Introduction
Phase transitions are among the most fundamental mechanisms by which materials acquire new structures, properties, and functional responses [1 ] . They arise in a broad range of physical systems, including magnetic ordering, ferroelectricity, crystallization, martensitic transformations, and structural instabilities in solids [2 , 3 , 4 ] . These phenomena are characterized by changes in an underlying order parameter, the emergence or disappearance of symmetry, and the reorganization of the free-energy landscape [5 ] . In solid materials, such changes alter the constitutive symmetry, generate transformation strains, modify elastic anisotropy, and change the propagation and polarization of elastic waves [6 ] . A fundamental objective is therefore to establish how structural transformations emerge from thermodynamic principles and how their signatures can be identified through measurable mechanical responses.
The classical thermodynamic theory of phase transitions provides the starting point for this problem. The Gibbs and Ehrenfest descriptions characterize equilibrium transformations through changes in thermodynamic potentials and their derivatives [7 ] . Landau theory subsequently introduced a systematic description of symmetry-breaking transitions in terms of order parameters and symmetry-admissible free-energy expansions [18 ] . Nevertheless, conventional Landau models are commonly formulated for a selected order parameter and a prescribed symmetry class. Different phases or material systems may require different representations, invariant expansions, and constitutive assumptions. Moreover, a change in the symmetry of an order parameter does not necessarily constitute a thermodynamic phase transition. Statistical mechanics extends this picture by connecting phase transitions to collective fluctuations and the thermodynamic limit. The theory of critical phenomena and the renormalization group explain why microscopically different systems may exhibit identical critical exponents, and why mean-field predictions can fail near continuous transitions [8 ] . Dynamic critical phenomena further classify relaxation and transport according to the conservation laws and reversible or dissipative couplings of the relevant slow variables [19 ] . However, a universality class does not by itself specify the instantaneous material point group, the complete elastic tangent, or the directional spectral response of a mechanically probed solid. In particular, knowledge of an order-parameter exponent does not automatically determine the exponent measured through an acoustic mode splitting.
A complementary line of development originates from continuum thermodynamics and phase-field modeling. Ginzburg–Landau functionals introduce gradient contributions that penalize spatial variations of the order parameter, while Allen–Cahn and Cahn–Hilliard-type evolution equations describe, respectively, representative nonconserved and conserved dissipative dynamics [9 , 10 ] . More detailed approaches, including phase-field-crystal models, retain aspects of crystalline periodicity and couple structural ordering to elastic and defect-related phenomena [11 , 12 , 13 ] . Their limitations are model dependent, and difficult to construct a general correspondence between the local symmetry of an evolving state, the symmetry of its elastic tangent, and experimentally observable acoustic degeneracies. In solid mechanics, structural phase transformations have also been investigated through nonlinear elasticity, crystallographic compatibility, energy minimization, and the theory of martensitic microstructures. Variational models characterize transformations through multiple energy wells and deformation gradients related by crystallographic symmetry [14 ] . They have provided fundamental insight into twinning, coherent interfaces, transformation compatibility, and the reversibility of shape-memory materials [16 , 17 , 4 ] . Group-theoretic classifications of ferroelastic transitions similarly identify admissible symmetry-breaking pathways and transformation variants [15 ] . However, crystallographic classification, energy-well geometry, and nonlinear compatibility do not independently establish the critical fluctuation dynamics or the spectral observability of a given structural transition. Furthermore, constitutive descriptions employing separate phase-specific elastic laws may reproduce different phases without explicitly deriving their changing material symmetries from a common evolving internal state. These developments suggest that several logically distinct questions are often addressed separately. Although substantial connections between these approaches already exist, a remaining challenge is to formulate their common symmetry structure without conflating their respective physical criteria.
The present work addresses these issues through a symmetry-stratified constitutive framework for phase transitions with local, symmetry-bearing internal variables (see Figure 1 ). Building on Landau-type thermodynamics, the theory of critical phenomena, and the internal-tensor description of evolving material symmetry developed in Ref. [26 ] , we introduce a finite-dimensional representation of the parent symmetry group in which different structural states belong to distinct group orbits and stabilizer strata. Two mechanical realizations are established, including an isotropic parent with a symmetric traceless structural tensor and a cubic-to-tetragonal ferroelastic transformation. The numerical simulation and an experimental NiTi measurements were used to validate the prediction capability of the proposed models.
Figure 1. Schematic illustration of the symmetry-stratified framework for structural phase transitions and their mechanical signatures The upper panels illustrate the heterogeneous microstructure of a transforming solid, in which local structural states q ( X , t ) q(\mathbf{X},t) q ( X , t ) belong to distinct symmetry strata characterized by their stabilizer subgroups G q G_q G q . The magnified region highlights a representative structural variant, while the accompanying configuration-space diagram depicts the mapping from the reference body B \mathcal{B} B to the current configuration C t \mathcal{C}_t C t through a symmetry-stratified state representation. The lower-left panel illustrates the thermodynamic description, in which a parent-invariant free-energy landscape governs the stability, coexistence, and evolution of structural states under thermal activation or external driving. The lower-right panel establishes the mechanical connection between structural symmetry and observable acoustic spectra: the evolving internal state determines the tangent elastic modulus C ( q ) \mathbb{C}(q) C ( q ) , from which the direction-dependent acoustic tensor Γ ( n ^ ; q ) \mathbf{\Gamma}(\hat{\mathbf{n}};q) Γ ( n ^ ; q ) and its polarization eigenpairs are obtained. Symmetry constraints and constitutive coupling determine whether initially degenerate acoustic branches remain protected or undergo splitting, providing a spectral signature of the underlying structural transformation. The schematic distinguishes structural classification, thermodynamic phase selection, and mechanical observability within a unified constitutive framework.
The central mechanical contribution is to connect this state-space description to the elastic and acoustic operators. For a parent-equivariant constitutive law, the structural state constrains the symmetry of the tangent stiffness, while the Christoffel operator translates these restrictions into direction-dependent wave speeds and polarizations. Representation theory identifies symmetry-protected degeneracies, and degenerate perturbation theory determines the first nonvanishing order at which a structural perturbation becomes spectrally observable.
State variables, thermodynamics, and evolution
The formulation uses two levels of description. The state representation determines the symmetry of a given configuration. The free energy and evolution law determine which configurations can occur under specified mechanical and thermal controls.
Parent symmetry and structural state
Let Ω ⊂ R d \Omega\subset\mathbb{R}^{d} Ω ⊂ R d be a reference body, and let G 0 G_{0} G 0 be a compact parent point group or compact internal symmetry group. For broken spatial translations, spatially modulated phases, and noncompact groups, additional structure must be supplied. The following construction concerns a compact group acting on a finite-dimensional local state [36 ] .
Choose a finite-dimensional real vector space with orthogonal representation
V = ⨁ a = 1 m V a , ρ : G 0 ⟶ O ( V ) , η ( X , t ) = ( η 1 , … , η m ) . (1) \mathcal{V} = \bigoplus_{a=1}^{m} V_{a}, \qquad\rho: G_{0} \longrightarrow O(\mathcal{V}), \qquad\boldsymbol{\eta}(X,t) = (\eta_{1},\ldots,\eta_{m}). \tag*{(1)} V = a = 1 ⨁ m V a , ρ : G 0 ⟶ O ( V ) , η ( X , t ) = ( η 1 , … , η m ) . (1)
For complex fields, take their realification, or equivalently allow a unitary representation and complex-valued invariants. Under a material transformation g g g , the state changes as
η ↦ ρ ( g ) η . (2) \boldsymbol{\eta} \mapsto\rho(g)\boldsymbol{\eta}. \tag*{(2)} η ↦ ρ ( g ) η . (2)
A structural tensor, magnetic vector, complex condensate, or coupled multiplet is a particular representation. In the mechanical formulations below, G 0 G_{0} G 0 is a material point group represented by orthogonal spatial matrices, and g g g denotes that spatial action in expressions such as g E g T g\mathbf{E}g^{T} g E g T and g n ^ g\hat{\mathbf{n}} g n ^ . An abstract internal symmetry group requires a separately specified action on the mechanical variables; its representation on V \mathcal{V} V alone does not define these transformations. Spatial transformation rules for ∇ η \nabla\boldsymbol{\eta} ∇ η must rotate both the order-parameter indices and the spatial derivative index. In a material description, a superposed spatial observer rotation leaves referential structural coordinates unchanged; frame indifference and parent-group covariance are distinct requirements.
