A mathematical study by Thomas A. Weber demonstrates that the nondifferentiable minimum function can be systematically approximated from below using an increasing sequence of smooth functions with exponential convergence. Published in the Journal of Inequalities and Applications, the theoretical research resolves a longstanding analytical difficulty in continuous mathematics and optimization, where minimum values across multiple variables create sharp, nondifferentiable corners. By constructing an iterative algorithm centered on the harmonic mean, Weber established that a strictly conservative lower bound can be smoothly improved at every stage. Unlike standard approximation techniques that approach the minimum from above, this recursive formulation generates an explicit sequence of infinitely differentiable functions that never exceed the true minimum. The resulting sequence establishes guaranteed pointwise convergence across positive real vectors while preserving essential regularity across the entire analytical domain.
In mathematical modeling, functions governed by minimum operators appear naturally across systems constrained by bottlenecks, where overall performance is dictated by its scarcest component. Within economic theory, Leontief production functions express firm output as the minimum of multiple input factors, an operational structure observed in selective assembly processes where mating parts must match precisely to yield finished goods. Complementary consumption dynamics, such as Leontief utility functions requiring fixed proportions of goods like left and right shoes, similarly depend on minimum evaluations. Beyond economics, minimum functions dictate lifetimes in series reliability models where the first failure terminates operation, govern throughput across serial communication channels, define constraint slackness in optimization, and evaluate worst-case outcomes in robust decision-making and max-min welfare models. In these environments, smooth lower approximations provide conservative estimates that avoid overstating capacity, output, or safety margins while enabling standard differential methods.
The research contrasts the proposed construction with classical mean-based approximations that have historically dominated mathematical literature on inequalities. Traditional means, including arithmetic, geometric, and generalized power means, satisfy what mathematicians define as the internality property, meaning their evaluated values necessarily lie strictly between the smallest and largest components of an input vector. Families of power means and counter-harmonic Lehmer means can approximate the minimum function in the limit as their power parameters approach negative infinity, but because of internality, every finite evaluation yields a value greater than or equal to the true minimum, thereby providing an approximation strictly from above. Producing a smooth approximation strictly from below requires departing from this conventional internality framework to construct an external approximation. Rather than using an existing mean directly as a proxy, the investigation devises an additive recursive mechanism that pushes smooth lower bounds monotonically upward toward the objective.
At the foundation of this recursive sequence is the harmonic increment, derived from the classical harmonic mean of the Pythagorean tradition. For an n-dimensional vector of positive real numbers, the harmonic increment is defined mathematically as the n-th part of the harmonic mean, equivalent to the reciprocal of the sum of the reciprocals of all vector components. Algebraically, this expression equals the product of all components divided by the sum of all partial products omitting one component at a time. The algorithm initiates from a baseline lower bound of zero for positive inputs, which guarantees a strict initial underestimate. At each subsequent iteration, the next approximating function is generated by adding the harmonic increment of the remaining component slack. Because each coordinate remains strictly greater than the current lower bound, the resulting argument remains positive, ensuring that the computed increment is strictly positive and infinitely differentiable at every step.
Through a ten-step mathematical proof in Proposition 1, the research rigorously proves that the constructed sequence is strictly increasing, infinitely differentiable, and convergent to the minimum function. Working initially on the positive orthant of Euclidean space, the proof verifies through mathematical induction on dimension and iteration index that each successive function remains strictly bounded between the previous function and the exact minimum value. Because the update at each iteration represents a composition of rational expressions whose denominators never vanish over the domain, every function in the sequence inherits infinite smoothness. Monotone convergence follows directly from fundamental real analysis principles, as any sequence that increases monotonically while bounded above must possess a well-defined limit. The proof demonstrates by contradiction that this limit coincides identically with the true minimum, since any persistent positive slack across all coordinates would produce a nonzero increment, violating convergence.
The analysis establishes exact mathematical bounds on both relative and absolute approximation errors across the entire positive domain. Weber defines the relative error as the difference between the minimum function and the smooth lower bound divided by the minimum function itself. In Proposition 2, the author proves that the supremum of this relative error across all positive vectors is given precisely by one minus the reciprocal of the dimension, raised to the power of the iteration index. This worst-case relative error is attained universally along the diagonal of the domain, where all coordinates are identical and the minimum function exhibits its characteristic nondifferentiability. Proposition 3 extends these findings to quantify absolute approximation error over any bounded hypercube where coordinates do not exceed a bound M, demonstrating that the absolute error is bounded by M multiplied by the maximum relative error, confirming exponential convergence.
