The Sum-of-Squares (SoS) hierarchy is a semi-definite programming meta-algorithm that captures state-of-the-art polynomial time guarantees for many optimization problems such as Max- k-CSPs and Tensor PCA. On the flip side, a SoS lower bound provides evidence of hardness, which is particularly relevant to average-case problems for which NP-hardness may not be available. In this paper, we consider the following average case problem, which we call the Planted Affine Planes (PAP) problem: Given m random vectors d1, ..., dmin Rn, can we prove that there is no vector v ∈ IRnsuch that for all u ∈ [m], 〈v, du〉2= 1? In other words, can we prove that m random vectors are not all contained in two parallel hyperplanes at equal distance from the origin? We prove that for m ≤ n3/2-ε, with high probability, degree- nΩ(ε)SoS fails to refute the existence of such a vector v. When the vectors d1, ..., dmare chosen from the multivariate normal distribution, the PAP problem is equivalent to the problem of proving that a random n-dimensional subspace of Rmdoes not contain a boolean vector. As shown by Mohanty-Raghavendra-Xu [STOC 2020], a lower bound for this problem implies a lower bound for the problem of certifying energy upper bounds on the Sherrington-Kirkpatrick Hamiltonian, and so our lower bound implies a degree- nΩ(ε)SoS lower bound for the certification version of the Sherrington-Kirkpatrick problem.

We study the problem of computing the p → q operator norm of a matrix A in ℝm×n, defined as ‖A‖p→q : = supx∊ℝn\{0} ‖Ax‖q/‖x‖p. This problem generalizes the spectral norm of a matrix (p = q = 2) and the Grothendieck problem (p = ∞, q = 1), and has been widely studied in various regimes. When p ≥ q, the problem exhibits a dichotomy: constant factor approximation algorithms are known if 2 is in [q, p], and the problem is hard to approximate within almost polynomial factors when 2 is not in [q,p]. For the case when 2 is in [q, p] we prove almost matching approximation and NP-hardness results. The regime when p < q, known as hypercontractive norms, is particularly significant for various applications but much less well understood. The case with p = 2 and q > 2 was studied by [Barak et. al., STOC’12] who gave sub-exponential algorithms for a promise version of the problem (which captures small-set expansion) and also proved hardness of approximation results based on the Exponential Time Hypothesis. However, no NP-hardness of approximation is known for these problems for any p < q. We prove the first NP-hardness result for approximating hypercontractive norms. We show that for any 1 < p < q < ∞ with 2 not in [p, q], ‖A‖p→q is hard to approximate within 2O(log1−∊ n) assuming NP is not contained in BPTIME(2logO(1) n)).

We study the approximability of constraint satisfaction problems (CSPs) by linear programming (LP) relaxations. We show that for every CSP, the approximation obtained by a basic LP relaxation, is no weaker than the approximation obtained using relaxations given by $Ω\left(\frac{\log n}{\log \log n}\right)$ levels of the Sherali-Adams hierarchy on instances of size $n$. It was proved by Chan et al. [FOCS 2013] that any polynomial size LP extended formulation is no stronger than relaxations obtained by a super-constant levels of the Sherali-Adams hierarchy.. Combining this with our result also implies that any polynomial size LP extended formulation is no stronger than the basic LP. Using our techniques, we also simplify and strengthen the result by Khot et al. [STOC 2014] on (strong) approximation resistance for LPs. They provided a necessary and sufficient condition under which $Ω(\log \log n)$ levels of the Sherali-Adams hierarchy cannot achieve an approximation better than a random assignment. We simplify their proof and strengthen the bound to $Ω\left(\frac{\log n}{\log \log n}\right)$ levels.

We consider the following basic problem: given an n-variate degree-d homogeneous polynomial f with real coefficients, compute a unit vector x in R̂n that maximizes abs(f(x)). Besides its fundamental nature, this problem arises in diverse contexts ranging from tensor and operator norms to graph expansion to quantum information theory. The homogeneous degree-2 case is efficiently solvable as it corresponds to computing the spectral norm of an associated matrix, but the higher degree case is NP-hard. We give approximation algorithms for this problem that offer a trade-off between the approximation ratio and running time: in n̂O(q) time, we get an approximation within factor (O(n)/q)̂(d/2-1) for arbitrary polynomials, (O(n)/q)̂(d/4-1/2) for polynomials with non-negative coefficients, and (m /q)̂(1/2) for sparse polynomials with m monomials. The approximation guarantees are with respect to the optimum of the level-q sum-of-squares (SoS) SDP relaxation of the problem (though our algorithms do not rely on actually solving the SDP). Known polynomial time algorithms for this problem rely on “decoupling lemmas.” Such tools are not capable of offering a trade-off like our results as they blow up the number of variables by a factor equal to the degree. We develop new decoupling tools that are more efficient in the number of variables at the expense of less structure in the output polynomials. This enables us to harness the benefits of higher level SoS relaxations. Our decoupling methods also work with “folded polynomials,” which are polynomials with polynomials as coefficients. This allows us to exploit easy substructures (such as quadratics) by considering them as coefficients in our algorithms. We complement our algorithmic results with some polynomially large integrality gaps for d-levels of the SoS relaxation. For general polynomials this follows from known results for random polynomials, which yield a gap of Omega(n)̂(d/4-1/2). For polynomials with non-negative coefficients, we prove an Omega(n̂(1/6) /polylogs) gap for the degree-4 case, based on a novel distribution of 4-uniform hypergraphs. We establish an n̂Omega(d) gap for general degree-d, albeit for a slightly weaker (but still very natural) relaxation. Toward this, we give a method to lift a level-4 solution matrix M to a higher level solution, under a mild technical condition on M. From a structural perspective, our work yields worst-case convergence results on the performance of the sum-of-squareshierarchy for polynomial optimization. Despite the popularity of SoS in this context, such results were previously only known for the case of q = Omega(n).