Throughout, ρ ( g 1 g 2 ) = ρ ( g 1 ) ρ ( g 2 ) \rho(g_{1}g_{2})=\rho(g_{1})\rho(g_{2}) ρ ( g 1 g 2 ) = ρ ( g 1 ) ρ ( g 2 ) is a left action. For a second-order tensor we accordingly use the active material action Q ↦ g Q g T \mathbf{Q}\mapsto g\mathbf{Q}g^{T} Q ↦ g Q g T and E ↦ g E g T \mathbf{E}\mapsto g\mathbf{E}g^{T} E ↦ g E g T . This convention makes successive rotations consistent with the conjugation rule for stabilizers used below.
Define the stabilizer and group orbit by
H η = Stab G 0 ( η ) = { g ∈ G 0 : ρ ( g ) η = η } , (3) H_{\eta}=\operatorname{Stab}_{G_{0}}(\eta)=\{g\in G_{0}:\rho(g)\eta=\eta\}, \tag*{(3)} H η = Stab G 0 ( η ) = { g ∈ G 0 : ρ ( g ) η = η } , (3)
O η = { ρ ( g ) η : g ∈ G 0 } ≃ G 0 / H η . (4) \mathcal{O}_{\eta}=\{\rho(g)\eta:g\in G_{0}\}\simeq G_{0}/H_{\eta}. \tag*{(4)} O η = { ρ ( g ) η : g ∈ G 0 } ≃ G 0 / H η . (4)
Stabilizers conjugated by the parent group correspond to the same intrinsic symmetry type. The space decomposes into orbit types,
V = ⋃ [ H ] V ( H ) , V ( H ) = { η : H η is conjugate to H } . (5) \mathcal{V}=\bigcup_{[H]}\mathcal{V}_{(H)},\qquad\mathcal{V}_{(H)}=\{\eta:H_{\eta}\text{ is conjugate to }H\}. \tag*{(5)} V = [ H ] ⋃ V ( H ) , V ( H ) = { η : H η is conjugate to H } . (5)
The orbit space V / G 0 \mathcal{V}/G_{0} V / G 0 encodes intrinsic amplitudes and invariant shape variables but suppresses the material orientation of an orbit. The symmetry class is unchanged under a parent-group reorientation: if η 2 = ρ ( g ) η 1 \eta_{2}=\rho(g)\eta_{1} η 2 = ρ ( g ) η 1 , then H η 2 = g H η 1 g − 1 H_{\eta_{2}}=gH_{\eta_{1}}g^{-1} H η 2 = g H η 1 g − 1 . Thus a state path contained in one orbit cannot change intrinsic symmetry type. For any h ∈ H η 1 h\in H_{\eta_{1}} h ∈ H η 1 , ρ ( g h g − 1 ) η 2 = ρ ( g ) ρ ( h ) η 1 = η 2 \rho(ghg^{-1})\eta_{2}=\rho(g)\rho(h)\eta_{1}=\eta_{2} ρ ( g h g − 1 ) η 2 = ρ ( g ) ρ ( h ) η 1 = η 2 . Conversely, conjugation by g − 1 g^{-1} g − 1 maps any stabilizer of η 2 \eta_{2} η 2 into the stabilizer of η 1 \eta_{1} η 1 . Consequently three paths must be distinguished: (1) orbit motion: η ( t ) = ρ ( g ( t ) ) η ∗ \eta(t)=\rho(g(t))\eta_{*} η ( t ) = ρ ( g ( t )) η ∗ , (2) within-stratum deformation: [ H η ( t ) ] = [ H ] [H_{\eta(t)}]=[H] [ H η ( t ) ] = [ H ] , η ( t ) ∉ O η ∗ \eta(t)\notin\mathcal{O}_{\eta_{*}} η ( t ) ∈ / O η ∗ , (3) symmetry-class change: [ H η ( t − ) ] ≠ [ H η ( t + ) ] [H_{\eta(t^{-})}]\ne[H_{\eta(t^{+})}] [ H η ( t − ) ] = [ H η ( t + ) ] . The group label can change abruptly when eigenvalues meet even if η ( t ) \eta(t) η ( t ) and all smooth constitutive functions remain continuous. It follows that an orbit-type crossing is not, by itself, a thermodynamic singularity.
