1: \begin{abstract}

2: %We present generalizations of Newton's method that incorporate derivatives of arbitrarily high order but maintain a polynomial dependence on dimension in their cost per iteration.

3: 

4: We present generalizations of Newton's method that incorporate derivatives of an arbitrary order $d$ but maintain a polynomial dependence on dimension in their cost per iteration. At each step, our $d$\textsuperscript{th}-order method uses semidefinite programming to construct and minimize a sum of squares-convex approximation to the $d$\textsuperscript{th}-order Taylor expansion of the function we wish to minimize. We prove that our $d$\textsuperscript{th}-order method has local convergence of order $d$. This results in lower oracle complexity compared to the classical Newton method. We show on numerical examples that basins of attraction around local minima can get larger as $d$ increases. Under additional assumptions, we present a modified algorithm, again with polynomial cost per iteration, which is globally convergent and has local convergence of order $d$.

5: 

6: %Finally, in the case of $d$ odd and under convexity assumptions, we show how our algorithm can be adapted to give global convergence while preserving polynomial work per iteration and local convergence of order $d$.

7: 

8: %our $d$\textsuperscript{th}-order method can have larger basins of attraction around minima compared to the classical Newton method. 

9: 

10: 

11: 

12: %We prove that our $d$\textsuperscript{th}-order method has local convergence of order $d$ with only one oracle per iteration.  

13: 

14: 

15: 

16: %We analyze the convergence rates of our higher-order Newton methods and compare their basins of attraction around local minima to those of the classical Newton method.

17: \end{abstract}

18: