Anchor-Based Gauge Map

• Anchor-based gauge maps are explicit linear reconstruction schemes that approximate global diffusion geometry using local anchor-to-node shortest-path distances.
• They employ a trilateration approach by solving a linear system with transformed distance measurements to efficiently estimate spectral coordinates.
• In practice, these methods achieve near-spectral performance in tasks like drug–drug interaction prediction while significantly reducing computational costs.

An anchor-based gauge map, or trilateration map, is an explicit linear reconstruction scheme that recovers truncated diffusion (spectral) coordinates of nodes in a graph from anchor-to-node shortest-path distance encodings. This method bridges two foundational positional encoding paradigms in graph machine learning: spectral coordinates from Laplacian or diffusion operators and anchor-based shortest-path distance encodings. Anchor-based gauge maps systematically approximate global diffusion geometry using local distance measurements to a fixed anchor set, yielding accuracy guarantees in both pointwise and matrix norms. In recent work, this framework has provided a scalable and theoretically rigorous alternative to expensive spectral eigendecompositions, showing practical success in molecular graph learning tasks such as drug–drug interaction (DDI) prediction (Yan et al., 8 Jan 2026).

1. Diffusion Geometry and Spectral Coordinates

Let $G=(V,E)$ be a connected, undirected graph of $n=|V|$ nodes. The normalized Laplacian operator is $L=I-(1/r)A$ for $r$-regular graphs ($A$: adjacency matrix). The diffusion operator at time $t>0$ is $K_t = e^{-tL}$, whose entries $k_t(u,v)$ define the graph’s heat kernel. The diffusion distance is

$$d_t(u,v)^2 = \sum_{j=1}^n e^{-2t\lambda_j}\left(\varphi_j(u)-\varphi_j(v)\right)^2 = k_{2t}(u,u) + k_{2t}(v,v) - 2k_{2t}(u,v),$$

where $(\lambda_j,\varphi_j)$ are the Laplacian eigenpairs. Truncating to the first $n=|V|$0 nontrivial modes ($n=|V|$1; $n=|V|$2) yields the $n=|V|$3-dimensional diffusion map:

$n=|V|$4

The truncated diffusion distance is $n=|V|$5.

2. Anchor-Based Distance Encodings

Anchor-based methods use a set $n=|V|$6 of $n=|V|$7 anchors (theoretical analysis uses i.i.d. uniform anchors on random regular graphs; practical implementations on molecular graphs use $n=|V|$8 anchors via farthest-point sampling in the shortest-path metric). For node $n=|V|$9, the anchor distance vector is

$L=I-(1/r)A$0

where $L=I-(1/r)A$1 is the shortest-path distance. Optionally, a monotone scalar transform $L=I-(1/r)A$2—such as $L=I-(1/r)A$3 or $L=I-(1/r)A$4—is applied elementwise to obtain $L=I-(1/r)A$5. This yields a low-dimensional anchor-based positional encoding.

3. Construction of the Trilateration (Gauge) Map

Given, for each anchor $L=I-(1/r)A$6, its spectral coordinates $L=I-(1/r)A$7, the true spectral coordinate $L=I-(1/r)A$8 for node $L=I-(1/r)A$9 satisfies $r$0 sphere equations:

$r$1

Let $r$2. Subtracting the equation for $r$3 from the first $r$4 yields the linear system

$r$5

with

$r$6

and

$r$7

Under affine-independence of the anchors, $r$8 is invertible and

$r$9

In practice, $A$0 is unavailable, and one substitutes transformed raw distances $A$1 into $A$2, defining the trilateration (gauge) map:

$A$3

which approximates $A$4. This "trilateration map" reconstructs the node's spectral embedding from its anchor distances.

4. Theoretical Guarantees on Reconstruction

On $A$5-regular graphs, the connection between shortest-path and diffusion distances is formalized as follows. For $A$6 with $A$7:

$A$8

(coupling argument), and

$A$9

(from truncation error). Thus, Theorem 2 states that for all $t>0$0 with $t>0$1 and any strictly increasing $t>0$2:

$t>0$3

Injecting $t>0$4 into the linear system yields

$t>0$5

with high probability (Theorem 5: Pointwise Reconstruction), and for the distance matrices $t>0$6:

$t>0$7

(Theorem 6: Frobenius-Gap). Thus, the trilateration map recovers leading $t>0$8 diffusion modes up to controlled errors. The conditioning of $t>0$9 governs stability, with heavy-tailed $K_t = e^{-tL}$0 but only $K_t = e^{-tL}$130% exceeding $K_t = e^{-tL}$2 on molecular graphs.

5. Practical Implementation and Performance

On DrugBank molecular graphs ($K_t = e^{-tL}$3), $K_t = e^{-tL}$4 anchors per molecule are selected by farthest-point sampling. The preferred transform is $K_t = e^{-tL}$5 (ablation supports its superiority). For exact theoretical trilateration, $K_t = e^{-tL}$6 and $K_t = e^{-tL}$7 anchors are sufficient, but diagnostics and Nyström approximations use the full $K_t = e^{-tL}$8 vectors. The Nyström kernel approximation $K_t = e^{-tL}$9 employs $k_t(u,v)$0, $k_t(u,v)$1, and Tikhonov regularization ($k_t(u,v)$2).