If separate variables η a \eta_{a} η a carry independent structural information, the stabilizer of their direct sum is
H η = ⋂ a H η a . (6) H_{\eta}=\bigcap_{a}H_{\eta_{a}}. \tag*{(6)} H η = a ⋂ H η a . (6)
The instantaneous symmetry of a mechanically loaded state can be even smaller, because the strain, prestress, gradients, and boundary conditions must also be preserved. For a bulk state with strain E \mathbf{E} E , for example,
H l o a d e d = H η ∩ H E ∩ H ξ , (7) H_{\mathrm{loaded}}=H_{\eta}\cap H_{\mathbf{E}}\cap H_{\xi}, \tag*{(7)} H loaded = H η ∩ H E ∩ H ξ , (7)
when ξ \xi ξ denotes all other symmetry-bearing internal variables. Equations (3) –(7) should not be confused with an assertion that a specimen has the same global symmetry as one of its local material points.
For the isotropic parent G 0 = S O ( 3 ) G_{0}=SO(3) G 0 = S O ( 3 ) , a transparent specialization is
Q ∈ Sym 0 ( 3 ) = { A ∈ R 3 × 3 : A T = A , tr A = 0 } , Q ↦ g Q g T . (8) \mathbf{Q}\in\operatorname{Sym}_{0}(3)=\{\mathbf{A}\in\mathbb{R}^{3\times3}:\mathbf{A}^{T}=\mathbf{A},\ \operatorname{tr}\mathbf{A}=0\},\qquad\mathbf{Q}\mapsto g\mathbf{Q}g^{T}. \tag*{(8)} Q ∈ Sym 0 ( 3 ) = { A ∈ R 3 × 3 : A T = A , tr A = 0 } , Q ↦ g Q g T . (8)
With
J 2 = tr Q 2 , J 3 = tr Q 3 , 6 J 3 2 ≤ J 2 3 , (9) J_{2}=\operatorname{tr}\mathbf{Q}^{2},\qquad J_{3}=\operatorname{tr}\mathbf{Q}^{3},\qquad6J_{3}^{2}\leq J_{2}^{3}, \tag*{(9)} J 2 = tr Q 2 , J 3 = tr Q 3 , 6 J 3 2 ≤ J 2 3 , (9)
its eigenvalue multiplicities lead to the well-known structure
Q = 0 : S O ( 3 ) , Q = s ( n ⊗ n − I / 3 ) , s ≠ 0 : O ( 2 ) , q i ≠ q j for all i ≠ j : D 2 . (10) \begin{aligned}
\mathbf{Q}=0 &: \ SO(3),\\
\mathbf{Q}=s(\mathbf{n}\otimes\mathbf{n}-\mathbf{I}/3),\quad s\ne0 &: \ O(2),\\
q_{i}\ne q_{j}\quad\text{for all }i\ne j &: \ D_{2}. \tag*{(10)}
\end{aligned} Q = 0 Q = s ( n ⊗ n − I /3 ) , s = 0 q i = q j for all i = j : S O ( 3 ) , : O ( 2 ) , : D 2 . (10)
Here O ( 2 ) O(2) O ( 2 ) denotes the appropriate embedded subgroup of proper three-dimensional rotations preserving an unoriented axis. The quantity
β Q 2 = 1 − 6 J 3 2 J 2 3 , J 2 > 0 , (11) \beta_{Q}^{2}=1-\frac{6J_{3}^{2}}{J_{2}^{3}},\qquad J_{2}>0, \tag*{(11)} β Q 2 = 1 − J 2 3 6 J 3 2 , J 2 > 0 , (11)
measures biaxiality but is undefined at the isotropic origin. This representation does not resolve cubic, chiral, and other symmetry classes which require different or higher-order descriptors. That limitation is a principal motivation for Eq. (1) .