A key analytical property proven in the study is that every function in the approximating sequence inherits the degree-one homogeneity of the original minimum function. Scaling all input coordinates by a positive scalar factor scales the output of every lower bound function by precisely that same factor. This property enables an explicit closed-form algebraic representation for two-dimensional inputs, which connects directly to the classical crossed-ladders problem. In that celebrated puzzle, two ladders crossing within an alley between vertical walls determine a crossing height equal to the harmonic increment of the wall contact heights. For distinct coordinates x and y, the k-th lower bound simplifies to the difference between x raised to the power two to the k times y and x times y raised to the power two to the k, divided by their difference. Applying this recursion to contact heights of three and six units yields successive bounds of two, two point eight, and two hundred fifty-four over eighty-five, reaching an error below zero point four percent after three iterations.
While extending explicit closed-form algebraic expressions directly to higher dimensions becomes algebraically cumbersome, the study introduces an alternative recursive variant that combines lower-dimensional smooth approximations with the two-dimensional closed form. By treating an n-dimensional smooth lower bound and an additional coordinate as two scalar inputs to the two-dimensional formula, the researcher constructs an explicit closed-form sequence for n plus one dimensions. The paper provides a formal error decomposition showing that the total absolute error of this hybrid formulation is bounded by the sum of the error from the lower-dimensional approximation and the contraction error from the two-dimensional step. When applying this variant across a bounded hypercube, the resulting uniform absolute error bound is explicitly given by the hypercube bound M multiplied by the sum of one minus one over n to the power k and two to the negative power k, ensuring analytical tractability in higher dimensions.
In addition to approximating the minimum function from below, the paper resolves the corresponding challenge of constructing smooth upper bounds that converge monotonically toward the minimum function. For any positive coordinate vector, Weber defines a complementary sequence of smooth upper bounds by adding the geometric mean of coordinate slacks to the previously computed smooth lower bounds. Proposition 4 demonstrates that because the smooth lower bounds are infinitely differentiable and strictly underestimate the minimum, their coordinate slacks remain strictly positive, ensuring that the resulting upper bound functions are smooth and evaluate strictly above or equal to the minimum. As the iteration index advances, the difference between the upper and lower bounds decreases strictly and monotonically toward zero. The paper computes that setting the iteration count to n times the threshold required for the lower bound guarantees that the upper bound satisfies any prescribed error tolerance over a bounded domain, yielding a coordinated pair of smooth envelopes.
The investigation concludes by generalizing these results beyond the positive orthant to the entire n-dimensional Euclidean space, establishing what Weber terms a universal harmonic improvement principle. The recursive formulation is not inherently restricted to starting at zero; rather, any smooth strict lower bound function can serve as an initial seed. When applied as an abstract operator, the harmonic increment strictly improves any non-tight lower bound while contracting the remaining approximation error by at least a factor of one minus one over n at every single step. This contracting property demonstrates that the harmonic increment can be combined synergistically with existing smooth approximations, such as the LogSumExp soft minimum family, by taking their analytical outputs as seeds and systematically contracting their remaining error while preserving infinite smoothness. As noted in the study, no empirical datasets were generated or analyzed during this research, but the resulting conservative approximations provide foundations for numerical optimization, reliability engineering, and economic bottleneck models.
Subject of Research: Smooth mathematical approximation of the nondifferentiable minimum function using recursive harmonic increments
Article Title: Smooth approximation of the minimum function from below (and above)
Article References: Weber, T. A. (2026). Smooth approximation of the minimum function from below (and above). Journal of Inequalities and Applications, 2026(1), Article 110. https://doi.org/10.1186/s13660-026-03519-3
Image Credits: AI Generated
DOI: 10.1186/s13660-026-03519-3
Keywords: minimum function, smooth approximation, harmonic increment, Pythagorean means, exponential convergence, monotone sequence, Leontief production function, bottleneck problems, crossed-ladders problem, relative error bound, absolute error bound, degree-one homogeneity, geometric mean, LogSumExp, series reliability, selective assembly, constrained optimization
Cite Scienmag News
Reid Dalton. (November 21, 2026). Harmonic Increment Sequence Enables Smooth Exponential Approximation of the Minimum Function. Scienmag. https://scienmag.com/harmonic-increment-sequence-enables-smooth-exponential-approximation-of-the-minimum-function/
Reid Dalton. "Harmonic Increment Sequence Enables Smooth Exponential Approximation of the Minimum Function." Scienmag, 21 November 2026, https://scienmag.com/harmonic-increment-sequence-enables-smooth-exponential-approximation-of-the-minimum-function/. Accessed 21 November 2026.
Reid Dalton. "Harmonic Increment Sequence Enables Smooth Exponential Approximation of the Minimum Function." Scienmag. November 21, 2026. https://scienmag.com/harmonic-increment-sequence-enables-smooth-exponential-approximation-of-the-minimum-function/








