平方和方法的数值精度?


13

我已经从Barak&Steurer调查和Barak讲义中读到一些平方和方法(SOS)。在这两种情况下,它们都会清除地毯下的数值精度问题。

根据我对方法的认识(有限),以下内容应该是正确的:

给定任何在实值变量的多项式等式系统,其中所有参数均为(,和每个约束的度),度为“ “()SOS方法找到了令人满意的变量分配,或证明在时间内不存在任何变量。 ExRnO(1)n|E|2n=O(1)O(1)

我的第一个问题是上述说法是否正确(是否有一个不使用SOS解决这一问题的幼稚论点?)。第二个问题是数值精度适合的位置。如果我想获得一个满足所有精度内的所有约束的赋值,那么运行时间如何取决于?特别是多项式吗?ε1/ε

例如,这样做的动机是在大型系统上应用分治法,直到基本情况是系统。O(1)

编辑:从巴拉克- Steurer,看来“度加总平方算法”的第9页(和段落导致它)都定义了问题的解决,,而实际上定义了第2.2节中的伪分发。现在,我从引理2.2中看到,没有二进制变量,不能保证在有解/反驳。lRR2n

因此,我可以稍微完善一下我的问题。如果您的变量不是二进制变量,则担心的是输出的顺序不是有限的(甚至可能不是单调递增的吗?)。所以问题是:还在增加?如果是这样,您必须走多远才能获得加性精度?φ(l)φ(l)ε

尽管这可能不会改变任何东西,但是我碰巧知道我的系统是令人满意的(没有任何程度的反驳),所以我真的只是在担心。最后,我对理论解决方案感兴趣,而不是数值求解器。l


Answers:


1

这是Boaz Barak 对这个问题的评论

我们确实在地毯下扫描数值精度-Parrilo,Lasserre等的更“传统” SOS文献。涉及这些问题(例如,请参见Monique Laurent的调查及其参考)。已知的是,该层次结构是单调(不难看出,一定程度的伪分布是特别度的一个),并且,将在有限的程度为任何固定的一组方程的收敛(这是(Positivstellensatz)。确切程度可能有所不同。通常,如果多项式的所有系数都是有界的,并且您试图区分存在一个解决方案的情况和在任何分配中一个方程式被抵消的情况,则可以将其离散化为ll1ϵδ-delta的 net 与变量的数量,方程式的程度和,然后(假设网络足够“好”和“像立方体”),所需的程度应大致记录网络的大小。δϵ


作为答案发布,以避免将来社区机器人再次提出这个问题。
卡夫

1

我认为我的答案可能不够,但出于完整性考虑,它仍然存在(尽管请参阅下面的Boaz评论,以获得更好的答案)

当我们将自己限制为布尔变量时,对于所有,当时都可以看到该声明,并且观察到伪分布是实际分布,也就是说,假设您在多项式解上具有伪分布,满足:(xi21)Ei[n]2nμ(x)xE

x{1,1}nμ(x) and x{1,1}nμ(x)p2(x)0 for all polynomials p with degree at most n

But degree n polynomials include the indicator polynomial (for example, x1=1,x2=1,x3=1 has 23(1+x1)(1x2)(1+x3) which is all-zero elsewhere and 1 on that assignment). So μ(x)0 for all x{1,1}n, so we conclude μ is an actual distribution over the solutions of E. Degree pseudo-distributions can be found by using semidefinite programming to find an associated degree pseudo-expectation operator in nO() time, so we can find the actual distribution μ in time nO(n) by using that pseudo-expectation (now an actual expectation) to find all the moments of μ.

So, if |E|=O(1), then you can find a distribution of solutions to E in O(1) time. Of course, brute force search guarantees the same.

However, if the solutions are not necessarily boolean, then degree-2n pseudo-expectations are not sufficient to find a distribution over solutions. As can be seen above, the proof that degree 2n pseudo-distributions are actual distributions depends on the fact that degree n polynomials are sufficient to 'pick out' individual assignments, which is not true more generally. Another way of viewing it is that boolean-variable polynomials are considered mod(xi2), so the degree of every monomial is at most n.

For example, one could consider replacing every binary variable with a 4-ary variable, say by including (xi21)(xi24)E. Then you would have to have a degree 4n pseudo-expectation in order to guarantee recovery of a distribution over solutions.

Now, for theoretical guarantees, it seems like approximating a root of a system of polynomals is also known as Smale's 17th problem, and apparently there is a randomized (Las Vegas) polynomial time algorithm that solves this - see http://arxiv.org/pdf/1211.1528v1.pdf. Note that this seems to be in the Blum-Shub-Smale model, so real operations are the primitive. I am not sure if this gives the guarantee that you need.


I think I may not have made this clear: my variables xi are in R, since otherwise I could just do a trivial O(2n)=O(1) search over the boolean hypercube. I have updated the question to reflect this. SDP/SOS also applies to real-input optimization problems, right?
Jeremy Kun

Oops, my mistake! Yes, it does apply to more general settings, although many times we just assume we are on the hypercube. I've updated my answer, though my answer will be less clear than I'd have hoped.
Joe Bebel

10
We do sweep numerical accuracy under the rug -- the more "traditional" SOS literature of Parrilo, Lasserre etc.. deals with these issues (e.g., see Monique Laurent's surveys and the references therein). It is known that that the hierarchy is monotone (it's not hard to see that a degree psuedo-distribution is in particular a degree 1 one), and that it will converge in finite degree for any fixed set of equations (this is the Positivstellensatz).
Boaz Barak

9
..The exact degree could vary. Generally, if all the coefficients of the polynomials are bounded and you are trying to distinguish between the case that there is a solution and the case that in any assignment one of the equations is off by ϵ, then one could discretize this to a δ-net for δ related to the number of variables, degree of equations, and ϵ, and then (assuming the net is sufficiently "nice" and "cube like") the degree required should be roughly log the size of the net.
Boaz Barak

4
@BoazBarak maybe this could be an answer ?
Suresh Venkat
By using our site, you acknowledge that you have read and understand our Cookie Policy and Privacy Policy.
Licensed under cc by-sa 3.0 with attribution required.