Critical Parameter: Theory & Applications

• Critical parameter is a value or set of values that marks the onset of phase transitions and singularities in system observables.
• They are determined using scaling laws, variational principles, and finite-size corrections across diverse disciplines.
• Their identification has practical implications for understanding AI robustness, PDE bifurcations, and non-equilibrium physical phenomena.

A critical parameter is a value or set of values in a physical, mathematical, computational, or engineering system at which the qualitative or universal behavior of the system undergoes a non-analytic change, initiates a phase transition, or exhibits a singularity in an appropriate observable or response function. In statistical physics and field theory, a critical parameter typically denotes a coupling, temperature, chemical potential, or control variable at which thermodynamic or quantum fluctuations become correlated on all scales, often corresponding to second-order or continuous phase transitions. The notion of "critical parameter" also arises in a wide range of contexts, from condensed matter and nuclear matter, to percolation processes, random networks, PDE bifurcations, AI parameter sensitivity, and systems with multiple control variables. Definitions and methods for identifying, characterizing, and exploiting critical parameters vary by domain but share deep connections via scaling, universality, and variational or extremal principles.

1. Theoretical Formulation and Universal Structure

The definition of a critical parameter is rooted in the detection of singularities or non-analytic behavior in observable quantities or the system’s state. In classical phase transitions, the critical parameter is often the value of temperature ($T_c$), chemical potential ($\mu_c$), pressure ($P_c$), or external field at which order–disorder transitions occur. Critical parameters are identified through thermodynamic inflection-point criteria or as unique solutions to systems of equations such as

$$\left.\frac{\partial P}{\partial \rho}\right|_{(\rho_c,T_c)}=0, \quad \left.\frac{\partial^2 P}{\partial \rho^2}\right|_{(\rho_c,T_c)}=0$$

as in the liquid–gas critical point of nuclear models (Lourenço et al., 2017).

In statistical field theory and the Ginzburg–Landau–Wilson paradigm, the approach to the critical parameter is characterized by universal scaling of correlation length ($\xi$), order-parameter cumulants, and effective interaction strengths, all diverging or changing sign at the critical point (Bluhm et al., 2016). The mapping between model variables (e.g., Ising model $r$, $h$) and physical observables is system-dependent but preserves the universality of critical phenomena.

In the context of percolation and random geometric models, the critical parameter often delineates the percolative and non-percolative regimes, as in the critical intensity $u_*$ controlling the density of random interlacements (Sznitman, 2010, Sznitman, 2010) and the critical rate $\lambda_c$ for infinite-loop formation in random loop models (Björnberg et al., 2016). In these cases, $u_*$ (or $\mu_c$0) is defined as the threshold at which an infinite connected component appears or vanishes.

Critical parameters also act as bifurcation points for solution branches of nonlinear PDEs (Quoirin et al., 2021), or as singularities in thermodynamic response functions such as the Grüneisen parameter $\mu_c$1, which signals diverging susceptibilities at phase transitions (Soares et al., 2024).

2. Mathematical Characterization and Determination

A broad set of techniques exists to determine critical parameters:

• Legendre and variational principles: Critical parameters are the unique values at which variational derivatives vanish or where the corresponding functional attains zero energy at a solution. For example, in the context of parameter-dependent functionals $\mu_c$2 on Banach spaces,

$\mu_c$3

This admits a sequence $\mu_c$4 of critical parameters via Ljusternik–Schnirelmann minimax theory and the nonlinear Rayleigh quotient method (Quoirin et al., 2021).

• Scaling and universality: Near a critical parameter, observables exhibit power-law divergence or specific sign structure, as in $\mu_c$5 or high-order cumulants scaling with $\mu_c$6 (Bluhm et al., 2016).
• Percolation thresholds: For probabilistic models, $\mu_c$7 or $\mu_c$8 is characterized by

$\mu_c$9

and can be estimated by renormalization, coupled Poisson process analysis, and large-deviation bounds (Sznitman, 2010, Björnberg et al., 2016).

