Interlacing families. II: Mixed characteristic polynomials and the Kadison-Singer problem (Q2352912): Difference between revisions

Let \(\ell^2\) be the Hilbert space of square-summable sequences, let \(B(\ell^2)\) be the space of bounded operators on \(\ell^2\), and let \(A\) be a closed, self-adjoint subalgebra of \(B(\ell^2)\) containing the identity operator \(I\). A functional on \(A\) is called a state if it has norm \(1\) and maps positive operators to non-negative real numbers. The set of all states of \(A\) is convex and \(w^*\)-closed and its extreme points are called pure states. Using the Hahn-Banach theorem, we can extend every state of \(A\) to a state of \(B(\ell^2)\). Furthermore, every pure state of \(A\) can be extended to a pure state of \(B(\ell^2)\). In the context of the mathematical formulation of quantum mechanics, \textit{R. V. Kadison} and \textit{I. M. Singer} [Am. J. Math. 81, 383--400 (1959; Zbl 0086.09704)] asked the question if the extension of a pure state of an abelian subalgebra \(A\) is unique. They showed that the extension is not unique if \(A\) can be identified with \(L^\infty[0,1]\). But the question remainded open if every pure state of the algebra of bounded diagonal operators on \(\ell^2\) has a unique extension to a pure state of \(B(\ell^2)\). This question became known as the Kadison-Singer problem. Over the years, several equivalent formulations of this problem were found, for example \textit{J. Anderson}'s paving conjecture [Trans. Am. Math. Soc. 249, 303--329 (1979; Zbl 0408.46049)]. Anderson's paving conjecture transfers the Kadison-Singer problem from operator theory to the paving of matrices and reads in the finite-dimensional version as follows: For every \(\varepsilon > 0\), there is an \(r \in \mathbb N\) such that for every \(n \times n\) Hermitian matrix \(T\) with zero diagonal, there are diagonal projections \(P_1, \dotsc, P_r\) with \(\sum_{k=1}^r P_k = I_n\) such that \[ \| P_k T P_k \| \leq \varepsilon \| T \| \quad \text{for } k = 1, \dotsc, r. \] Other examples of equivalent formulations of the Kadison-Singer problem are Weaver's \(\text{KS}_r\) and \(\text{KS}'_r\) conjectures [\textit{N. Weaver}, Discrete Math. 278, No. 1--3, 227--239 (2004; Zbl 1040.46040)], the Feichtinger conjecture and the \(R_\varepsilon\) conjecture [\textit{P. G. Casazza} et al., Proc. Am. Math. Soc. 133, No. 4, 1025--1033 (2005; Zbl 1082.46018)]. For more details about the Kadison-Singer problem, see [\textit{P. G. Casazza} et al., Contemporary Mathematics 414, 299--355 (2006; Zbl 1110.46038)]. The authors solve the Kadison-Singer problem that was open for more than fifty years by proving the following theorem: If \(\varepsilon > 0\) and \(v_1, \dotsc, v_m\) are independent random vectors in \(\mathbb C^d\) with finite support such that \[ \sum_{k=1}^m \mathbb E v_k v_k^* = I_d \] and \[ \mathbb \| v_k \|^2 \leq \varepsilon \quad \text{for } k= 1, \dotsc, m, \] then \[ \mathbb P \left [ \left \| \sum_{k=1}^m v_k v_k^* \right \| \leq (1 + \sqrt \varepsilon ) ^2 \right ] > 0. \] This theorem implies Anderson's paving conjecture and therefore provides a positive solution to the Kadison-Singer problem. In order to prove the theorem, the authors study the eigenvalues and characteristic polynomials of \(\sum_{k=1}^m v_k v_k^*\). In the first step, they show that the characteristic polynomials of the matrices that arise in the theorem form a so-called interlacing family of polynomials, i.e., the roots of the polynomials are all real and have a special ordering. In the second step, they use a multivariate barrier argument to show that the largest root of the expected characteristic polynomial of \(\sum_{k=1}^m v_k v_k^*\) is at most \((1+ \sqrt \varepsilon)^2\). Combining the results of these two steps finishes the proof of the theorem. The line of reasoning is rather clear and easy to follow, not least because the paper is well-written and the proof uses quite elementary tools -- which makes it even more astonishing that the Kadison-Singer problem can be solved in this way. For Part I see [ibid. (2) 182, No. 1, 307--325 (2015; Zbl 1316.05066)].