The constraint in Eq. (9) follows directly from the characteristic polynomial, rather than from an additional constitutive assumption. Since tr Q = 0 \operatorname{tr}\mathbf{Q}=0 tr Q = 0 , the Cayley–Hamilton identity is
Q 3 − J 2 2 Q − J 3 3 I = 0 , ∏ i < j ( q i − q j ) 2 = 1 2 ( J 2 3 − 6 J 3 2 ) ≥ 0. (12) \mathbf{Q}^{3}-\frac{J_{2}}{2}\mathbf{Q}-\frac{J_{3}}{3}\mathbf{I}=0,\qquad\prod_{i<j}(q_i-q_j)^2=\frac{1}{2}\left(J_2^3-6J_3^2\right)\geq0. \tag*{(12)} Q 3 − 2 J 2 Q − 3 J 3 I = 0 , i < j ∏ ( q i − q j ) 2 = 2 1 ( J 2 3 − 6 J 3 2 ) ≥ 0. (12)
The second identity is the discriminant of x 3 − J 2 x / 2 − J 3 / 3 x^3-J_2x/2-J_3/3 x 3 − J 2 x /2 − J 3 /3 . Equality corresponds to a repeated eigenvalue; for a nonzero traceless tensor this is the uniaxial case. Three distinct eigenvalues leave only the proper rotations preserving the three principal axes, giving D 2 D_2 D 2 . At the origin every rotation is admissible. This also explains why biaxiality loses its definition as J 2 → 0 J_2\to0 J 2 → 0 without a prescribed approach direction [37 ] .
Free energy and equilibrium phase transitions
Consider a small-strain representative for which a general local functional can be written
F [ η , u ; T , λ ] = ∫ Ω { f 0 ( η ; T , λ ) + 1 2 K a b i j ( η ) ∂ i η a ∂ j η b + W ( E ( u ) , η ; T , λ ) } d V − W e x t , (13) \mathcal{F}[\boldsymbol{\eta},\mathbf{u};T,\lambda]=\int_{\Omega}\left\{f_0(\boldsymbol{\eta};T,\lambda)+\frac{1}{2}K_{ab}^{ij}(\boldsymbol{\eta})\partial_i\eta_a\partial_j\eta_b+W(\mathbf{E}(\mathbf{u}),\boldsymbol{\eta};T,\lambda)\right\}\,\mathrm{d}V-\mathcal{W}_{\mathrm{ext}}, \tag*{(13)} F [ η , u ; T , λ ] = ∫ Ω { f 0 ( η ; T , λ ) + 2 1 K ab ij ( η ) ∂ i η a ∂ j η b + W ( E ( u ) , η ; T , λ ) } d V − W ext , (13)
where E = sym ∇ u \mathbf{E}=\operatorname{sym}\nabla\mathbf{u} E = sym ∇ u and λ \lambda λ is an externally controlled parameter (temperature, pressure, applied field, or loading). The local coefficients are subject to the parent covariance condition
f 0 ( ρ ( g ) η ; T , λ ) = f 0 ( η ; T , λ ) , W ( g E g T , ρ ( g ) η ; T , λ ) = W ( E , η ; T , λ ) , (14) f_0(\rho(g)\boldsymbol{\eta};T,\lambda)=f_0(\boldsymbol{\eta};T,\lambda),\qquad W(g\mathbf{E}g^{\mathsf{T}},\rho(g)\boldsymbol{\eta};T,\lambda)=W(\mathbf{E},\boldsymbol{\eta};T,\lambda), \tag*{(14)} f 0 ( ρ ( g ) η ; T , λ ) = f 0 ( η ; T , λ ) , W ( g E g T , ρ ( g ) η ; T , λ ) = W ( E , η ; T , λ ) , (14)
with the corresponding transformed-gradient condition.
