How can I distinguish a genuine solution of polynomial equations from a near miss?

Suppose I have a large system of polynomial equations in a large number of real-valued variables. \begin{align} f_1(x_1, x_2, &\dots, x_m) = 0 \\ f_2(x_1, x_2, &\dots, x_m) = 0 \\ &\vdots \\ f_n(x_1, x_2, &\dots, x_m) = 0 \\ \end{align} (In my particular case, I have about $n \approx 1000$ equations of degree $10$ in about $m \approx 200$ variables.) By numerical means, I've found an approximate solution vector $(\tilde{x}_1, \tilde{x}_2, \dots, \tilde{x}_m)$ at which the value of each $f_j$ is very small: $$\lvert f_j(\tilde{x}_1, \tilde{x}_2, \dots, \tilde{x}_m) \rvert < 10^{-16} \quad \forall j = 1, \dots, n.$$ This leads me to believe that a genuine solution of my system exists somewhere in a small neighborhood of $(\tilde{x}_1, \tilde{x}_2, \dots, \tilde{x}_m)$, and that the small residuals I see are due to round-off error in finite (IEEE double) precision arithmetic. However, it could conceivably be the case that the zero loci of my polynomials $f_j$ come very close to each other (within $10^{-16}$) but do not mutually intersect. How can I rigorously tell which is the case?

I could, of course, further refine my solution using quadruple- or extended-precision arithmetic to push the residuals even closer to zero, but this would only provide supporting empirical evidence. Even an approximate solution accurate to one million decimal places may not guarantee the existence of a true mathematical solution; the next digit could always turn out to be irreconcilably wrong.

In principle, there are methods in computational algebraic geometry (Groebner bases, cylindrical decomposition, etc.) that can algorithmically prove the existence of a true mathematical solution to a polynomial system, but my system is completely out of reach of all such algorithms I know. Buchberger's algorithm, for example, has doubly-exponential time complexity in the number of input variables.

If it helps, all of my polynomials have integer coefficients and can be evaluated exactly on integer and rational inputs. However, my approximate solution $(\tilde{x}_1, \tilde{x}_2, \dots, \tilde{x}_m)$ is probably irrational. Note that interval/ball arithmetic won't help, because even if I can show that each $f_j$ exactly assumes the value $0$ in a small neighborhood of $(\tilde{x}_1, \tilde{x}_2, \dots, \tilde{x}_m)$, it could be the case that a different point zeroes out each $f_j$, and no single point simultaneously zeroes out all of them.

• You probably cannot. If you're working with floating point numbers you can't even exactly evaluate (or even represent) your polynomials. – flawr Mar 26 '17 at 21:05
• @flawr That's true, but I can certainly use e.g. rational or interval arithmetic to obtain exact/certified results. Can these techniques be applied to my problem? – David Zhang Mar 26 '17 at 21:08
• You would probably need to know something about the regularity of the polynomials, but if they are just given as a "black box" I think it is impossible. – flawr Mar 26 '17 at 21:14
• @flawr I know my polynomials $f_j$ exactly (they have integer coefficients), so I can supply regularity information. What do you have in mind? – David Zhang Mar 26 '17 at 21:20
• Ok that changes a bit: I don't know how usefull it is in practice but from a theoretical standpoint you might want to look in to the resultant of polynomials. – flawr Mar 26 '17 at 21:28