proximity kernels and GNN propagation rules, gives a rigorous relative-error guarantee, and runs in time proportional to the nonzero output size up to logarithmic factors [2106.03058]. The 2025 revisit retains that unifying role while tightening the static theory and extending the framework to dynamic settings, specifically by removing an \(O(\log^2 n)\) query overhead relative to the earlier practical implementation, eliminating one factor of \(L\) in the runtime dependence, and achieving \(O(1)\) amortized update time in AGP-Dynamic [2509.10036].

Empirically, the revised framework reports that AGP-Static++ is up to \(10\times\) faster than AGP-Static, AGP-Dynamic matches AGP-Static++ within \(20\%\) overhead, and AGP-Dynamic achieves update-time speedups of up to \(177\times\) over the \(O(n)\) per-update maintenance of AGP-Static++ [2509.10036]. These are implementation-level consequences of tighter subset sampling, improved variance analysis, and lazy bucket maintenance rather than changes to the underlying propagation definition.

A common misconception is to identify AGP exclusively with Personalized PageRank or with a single local-push routine. The formalism and specialization table show that AGP is instead a parameterized framework for truncated graph diffusions, spanning PageRank-style kernels, heat kernels, Katz scores, transition probabilities, and decoupled GNN propagators [2106.03058]. Another potential misunderstanding is to treat AGP as inherently static; the dynamic extensions demonstrate that the same approximation model can be maintained under edge insertions and deletions without changing its query guarantees [2509.10036].

Taken together, the 2021 and 2025 papers define AGP as a general-purpose methodology for approximate graph diffusion under relative-error guarantees, with subsequent work clarifying both its strengths and its boundaries. Its role in the literature is therefore dual: it is both a practical algorithmic toolkit for node proximity and GNN propagation, and a baseline against which newer approximation schemes for graph propagation are now measured [2106.03058] [2509.10036] [2412.10789].

Source: https://www.emergentmind.com/topics/approximate-graph-propagation-agp