Here a control held fixed in a parent-invariance statement must itself preserve the parent symmetry, as temperature or a scalar coefficient does. An oriented applied field or stress instead transforms jointly with the material variables; if it is held fixed, the relevant invariance group is its residual subgroup. This avoids attributing full-parent invariance to an explicitly biased specimen. Terms linear in strain and terms quadratic in strain describe different physics:
W ( E , η ) = 1 2 ( E − E ∗ ( η ) ) : C ( η ) : ( E − E ∗ ( η ) ) + W h i g h e r . (15) W(\mathbf{E},\boldsymbol{\eta})=\frac{1}{2}\left(\mathbf{E}-\mathbf{E}^{*}(\boldsymbol{\eta})\right):\mathbb{C}(\boldsymbol{\eta}):\left(\mathbf{E}-\mathbf{E}^{*}(\boldsymbol{\eta})\right)+W_{\mathrm{higher}}. \tag*{(15)} W ( E , η ) = 2 1 ( E − E ∗ ( η ) ) : C ( η ) : ( E − E ∗ ( η ) ) + W higher . (15)
The transformation strain E ∗ \mathbf{E}^{*} E ∗ can produce ferroelastic variants; C ( η ) \mathbb{C}(\boldsymbol{\eta}) C ( η ) controls incremental elasticity. In applications with finite strain one instead constructs a frame-indifferent energy Ψ ( C , η , ∇ η ) \Psi(\mathbf{C},\boldsymbol{\eta},\nabla\boldsymbol{\eta}) Ψ ( C , η , ∇ η ) with C = F T F \mathbf{C}=\mathbf{F}^{\mathsf{T}}\mathbf{F} C = F T F and computes the relevant push-forward tangent.
