# Robust polynomial regression up to the information ... ecprice/papers/robustpoly.pdfآ regression,...

date post

03-Aug-2020Category

## Documents

view

1download

0

Embed Size (px)

### Transcript of Robust polynomial regression up to the information ... ecprice/papers/robustpoly.pdfآ regression,...

Robust polynomial regression up to the information theoretic limit

Daniel Kane Sushrut Karmalkar Eric Price

August 16, 2017

Abstract

We consider the problem of robust polynomial regression, where one receives samples (xi, yi) that are usually within σ of a polynomial y = p(x), but have a ρ chance of being arbitrary adversarial outliers. Previously, it was known how to efficiently estimate p only when ρ < 1log d .

We give an algorithm that works for the entire feasible range of ρ < 1/2, while simultaneously improving other parameters of the problem. We complement our algorithm, which gives a factor 2 approximation, with impossibility results that show, for example, that a 1.09 approximation is impossible even with infinitely many samples.

1 Introduction

Polynomial regression is the problem of finding a polynomial that passes near a collection of input points. The problem has been studied for 200 years with diverse applications [Ger74], including computer graphics [Pra87], machine learning [KKMS08], and statistics [Mac78]. As with linear regression, the classic solution to polynomial regression is least squares, but this is not robust to outliers: a single outlier point can perturb the least squares solution by an arbitrary amount. Hence finding a “robust” method of polynomial regression is an important question.

A version of this problem was formalized by Arora and Khot [AK03]. We want to learn a degree d polynomial p : [−1, 1]→ R, and we observe (xi, yi) where xi is drawn from some measure χ (say, uniform) and yi = p(xi) +wi for some noise wi. Each sample is an “outlier” independently with probability at most ρ; if the sample is not an outlier, then |wi| ≤ σ for some σ. Other than the random choice of outlier locations, the noise is adversarial. One would like to recover p̂ with ‖p− p̂‖∞ ≤ Cσ using as few samples as possible, with high probability.

In other contexts an `∞ requirement on the input noise would be a significant restriction, but outlier tolerance makes it much less so. For example, independent bounded-variance noise fits in this framework: by Chebyshev’s inequality, the framework applies with outlier chance ρ = 1100 and

σ = 10E[w2i ]1/2, which gives the ideal result up to constant factors. At the same time, the `∞ requirement on the input allows for the strong `∞ bound on the output.

Arora and Khot [AK03] showed that ρ < 1/2 is information-theoretically necessary for non- trivial recovery, and showed how to solve the problem using a linear program when ρ = 0. For ρ > 0, the RANSAC [FB81] technique of fitting to a subsample of the data works in polynomial time when ρ . log dd . Finding an efficient algorithm for ρ�

log d d remained open until recently.

Last year, Guruswami and Zuckerman [GZ16] showed how to solve the problem in polynomial time for ρ larger than log dd . Unfortunately, their result falls short of the ideal in several ways: it

needs a low outlier rate (ρ < 1log d), a bounded signal-to-noise ratio ( ‖p‖∞ σ < d

O(1)), and has a

super-constant approximation factor (‖p − p̂‖∞ . σ (

1 + ‖p‖∞σ

)0.01 ). It uses O(d2 logc d) samples

from the uniform measure, or O(d logc d) samples from the Chebyshev measure 1√ 1−x2 .

1

These deficiencies mean that the algorithm doesn’t work when all the noise comes from outliers (σ = 0); the low outlier rate required also leads to a superconstant factor loss when reducing from other noise models.

In this work, we give a simple algorithm that avoids all these deficiencies: it works for all ρ < 1/2; it has no restrictions on σ; and it gets a constant approximation factor C = 2 + �. Additionally, it only uses O(d2) samples from the uniform measure, without any log factors, and O(d log d) samples from the Chebyshev measure. We also give lower bounds for the approximation factor C, indicating that one cannot achieve a 1 + � approximation by showing that C > 1.09. Our lower bounds are not the best possible, and it is possible that the true lower bounds are much better.

1.1 Algorithmic results

The problem formulation has randomly placed outliers, which is needed to ensure that we can estimate the polynomial locally around any point x. We use the following definition to encapsulate this requirement, after which the noise can be fully adversarial:

Definition 1.1. The size m Chebyshev partition of [−1, 1] is the set of intervals Ij = [cos πjm , cos π(j−1) m ]

for j ∈ [m]. We say that a set S of samples (xi, yi) is “α-good” for the partition if, in every interval Ij,

strictly less than an α fraction of the points xi ∈ Ij are outliers.

The size m Chebyshev partition is the set of intervals between consecutive extrema of the Chebyshev polynomial of degree m. Standard approximation theory recommends sampling func- tions at the roots of the Chebyshev polynomial, so intuitively a good set of samples will allow fairly accurate estimation near each of these “ideal” sampling points. All our algorithms will work for any set of α- good samples, for α < 12 and sufficiently large m. The sample complexities are then what is required for S to be good with high probability.

We first describe two simple algorithms that do not quite achieve the goal. We then describe how to black-box combine their results to get the full result.

L1 regression. Our first result is that `1 regression almost solves the problem: it satisfies all the requirements, except that the resulting error ‖p̂− p‖ is bounded in `1 not `∞:

Lemma 1.2. Suppose the set S of samples is α-good for the size m = O(d) Chebyshev partition, for constant α < 12 . Then the result p̂ of `1 regression

arg min degree-d

polynomial p̂

n∑ i=1

|Ij |mean xi∈Ij

|yi − p̂(xi)|

satisfies ‖p̂− p‖1 ≤ Oα(1) · σ.

This has a nice parallel to the d = 1 case, where `1 regression is the canonical robust estimator. As shown by the Chebyshev polynomial in Figure 1, `1 regression might not give a good solution

to the `∞ problem. However, converting to an `∞ bound loses at most an O(d 2) factor. This

means this lemma is already useful for an `∞ bound: one of the requirements of [GZ16] was that ‖p‖∞ ≤ σdO(1). We can avoid this by first computing the `1 regression estimate p̂(`1) and then applying [GZ16] to the residual p− p̂(`1), which always satisfies the condition.

2

−1.0 −0.5 0.0 0.5 1.0 x

−5 0

5

10

15

20

25

30

y

`1 regression may yield high `∞ error

Observation Scaled Chebyshev T8 Zero polynomial

Figure 1: The blue observations are within σ = 1 of the red zero polynomial at each point, but are closer in `1 to the green Chebyshev polynomial. This demonstrates that `1 regression does not solve the `∞ problem even for ρ = 0.

Median-based recovery. Our second result is for a median-based approach to the problem: take the median ỹj of the y values within each interval Ij . Because the median is robust, this will lie within the range of inlier y values for that interval. We don’t know which x ∈ Ij it corresponds to, but this turns out not to matter too much: assigning each ỹj to an arbitrary x̃j ∈ Ij and applying non-robust `∞ regression gives a useful result.

Lemma 1.3. Let �, α < 12 , and suppose the set S of samples is α-good for the size m = O(d/�) Chebyshev partition. Let x̃j ∈ Ij be chosen arbitrarily, and let ỹj = medianxi∈Ij yi. Then the degree d polynomial p̂ minimizing

max j∈[m]

|p̂(x̃j)− ỹj |

satisfies ‖p̂− p‖∞ ≤ (2 + �)σ + �‖p‖∞.

Without the additive �‖p‖∞ term, which comes from not knowing the x ∈ Ij for ỹj , this would be exactly what we want. If σ � ‖p‖∞, however, this is not so good—but it still makes progress.

Iterating the algorithms. While neither of the above results is sufficient on its own, simply applying them iteratively on the residual left by previous rounds gives our full theorem. The result of `1 regression is a p̂

(0) with ‖p − p̂(0)‖∞ = O(d2σ). Applying the median recovery algorithm to the residual points (xi, yi − p̂(0)(xi)), we will get p̂′ so that p̂(1) = p̂(0) + p̂′ has

‖p̂(1) − p‖∞ ≤ (2 + �)σ + � ·O(d2σ)

If we continue applying the median recovery algorithm to the remaining residual, the `∞ norm of the residual will continue to decrease exponentially. After r = O(log1/� d) rounds we will reach

‖p̂(r) − p‖∞ ≤ (2 + 4�)σ

as desired1. Since each method can be computed efficiently with a linear program, this gives our main theorem:

1We remark that one could skip the initial `1 minimization step, but the number of rounds would then depend on log( ‖p‖∞

σ ), which could be unbounded. Note that this only affects running time and not the sample complexity.

3

Theorem 1.4. Let �, α < 12 , and suppose the set S of samples is α-good for the size m = O(d/�) Chebyshev partition. The result p̂ of Algorithm 2 is a degree d polynomial satisfying ‖p̂ − p‖∞ ≤ (2 + �)σ. Its running time is that of solving O(log1/� d) linear programs, each with O(d) variables and O(|S|) constraints.

We now apply this to the robust polynomial regression problem, where we receive points (xi, yi) such that each xi is drawn from some distribution, and with probability ρ it is a

*View more*