• Trap-size scaling (TSS) and finite-size scaling (FSS): In numerical and experimental analyses, critical parameters are determined by extracting pseudo-critical points from observable crossings or peak shifts in the trap size $P_c$0 or system size $P_c$1, and extrapolating to the thermodynamic or infinite-trap-size limit using scaling corrections governed by irrelevant exponents (Ceccarelli et al., 2012).
• Parameter sensitivity and model selection: In machine learning and statistics, "critical" parameters are those whose small perturbations lead to large losses or prediction shifts, identified via loss-sensitivity or gradient-based metrics, and are prioritized for robust estimation or protection in continual learning and federated settings (Ling et al., 18 Apr 2025, Han et al., 2023).

3. Prototypical Critical Parameters Across Disciplines

Domain	Critical Parameter(s)	Definition/Location
QCD phase diagram	$P_c$2	Divergence of $P_c$3, cumulants; mapped from Ising universality
Nuclear liquid–gas transition	$P_c$4	Inflection point of $P_c$5 isotherm
Random interlacement	$P_c$6	Percolation threshold for infinite vacant set
Random loop model	$P_c$7	Emergence of infinite loops in Poisson process on trees
Trapped Bose–Hubbard	$P_c$8	Trap-size scaling of susceptibility, correlation length
Parameter-tuned PDEs	$P_c$9	Minimax (LS genus) levels of zero-energy critical points
Grüneisen ratio	$$\left.\frac{\partial P}{\partial \rho}\right|_{(\rho_c,T_c)}=0, \quad \left.\frac{\partial^2 P}{\partial \rho^2}\right|_{(\rho_c,T_c)}=0$$0 or $$\left.\frac{\partial P}{\partial \rho}\right|_{(\rho_c,T_c)}=0, \quad \left.\frac{\partial^2 P}{\partial \rho^2}\right|_{(\rho_c,T_c)}=0$$1	Point where entropy scaling changes; regularizes divergences
AI robustness/Federated learning	Top-k parameters	Highest loss-gradient sensitivity in network weight matrices

The specific formalism and linguistic convention (e.g., "critical point," "critical level," "critical coupling," "critical matrix") depend on the mathematical structure and application domain but consistently reflect the locus of qualitative transition.

4. Physical and Computational Implications

Physical systems at or near a critical parameter exhibit:

• Diverging fluctuations: Correlation length and higher-order cumulants (susceptibility, skewness, kurtosis) diverge at the critical parameter, but their observable signals (e.g., in finite-size systems or non-equilibrium settings) are softened by finite-size/time effects or entropic regularization schemes (Bluhm et al., 2016, Li et al., 2018, Soares et al., 2024).
• Universality and scaling: The scaling exponents and functional forms near the critical parameter are universal, depending only on symmetry and dimensionality, not on microscopic details.
• Nonanalyticity and bifurcation: The critical parameter marks nonanalytic behavior in the thermodynamic potentials (free energy, pressure) or bifurcation in the solution set of a PDE or functional (Quoirin et al., 2021).
• Parameter estimation challenges: In inference for critical models (e.g., affine processes at the critical mean-reversion), standard statistical rates and limit theorems are invalid; "slow" convergence and non-standard (non-Gaussian) limiting distributions require specialized inference and test procedures (Barczy et al., 2012).
• Sensitivity "hot spots" in optimization: In high-dimensional learning, critical parameters often correspond to subnetworks or weights whose integrity is essential for robustness to adversarial or catastrophic forgetting, and are explicitly detected and protected (Ling et al., 18 Apr 2025, Han et al., 2023).

