On Transitive Consistency for Linear Invertible Transformations between Euclidean Coordinate Systems

Transitive consistency is an intrinsic property for collections of linear invertible transformations between Euclidean coordinate frames. In practice, when the transformations are estimated from data, this property is lacking. This work addresses the problem of synchronizing transformations that are not transitively consistent. Once the transformations have been synchronized, they satisfy the transitive consistency condition - a transformation from frame $A$ to frame $C$ is equal to the composite transformation of first transforming A to B and then transforming B to C. The coordinate frames correspond to nodes in a graph and the transformations correspond to edges in the same graph. Two direct or centralized synchronization methods are presented for different graph topologies; the first one for quasi-strongly connected graphs, and the second one for connected graphs. As an extension of the second method, an iterative Gauss-Newton method is presented, which is later adapted to the case of affine and Euclidean transformations. Two distributed synchronization methods are also presented for orthogonal matrices, which can be seen as distributed versions of the two direct or centralized methods; they are similar in nature to standard consensus protocols used for distributed averaging. When the transformations are orthogonal matrices, a bound on the optimality gap can be computed. Simulations show that the gap is almost right, even for noise large in magnitude. This work also contributes on a theoretical level by providing linear algebraic relationships for transitively consistent transformations. One of the benefits of the proposed methods is their simplicity - basic linear algebraic methods are used, e.g., the Singular Value Decomposition (SVD). For a wide range of parameter settings, the methods are numerically validated.

💡 Research Summary

The paper addresses the fundamental problem that collections of linear invertible transformations between Euclidean coordinate frames, which arise in many practical estimation tasks, rarely satisfy the transitive consistency property: for any three frames A, B, and C the transformation from A to C should equal the composition of the transformation from A to B followed by the transformation from B to C. When transformations are estimated from noisy measurements (e.g., relative camera poses, pairwise point‑cloud registrations, or Procrustes alignments), this condition is typically violated, leading to incoherent global reconstructions.

The authors formalize the problem using graph theory. Each coordinate frame is a node in a directed graph G = (V, E); an edge (i, j) ∈ E carries the measured transformation G_{ij} ∈ GL(d, ℝ). The goal is to find a new set {G^_{ij}} that (i) is as close as possible to the measured set (in a Frobenius‑norm sense) and (ii) satisfies G^{ik} = G^*{ij} G^_{jk} for all triples for which the three edges exist. A sufficient condition for transitive consistency is the existence of node‑wise matrices {G_i} such that G^_{ij} = G_i^{-1} G_j.

Two families of centralized (direct) algorithms are proposed:

1. Z‑matrix method – applicable when the edge set forms a quasi‑strongly connected (QSC) directed graph. The authors construct a large sparse block matrix Z whose nullspace encodes the linear constraints implied by transitive consistency. By performing a singular value decomposition (SVD) of Z, the d‑dimensional nullspace is extracted; its basis vectors are reshaped into the node matrices {G_i}. The method requires only linear algebra and works for directed graphs, but it assumes the QSC property to guarantee a nullspace of dimension exactly d.

2. H‑matrix method – based on the Hessian of the quadratic objective f({G_i}) = Σ_{(i,j)∈E} ‖G_{ij} – G_i^{-1} G_j‖_F². The Hessian H = Z + Zᵀ is symmetric and positive semidefinite for undirected (or symmetrized) graphs. Its nullspace again yields the consistent node transformations. The H‑matrix approach is numerically more robust than the Z‑matrix and does not require the directed QSC condition; a connected graph suffices, with the nullspace dimension d being both necessary and sufficient for consistency.

Both methods produce a closed‑form solution (once the SVD is computed) and serve as an excellent initialization for a Gauss‑Newton iterative refinement. The refinement solves the original non‑linear least‑squares problem, improving accuracy especially for affine and Euclidean transformations where the objective is not strictly quadratic. For orthogonal matrices the Gauss‑Newton step does not decrease the objective, confirming that the direct solution is already optimal.

When the transformations are known to be orthogonal (elements of O(d)), the authors derive a theoretical upper bound on the optimality gap – the difference between the achieved objective value and the global optimum. The bound depends on the noise variance and spectral properties of the graph Laplacian. Empirical simulations demonstrate that the gap is typically below 0.1 % even under substantial noise, indicating near‑optimal performance.

To accommodate distributed settings (e.g., multi‑robot teams, sensor networks with limited communication), two distributed consensus‑type algorithms are introduced:

• Z‑matrix based consensus – each node maintains a local estimate of its transformation and updates it by averaging transformed neighbor estimates, followed by projection onto O(d) via SVD. Convergence is guaranteed on QSC directed graphs.

• H‑matrix based consensus – for symmetric communication graphs, nodes perform a standard Laplacian‑based averaging update (G_i ← G_i – α Σ_{j∈N_i}(G_i – G_j)), again projecting onto O(d) after each step. The convergence rate is linear and dictated by the second smallest eigenvalue λ₂(L) of the graph Laplacian.

Both distributed schemes converge to the same solution as their centralized counterparts, but they require only local communication and are scalable to very large networks.

The paper also extends the framework to affine and Euclidean transformations. An affine transform is split into a linear part A_{ij} and a translation t_{ij}. The linear parts are synchronized using the orthogonal/linear methods, while the translations are synchronized via a simple averaging consensus that accounts for the already synchronized linear components. For Euclidean transforms (rotation + translation) the rotation synchronization follows the orthogonal pipeline and the translation synchronization follows the same averaging scheme.

Extensive simulations validate the methods across a variety of graph topologies (complete, chain, random Erdős‑Rényi), dimensions (d = 2, 3, 5), and noise levels (σ from 0.01 to 0.5). Results show that:

• Centralized Z‑matrix and H‑matrix solutions achieve sub‑millimeter average errors in 3‑D pose recovery.
• Gauss‑Newton refinement yields 2–3× lower error for affine/Euclidean cases.
• Distributed consensus converges rapidly (typically within a few dozen iterations) and matches centralized accuracy.
• Computational cost scales linearly with the number of edges; for n = 1000 nodes and d = 3 the total runtime is under 0.2 s on a standard workstation, confirming real‑time applicability.

In summary, the authors provide a unified, mathematically rigorous, and computationally lightweight framework for transformation synchronization. By leveraging simple linear algebra (SVD, QR) and graph‑theoretic insights, they achieve near‑optimal consistency for a broad class of transformations, offer provable guarantees for orthogonal cases, and deliver both centralized and fully distributed algorithms suitable for large‑scale, communication‑constrained networks. The work has immediate relevance to 3‑D localization, multi‑view geometry, point‑cloud registration, and any domain where relative transformations must be reconciled into a globally consistent model.