We settle the complexity of dynamic least-squares regression (LSR), where rows and labels $(\mathbf{A}^{(t)}, \mathbf{b}^{(t)})$ can be adaptively inserted and/or deleted, and the goal is to efficiently maintain an $\epsilon$-approximate solution to $\min_{\mathbf{x}^{(t)}} \| \mathbf{A}^{(t)} \mathbf{x}^{(t)} - \mathbf{b}^{(t)} \|_2$ for all $t\in [T]$. We prove sharp separations ($d^{2-o(1)}$ vs. $\sim d$) between the amortized update time of: (i) Fully vs. Partially dynamic $0.01$-LSR; (ii) High vs. low-accuracy LSR in the partially-dynamic (insertion-only) setting. Our lower bounds follow from a gap-amplification reduction -- reminiscent of iterative refinement -- rom the exact version of the Online Matrix Vector Conjecture (OMv) [HKNS15], to constant approximate OMv over the reals, where the $i$-th online product $\mathbf{H}\mathbf{v}^{(i)}$ only needs to be computed to $0.1$-relative error. All previous fine-grained reductions from OMv to its approximate versions only show hardness for inverse polynomial approximation $\epsilon = n^{-\omega(1)}$ (additive or multiplicative) . This result is of independent interest in fine-grained complexity and for the investigation of the OMv Conjecture, which is still widely open.
翻译:摘要: 我们解决了动态最小二乘回归 (LSR) 的复杂度问题,其中可以自适应地插入和/或删除行和标签 $(\mathbf{A}^{(t)}, \mathbf{b}^{(t)})$,目标是对于所有 $t\in [T]$,有效地维护 $\min_{\mathbf{x}^{(t)}} \| \mathbf{A}^{(t)} \mathbf{x}^{(t)} - \mathbf{b}^{(t)} \|_2$ 的 $\epsilon$-近似解。我们证明了以下情况的摊销更新时间之间的尖锐分离 ($d^{2-o(1)}$ vs. $\sim d$):(i) 完全动态与部分动态的 $0.01$-LSR;(ii) 部分动态 (仅插入) 设置下高精度和低精度 LSR。我们的下界来自于一个 LSR 的 gap-amplification 级联带来的 Online Matrix Vector Conjecture (OMv) [HKNS15] 的精确版本到实数上的常数近似 OMv,其中第 $i$ 个在线乘积 $\mathbf{H}\mathbf{v}^{(i)}$ 只需要计算到相对误差为 $0.1$。所有之前的从 OMv 到其近似版本的精细化约简都仅显示了 $\epsilon = n^{-\omega(1)}$ (加法或乘法) 的逆多项式近似的难度。这个结果在精细化复杂度和 OMv Conjecture 的研究中都有独立的意义,OMv Conjecture 仍然广泛存在挑战性。