The use of fixed-point theory in solving mathematical modeling problems has gained increasing attention in recent literature, ²¹ further validating its suitability for stabilizing fuzzy regression models under uncertainty. The methodological framework is anchored in three rigorously defined mathematical constructs that establish the foundation for robust fuzzy regression analysis. A trapezoidal fuzzy number (TrFN) A ~ is formally characterized by a quadruple ( a , b , c , d ) ∈ ℝ 4 where a ≤ b ≤ c ≤ d , with the membership function: μ A ~ ( x ) = { x − a b − a a ≤ x < b 1 b ≤ x ≤ c d − x d − c c < x ≤ d } (1)

This representation distinguishes the core [ b , c ] (values with full membership) from the support [ a , d ] (values with non-zero membership), enabling nuanced encoding of epistemic uncertainty. The d ∞ -metric between two TrFNs A ~ 1 = ( a 1 , b 1 , c 1 , d 1 ) and A ~ 2 = ( a 2 , b 2 , c 2 , d 2 ) is defined as: d ∞ ( A ~ 1 , A ~ 2 ) = max ( | a 1 − a 2 | , | b 1 − b 2 | , | c 1 − c 2 | , | d 1 − d 2 | ) (2) Theorem 1.

²² The space ( F ( ℝ ) , d ∞ ) of TrFNs equipped with the d ∞ -metric forms a complete metric space.

The parameter space Θ = F ( ℝ ) p + 1 of TrFN vectors θ = ( A ~ 0 , … , A ~ p ) is equipped with the extended d ∞ -metric: ‖ θ 1 − θ 2 ‖ ∞ = max 0 ≤ j ≤ p d ∞ ( A ~ j ( 1 ) , A ~ j ( 2 ) ) (4) Lemma 1.

( Θ , ∥ ⋅ ∥ ∞ ) is a Banach space.

The cost function J ( θ ) = ∑ i d ∞ 2 ( y ~ i , y ~ i est ( θ ) ) is defined within the complete metric space ( T , d ∞ ) of trapezoidal fuzzy numbers. Because d ∞ represents a metric rather than an inner product, J ( θ ) is not differentiable in the classical sense. Instead, it possesses Fréchet continuity with respect to the metric topology, ensuring smooth variation of J ( θ ) under infinitesimal perturbations of fuzzy parameters. The iterative operator T ( θ ) associated with this cost function satisfies the contraction property: ∥ T ( θ 1 ) − T ( θ 2 ) ∥ d ∞ ≤ L ∥ θ 1 − θ 2 ∥ d ∞ , L < 1 , guaranteeing convergence toward a unique fixed point that minimizes J ( θ ) .

Therefore, the estimation procedure avoids explicit gradient computation and instead relies on Banach’s Fixed-Point Theorem as a theoretically sound alternative to gradient-based optimization in fuzzy regression analysis. The empirical Lipschitz constants computed for the key predictors are presented in Table 2. Theorem 2.

(Contraction). Under the Lipschitz condition ∥ ∇ J ( θ 1 ) − ∇ J ( θ 2 ) ∥ ∞ ≤ K ∥ θ 1 − θ 2 ∥ ∞ with K ⋅ max ∥ X ~ ∥ < 1 , T satisfies: ‖ T ( θ 1 ) − T ( θ 2 ) ‖ ∞ ≤ L ‖ θ 1 − θ 2 ‖ ∞ , L = 1 − η + ηK max ‖ X ~ ‖ < 1 (6)

Derived via mean value theorem and bounded gradient variation. Empirical verification:

K quantifies gradient sensitivity. The max ∥ X ~ j ∥ values derive from the UCI dataset. All L < 1 confirm contractivity.

The parameter estimation procedure is formalized as:

Initialization: θ ( 0 ) ← Tanaka ’ s solution

Iteration: For k = 0 , 1 , … a.

Compute predictions: Y ~ i ( k ) = ⨁ j A ~ j ( k ) ⊗ X ~ ij

Termination: ∥ θ ( k + 1 ) − θ ( k ) ∥ ∞ < 10 − 6

The per-iteration computational requirements of the proposed method are listed in Table 3.

Computational Implementation: •

TrFNs represented as 4D vectors ( a , b , c , d )

Linear complexity enables scalability. Benchmarks performed on Intel i7-12700H.

For the UCI Concrete Dataset, crisp values x i are transformed to TrFNs X ~ i via: X ~ i = { ( x i ( 1 − δ 2 ) , x i ( 1 − δ 1 ) , x i ( 1 + δ 1 ) , x i ( 1 + δ 2 ) ) ( symmetric ) ( x i − δ 2 − , x i − δ 1 − , x i + δ 1 + , x i + δ 2 + ) ( asymmetric ) } (7)

Core ( δ 1 ) and support ( δ 2 ) widths derive from material uncertainty sources:

Asymmetric fuzzification applied where physical constraints exist (e.g., strength ≥0).

Robustness validation

Noise sensitivity is quantified via perturbations: X ~ i ϵ = ( x i ( 1 − ϵ ) , x i ( 1 − ϵ 2 ) , x i ( 1 + ϵ 2 ) , x i ( 1 + ϵ ) ) , ϵ ∼ U ( 0,0.05 ) (8)

The fuzzification parameters used to model input uncertainty are shown in Table 4. While a deeper justification of the fuzzification intervals is provided in Table 5.

