Quantification of Non-Monotonicity

• Quantification of non-monotonicity is a framework that defines graded indices like LOI, LOD, and LOM to measure deviations from strict monotonic behavior.
• It employs rigorous grid-based computations and L1-type metrics to assess monotonicity in functions, measures, and optimization procedures.
• The approach is pivotal in various fields—econometrics, quantum information, and frictional system analysis—by providing actionable, operational metrics.

Quantification of non-monotonicity addresses the measurement and analysis of deviations from monotonic behavior in mathematical objects such as functions, measures, optimization procedures, and physical or algorithmic systems. Unlike binary monotonicity tests, quantification provides a graded assessment of how far a given entity lies from satisfying monotonicity constraints, with rigorous indices and operational metrics. Such analysis is critical in mathematical analysis, probability, optimization, auction theory, quantum information, and engineering, where monotonicity assumptions often underlie theoretical results and practical algorithms.

1. Indices for Lack of Monotonicity in Functions and Measures

Several indices have been developed to measure the extent of non-monotonicity for real functions and signed measures. Let $g_0:[a, b]\to\mathbb{R}$ be absolutely continuous and $g(x) = g_0(x+a)-g_0(a)$ on $[0, y]$, $y = b-a$. Writing the derivative as $g'(x) = g'^+(x) - g'^-(x)$ with $g'^+ = \max\{g', 0\}$, $g'^- = \max\{-g', 0\}$, the following indices are central (Davydov et al., 2017):

• Lack of Increase (LOI):

$LOI_{[a,b]}(g_0) = \int_a^b (g_0'(x))^- \, dx$

• Lack of Decrease (LOD):

$LOD_{[a,b]}(g_0) = \int_a^b (g_0'(x))^+ \, dx$

• Lack of Monotonicity (LOM):

$$LOM_{[a,b]}(g_0) = 2\cdot\min(LOI_{[a,b]}(g_0), LOD_{[a,b]}(g_0))$$

Normalized forms in $g(x) = g_0(x+a)-g_0(a)$0 are defined for comparability across scales:

• $g(x) = g_0(x+a)-g_0(a)$1
• $g(x) = g_0(x+a)-g_0(a)$2

For finite signed measures $g(x) = g_0(x+a)-g_0(a)$3, the corresponding indices are:

• Lack of Positivity (LOP): $g(x) = g_0(x+a)-g_0(a)$4
• Lack of Sign-Constancy (LOS): $g(x) = g_0(x+a)-g_0(a)$5

Projection theorems ensure that LOI, LOD, and LOM are always attained at the "closest" monotonic function. These indices are translation invariant, positive-scale invariant (for normalization), and allow ranking by monotonicity—$g(x) = g_0(x+a)-g_0(a)$6 if $g(x) = g_0(x+a)-g_0(a)$7. Practical computation reduces to summing over derivative signs on prescribed grids (Davydov et al., 2017).

Qoyyimi and Zitikis introduced complementary $g(x) = g_0(x+a)-g_0(a)$8-type indices based on the $g(x) = g_0(x+a)-g_0(a)$9-distance from a function to its non-decreasing rearrangement on $[0, y]$0 (Qoyyimi et al., 2014):

• Index $[0, y]$1: $[0, y]$2
• Linear Index $[0, y]$3: $[0, y]$4

Here, $[0, y]$5 is the non-decreasing rearrangement. Both indices vanish iff $[0, y]$6 is non-decreasing almost everywhere. Efficient computation is via grid discretization and sorting. The linear index $[0, y]$7 is comonotonic-additive and both indices are robust under perturbations in $[0, y]$8 (Lipschitz continuity).

These indices have proven applicability in econometrics (e.g., monotonicity of demand/supply curves), finance (utility/distortion functions), insurance (premium principles), and reliability (hazard rates) (Davydov et al., 2017, Qoyyimi et al., 2014).

2. Non-monotonicity in Optimization and Algorithmic Procedures

Non-monotonicity also arises concretely in optimization routines, notably in mixed-integer programming (MIP) and auction theory. Quantifiable failure of monotonicity manifests as paradoxical increases in computational resource usage or output sub-optimality upon "improvement" of model relaxations or data.

In the branch-and-bound framework for MIP, monotonicity of a branching rule $[0, y]$9 is defined such that for any relaxations $y = b-a$0 with matched integer points, and objective $y = b-a$1, the branch-and-bound tree size satisfies $y = b-a$2 (Shah et al., 2024). Non-monotonicity occurs when $y = b-a$3.

Theoretical results show:

• Any fractional-only branching rule (including full strong branching, reliability branching) is non-monotonic. Explicit constructions exist where adding a single valid cut increases the tree size from $y = b-a$4 to $y = b-a$5 for $y = b-a$6-dimensional instances (Shah et al., 2024).
• Empirically, in multi-dimensional knapsacks, about 20% of single cover cuts increase branch-and-bound tree size; with all cuts, 8% of instances still increase, but only for small gap closures ($y = b-a$7). For general MIPLIB instances, non-monotonicity is prevalent for cuts closing less than 10% optimality gap, but vanishes for higher gap closure (Shah et al., 2024).

These findings establish quantitative thresholds for non-monotonicity in practical solver performance.

In single-parameter auction environments, strong revenue monotonicity (SRM) holds iff feasible allocations form a matroid: running a Myerson auction designed for an underestimating prior cannot reduce expected revenue if the true distribution dominates stochastically. Outside the matroid setting, explicit counterexamples achieve constant-factor non-monotonicity; however, an approximate SRM is established under quantile-space proximity bounds between prior and true distributions, yielding tight additive guarantees, which directly inform sample complexity bounds for revenue learning (Chen et al., 2022).