5. Methods for Experimental and Numerical Identification

Accurate determination of critical parameters employs several methodologies:

• Scaling and collapse: Plotting suitably rescaled observables as functions of control parameters across multiple system or trap sizes and identifying crossing points, then extrapolating using known correction-to-scaling exponents (Ceccarelli et al., 2012).
• Precision measurement and parameter ranking: In transport or device models, critical parameters are those whose experimental uncertainty produces the largest spread in predictions (e.g., protonic conductivity, water diffusivity, electro-osmotic drag in PEMFCs) (Vetter et al., 2018).
• Finite-size corrections and binning: Accounting for corrections of the form $$\left.\frac{\partial P}{\partial \rho}\right|_{(\rho_c,T_c)}=0, \quad \left.\frac{\partial^2 P}{\partial \rho^2}\right|_{(\rho_c,T_c)}=0$$2 or $$\left.\frac{\partial P}{\partial \rho}\right|_{(\rho_c,T_c)}=0, \quad \left.\frac{\partial^2 P}{\partial \rho^2}\right|_{(\rho_c,T_c)}=0$$3 to extrapolate to the infinite-system limit.
• Sparse identification and gradient analysis: In ML, importance scores based on the local gradient or parameter-update size quantify criticality per parameter or per block, informing freezing or robust aggregation (Ling et al., 18 Apr 2025, Han et al., 2023).
• Variational minimax search: For functionals, critical parameters arise from the infimum/supremum over topological genus classes and are detected via critical-point theory (Quoirin et al., 2021).

6. Extensions to Multi-Parameter and Out-of-Equilibrium Contexts

In multi-parameter systems (e.g., quantum metrology, Dicke models), "multi-critical" points arise when two or more parameters are tuned such that multiple excitation gaps close simultaneously, enhancing parameter sensitivity and potentially overcoming rank-deficiency ("sloppiness") in the Fisher information metric. The variance bounds for simultaneous estimation often scale as distinct powers of the distance to the multi-critical parameter, offering tradeoffs between precision and state preparation time (Previdi et al., 3 Mar 2026). Out-of-equilibrium transitions, such as steady-state phase transitions in quenched Ising models, may feature "critical parameters" not evident in equilibrium statistics (e.g., logarithmic vanishing of the order parameter rather than power-law) (Li et al., 2018).

7. Universality, Regularization, and Open Challenges

The behavior near a critical parameter is fundamentally dictated by universality classes: systems with different microscopic details but identical symmetries and dimensions exhibit quantitatively identical scaling of observables at criticality (Bluhm et al., 2016, Ceccarelli et al., 2012). However, classical formulations often predict illusory infinite responses (susceptibilities, Grüneisen ratios) that are regularized via generalized entropic frameworks (e.g., Tsallis entropy with system-dependent $$\left.\frac{\partial P}{\partial \rho}\right|_{(\rho_c,T_c)}=0, \quad \left.\frac{\partial^2 P}{\partial \rho^2}\right|_{(\rho_c,T_c)}=0$$4), restoring finite, system-size-independent maxima at the critical point (Soares et al., 2024).

Open challenges include full characterization of subleading corrections (e.g., in percolation critical levels), extension of scaling concepts to non-equilibrium or disordered systems, and robust algorithmic identification and protection of critical parameters in high-dimensional learning under adversarial or distributed conditions.

---

References:

• Order-parameter and cumulant scaling in QCD criticality (Bluhm et al., 2016)
• Grüneisen parameter regularization and entropy extensivity (Soares et al., 2024)
• Continual learning and critical parameter freezing in vision transformers (Ling et al., 18 Apr 2025)
• Federated learning robust aggregation using critical parameter analysis (Han et al., 2023)
• Critical points in parameter-dependent functionals and PDEs (Quoirin et al., 2021)
• High-dimensional percolation critical threshold (Sznitman, 2010, Sznitman, 2010)
• Random loop models on trees and critical rate expansion (Björnberg et al., 2016)
• Nuclear matter criticality in RMF models (Lourenço et al., 2017)
• Trap-size scaling for critical parameter extraction in cold atomic systems (Ceccarelli et al., 2012)
• Non-equilibrium critical scaling in Ising quenches (Li et al., 2018)
• Statistical inference for critical affine processes (Barczy et al., 2012)
• Experimental prioritization of critical transport parameters in PEMFCs (Vetter et al., 2018)
• Multi-critical quantum metrology and sloppiness (Previdi et al., 3 Mar 2026)