Point-Wise Robust Reference Paths

• Point-Wise Robust Reference Paths are defined as trajectories constructed to withstand minor perturbations, serving as stable comparators across hardware Trojan detection, robust optimization, stochastic filtering, and algebraic inclusions.
• They are established through domain-specific methodologies such as SAT-ATPG extraction, Bregman projection, Taylor expansion, and Newton-corrector steps, each tailored to preserve stability against uncertainties.
• Empirical validations across different applications demonstrate high detection rates in hardware and provable error bounds in optimization and filtering, underscoring their practical importance for system integrity.

A point-wise robust reference path is a solution concept defined and operationalized across multiple areas including hardware Trojan detection, robust optimization, stochastic filtering, and algebraic inclusion problems. The unifying theme is pointwise robustness: the reference path is constructed or selected such that small perturbations or model uncertainties do not degrade its utility as a stable comparator or trajectory. This article surveys the foundations, mathematical formalism, algorithmic methods, and empirical validations associated with point-wise robust reference paths.

1. Formal Definitions and Mathematical Foundations

The definition of a point-wise robust reference path depends on the domain but typically involves two key aspects—pointwise comparison and robustness to perturbations. In hardware Trojan detection, a reference path is one member of a topologically symmetric pair of sensitizable paths in a circuit; their delays respond identically to inter-die process variation but can reveal localized anomalies such as Trojan insertion (Vaikuntapu et al., 2022). For robust linear optimization, the robust path is the trajectory $\{x^*(\tau):\tau\in[0,1]\}$ where $x^*(\tau)$ solves the robustified problem for uncertainty set parameter $\tau$ (Hao et al., 27 Aug 2025). In path-following for variational inclusions, one seeks a continuous mapping $x:[0,T]\rightarrow\mathbb{R}^n$ satisfying $0\in F(t,x(t))$ at every parameter value $t$, with pointwise convergence guarantees under semismooth$^*$ and subregularity conditions (Roubal et al., 2024).

For nonlinear filtering, the meaning is specialized. Given observation path $y$ and partition $\tau$, the discrete filter is encoded as a Lipschitz functional $F^{\tau,m}_\varphi(y)$, furnishing a deterministic, pathwise value robust against small changes in the input trajectory (Crisan et al., 2021).

The generic robustness criterion is often expressed via normalized deviations. For circuit paths, the detection metric

$DM = \frac{d}{\sqrt{D_{s,nom}^2 + D_{r,nom}^2}}$

reveals if a delay anomaly exceeds a Monte Carlo-calibrated threshold.

2. Domain-Specific Construction and Selection Procedures

The operationalization of point-wise robust reference paths varies with application:

• Hardware Trojan Detection: The central procedure involves selecting (or constructing) topologically symmetric path pairs that traverse the same number and types of gates. If no reference path exists for a vulnerable net, extra logic gates are inserted to fabricate a symmetric counterpart (Type-1 symmetry), ensuring process variation cancellation. The selection further prioritizes physical proximity in layout to exploit spatial correlation, minimizing sensitivity to intra-die variation (Vaikuntapu et al., 2022).
• Robust Optimization: Here, the robust path $\{x^*(\tau)\}$ is realized as a Bregman projection

$x^*(\tau) = \Pi^H_X( y(\tau) )$

where $y(\tau)$ encodes the dual geometry of the uncertainty set. Computationally, proximal or mirror descent trajectories

$$x_{k+1} = \arg\min_{x\in X} \{ c^Tx + \lambda_k D_H(x, x_k) \}$$

approximate the robust path, with provable error bounds tied to geometric mismatch parameters (Hao et al., 27 Aug 2025).

• Filtering Functionals: For time-discretized stochastic filtering, the reference functional $F^{\tau,m}_\varphi(y)$ is explicitly constructed by stochastic Taylor expansion, Riemann–Stieltjes integrals, and careful truncation. Its Lipschitz continuity in the sup-norm ensures pathwise robustness (Crisan et al., 2021).
• Algebraic Inclusions: In semismooth$^*$ path-following, the reference path is tracked by repeated Newton-corrected steps, with the one-step convergence quantified under local coderivative and strong metric subregularity conditions (Roubal et al., 2024).

