Two-Piece Linear Fit in Data Analysis

• The two-piece linear fit is a segmented regression technique that divides data into two regimes with a continuity constraint at the breakpoint.
• It employs methods like grid search, dynamic programming, and total variation regularization to achieve robust and sparse model estimation.
• Applications span mechanical stress analysis, signal processing, and computer vision where quickly changing data behaviors need accurate quantification.

A two-piece linear fit, often referred to as a segmented linear or piecewise-linear regression with one breakpoint, comprises two connected linear segments parameterized to optimally approximate a real-valued dataset. This structure is fundamental in settings where the underlying data displays distinct regimes or transitions in behavior—such as mechanical yielding in materials or functional changes in empirical measurements. Applications range from statistical modeling of physical systems (Dubey et al., 2015), robust curve fitting in engineering and computer vision (Shklyar, 2016), to sparse adaptive spline construction in signal processing (Debarre et al., 2020). The fitting procedure is strongly linked to the choice of constraints (such as continuity at the breakpoint), regularization (e.g., total variation penalties), and statistical properties (such as equivariance and consistency under realistic noise models).

1. Theoretical Motivation and Problem Definition

A two-piece linear fit models observations $y_i$ versus predictors $x_i$ with the form

$$f(x) = \begin{cases} m_1 x + b_1, & x \leq t \ m_2 x + b_2, & x \geq t \end{cases}$$

subject to a continuity constraint at the (a priori unknown) breakpoint $t$: $m_1 t + b_1 = m_2 t + b_2$. In stress-strain analysis of amorphous solids, these two segments naturally encode elastic and post-yield regimes (Dubey et al., 2015). More generally, the model captures regimes where changing system properties or responses produce abrupt variation in slope or intercept.

This approach is theoretically justified in settings where nonlinear power-series expansions (e.g., $$\sigma(\gamma) = \mu_0 \gamma + B_2 \gamma^2 + \ldots$$) are either ill-posed or fail to capture regime phenomena—as happens in low-temperature, strain-controlled measurements where the elastic response is in fact piecewise linear and not globally nonlinear (Dubey et al., 2015). In statistics and signal processing, promoting sparsity in the second derivative via total variation leads to solutions with a minimal number of knots, with the two-piece linear fit representing the sparsest nontrivial case (Debarre et al., 2020).

2. Model Formulations and Key Equations

2.1 Least Squares (Design Matrix Formulation)

For data sorted such that $x_1 < x_2 < ... < x_N$, the least squares objective, with continuity enforced at $t$, is

$$\min_{m_1, b_1, m_2, b_2, t} \sum_{i=1}^N \left(y_i - f(x_i)\right)^2, \quad \text{s.t.} \quad m_1 t + b_1 = m_2 t + b_2$$

Fixing $t$ reduces the problem to linear regression in $x_i$0, where $x_i$1, with $x_i$2 and $x_i$3. The solution satisfies the normal equations

$x_i$4

with $x_i$5 as the $x_i$6 design matrix $x_i$7 (Troeng et al., 2018).

2.2 Total Variation Regularization

The penalized formulation with total variation on the second derivative yields the "generalized Beurling-LASSO": $x_i$8 where $x_i$9 denotes the space of functions with bounded second derivative in the sense of measures (Debarre et al., 2020). This always yields a piecewise linear solution with the number of knots controlled by $$f(x) = \begin{cases} m_1 x + b_1, & x \leq t \ m_2 x + b_2, & x \geq t \end{cases}$$0.

2.3 Two-Line Fitting via Degenerate Conic Sections

For bivariate ($$f(x) = \begin{cases} m_1 x + b_1, & x \leq t \ m_2 x + b_2, & x \geq t \end{cases}$$1) data possibly sampled from two distinct, unlabelled lines, a degenerate conic formulation captures all points $$f(x) = \begin{cases} m_1 x + b_1, & x \leq t \ m_2 x + b_2, & x \geq t \end{cases}$$2 such that

$$f(x) = \begin{cases} m_1 x + b_1, & x \leq t \ m_2 x + b_2, & x \geq t \end{cases}$$3

which expands into a second-degree polynomial in $$f(x) = \begin{cases} m_1 x + b_1, & x \leq t \ m_2 x + b_2, & x \geq t \end{cases}$$4 whose coefficients can be fit via adjusted least squares, maximum likelihood, or robust moment methods (Shklyar, 2016).

3. Algorithmic Approaches and Computational Complexity

The two-piece linear fit can be computed by:

• Grid Search on Breakpoint: For every candidate $$f(x) = \begin{cases} m_1 x + b_1, & x \leq t \ m_2 x + b_2, & x \geq t \end{cases}$$5, solve the reduced 3-parameter least squares problem and compute the residual sum of squares (RSS). The minimizer $$f(x) = \begin{cases} m_1 x + b_1, & x \leq t \ m_2 x + b_2, & x \geq t \end{cases}$$6 gives the optimal breakpoint (Troeng et al., 2018).
• Dynamic Programming: For $$f(x) = \begin{cases} m_1 x + b_1, & x \leq t \ m_2 x + b_2, & x \geq t \end{cases}$$7 pieces, dynamic programming maintains value functions representing the best achievable fit with $$f(x) = \begin{cases} m_1 x + b_1, & x \leq t \ m_2 x + b_2, & x \geq t \end{cases}$$8 segments up to each point, but the two-piece case reduces to a direct scan over $$f(x) = \begin{cases} m_1 x + b_1, & x \leq t \ m_2 x + b_2, & x \geq t \end{cases}$$9 points with $t$0 time per candidate (Troeng et al., 2018).
• Regularized Regression: For total variation-regularized problems, one first solves a discrete LASSO ($t$1-penalized second differences) to obtain smoothed targets, then fits a canonical interpolant and locates dual certificate saturation regions to place the knot efficiently (Debarre et al., 2020).