To identify the field equations, write f f f for the complete local integrand in Eq. (13) , assume K a b i j = K b a j i K_{ab}^{ij}=K_{ba}^{ji} K ab ij = K ba j i , and set π a i = ∂ f / ∂ ( ∂ i η a ) \pi_{ai}=\partial f/\partial(\partial_i\eta_a) π ai = ∂ f / ∂ ( ∂ i η a ) . Integration by parts gives the first variation
δ F = ∫ Ω [ ( ∂ η a f − ∂ i π a i ) δ η a + σ : sym ∇ δ u ] d V + ∫ ∂ Ω n i π a i δ η a d S − δ W e x t , σ = ∂ E W . (16) \delta\mathcal{F}=\int_{\Omega}\left[(\partial_{\eta_a}f-\partial_i\pi_{ai})\delta\eta_a+\boldsymbol{\sigma}:\operatorname{sym}\nabla\delta\mathbf{u}\right]\,\mathrm{d}V+\int_{\partial\Omega}n_i\pi_{ai}\delta\eta_a\,\mathrm{d}S-\delta\mathcal{W}_{\mathrm{ext}},\qquad\boldsymbol{\sigma}=\partial_{\mathbf{E}}W. \tag*{(16)} δ F = ∫ Ω [ ( ∂ η a f − ∂ i π ai ) δ η a + σ : sym ∇ δ u ] d V + ∫ ∂ Ω n i π ai δ η a d S − δ W ext , σ = ∂ E W . (16)
Thus the natural structural boundary condition is n i π a i = 0 n_i\pi_{ai}=0 n i π ai = 0 when no structural surface power is prescribed. For a body force b \mathbf{b} b and an applied traction t \mathbf{t} t , mechanical stationarity gives ∇ ⋅ σ + b = 0 \nabla\cdot\boldsymbol{\sigma}+\mathbf{b}=0 ∇ ⋅ σ + b = 0 and σ n = t \boldsymbol{\sigma}\mathbf{n}=\mathbf{t} σ n = t on the traction boundary. These are the usual variational and microforce ingredients of gradient theories [9 , 10 , 38 ] . In particular, a state-dependent gradient stiffness contributes
δ F δ η c = ∂ η c ( f 0 + W ) + 1 2 ∂ η c K a b i j ∂ i η a ∂ j η b − ∂ i ( K c b i j ∂ j η b ) . (17) \frac{\delta\mathcal{F}}{\delta\eta_c}=\partial_{\eta_c}(f_0+W)+\frac{1}{2}\partial_{\eta_c}K_{ab}^{ij}\partial_i\eta_a\partial_j\eta_b-\partial_i(K_{cb}^{ij}\partial_j\eta_b). \tag*{(17)} δ η c δ F = ∂ η c ( f 0 + W ) + 2 1 ∂ η c K ab ij ∂ i η a ∂ j η b − ∂ i ( K c b ij ∂ j η b ) . (17)
The derivative of W W W is taken at fixed strain. If the displacement is first minimized, the derivative of the reduced functional is evaluated at that mechanical solution; its second derivative generally contains an additional relaxation correction.
The mean-field equilibrium prediction is given by
F M F ( T , λ ) = inf η , u F [ η , u ; T , λ ] . (18) F_{\mathrm{MF}}(T,\lambda)=\inf_{\eta,\mathbf{u}}\mathcal{F}[\eta,\mathbf{u};T,\lambda]. \tag*{(18)} F MF ( T , λ ) = η , u inf F [ η , u ; T , λ ] . (18)
Comparing the globally minimizing branches supplies mean-field transition and coexistence points. Metastability requires retaining the other local minima and their Hessians, which are not recorded by the infimum alone. At nonzero temperature, the equilibrium partition function of a specified finite, regularized statistical model can be written schematically as
Z L ( T , λ , h ) = ∫ D η D u exp [ − F L [ η , u ; T , λ ] − h ⋅ ∫ Ω L η d V k B T ] . (19) Z_L(T,\lambda,h)=\int\mathcal{D}\eta\,\mathcal{D}\mathbf{u}\,\exp\left[-\frac{\mathcal{F}_L[\eta,\mathbf{u};T,\lambda]-h\cdot\int_{\Omega_L}\eta\,\mathrm{d}V}{k_{\mathrm{B}}T}\right]. \tag*{(19)} Z L ( T , λ , h ) = ∫ D η D u exp [ − k B T F L [ η , u ; T , λ ] − h ⋅ ∫ Ω L η d V ] . (19)
A spatial discretization or short-distance cutoff, together with the mechanical ensemble and treatment of rigid displacement modes, is implicit in this measure. The functional in the Boltzmann weight must be a Hamiltonian at that specified coarse-graining scale. An already fully fluctuation-renormalized thermodynamic potential cannot be used as a new bare Hamiltonian without controlling the resulting double counting of fluctuations [39 ] . The thermodynamic free-energy density is defined by
