$ \newcommand{\R}{\mathbb{R}} \newcommand{\C}{\mathbb{R}} \newcommand{\ve}[1]{{\boldsymbol{\mathbf{#1}}}} \newcommand{\trans}{\top} \newcommand{\ma}[1]{\ve{#1}} \newcommand{\tve}[1]{{\widetilde{\boldsymbol{\mathbf{#1}}}}} \DeclareMathOperator*{\minimize}{minimize} \newcommand{\set}[1]{{\mathcal{#1}}} $

Research interests

I am an applied mathematician interested in the application of optimization to problems in system identification and design under uncertainty. My current work with Paul focuses on constructing ridge approximations of expensive quantities of interest using only function values and then exploiting the structure of these approximations to solve chance constrainted design problems. My thesis work centered on solving a large nonlinear least squares problem by projecting onto a low-dimensional subspace of the measurements. I am currently working on extending this projection approach to other nonlinear least squares problems.

Publications

Data-driven polynomial ridge approximation using variable projection
Jeffrey M. Hokanson and Paul G. Constantine

A ridge approximation mimics a scalar function of many variables $f: \R^m\to \R$ by another function $g:\R^n \to \R$ of a few linear combinations of inputs such that $$ f(\ve x) \approx g(\ma U^\trans \ve x) \quad \ma U\in \R^{m\times n} \quad \ma U^\trans \ma U = \ma I. $$ In this paper we find the ridge function $g$ and the subspace spanned by $\ma U$ by minimizing the 2-norm mismatch between $f(\ve x_i)$ and $g(\ma U^\trans \ve x_i)$ over a user provided set of points $\lbrace \ve x_i\rbrace_{i=1}^N$. By assuming $g$ is a multidimensional polynomial, we can implicitly solve for the coefficients using variable projection and solve the remaining optimization problem over the Grassmann manifold of $n$-dimensional subspaces of $\R^m$. The resulting algorithm allows for the rapid construction of ridge approximations using only function values.

Projected nonlinear least squares for exponential fitting
Jeffrey M. Hokanson

The exponential fitting problem seeks to minimize the mismatch between a noisy vector of measurements $\tve y\in \R^n$ and a sum of complex exponentials in a least squares sense: $$ \minimize_{\ve \omega, \ve a\in \C^p} \| \ve f(\ve\omega, \ve a) - \tve y\|_2^2 \qquad [\ve f(\ve\omega, \ve a)]_k = \sum_{j=1}^p a_j e^{\omega_j k}. $$ To speed optimization when vast quantities of data are present, we project the problem onto a low dimensional subspace $$ \|\ma W \ma W^*\ve f(\ve\omega, \ve a) - \ma W^*\tve y\|_2^2= \|\ma W^*\ve f(\ve\omega, \ve a) - \ma W^*\tve y\|_2^2, \quad \ma W^*\ma W = \ma I. $$ By a careful selection of $\ma W$ that allows us to efficiently evaluate $\ma W^*\ve f(\ve\omega, \ve a)$ we are able to outperform subspace-based approaches for exponential fitting.

One can hear the composition of a string: experiments with an inverse eigenvalue problem
Steven J. Cox, Mark Embree, and Jeffrey M. Hokanson

An inverse eigenvalue problem seeks to infer properties of a system through its eigenvalues. A classic example is Kac's question: Can one hear the shape of a drum?; the answer turns out to be no in two dimensions. In this paper we develop a version of this problem accessible to undergraduates in a matrix analysis course. Here we consider the case of a beaded string where the beads are symmetrically distributed around the mid-point of the string. Using classical results of Gantmacher, Krein, and others, we develop an algorithm to recover the masses and locations of these beads given the resonant frequencies of the beaded string. We apply this algorithm to datasets generated from an actual beaded string. This example is used as component of a teaching lab suitable for juniors and above.

Applied Math Laboratory

During the summer of 2007 I was supported by a Brown teaching grant to help build a teaching laboratory to accompany the CAAM department's undergraduate linear algebra class. Together with Steve Cox and Mark Embree I helped devise the experiments and built the accompanying apparatus. Below, I highlight a few of these experiments.

Force table

As part of labs 3 and 4, the students would work with a spring network resembling problems in the course textbook. In lab 3, the students would verify the forward problem: knowing the stiffnesses of the springs, the topology, and the applied loads they would compute the expected displacement and check that it matched the model. In lab 4, the students would solve the inverse problem: knowing the topology of the spring network and the displacements they would attempt to infer stiffness of each spring.

Multiple pendula

In lab 8, the students would recreate the 1733 experiment of Daniel Bernoulli illustrating the superposition of eigenfunctions on a multiple pendula. First the students would displace the each of the masses in an eigenvector, and observe that the motion was always a scalar mulitple of the initial displacement. Then the students displace the masses randomly and measure their displacement as a function of time using video tracking software. By taking the discrete Fourier transform of this displacement the students would note the same three resonant frequencies of the pendula seen when the masses were displaced in an eigenvector configuration.

Beaded string

In labs 9 and 10 the students would work with a beaded string. In the first lab, the students would study the forward problem: placing beads on the string knowing their displacement and mass and verify the eigenvalues of the string matched a discrete model In the second lab, the students would place the beads symmetrically around the mid-point of the string and estimate the masses and location of the beads from the measured eigenvalues.