1: \begin{abstract}

2: We propose faster methods for unconstrained optimization of \emph{structured convex quartics}, which are convex functions of the form

3: \begin{equation*}

4: f(x) = c^\top x + x^\top \bG x + \bT[x,x,x] + \frac{1}{24}\norm{\bA x}_4^4

5: \end{equation*}

6: for  $c \in \reals^d$, $\bG \in \reals^{d \times d}$, $\bT \in \reals^{d \times d \times d}$, and $\bA \in \reals^{n \times d}$ such that $\bA^\top \bA \succ 0$. In particular, we show how to achieve an $\eps$-optimal minimizer for such functions with only $O(n^{1/5}\log^{O(1)}(\Zcal/\eps))$ calls to a gradient oracle and linear system solver, where $\Zcal$ is a problem-dependent parameter.

7: Our work extends recent ideas on efficient tensor methods and higher-order acceleration techniques to develop a descent method for optimizing the relevant quartic functions. As a natural consequence of our method, we achieve an overall cost of $O(n^{1/5}\log^{O(1)}(\Zcal / \eps))$ calls to a gradient oracle and (sparse) linear system solver for the problem of $\ell_4$-regression when $\bA^\top \bA \succ 0$, providing additional insight into what may be achieved for general $\ell_p$-regression. Our results show the benefit of combining efficient higher-order methods with recent acceleration techniques for improving convergence rates in fundamental convex optimization problems.

8: \end{abstract}

9: