THE GROTHENDIECK CONSTANT IS STRICTLY SMALLER THAN KRIVINE’S BOUND

The (real) Grothendieck constant ${K}_{G} $ is the infimum over those $K\in (0, \infty )$ such that for every $m, n\in \mathbb{N} $ and every $m\times n$ real matrix $({a}_{ij} )$ we have $$\begin{eqnarray*}\displaystyle \max _{\{ x_{i}\} _{i= 1}^{m} , \{ {y}_{j} \} _{j= 1}^{n} \subseteq {S}^{n+ m- 1} }\sum _{i= 1}^{m} \sum _{j= 1}^{n} {a}_{ij} \langle {x}_{i} , {y}_{j} \rangle \leqslant K\max _{\{ \varepsilon _{i}\} _{i= 1}^{m} , \{ {\delta }_{j} \} _{j= 1}^{n} \subseteq \{ - 1, 1\} }\sum _{i= 1}^{m} \sum _{j= 1}^{n} {a}_{ij} {\varepsilon }_{i} {\delta }_{j} . &&\displaystyle\end{eqnarray*}$$ The classical Grothendieck inequality asserts the nonobvious fact that the above inequality does hold true for some $K\in (0, \infty )$ that is independent of $m, n$ and $({a}_{ij} )$. Since Grothendieck’s 1953 discovery of this powerful theorem, it has found numerous applications in a variety of areas, but, despite attracting a lot of attention, the exact value of the Grothendieck constant ${K}_{G} $ remains a mystery. The last progress on this problem was in 1977, when Krivine proved that ${K}_{G} \leqslant \pi / 2\log (1+ \sqrt{2} )$ and conjectured that his bound is optimal. Krivine’s conjecture has been restated repeatedly since 1977, resulting in focusing the subsequent research on the search for examples of matrices $({a}_{ij} )$ which exhibit (asymptotically, as $m, n\rightarrow \infty $) a lower bound on ${K}_{G} $ that matches Krivine’s bound. Here, we obtain an improved Grothendieck inequality that holds for all matrices $({a}_{ij} )$ and yields a bound ${K}_{G} \lt \pi / 2\log (1+ \sqrt{2} )- {\varepsilon }_{0} $ for some effective constant ${\varepsilon }_{0} \gt 0$. Other than disproving Krivine’s conjecture, and along the way also disproving an intermediate conjecture of König that was made in 2000 as a step towards Krivine’s conjecture, our main contribution is conceptual: despite dealing with a binary rounding problem, random two-dimensional projections, when combined with a careful partition of ${ \mathbb{R} }^{2} $ in order to round the projected vectors to values in $\{ - 1, 1\} $, perform better than the ubiquitous random hyperplane technique. By establishing the usefulness of higher-dimensional rounding schemes, this fact has consequences in approximation algorithms. Specifically, it yields the best known polynomial-time approximation algorithm for the Frieze–Kannan Cut Norm problem, a generic and well-studied optimization problem with many applications.