The computational costs are dominated by $t$2 candidate breakpoint evaluations for standard ordinary least squares. TV-based methods scale as $t$3 per iteration for dense solvers, or $t$4 with matrix-free algorithms exploiting the banded structure of the difference operator (Debarre et al., 2020).

4. Statistical and Mechanical Interpretations

The fitted parameters of the two-piece linear model possess distinct mechanical and statistical meanings:

• Slope of First Segment ($t$5 or $t$6): Represents the pre-breakpoint or "small-strain" modulus, directly linked to the Born-plus-fluctuation shear modulus in amorphous solids (Dubey et al., 2015).
• Breakpoint ($t$7 or $t$8): Physically, marks the location of a plastic threshold, yield event, or systemic change; statistically, partitions data into two regimes for distinct linear behavior.
• Slope of Second Segment ($t$9 or $m_1 t + b_1 = m_2 t + b_2$0): Represents post-break, "softened" modulus. Steady-state regimes in mechanical data are characterized by non-vanishing $m_1 t + b_1 = m_2 t + b_2$1 when averaged in the "quenched" sense (Dubey et al., 2015).
• Continuity Enforcement: Imposing $m_1 t + b_1 = m_2 t + b_2$2 reduces degrees of freedom, restricting to parsimonious, physically interpretable fits (Troeng et al., 2018).
• Total-Variation Dual Certificate: Satisfying Karush-Kuhn-Tucker conditions ensures that the solution is sparsest in knots, and characterizes uniqueness/nonuniqueness of the fit (Debarre et al., 2020).

Contrasts between "quenched" (fit each realization, then average) and "annealed" (average curves, then differentiate) interpretations in amorphous solids lead to different modulus estimations, with only the quenched average matching theoretical predictions (Dubey et al., 2015).

5. Robustness, Uniqueness, and Alternatives

Uniqueness Properties

• In total-variation regularized regression, the solution with exactly one knot is unique iff the dual certificate saturates only at a point; otherwise, there exists a one-parameter family of fits with the knot in an interval, all with identical minimal TV norm (Debarre et al., 2020).
• For bivariate two-line fitting, the adjusted least squares (ALS) projection onto the degenerate conic submanifold is consistent and asymptotically normal under both "functional" and "structural" noise models. Maximum likelihood achieves lower bias/variance when Gaussian mixture assumptions apply, but ALS and robust moment estimators outperform for non-Gaussian or high-noise regimes (Shklyar, 2016).

Methodological Alternatives

Estimator	Consistency Guarantees	Practical Notes
Adjusted Least Squares	Functional/Structural, proven	Fast—eigenvalue computations
Orthogonal Regression	Functional (not always structural)	ML for functional, can break down
Maximum Likelihood	If model exactly Gaussian mixture	EM algorithm required, best for ML case
RBAN Moment	Moment-based, always	Small least-squares, robust
TV-Regularization (sparse)	Minimally knots, unique when possible	$m_1 t + b_1 = m_2 t + b_2$3 solvers, sparsity control

• All canonical estimators (ALS, ML, RBAN, Orthogonal Regression) are similarity equivariant (rotation, translation, scaling invariant) (Shklyar, 2016).
• Bootstrap, weighted least squares, and dual-certificate procedures allow uncertainty quantification and finite-sample adaptive analysis (Dubey et al., 2015, Debarre et al., 2020).

6. Implementation and Applications

Pseudocode and efficient updates for the two-piece linear fit are available for both regression and mechanical data contexts (Dubey et al., 2015, Troeng et al., 2018). For low-noise elasticity applications, median or low-pass filtering may be employed before fitting to suppress high-frequency fluctuations, and intervals directly following large plastic drops can be masked for enhanced sensitivity (Dubey et al., 2015). Confidence intervals for model parameters can be established via bootstrapping or analytic propagation of error.

Key applications include:

• Quantitative analysis of elastic-plastic transitions in amorphous solids under strain (Dubey et al., 2015)
• Robust (sparse) regression and denoising of one-dimensional signals (Debarre et al., 2020)
• Geometric fitting of mixed data from multiple lines, such as track reconstruction or vision tasks (Shklyar, 2016)
• Curve approximation and change-point detection in time-series (Troeng et al., 2018)

The methodology is compatible with statistical model comparison tools, such as F-tests, AIC, or BIC, to evaluate the improvement of two-piece models over single-line fits (Dubey et al., 2015). The rigorous handling of regularization, dual certificates, and equivariant estimation makes two-piece linear fits fundamental in both theory and applied data analysis.

---

References:

• (Dubey et al., 2015) Elasticity in Amorphous Solids: Nonlinear or Piece-Wise Linear?
• (Shklyar, 2016) Equivariant adjusted least squares estimator in two-line fitting model
• (Troeng et al., 2018) A Refined Algorithm for Curve Fitting by Segmented Straight Lines
• (Debarre et al., 2020) Sparsest Piecewise-Linear Regression of One-Dimensional Data