3. Robustness Criteria and Error Quantification

Robustness is validated through analytical bounds and statistical calibration:

• Hardware Circuits: Robustness is asserted when, under process variation alone, the reference metric $DM$ remains below the detection threshold $DT$. The threshold is set so $\Pr\{DM > DT \mid \text{no Trojan}\}\leq \epsilon$ (e.g., $3\%$ false positives) (Vaikuntapu et al., 2022).
• Optimization and Filtering: In robust optimization, a sharp error bound exists

$$D_H(x_{CP}(\omega), x^*(1/\omega)) \leq \kappa^2 D_H(\Pi^H_X(0), \Pi^H_{\text{Aff}(X)}(0)),$$

while in filtering, the discretized filter achieves $O(\delta^m)$ mean-square error for mesh size $\delta$ of discretization (Crisan et al., 2021).

• Algebraic Inclusions: The error of path-tracking via Newton-corrector is quantified by

$$\|\bar u - x(s)\| \leq \frac{1}{2} \max\{a, s-t\} + \ell (s-t)$$

where $a$ is the initial deviation and $\ell$ is the Lipschitz constant on $x(\cdot)$ (Roubal et al., 2024).

A plausible implication is that, by controlling problem geometry and algorithmic step size, pointwise robustness can be engineered to persist even under adversarial or stochastic perturbations.

4. Algorithmic Procedures and Implementation Aspects

Tables summarizing main algorithmic elements from each domain:

Domain	Selection/Construction Method	Robustness Metric/Error Bound
HW Trojan Detection (Vaikuntapu et al., 2022)	SAT-ATPG path extraction; gate insertion for symmetry	$DM < DT$, calibrated by Monte Carlo
Robust Optimization (Hao et al., 27 Aug 2025)	Bregman projection of dual curve; proximal method	$D_H$-gap between central and robust path
Filtering (Crisan et al., 2021)	Taylor expansion; integration by parts; functional construction	$O(\delta^m)$ error, Lipschitz continuity
Algebraic Inclusions (Roubal et al., 2024)	Newton-corrector; step-size regulation	$O(h)$ local path-following error

Practical implementation involves netlist scanning and logic synthesis (HW), proximal steps with Bregman updates (optim.), pathwise functional calculation (filtering), and pointwise Newton prediction plus coderivative algebra (inclusions).

5. Modeling and Simulation Validation

Empirical validation is domain-dependent:

• Hardware: ISCAS-85 benchmarks (c432–c7552) with 32 nm PTM models report 100% true positive rates and $\leq3\%$ false positive rates under up to $15\%$ intra-die and $20\%$ inter-die process variation. Area overhead due to gate insertion for symmetric reference path creation is $0$–$14\%$ (Vaikuntapu et al., 2022).
• Optimization: Portfolio optimization under ellipsoidal uncertainty demonstrates exact tracing of the mean-variance frontier by proximal trajectories; simplex feasible sets and dual uncertainty structures deliver zero discrepancy between computed and true robust paths (Hao et al., 27 Aug 2025).
• Filtering: Theoretical analysis confirms robustness but awaits experimental realization. The Lipschitz property of the reference functional underpins stability necessary for ML-based surrogate construction (Crisan et al., 2021).
• Algebraic Inclusions: Numerical examples (e.g., diode-resistor circuit) validate that Newton-corrected path tracking incurs only linear local errors with grid size (Roubal et al., 2024).

Uniform high detection rates and consistent error control attest to the operational validity of the point-wise robust reference path framework in each domain.

6. Applications, Limitations, and Implications

Point-wise robust reference paths are foundational for:

• Hardware security, critical for “golden-free” detection of circuit Trojans and adversarial modifications.
• Robust decision-making and adversarial learning, where optimization under uncertainty and geometric control are paramount.
• Filtering and estimation in stochastic systems, especially where data-driven surrogates require stability with respect to input path fluctuations.
• Path-following in nonsmooth algebraic inclusions, providing rigorous local error guarantees and enabling adaptive algorithms.

Common limitations concern scalability, dependence on precise geometry or topology, and the necessity for careful calibration of robustness parameters and error thresholds.

A plausible implication is that widespread adoption of point-wise robust reference path constructions can improve the integrity and verifiability of digital systems, optimization pipelines, and learning algorithms under variable or uncertain operating conditions.