Here "nice" would be some property that is satisfied for common rings like $\mathbb Z$ or $\mathbb Z/n\mathbb Z$. For example being a subring of the direct product of a family of fields might work for square-free $n$. (So "nice" commutative ring might be replaced by subring of a commutative (von Neumann) regular ring.) The condition "$\forall s, t\in S:(\dots)$" tries to ensure that one can apply the normal Schwartz-Zippel lemma independently to each field from the underlying family of fields. This condition is automatically satisfied for a subring of a field (like $\mathbb Z$), and can be checked easily by verifying $\operatorname{gcd}(s-t,n)=1$ for $\mathbb Z/n\mathbb Z$, hence it is no real limitation.
Question 1 Does this work, or am I overlooking something? (Like "$\exists u\in R$ mustThis question also refers to the implicit assumption that the reformulation is more useful than the original lemma, because it doesn't exclude "common rings like $\mathbb Z$ or $\mathbb Z/n\mathbb Z$". It applies to $(1+x^n)=1+x^n(\operatorname{mod}n)$, because "This condition ... can be replacedchecked easily by verifying $\exists u\in \prod_iF_i$"$\operatorname{gcd}(s-t,n)=1$ for $\mathbb Z/n\mathbb Z$, which would nearly annihilate the usefulness of the reformulationhence it is no real limitation.")
Edit Note that Question 1 has been answered in the comments, i.e. I am indeed overlooking something. The reformulation is true for arbitrary commutative rings, as shown by Emil Jeřábek's nice proof. However, it is not at all clear whether it is really more useful than the original lemma:
- $q\prod^{p−1}_{i=0}(x−i)$ in $\mathbb Z/pq \mathbb Z$ is a non-zero polynomial of total degree $p$, but it evaluates to $0$ for any $x\in\mathbb Z/pq \mathbb Z$. So even the reformulation can't be used directly to probabilistically check $(1+x^n)=1+x^n(\operatorname{mod}n)$. What I overlooked here is that the ability to check whether $s-t$ is a zero-divisor for the individual witnesses from $S$ used to evaluate $P$ is not enough to ensure $|S|>d$ with sufficiently high probability to efficiently test the identity $P=0$.
- The reformulation also can't be used directly for identity testing over $\mathbb Z$. The problem here is that the intermediate results during evaluation of $P$ can become too big for efficient computation. The original lemma can be applied to this problem by first randomly drawing a sufficiently big prime number $p$, and then doing the computation in the field $\mathbb Z/p \mathbb Z$. The probability to draw a divisor of $P$ is quite small, if $p$ was drawn from a sufficiently big collection. One could hope now that the reformulation would allow to randomly draw a sufficiently big $p^\alpha$ instead, and then do the computation in the ring $\mathbb Z/p^\alpha \mathbb Z$. But the fact that $|S|\leq p$ makes this inapplicable for $d>=p$. What one can do is apply the original lemma in $\mathbb Q$ (or the reformulation in $\mathbb Z$), and then draw sufficiently big random numbers $n_i$ for each individual evaluation of $P$. This seems to require more random bits, and doesn't need the reformulation either.
