Wiki
Wiki

Research notes on every problem, and a library of the papers behind them. Built from the open erdos repository.

Updated

The pointwise interpolation extremum


Source. Bernstein 1931, equation (1), printed p. 1025 / PDF p. 1, and its explanation on printed p. 1026 / PDF p. 2, in the complete source. The argument below uses degree at most dd and d+1d+1 nodes; Bernstein calls the degree nn.

Let d≥0d\ge0, let a0<⋯<ada_0<\cdots<a_d be real, and put

A(x)=∏j=0d(x−aj),ℓj(x)=∏i≠jx−aiaj−ai,F(x)=∑j=0d∣ℓj(x)∣.A(x)=\prod_{j=0}^d(x-a_j),\qquad \ell_j(x)=\prod_{i\ne j}\frac{x-a_i}{a_j-a_i},\qquad F(x)=\sum_{j=0}^d|\ell_j(x)|.

These are polynomials, including at their nodes. Away from the nodes, ℓj(x)=A(x)/((x−aj)A′(aj))\ell_j(x)=A(x)/((x-a_j)A'(a_j)). For every real xx,

F(x)=max⁡{∣P(x)∣:deg⁡P≤d, ∣P(aj)∣≤1 (0≤j≤d)}.(E)F(x)= \max\left\{|P(x)|:\deg P\le d,\ |P(a_j)|\le1\ (0\le j\le d)\right\}. \tag{E}

The maximum is the same for real or complex coefficients. In particular, F(x)≥1F(x)\ge1, F(aj)=1F(a_j)=1, and FF is continuous.

Proof. The polynomial P−∑jP(aj)ℓjP-\sum_jP(a_j)\ell_j has degree at most dd and vanishes at d+1d+1 distinct points, so it is zero. Thus

∣P(x)∣≤∑j∣P(aj)∣ ∣ℓj(x)∣≤F(x).|P(x)|\le\sum_j|P(a_j)|\,|\ell_j(x)|\le F(x).

For fixed real xx, assign P(aj)=sgn⁡ℓj(x)P(a_j)=\operatorname{sgn}\ell_j(x) when ℓj(x)≠0\ell_j(x)\ne0, and assign any value of modulus at most one to the remaining node data. The interpolating polynomial has real coefficients and value F(x)F(x) at xx. This proves attainment and (E). Interpolating the constant polynomial gives ∑jℓj=1\sum_j\ell_j=1, hence F≥1F\ge1; evaluation at a node and continuity give the other assertions.

Equality and endpoints. At a non-node xx, every ℓj(x)\ell_j(x) is nonzero. Equality ∣P(x)∣=F(x)|P(x)|=F(x) holds precisely when all node values have modulus one and the numbers P(aj)ℓj(x)P(a_j)\ell_j(x) have one common complex argument. For real coefficients this allows a common sign. At x=ajx=a_j, equality requires only ∣P(aj)∣=1|P(a_j)|=1. Also, F(x)=1F(x)=1 precisely when all the real numbers ℓj(x)\ell_j(x) are nonnegative. All statements apply at ±1\pm1 when the nodes lie in [−1,1][-1,1]; the rational expression is never evaluated through a zero denominator.

Dependencies. Polynomial interpolation and the triangle inequality, both proved or applied explicitly above.

Proof scope. Complete rewritten elementary argument; independently reviewed on 6 September 2026 (component C1 of the local-chain review). No formal verification is claimed.

Bears on. Problem 1153; Problem 1129, common global minimax quantity.