A mathematical preprint reports a strong local result for an iterative construction of the Malfatti configuration, the target arrangement studied in the model: for every nondegenerate triangle, the target is a locally asymptotically stable fixed point when the contraction parameter q is set to 0.9. In plain terms, configurations that begin sufficiently close to the target move toward it at a geometric rate. The analysis does not say how large that starting neighborhood is, and it does not establish convergence from arbitrary starting configurations.
The work, labeled arXiv version 1 and dated 25 August 2026, examines three linked questions: whether the Malfatti configuration is a stable fixed point, which values of the module parameter permit local convergence, and whether the global initial-value problem can be solved. The manuscript also states that ChatGPT generated all of its mathematical proofs.
How the iteration works
The model uses a nine-dimensional state. Six line-circle correction modules and three circle-circle modules act simultaneously on the current configuration at each iteration. Each correction is designed to multiply its contact error by q, and the paper notes that repeated isolated updates converge when the absolute value of q is less than 1.
The simultaneous step is written as unit-step gradient descent on a weighted sum of squared contact errors. In ordinary terms, it uses the negative gradient of that combined error measure to choose its update. At the Malfatti configuration, the proof asserts that the six line-circle and three circle-circle constraint gradients are linearly independent, which makes the associated matrix H positive definite in the nine-dimensional state space.
A parameter range that works across triangles
For a particular fixed triangle, the stated local-stability condition is 1 − 2/λmax(M) < q < 1, where λmax(M) is the largest eigenvalue used in the triangle-specific calculation. The paper also gives a simpler triangle-independent sufficient range: one-half < q < 1 is enough for local stability for every nondegenerate triangle. The result is sufficient, not a demonstration that the interval is the only possible or best one.
A regular-triangle example illustrates how the geometry-specific calculation can produce a lower threshold. The paper reports a largest eigenvalue of about 2.11965684 and a critical q of about 0.056451 for that example. Those figures illustrate the stated fixed-triangle criterion, not a replacement for the uniform guarantee.
For q = 0.9, the stability proof bounds the largest eigenvalue of H by 0.9, combining a line-circle contribution bounded by 0.6 with a circle-circle contribution bounded by 0.3. Together with the asserted independence of the constraint gradients, that bound supplies the paper's local stability argument.
The hard part begins outside the local neighborhood
The central gap is global convergence. The paper does not provide an invariant region or a global Lyapunov argument that would guarantee convergence from arbitrary initial configurations, and it says that a satisfactory global initial-value solution remains out of reach. The local theorem therefore cannot be read as a guarantee for every starting arrangement.
Three circles that initially lie in the triangle with pairwise disjoint interiors can overlap after a single update, even when the starting state is arbitrarily close to the Malfatti configuration. Initial containment and non-overlap are therefore not preserved by the iteration.
The analysis reports another warning sign in the form of an explicit non-Malfatti critical point. At that point, the gradient of H is zero, the pairwise contact errors are −3/4, and the configuration is stated to be unstable. The example shows that stationary behavior in the model is not automatically the desired Malfatti solution, although it does not classify every possible critical point.
Numerical evidence, not a proof
A separate numerical experiment tested 10,000 randomly generated initial configurations satisfying the paper's stated positive-error conditions. Every reported trajectory converged to the Malfatti configuration. The authors present that result as evidence for their convergence conjecture, but as a conjecture rather than a proof.
The numerical finding comes with important unanswered questions: the supplied analysis does not report the generation procedure, convergence tolerance, stopping criteria, or reproducible outputs. It also says that independent mathematical verification of the ChatGPT-generated proofs was not reported. The most secure takeaway is therefore narrow but useful: the iteration has an analytical local-stability result with explicit q bounds, while its behavior from broad classes of initial configurations remains open.
Paper data and sources
Original title: ChatGPT solved the dynamic construction problem of Malfatti circles
Authors: Kazushi Ahara
Journal/Repository: arXiv
Status: Preprint, not yet peer-reviewed
First online: 2026-08-25
DOI: Not available
Original paper · Full text