Skip to content

problems #4

Description

@bertrandchatelet-rgb

Below is a detailed analysis of the main errors and weaknesses in the text:

1. Major Inconsistency in Numerical Integration ("Marching Squares")

  • Sections 3.3 & 4.3: The author claims to provide a certified proof using interval arithmetic via the inari library.

  • The Flaw: Section 4.3 states that lemniscate length sampling relies on the standard double-precision floating-point (f64) marching squares algorithm, with a "conservative 5% grid error envelope absorbed into the interval arithmetic bounds".

  • Why it is incorrect: The marching squares algorithm relies on spatial discretization and linear interpolation, which offer no rigorous bounds on the true arc length of the level set. Arbitrarily adding a 5% margin completely breaks the IEEE 1788 certification chain. A true rigorous proof would require bounding the curve length analytically or via rigorous interval integration techniques (e.g., Taylor model / Krawczyk methods).

2. Non-Rigorous Estimation of Lipschitz Constants

  • Section 3.2: The certified upper bound on each box relies on the formula:

    $$u_{i} \le L(p_{\text{center}}) + \text{Lip}(\nabla L) \cdot \text{diam}(B_i) + \epsilon_{\text{grid}}$$

  • The Flaw: The text specifies that the Lipschitz constant $\text{Lip}(\nabla L)$ is "estimated via finite differences at multiple scales".

  • Why it is incorrect: Estimating a Lipschitz constant using finite differences is a heuristic sampling technique, not a certified mathematical proof. It offers no guarantee that the function $L(p)$ does not exhibit steep derivatives or singularities between sample points. In a certified branch-and-bound proof, the Lipschitz constant must be analytically bounded or evaluated over the entire interval using automatic interval differentiation.

3. Flawed Hessian Analysis (Local Optimality)

  • Sections 3.4 & 4.4: The author computes the Hessian matrix of $L(p)$ at $z^n - 1$ via central differences at various step sizes ($\epsilon \in \{10^{-4}, \dots, 10^{-7}\}$) and finds that all eigenvalues are strictly negative.

  • Why it is incorrect: Computing the Hessian via central differences is a numerical approximation method. While it provides an indication of a local maximum, it does not constitute a certified proof.

4. Claim of a "New" Closed-Form Expression (Theorem 1)

  • Section 2.2: The author presents the closed-form expression:

    $$L(z^n - 1) = 2^{1/n} \sqrt{\pi} \frac{\Gamma(1/(2n))}{\Gamma(1/(2n) + 1/2)}$$

    as a "new closed-form expression" (Contribution 2).

  • The Reality: This exact formula for the arc length of the lemniscate $z^n - 1$ is a standard, well-known result in complex analysis and geometry (derived directly via polar coordinate substitution and reduction to the Beta/Gamma functions). Claiming it as a novel contribution reveals a lack of proper literature review.

5. Paradox in Complexity Jump ($n=10$ to $n=11$)

  • Table 1: For $n=10$ (15 dimensions), the algorithm requires 32,768 evaluations and completes in 12 seconds.

  • For $n=11$ (17 dimensions), the evaluation count surges to 10,223,616 taking 33 minutes.

  • The Inconsistency: The author claims that "all non-extremizer boxes are eliminated at level 0 for $n \ge 5$". If boxes were truly eliminated at level 0 without further subdivision, the growth in evaluations would simply follow the increase in initial root boxes ($4^{n-2}$ or similar). The explosive ~300x jump between $n=10$ and $n=11$ indicates that deep box subdivisions occurred ($level > 0$), contradicting the claim that margins allow immediate elimination of the search space at level 0.

6 number of evaluations and computation time

Additional anomaly I found (not mentioned in the report): for n=4 (dim=3), the table shows 13,504 evaluations in 254 seconds, whereas 2³=8, and all other small values of n (3, 5, 6, 7...) resolve in under a second with evaluation counts matching 2^dim. Such a low dimension (3) justifies neither this number of evaluations nor this computation time — it's an isolated outlier in the table.

Summary

The paper attempts to blend formal verification terminology (IEEE 1788, Rust inari, branch-and-bound) with non-rigorous heuristic approximations (marching squares, finite-difference Lipschitz estimates, approximated Hessians). The certificate chain is broken, meaning the document does not constitute a valid mathematical verification of the conjecture for $n \in \{3, \dots, 12\}$.

errors_Mendoza_EHP_n3-12_March2026.pdf

Metadata

Metadata

Assignees

No one assigned

    Labels

    No labels
    No labels

    Type

    No type

    Projects

    No projects

    Milestone

    No milestone

    Relationships

    None yet

    Development

    No branches or pull requests

    Issue actions