The diagonal_op_ function performs an in-place multiplication on the matrix A. Since this is inside the while loop, if a step is rejected and the loop repeats, the damping will be applied cumulatively (e.g., $(1+\lambda_1)(1+\lambda_2)...$) instead of being applied to the original $J^T J$ diagonal. This deviates from the standard Levenberg-Marquardt algorithm and can lead to excessively aggressive damping. Consider cloning the matrix or resetting the diagonal before applying damping in each iteration.
Originally posted by @gemini-code-assist[bot] in #35
Originally posted by @gemini-code-assist[bot] in #35