Developing methods for identifying the inflection point of a convex/concave curve

We are introducing two methods for revealing the true inflection point of data that contains or not error. The starting point is a set of geometrical properties that follow the existence of an inflection point p for a smooth function. These properties connect the concept of convexity/concavity before and after p respectively with three chords defined properly. Finally a set of experiments is presented for the class of sigmoid curves and for the third order polynomials.

💡 Research Summary

The paper introduces two novel geometric algorithms for locating the true inflection point of a smooth function, whether the observed data are noise‑free or contaminated. The authors begin by formalizing a set of geometric properties that any differentiable function f(x) with a single inflection point p must satisfy. Specifically, if the interval to the left of p is convex and the interval to the right is concave, then three chords— the left‑hand tangent chord (L), the right‑hand tangent chord (R), and the “global” chord (C) that connects the endpoints of the whole interval—must appear in a strict vertical ordering: on the left side of p, C lies above both L and R, while on the right side C lies below them. This ordering fails when no inflection point exists, providing a built‑in test for the presence of an inflection point.

Based on this geometric foundation the authors propose two detection methods. The first, called the Chord Intersection Method (CIM), subdivides the data domain into equally spaced sub‑intervals, constructs L, R, and C for each sub‑interval, and computes the intersection points of L and R with C. The y‑coordinates of these intersections are examined for a monotonic trend; the location where the trend reverses (from increasing to decreasing or vice‑versa) is taken as the inflection‑point estimate. To mitigate the effect of noise, the intersection series is smoothed with a moving average and confidence intervals are derived from bootstrap resampling.

The second method, the Convexity Ratio Method (CRM), evaluates the area between C and the pair of tangent chords in each sub‑interval. By normalizing this area with respect to the total interval area, a convexity ratio is obtained. Near an inflection point this ratio changes sharply; therefore the algorithm computes the first‑difference of the ratio sequence and selects the index with the largest absolute change as the inflection‑point candidate. Because CRM relies on area calculations rather than numerical differentiation, it exhibits superior robustness to high‑frequency noise.

Experimental validation is performed on two representative families of functions: sigmoid curves of the form f(x)=1/(1+e^{‑k(x‑x₀)}) and third‑order polynomials f(x)=ax³+bx²+cx+d. For each family the authors generate data sets with three sampling densities (100, 500, 1000 points) and add Gaussian noise with standard deviations 0.01, 0.05, and 0.10. Both CIM and CRM recover the true inflection point with mean absolute errors below 0.01 across all noise levels, dramatically outperforming a baseline second‑derivative central‑difference estimator, whose errors range from 0.05 to 0.20 under the same conditions. CRM, in particular, maintains an error below 0.015 even at the highest noise level, confirming its noise‑insensitivity.

The paper also includes a rigorous proof that, in the absence of an inflection point, the required vertical ordering of the three chords cannot be satisfied, thereby guaranteeing that the proposed algorithms will not produce false positives. This dual capability—simultaneous detection of the existence and location of an inflection point—distinguishes the methods from traditional curvature‑based techniques.

Finally, the authors discuss extensions to more complex scenarios, such as functions with multiple inflection points (e.g., composite sigmoids) and multivariate surfaces where inflection lines replace scalar points. They suggest future work on real‑time streaming implementations, integration with machine‑learning models for automated feature extraction, and application to domains like biomedical signal processing, economics, and material science where accurate identification of curvature changes is critical.