We consider the robust curve fitting problem, for both algebraic and Fourier (trigonometric) polynomials, in the presence of outliers. In particular, we study the model of Arora and Khot (STOC 2002), who were motivated by applications in computer vision. In their model, the input data consists of ordered pairs (x<sub>i</sub>, y<sub>i</sub>) ε [-1, 1] × [-1, 1], i = 1, 2,..., N, and there is an unknown degree-d polynomial p such that for all but ρ fraction of the i, we have |p(x<sub>i</sub>) - y<sub>i</sub>|≤ δ. Unlike Arora-Khot, we also study the trigonometric setting, where the input is from T × [-1, 1], where T is the unit circle. In both scenarios, the i corresponding to errors are chosen randomly, and for such i the errors in the yi can be arbitrary. The goal is to output a degree-d polynomial q such that ||p - q||<sub>∞</sub> is small (for example, O(δ)). Arora and Khot could achieve a polynomial-time algorithm only for ρ = 0. Daltrophe et al. observed that a simple median-based algorithm can correct errors if the desired accuracy δ is large enough. (Larger δ makes the output guarantee easier to achieve, which seems to typically outweigh the weaker input promise.) We dramatically expand the range of parameters for which recovery of q is possible in polynomial time. Specifically, we show that there are polynomial-time algorithms in both settings that recover q up to l∞ error O(δ.99) provided 1) ρ ≤/c1log d and δ ≥ 1/(log d)c, or 2) ρ ≤ c1/log log d/log2 d and δ ≥ 1/dc. Here c is any constant and c1 is a small enough constant depending on c. The number of points that suffices is N = Õ(d) in the trigonometric setting for random x<sub>i</sub> or arbitrary x<sub>i</sub> that are roughly equally spaced, or in the algebraic setting when the x<sub>i</sub> are chosen according to the Chebyshev distribution, and N = Õ(d2) in the algebraic setting with random (or roughly equally spaced) x<sub>i</sub>.
[1]
T. J. Rivlin.
An Introduction to the Approximation of Functions
,
2003
.
[2]
Karl Zeller,et al.
Schwankung von Polynomen zwischen Gitterpunkten
,
1964
.
[3]
Shie Mannor,et al.
Robust Regression and Lasso
,
2008,
IEEE Transactions on Information Theory.
[4]
Erich Kaltofen,et al.
Sparse multivariate function recovery from values with noise and outlier errors
,
2013,
ISSAC '13.
[5]
Annie A. M. Cuyt,et al.
Approximation Theory
,
2008,
Wiley Encyclopedia of Computer Science and Engineering.
[6]
Radakovič.
The theory of approximation
,
1932
.
[7]
Robert C. Bolles,et al.
Random sample consensus: a paradigm for model fitting with applications to image analysis and automated cartography
,
1981,
CACM.
[8]
Erich Kaltofen,et al.
Sparse multivariate function recovery with a high error rate in the evaluations
,
2014,
ISSAC.
[9]
Shlomi Dolev,et al.
Big Data Interpolation an Efficient Sampling Alternative for Sensor Data Aggregation
,
2012,
ALGOSENSORS.
[10]
Boris Shekhtman,et al.
On discrete norms of polynomials
,
2006,
J. Approx. Theory.
[11]
Sanjeev Arora,et al.
Fitting algebraic curves to noisy data
,
2002,
STOC '02.
[12]
T. J. Rivlin,et al.
The growth of polynomials bounded at equally spaced points
,
1992
.
[13]
G. Alexits.
Approximation theory
,
1983
.