3. Quantifying Non-monotonicity in Quantum Distinguishability

In quantum information, monotonicity of distance measures under tensor product extension is desirable but not guaranteed. The trace distance between density operators $y = b-a$8 is defined as $y = b-a$9 (Maziero, 2015). Non-monotonicity under tensor products (NMuTP) occurs when the order of distinguishability between two pairs of states reverses upon tensoring:

$g'(x) = g'^+(x) - g'^-(x)$0

Analytic results show that NMuTP does not occur for pure states or special collinear mixed states, but numerical sampling of one-qubit mixed states reveals that up to 8.85% of quartets exhibit non-monotonicity, with characteristic strength $g'(x) = g'^+(x) - g'^-(x)$1 out of 2. In higher $g'(x) = g'^+(x) - g'^-(x)$2, the fraction and magnitude of non-monotonicity fall rapidly, dropping below 1% for $g'(x) = g'^+(x) - g'^-(x)$3 (Maziero, 2015). The metric $g'(x) = g'^+(x) - g'^-(x)$4—the sum of differences in pre- and post-tensor distance—quantifies strength.

This property mandates caution in applying trace distance for mixed-state discrimination, especially in low-dimensional systems.

4. Physical Quantification: Non-monotonicity in Frictional Bimaterial Systems

Non-monotonic dependence of physical observables on system contrast parameters arises in frictional sliding across bimaterial interfaces. The maximal growth rate $g'(x) = g'^+(x) - g'^-(x)$5 of unstable interfacial perturbations (under regularized Coulomb or rate-and-state friction law) is shown to be a non-monotonic function of the contrast parameter $g'(x) = g'^+(x) - g'^-(x)$6 (shear modulus ratio) (Aldam et al., 2017). Analytically, for fixed physical parameters:

• $g'(x) = g'^+(x) - g'^-(x)$7 increases from zero at $g'(x) = g'^+(x) - g'^-(x)$8, peaks near $g'(x) = g'^+(x) - g'^-(x)$9, then declines, with maximal non-monotonicity $g'^+ = \max\{g', 0\}$0 (dimensionless units).
• The origin lies in the monotonic growth of elastodynamic bimaterial coupling $g'^+ = \max\{g', 0\}$1 versus simultaneous growth of Rayleigh-wave-related denominator $g'^+ = \max\{g', 0\}$2 so that their ratio peaks at intermediate contrast.
• Numerical simulations of dynamic rupture (slip-weakening law) confirm the same peak in observable measures versus contrast.

This quantification demonstrates that coupled material-friction systems can exhibit sharp, predictable non-monotonic effects, crucial for geophysical and engineering applications.

5. Summary Table: Indices and Manifestations Across Domains

Domain	Non-monotonicity Metric	Empirical/Theoretical Quantification
Real functions, measures	$g'^+ = \max\{g', 0\}$3, $g'^+ = \max\{g', 0\}$4, $g'^+ = \max\{g', 0\}$5, $g'^+ = \max\{g', 0\}$6, $g'^+ = \max\{g', 0\}$7	Values in $g'^+ = \max\{g', 0\}$8 (normalized), analytic, grid-based computation (Davydov et al., 2017, Qoyyimi et al., 2014)
MIP/Optimization	$g'^+ = \max\{g', 0\}$9 (relative tree size), worst-case gap	Up to $g'^- = \max\{-g', 0\}$0 blowup, 20% single-cut freq. (Shah et al., 2024)
Auctions	Revenue shortfall; tightness via $g'^- = \max\{-g', 0\}$1-sample	Constant-factor failure outside matroids, additive/proximity bounds (Chen et al., 2022)
Quantum states	$g'^- = \max\{-g', 0\}$2 (distance order-reversal measure)	8.85% in qubits, $g'^- = \max\{-g', 0\}$3 in $g'^- = \max\{-g', 0\}$4 (Maziero, 2015)
Bimaterial friction	$g'^- = \max\{-g', 0\}$5, $g'^- = \max\{-g', 0\}$6 (peak minus endpoints)	$g'^- = \max\{-g', 0\}$7, peak at $g'^- = \max\{-g', 0\}$8 (Aldam et al., 2017)

6. Applications and Significance Across Fields

The ability to quantify non-monotonicity enables:

• Model selection and fit assessment in statistics, econometrics, insurance premium calculation, and reliability theory by ranking candidates via lack-of-monotonicity indices (Davydov et al., 2017, Qoyyimi et al., 2014).
• Decision-theoretic calibration of utility and distortion functions in economics, behavioral finance, and actuarial science through explicit normalized measures (Davydov et al., 2017).
• Algorithmic diagnosis in computational optimization, informing branching rule design, cut selection, and solver benchmarking, based on empirically grounded non-monotonicity statistics (Shah et al., 2024).
• Theoretical analysis of auction mechanisms and revenue guarantees, sharply identifying the structural boundaries (matroidality) for strong monotonicity results and guiding empirical design via approximate controls (Chen et al., 2022).
• Physical system design and interpretation in geophysics and materials engineering, predicting and leveraging non-monotonic responses in complex coupled systems (Aldam et al., 2017).

In all these applications, quantification of non-monotonicity offers practitioners and theorists a principled alternative to binary monotonicity checks, enabling gradient-like reasoning and precise characterization of deviation from ideal structural or behavioral assumptions.