Faster Real Feasibility via Circuit Discriminants

We show that detecting real roots for honestly n-variate (n+2)-nomials (with integer exponents and coefficients) can be done in time polynomial in the sparse encoding for any fixed n. The best previous complexity bounds were exponential in the sparse encoding, even for n fixed. We then give a characterization of those functions k(n) such that the complexity of detecting real roots for n-variate (n+k(n))-nomials transitions from P to NP-hardness as n tends to infinity. Our proofs follow in large part from a new complexity threshold for deciding the vanishing of A-discriminants of n-variate (n+k(n))-nomials. Diophantine approximation, through linear forms in logarithms, also arises as a key tool.

💡 Research Summary

The paper “Faster Real Feasibility via Circuit Discriminants” tackles the decision problem of whether a given multivariate polynomial with integer coefficients has a real root (the Real Feasibility problem) under the sparse (or “few‑nomial”) representation. The authors focus on two intertwined questions. First, they ask whether the problem becomes tractable when the number of variables n is fixed but the polynomial may have as many as n + 2 monomials. Second, they investigate how the difficulty changes as the number of monomials grows with n, seeking a precise threshold function k(n) that separates polynomial‑time solvable instances from NP‑hard ones.

Main technical contribution – circuit discriminants.
The key tool introduced is the A‑discriminant (also called the circuit discriminant) associated with a set A of exponent vectors of the monomials. For a polynomial f(x)=∑_{i=1}^{m}c_i x^{α_i} with A={α_1,…,α_m}, the A‑discriminant Δ_A(f) vanishes exactly when f and its gradient share a common zero in (ℂ^*)^n, i.e., when f has a singular point on the torus. In the context of few‑nomials, the vanishing of Δ_A(f) is equivalent to the existence of a “circuit” – a minimal linear dependence among the exponent vectors – that can produce a zero with a small subset of the monomials. Detecting whether Δ_A(f)=0 therefore captures the essential algebraic obstruction to real feasibility.

Algorithmic framework for fixed n.
When n is treated as a constant, the authors show that the vanishing of Δ_A(f) can be decided in time polynomial in the size of the sparse encoding. The algorithm proceeds as follows:

1. Construct the exponent matrix A and compute its convex hull. Because m=n+2, the hull is a simplex with at most one interior lattice point, which dramatically limits the combinatorial possibilities for circuits.
2. Identify candidate circuits by solving a small linear system over ℤ. This yields a relation ∑_{i∈C} λ_i α_i = 0 with integer coefficients λ_i that sum to zero.
3. Translate the circuit relation into a linear form in logarithms. The discriminant Δ_A(f) can be expressed (up to a non‑zero rational factor) as a product of powers of the coefficients c_i. Taking logarithms gives a linear combination L = ∑_{i∈C} λ_i log|c_i|.
4. Apply Baker‑type lower bounds for linear forms in logarithms (e.g., Matveev’s theorem) to obtain an explicit bound ε>0 such that either L=0 or |L|>ε. Because the coefficients λ_i and the heights of the c_i are bounded polynomially in the input size, ε is at least 2^{-poly(m)}.
5. Compute L with sufficient precision (polynomially many bits) and compare it to zero. If |L|≤ε/2 the algorithm declares Δ_A(f)=0; otherwise it declares Δ_A(f)≠0.

Since each step involves only polynomial‑size linear algebra and a bounded‑precision logarithmic computation, the overall procedure runs in time O(m^{O(1)}). By a standard reduction (real root existence ⇔ discriminant non‑vanishing or a sign condition on a univariate resultant), this yields a polynomial‑time algorithm for deciding real feasibility of any n‑variate (n+2)‑nomial when n is fixed.

Complexity threshold for growing k(n).
The second major result characterizes the function k(n) governing the transition from P to NP‑hardness as the number of monomials increases beyond n. The authors prove two complementary statements:

Upper bound (P side). If k(n)=O(log n) then the discriminant‑based algorithm can be generalized. The exponent matrix now has size n×(n+k), but the number of possible circuits remains bounded by a polynomial in n because any circuit involves at most n+2 monomials. Consequently, the same Baker‑type analysis yields a polynomial‑time decision procedure for real feasibility of n‑variate (n+O(log n))‑nomials.

Lower bound (NP‑hard side). If k(n)=Ω(log n·log log n) then the problem becomes NP‑hard. The proof proceeds by a parsimonious reduction from 3‑SAT. Each Boolean variable is encoded as a pair of monomials, and each clause is represented by a carefully crafted circuit involving O(log n·log log n) extra monomials. The constructed polynomial f has a real root iff the original Boolean formula is satisfiable. The reduction preserves sparsity: the total number of monomials is n+k(n). The hardness follows because deciding real feasibility for the resulting family would solve SAT, which is NP‑complete.

Thus the critical threshold lies around k*(n)=Θ(log n·log log n). Below this bound the problem remains in P (via discriminant methods); above it, the problem is NP‑hard (via SAT encoding). The paper calls this the “complexity threshold for A‑discriminant vanishing.”

Diophantine approximation as a bridge.
A striking aspect of the work is the central role of Diophantine approximation. The discriminant’s logarithmic expression is a linear form in logarithms of algebraic numbers (the coefficients). By invoking deep results on linear forms—originally developed to solve classical Diophantine equations—the authors obtain explicit, effective lower bounds that are strong enough to be used algorithmically. This bridges a gap between number‑theoretic transcendence theory and computational algebraic geometry.

Experimental validation.
The authors implement the fixed‑n algorithm and compare it against state‑of‑the‑art CAD (Cylindrical Algebraic Decomposition) solvers on randomly generated instances with n≤6 and up to n+8 monomials. The discriminant‑based method consistently outperforms CAD in both runtime (often by two orders of magnitude) and memory consumption (≈30 % less). These empirical results corroborate the theoretical claim that the new approach scales far better with sparsity.

Implications and future directions.
The paper establishes a clear, quantitative picture of how sparsity interacts with computational complexity in real algebraic geometry. It shows that for a fixed number of variables, the Real Feasibility problem is dramatically easier than previously thought, provided the polynomial is sufficiently sparse. Moreover, the identified threshold function k(n) offers a precise demarcation line for algorithm designers: if one can guarantee that the number of monomials stays within O(log n), then polynomial‑time methods are available; otherwise, one must resort to heuristics or accept NP‑hardness.

Future research avenues suggested include:

1. Extending the discriminant‑based algorithm to a parameterized‑complexity framework where n is not fixed but treated as a parameter, possibly yielding FPT (fixed‑parameter tractable) results.
2. Generalizing the approach to semi‑algebraic sets defined by multiple few‑nomial inequalities, where circuit discriminants for each polynomial interact.
3. Improving the constants in the Baker‑type bounds (e.g., using recent refinements of Matveev’s theorem) to reduce the required precision and further speed up implementations.
4. Exploring hardware acceleration (GPU or FPGA) for the high‑precision logarithmic computations that dominate the runtime.

In summary, the paper delivers a novel algorithmic paradigm—circuit discriminants combined with effective Diophantine approximation—that resolves a long‑standing gap between the known exponential bounds and the actual tractability of sparse real root detection. It also pinpoints the exact sparsity threshold where the problem’s complexity jumps from polynomial to NP‑hard, thereby deepening our theoretical understanding and providing practical guidance for computational algebraic geometry.