The Bojanov--Naidenov inequality for quartics and second derivatives

Technical Summary: The Bojanov–Naidenov Inequality for Quartics and Second Derivatives

Problem Statement
This paper addresses the Bojanov–Naidenov problem for algebraic polynomials on the interval $[-1, 1]$. Specifically, it settles the case where the degree of the polynomial is $n=4$ and the order of the derivative is $k=2$. The central question is whether the Chebyshev polynomial of the first kind, $T_n(x)$, maximizes the integral of a strictly increasing convex function applied to the absolute value of the $k$-th derivative, among all polynomials of degree at most $n$ with a uniform norm bounded by 1.

Formally, let $\pi_n$ be the space of real algebraic polynomials of degree at most $n$, and $\pi^\circ_n = \{P \in \pi_n : \|P\|_{C[-1,1]} \le 1\}$. The problem asks if, for every $P \in \pi^\circ_n \setminus \{\pm T_n\}$ and every strictly increasing convex function $\phi: [0, \infty) \to \mathbb{R}$, the inequality
$$\int_{-1}^1 \phi(|P^{(k)}(x)|) \, dx < \int_{-1}^1 \phi(|T_n^{(k)}(x)|) \, dx$$
holds. While the endpoint cases ($k=n$ and $k=1$) and certain special classes of polynomials were previously resolved, the general intermediate case for $n=4, k=2$ remained open.

Methodology
The proof is described as "elementary and finite," relying on the specific properties of quartic polynomials rather than general functional analytic arguments. The methodology proceeds through the following steps:

1. Interpolation and Convexity Reduction:
   A quartic polynomial is uniquely determined by its values at the five extremal points of $T_4(x)$, denoted $x_j = \cos(j\pi/4)$ for $j=0, \dots, 4$. The authors define a map from the vector of nodal values $y = (y_0, \dots, y_4)$ to the integrated tail of the second derivative:
   $$F_t(y) = \int_{-1}^1 (|P_y''(x)| - t)_+ \, dx,$$
   where $(u)_+ = \max(u, 0)$. By observing that interpolation is linear and the function $u \mapsto (|u|-t)_+$ is convex, they prove that $F_t(y)$ is a convex function on $\mathbb{R}^5$. Consequently, the maximum of $F_t(y)$ over the feasible set (the image of polynomials with $\|P\|_\infty \le 1$) is bounded by the maximum over the full hypercube $[-1, 1]^5$, which is attained at one of the $2^5 = 32$ vertices.

2. Symmetry and Reduction to Quadratic Forms:
   At the vertices of the hypercube, the nodal values $y_j$ are $\pm 1$. The second derivative $P_y''(x)$ becomes a quadratic polynomial. Using symmetries (sign changes and reflection $x \to -x$), the authors reduce the 32 possible second derivatives to a finite list of distinct quadratic forms (including a linear case and the zero case).

3. Level Set Comparison (Tail Estimates):
   The core of the proof involves comparing the "integrated tails" of these vertex quadratics against the tail of the Chebyshev second derivative, $T_4''(x) = 96x^2 - 16$. The authors define the level-set function $\Lambda_q(t) = \text{meas}\{x \in [-1, 1] : |q(x)| > t\}$. Since the integrated tail is the integral of the level set function, proving $\Lambda_q(t) \le \Lambda_{T_4''}(t)$ for all $t \ge 0$ establishes the tail inequality.
   The authors perform an explicit, case-by-case analysis of the level sets for each reduced quadratic form. They utilize elementary calculus to compare the roots of the quadratics and verify the inequality across different intervals of $t$.

4. Extension to General Convex Functions:
   Once the tail inequality is established for all $t \ge 0$, the authors invoke a standard representation of convex functions (via the Choquet integral or similar representation involving $(u-t)_+$) to extend the result to any nondecreasing convex function $\phi$.

Key Contributions and Results

• Theorem 1: For any $P \in \pi_4$ with $\|P\|_{C[-1,1]} \le 1$ and $T_4(x) = 8x^4 - 8x^2 + 1$, the following tail inequality holds for all $t \ge 0$:
  $$\int_{-1}^1 (|P''(x)| - t)_+ \, dx \le \int_{-1}^1 (|T_4''(x)| - t)_+ \, dx.$$
• Corollary: This tail dominance implies the integral inequality for every nondecreasing convex function $\phi$:
  $$\int_{-1}^1 \phi(|P''(x)|) \, dx \le \int_{-1}^1 \phi(|T_4''(x)|) \, dx.$$
• Strictness: If $\phi$ is strictly increasing and convex, and $P \neq \pm T_4$, the inequality is strict.
• Norm Bound: The paper also establishes that $\|P''\|_{C[-1,1]} \le 80 = \|T_4''\|_{C[-1,1]}$, with equality if and only if $P = \pm T_4$.

Significance and Limitations
The paper claims to provide a complete solution to the Bojanov–Naidenov problem for the specific intermediate case of quartics ($n=4, k=2$). The significance lies in resolving a long-standing open problem in approximation theory for this specific degree.

However, the authors explicitly note the limitations of their approach. The proof relies heavily on the fact that a quartic is determined by five points, reducing the infinite-dimensional optimization problem to a finite set of 32 vertices. The authors state that this "finite character of the proof is also its limitation": while it solves the concrete case of $n=4$, it does not provide a direct argument or generalizable mechanism for higher degrees ($n > 4$). The work is presented as a definitive solution to this specific instance rather than a new general theory for all $n$.