13
$\begingroup$

I am studying the BSS model of computation recently (cf. for instance Complexity and Real Computation; Blum, Cucker, Shub, Smale.)

For reals $R$, it is shown that, given a system of polynomials $f_1,\cdots, f_m\in R[x_1, \cdots, x_n]$, the existence of zeros is $NP_R$-complete. However, I am wondering, if those $f$'s are polynomials only have integer coefficients, i.e., $f_1,\cdots, f_m\in Z[x_1, \cdots, x_n]$, is still the problem $NP_R$-hard? (it is obviously in $NP_R$).

I suspect yes, but is there a simple proof?

$\endgroup$
3
$\begingroup$

I think the answer is no, assuming $\mathsf{P}_{\mathbb{R}} \neq \mathsf{NP}_{\mathbb{R}}$ (I believe I give a proof below, but there are enough potentially nitpicky definitional issues here that I'm being cautious about my claims).

Proof that the answer is no assuming $\mathsf{P}_\mathbb{R} \neq \mathsf{NP}_{\mathbb{R}}$: In fact, I believe the following stronger statement holds:

Lemma: For any BSS decision problem $L$ over $\mathbb{R}$, if $L$ poly-time-BSS$_{\mathbb{R}}$ reduces to a problem on integer inputs, then $L \in \mathsf{P}_\mathbb{R}$.

Proof of lemma: Suppose there were a polynomial-time BSS$_{\mathbb{R}}$ reduction from $L$ to a problem on integer inputs, given by a machine $M$. For inputs consisting of $n$ real parameters, unroll the computation of $M$ into an algebraic computation tree. There are only finitely many leaves, and the result at each leaf is a single rational function in the input parameters. In order for a rational function of real inputs to always output an integer value, it must be a constant function, and therefore not depend on the input. However, which constant function is used at each leaf can, of course, depend on the branches. However, since $M$ is a uniform machine, there can be only $O(1)$ output nodes, and thus only $O(1)$ output values. Thus $M$ can be trivially modified to in fact decide $L$ in polynomial time. QED

Now, take $L$ to be real feasibility of real polynomials. If $\mathsf{P}_{\mathbb{R}} \neq \mathsf{NP}_{\mathbb{R}}$, then $L \notin \mathsf{P}_{\mathbb{R}}$, and by the Lemma there is no reduction from $L$ to any problem on integer inputs (in particular, to real feasibility of integer polynomials).

Promise problem issue?: Another potential issue with your question is that real feasibility of integer polynomials may not be in $\mathsf{NP}_{\mathbb{R}}$, but only in its promise version. The issue here is that to verify that an input (such as the coefficient of a polynomial $f_i$) is an integer takes time that depends on the magnitude of $x$, whereas the set of instances (all instances, not just yes-instances) for an $\mathsf{NP}_{\mathbb{R}}$ decision problem should be decidable in $\mathsf{P}_\mathbb{R}$, the latter meaning that it takes polynomial time in the number of parameters, and not their magnitudes. This is, I believe, closely related to the fact that the integers are not first-order definable within the reals. (Essentially the best a BSS$_{\mathbb{R}}$-machine can do to test if an input $x$ is an integer is to compute the integer part of $x$ by computing powers of $2$ and doing "binary search." Once it's computed the integer part of $x$, it just checks whether that is equal to $x$.) So I think the probleam of real feasibility of integer equations is in $\mathsf{PromiseNP}_{\mathbb{R}}$ but probably not in $\mathsf{NP}_{\mathbb{R}}$ (or at least it seems nontrivial to prove that it is in $\mathsf{NP}_{\mathbb{R}}$).

$\endgroup$

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.