A Hierarchical Online-Compression Approach to Proportional Subsets of Bounded Orthonormal Systems

Let $(ϕ_i)_{i=1}^n$ be an orthonormal system on a probability space with $\|ϕ_i\|_\infty\leq L$. We prove that, for every $2\leq k<n$, there is a set $I$ of at least $n-k$ indices such that $\|f\|_2\leq CM_k\sqrt{\log(e+M_k)}\|f\|_1$ for every $f$ in the span of $(ϕ_i)_{i\in I}$, where $M_k=L\sqrt{\frac{n}{k}\log(2k)}$. This replaces the logarithmic exponent $\frac{5}{2}$ in the corresponding theorem of Guédon, Mendelson, Pajor, and Tomczak-Jaegermann by $\frac{1}{2}$. We also obtain a partition into two sets of cardinality $\frac{n}{2}+O(\sqrt n)$ for which both spans satisfy the resulting norm comparison, and for uniformly bounded systems the proportional estimate is $C\sqrt{\log n\log\log n}$ for sufficiently large $n$. This is the third paper in a series systematically using gradient descent as a proof mechanism in harmonic analysis, following the $Λ(p)$ and Bernoulli restricted isometry papers of Burstein, Iosevich, and Krause, and the descent here is entropic mirror descent. We represent each function in an $L_1$ ball by a probability measure of controlled relative entropy, and this entropy pays for the corrections to the predictions of its coefficients. Processing the thresholds from large to small, we obtain a single weighted bound for all the corrections, which we turn into a finite description of approximate coefficient squares and then into a conditional Rademacher estimate, and symmetrization completes the argument directly on the $L_1$ ball. We also record consequences for concentration and for recovery from coefficient measurements.