This methodology establishes a theoretically rigorous fusion of fixed-point theory and fuzzy regression within a Banach space. The contraction operator ensures existence, uniqueness, and convergence—resolving foundational limitations in prior approaches. The fuzzification protocol embeds real-world uncertainty, while computational design ensures practical feasibility.

The empirical validation employs the UCI Concrete Compressive Strength Dataset, ¹⁹ comprising 1,030 observations of high-performance concrete formulations. Key variables include: •

Predictors:

Fuzzification transformed crisp values into trapezoidal fuzzy numbers (TrFNs) using industry-derived uncertainty parameters:

For example, a 35 MPa strength measurement becomes: Strength ~ = ( 35 × 0.935 , 35 × 0.982 , 35 × 1.018 , 35 × 1.065 ) = ( 32.7 , 34.4 , 35.6 , 37.3 )

Convergence analysis

The fixed-point algorithm demonstrated consistent convergence across 100 trials with randomized initial parameters. The convergence behavior across repeated runs is summarized in Table 6: •

Geometric decay of error ∥θ ^(k+1) − θ ^(k)∥∞ ( Figure 1) aligns with Theorem 3 predictions

Computational stability

The comparative ambiguity widths of model coefficients are given in Table 7. •

Zero overflow/underflow occurrences in 1.03 × 10 ⁵ arithmetic operations

Accuracy (MSE)

Mean Squared Error evaluated at α = 0 cut (support boundaries) and α = 1 (core boundaries):

MSE Reduction = 7.68 − 5.75 7.68 × 100 % = 25.1 % ( support : 12.5 % )

Solution ambiguity

Average width of fuzzy parameters (measure of model uncertainty):

Critical Insight: Tighter parameter distributions indicate enhanced estimation precision, reducing epistemic uncertainty in strength predictions.

Noise perturbation protocol

The robustness of all methods under ±5% noise is reported in Table 8. •

Training data contaminated with ±5% uniform noise:

ΔMSE = Percentage increase in test MSE after noise exposure

Failure Rate: Instances where ∥θ∥ → ∞ or MSE > 50 MPa²

Stability visualization

Fixed-point method maintains prediction coherence ( Figure 3) due to contractive properties 18.3% lower ambiguity persists under noise (validating Theorem 2)

Failure cases in benchmark methods linked to unbounded error growth

The fixed-point approach demonstrates triple superiority:

Accuracy: 25.1% lower MSE than Tanaka’s method (p < 0.001)

Precision: 18.3% narrower solution ambiguity bands

Robustness: 4× lower MSE degradation under noise vs. LS-based methods

These empirical outcomes directly validate the theoretical framework: contractive operators suppress error propagation, while the d ∞ -metric’s completeness ensures solution stability. The convergence profile ( Figure 1) further confirms the geometric decay rate predicted by Banach’s theorem.

The existential fragility of traditional fuzzy regression techniques is evident when systematically considered. Tanaka’s minimum fuzziness principle fails to guarantee solutions on multicollinear correlative cases, ¹² noted; while in traditional LS-based approaches propagation of noise also results in unbounded error, because algebraic inversions do not consider topological constraints. Our approach avoids these issues by replacing deterministic optimization with a contractive mapping that converges uniquely (as proven) under certain conditions ( Theorem 2). The mathematics here underscores the statistically greater 4 × robustness to ±5 % input noise that we witness ( Table 8): where Tanaka’s and LS approaches are subject to error propagation, and unhelpful to degrees of freedom given that there is mathematical inherent failure, when the Lipschitz contraction condition L < 1 is met, perturbation to convergence is limited by default.

These advancements have significant ramifications for uncertainty-aware modeling in material science and structural engineering. In high-stakes applications like concrete strength prediction, our method not only yields useful point estimates, but also estimate quantifiable uncertainty bounds based on the widths of the TrFNs (e.g., 32.7-37.3 MPa for a given 35 MPa specimen). This enables engineers to propagate imprecision through design calculations, supporting reliability-based decision-making. For instance, the 18.3% ambiguity reduction directly translates to narrower safety margins in load-bearing calculations, potentially reducing material overdesign by 12–15% while maintaining safety standards (ACI 318-19).

Three limitations warrant emphasis. First, the contraction property hinges on the Lipschitz condition K ⋅ max ∥ X ~ ∥ < 1 , which may require data scaling for high-magnitude predictors (e.g., aggregate content exceeding 1,000 kg/m³). Second, while TrFNs efficiently model symmetric uncertainty, they may require extensions to Gaussian or LR-type fuzzy numbers for skewed distributions. Third, computational complexity scales linearly with predictors ( O ( np ) ), but for p > 50 , the iterative gradient updates ( Equation 7) may benefit from quasi-Newton acceleration.

Critically, performance depends on appropriate fuzzification parameters ( Table 5). Overly narrow supports (e.g., δ 2 < 2 % ) artificially suppress uncertainty, while excessive widths inflate ambiguity. We recommend deriving δ 1 , δ 2 from domain-specific standards (e.g., ASTM tolerances) or bootstrap resampling. Future work should explore automated fuzzification via uncertainty quantification techniques like Monte Carlo dropout.

This study demonstrates that fixed-point theory transcends theoretical elegance to deliver tangible improvements in fuzzy regression’s practicality. By embedding estimation within a complete metric space and leveraging Banach’s contractive principles, we resolve the existence-uniqueness-stability trilemma that has hindered the field since Tanaka’s pioneering work. The resulting framework bridges mathematical rigor with engineering utility—a critical step toward trustworthy uncertainty quantification in data-driven design.