Certified surface approximations using the interval Krawczyk test

We propose an algorithm to construct a certified approximation of a surface by generalizing the Krawczyk test. The Krawczyk test is based on interval arithmetic, and confirms the existence and uniqueness of a solution to a square system of analytic equations in a region. By generalizing this test, we extend the reach of this technique to non-square systems and higher-dimensional varieties. We provide a prototype implementation and illustrate its use on several examples.

💡 Research Summary

The paper introduces a certified algorithm for approximating algebraic surfaces by extending the classical Krawczyk test, which is traditionally limited to square (equal‑number‑of‑equations‑and‑variables) analytic systems. The authors observe that interval arithmetic, the foundation of the Krawczyk test, enables reasoning about entire regions rather than single points. By formulating a “interval Krawczyk operator” that works on a product of intervals for the first d variables (the projection domain) and the remaining n‑d variables (the fiber), they prove a generalized existence‑and‑uniqueness theorem (Theorem 2.1). The theorem states that if the interval operator K(F, ẑ, r, A) is contained in a scaled version of the interval for the fiber variables, then for every point in the domain interval there exists a unique fiber point solving the system, and the solution lies within a provably small radius.

The theoretical development is followed by a concrete algorithmic framework. The authors first treat the case where the solution set X is a graph over a d‑dimensional box U (i.e., each projection x∈U lifts to a single point y). Algorithm 2 (GraphApproximation) proceeds by a breadth‑first subdivision of U: for each box B, the midpoint ẋ is used to compute an approximate fiber solution ẏ via a multivariate root‑isolation routine, and an approximate inverse Jacobian A is obtained. The interval Krawczyk test is then applied with radii r₁ (for the domain) and r₂ (for the fiber). If the test succeeds, the box B together with a small interval around ẏ is recorded as a certified piece of the surface; otherwise B is split into 2ᵈ sub‑boxes and the process repeats. The authors prove (Theorem 3.1) that this subdivision must terminate because, as the boxes shrink, the interval quantities in the Krawczyk test converge to zero, guaranteeing eventual success. Moreover, the union of all certified boxes deformation‑retracts onto the true surface, providing a topologically faithful approximation.

The paper also discusses extensions. For multiple sheets (k‑fold graphs) the algorithm is modified to isolate all k fiber solutions at each midpoint, and additional subdivision is performed until boxes belonging to different sheets no longer intersect (except possibly on boundaries). Corollary 3.2 shows that the same deformation‑retraction guarantee holds. The authors further explain how to tighten the Hausdorff distance by shrinking the output fiber intervals or by reducing the contraction factor ρ.

Section 4 critically examines practical limitations. Subdivision into 2ᵈ boxes becomes costly for higher dimensions (the paper focuses on surfaces, d = 2, where 2ᵈ = 4). Large slopes in the surface can produce highly elongated boxes, and the method cannot directly handle closed varieties because regularity fails at extreme points. The multivariate root‑isolation step (Step 8) is identified as a performance bottleneck; it currently relies on generic interval‑based isolation, which may be inefficient for high‑dimensional fibers.

To mitigate these issues, Section 4.1 proposes coordinate transformations (e.g., unitary rotations) that align the surface with the coordinate axes, thereby reducing slope magnitudes and improving box aspect ratios. Such transformations preserve the validity of Theorem 2.1 and Theorem 3.1, though they add complexity to the proof. The authors also hint at a patch‑based strategy, where the surface is covered by several locally regular charts, each processed independently and then stitched together.

Section 5 consolidates the full certified surface approximation pipeline: (1) scale the domain to a unit cube, (2) initialize a queue with the domain box, (3) iteratively pop boxes, compute midpoint approximations, evaluate the interval Krawczyk test, (4) record successful boxes, (5) subdivide failing boxes, and repeat until the queue empties. The output is a finite collection of axis‑aligned boxes that together contain the surface and are certified to contain exactly one solution per domain point.

Implementation details are provided in Section 6. A prototype in C++ using the INTLAB interval library demonstrates the method on several benchmark surfaces, including tori, paraboloids, and more intricate algebraic varieties. Experiments show that with modest choices of ρ and subdivision depth, the algorithm yields a few hundred boxes that tightly enclose the surface, with certification times on the order of milliseconds per box. However, for higher‑dimensional problems or surfaces with steep gradients, the number of boxes—and thus runtime—grows substantially.

In summary, the paper makes a substantial contribution by generalizing the Krawczyk test to non‑square systems and applying it to certified surface approximation. The combination of rigorous interval analysis, a provably terminating subdivision scheme, and practical implementation results bridges the gap between theoretical existence‑and‑uniqueness guarantees and usable numerical tools for algebraic geometry. Future work is suggested in extending the approach to higher dimensions, improving the efficiency of multivariate root isolation, and handling complex (as opposed to real) varieties.