The coefficients θj of a periodic function with ‖Dmf‖ ≤ c satisfy Σ aj²θj² ≤ c² with aj ≈ (πj)m (Theorem 7.11). Observing Zj = θj + (σ/√n)εj, the best linear estimator in the worst case keeps Zj with weight wj = (1 − aj/μ)+, where μ solves (σ²/n) Σ aj(μ − aj)+ = c², and its risk (σ²/n) Σ wj is asymptotically the minimax risk over all estimators (Theorem 7.36). Panel (a) shows the weights; panel (b) shows the rescaled risk n2m/(2m+1)Rn converging to Pinsker's constant.
The basis is the Fourier basis of Theorem 7.11 (with √2 in front of the cosines and sines): a1 = 0, aj = (πj)m for even j and (π(j − 1))m for odd j ≥ 3. μ is found by bisection. The best projection keeps the first J coefficients, with worst-case risk Jσ²/n + c²/aJ+1² minimised over J. The dashed line is the limit with (σ²/π)2m/(2m+1); the dotted line is the book's printed form with (σ/π), which agrees only when σ = 1.
code/ch07_normal_means_minimax.py (sobolev_a, pinsker_linear, pinsker_n_risk, pinsker_limit).