Diagnostic results on 80 graphs include:

• Relative Frobenius error of $k_t(u,v)$3 vs $k_t(u,v)$4: $k_t(u,v)$5 (median $k_t(u,v)$6)
• Coordinate MSE after Procrustes ($k_t(u,v)$7 vs $k_t(u,v)$8): $k_t(u,v)$9
• Pearson correlation ($$d_t(u,v)^2 = \sum_{j=1}^n e^{-2t\lambda_j}\left(\varphi_j(u)-\varphi_j(v)\right)^2 = k_{2t}(u,u) + k_{2t}(v,v) - 2k_{2t}(u,v),$$0 vs $$d_t(u,v)^2 = \sum_{j=1}^n e^{-2t\lambda_j}\left(\varphi_j(u)-\varphi_j(v)\right)^2 = k_{2t}(u,u) + k_{2t}(v,v) - 2k_{2t}(u,v),$$1): $$d_t(u,v)^2 = \sum_{j=1}^n e^{-2t\lambda_j}\left(\varphi_j(u)-\varphi_j(v)\right)^2 = k_{2t}(u,u) + k_{2t}(v,v) - 2k_{2t}(u,v),$$2 average

For DDI prediction (shared GNP backbone), various positional encoding schemes were benchmarked:

Method (DrugBank)	AUROC / F1
NoPE (no PE)	0.890 / 0.820
DE ($$d_t(u,v)^2 = \sum_{j=1}^n e^{-2t\lambda_j}\left(\varphi_j(u)-\varphi_j(v)\right)^2 = k_{2t}(u,u) + k_{2t}(v,v) - 2k_{2t}(u,v),$$3=16, $$d_t(u,v)^2 = \sum_{j=1}^n e^{-2t\lambda_j}\left(\varphi_j(u)-\varphi_j(v)\right)^2 = k_{2t}(u,u) + k_{2t}(v,v) - 2k_{2t}(u,v),$$4)	0.976 / 0.928
LapPE ($$d_t(u,v)^2 = \sum_{j=1}^n e^{-2t\lambda_j}\left(\varphi_j(u)-\varphi_j(v)\right)^2 = k_{2t}(u,u) + k_{2t}(v,v) - 2k_{2t}(u,v),$$5=8 Lap eigvecs)	0.980 / 0.934
RWSE (5-step walk)	0.892 / 0.812
HKS (5 heat times)	0.863 / 0.787

The anchor-based distance encoding with Nyström closely matches full spectral performance and consistently exceeds the no-encoding baseline. Similar results hold on ChCh-Miner. Ablation studies confirm $$d_t(u,v)^2 = \sum_{j=1}^n e^{-2t\lambda_j}\left(\varphi_j(u)-\varphi_j(v)\right)^2 = k_{2t}(u,u) + k_{2t}(v,v) - 2k_{2t}(u,v),$$6 and increased $$d_t(u,v)^2 = \sum_{j=1}^n e^{-2t\lambda_j}\left(\varphi_j(u)-\varphi_j(v)\right)^2 = k_{2t}(u,u) + k_{2t}(v,v) - 2k_{2t}(u,v),$$7 improve or maintain performance.

6. Connection to Nyström Approximation and Scalability

The same anchor set used for gauge mapping naturally enables a Nyström-type approximation of the full diffusion kernel $$d_t(u,v)^2 = \sum_{j=1}^n e^{-2t\lambda_j}\left(\varphi_j(u)-\varphi_j(v)\right)^2 = k_{2t}(u,u) + k_{2t}(v,v) - 2k_{2t}(u,v),$$8, leveraging anchor-to-node affinities. Here, anchor-based encodings not only reconstruct coordinates via trilateration but also generate a scalable approximate kernel with strong empirical alignment to spectral embeddings. On a 155-node DrugBank molecule, DE-Nyström (with $$d_t(u,v)^2 = \sum_{j=1}^n e^{-2t\lambda_j}\left(\varphi_j(u)-\varphi_j(v)\right)^2 = k_{2t}(u,u) + k_{2t}(v,v) - 2k_{2t}(u,v),$$9) matches the exact Gaussian-kernel eigenmap with mean nodewise error $(\lambda_j,\varphi_j)$0 and maximum $(\lambda_j,\varphi_j)$1.

A plausible implication is that anchor-based gauge maps, coupled with the Nyström scheme, provide a principled and computationally tractable alternative to eigendecomposition in large-scale graph learning, especially where spectral coordinates are costly or impossible to obtain exactly.

7. Summary and Significance

The anchor-based gauge map, as formalized by Yan et al. (Yan et al., 8 Jan 2026), delivers a rigorous and practical solution for approximating spectral positional signals from local distance measurements to anchors. The explicit trilateration reconstruction guarantees recovery of leading diffusion modes with bounded error both pointwise and in matrix norms. On molecular graphs, this yields nearly equivalent downstream performance to full Laplacian spectral encodings in DDI prediction tasks, while offering orders-of-magnitude gains in scalability. These findings establish anchor-based distance encodings, interpreted through the lens of diffusion geometry, as a powerful and flexible tool for geometric deep learning on graphs.