f e q ( T , λ , h ) = − lim L → ∞ k B T ∣ Ω L ∣ log Z L ( T , λ , h ) , (20) f_{\mathrm{eq}}(T,\lambda,h)=-\lim_{L\to\infty}\frac{k_{\mathrm{B}}T}{|\Omega_L|}\log Z_L(T,\lambda,h), \tag*{(20)} f eq ( T , λ , h ) = − L → ∞ lim ∣ Ω L ∣ k B T log Z L ( T , λ , h ) , (20)
when the limit exists. Genuine equilibrium phase transitions correspond to singularities or nonanalyticities of the thermodynamic potential or its appropriate derivatives; a sharp singularity generally emerges only in an infinite-system limit. Spontaneous symmetry breaking is obtained, where appropriate, using a prescribed order of limits, for instance
η s p = lim h → 0 + lim L → ∞ ⟨ 1 ∣ Ω L ∣ ∫ Ω L η d V ⟩ L , h . (21) \eta_{\mathrm{sp}}=\lim_{h\to0^+}\lim_{L\to\infty}\left\langle\frac{1}{|\Omega_L|}\int_{\Omega_L}\eta\,\mathrm{d}V\right\rangle_{L,h}. \tag*{(21)} η sp = h → 0 + lim L → ∞ lim ⟨ ∣ Ω L ∣ 1 ∫ Ω L η d V ⟩ L , h . (21)
The small conjugate field selects a branch and prevents a finite-volume symmetric mixture from artificially averaging a broken-symmetry order parameter to zero. Changes in the stabilizer of the structural state and thermodynamic phase transitions are distinct phenomena, and neither necessarily implies the other. A change in structural symmetry does not, by itself, establish a nonanalyticity in the thermodynamic potential. Conversely, a thermodynamic phase transition may occur without a change in the stabilizer of the structural state. For the first implication, a smooth externally prescribed path η ( t ) = ( t , 0 ) \eta(t)=(t,0) η ( t ) = ( t , 0 ) for a vector representation has an enhanced stabilizer at t = 0 t=0 t = 0 , whereas a smooth analytic energy evaluated on that prescribed path need not possess a thermodynamic singularity. The second implication fails for a liquid–gas first-order transition, whose two homogeneous phases can possess the same spatial rotational symmetry and differ by scalar density. Therefore the two criteria cannot be identified.
A complete characterization of an equilibrium branch therefore requires more than its structural symmetry. In addition to the stabilizer of the order parameter, the associated free-energy branch, coexistence and stability conditions, and relevant material response must be considered. These quantities collectively distinguish thermodynamically different states that may share the same residual symmetry. A thermodynamic phase and an orbit-type stratum are thus distinct mathematical objects whose relationship is physically informative.
Let I 1 ( η ) , … , I r ( η ) I_1(\eta),\ldots,I_r(\eta) I 1 ( η ) , … , I r ( η ) denote a symmetry-admissible collection of scalar invariants. A smooth parent-invariant polynomial potential can be written to specified order as
f 0 ( η ; λ ) = a 0 ( λ ) + ∑ i a i ( λ ) I i ( η ) + ∑ i j a i j ( λ ) I i ( η ) I j ( η ) + ⋯ . (22) f_0(\eta;\lambda)=a_0(\lambda)+\sum_i a_i(\lambda)I_i(\eta)+\sum_{ij}a_{ij}(\lambda)I_i(\eta)I_j(\eta)+\cdots. \tag*{(22)} f 0 ( η ; λ ) = a 0 ( λ ) + i ∑ a i ( λ ) I i ( η ) + ij ∑ a ij ( λ ) I i ( η ) I j ( η ) + ⋯ . (22)
An arbitrarily selected finite list is not claimed to be a complete integrity basis. Cubic invariants can permit first-order transitions, while a vanishing cubic term often allows a continuous mean-field onset. Fluctuations, strain accommodation, or competing branches can change the order of the actual transition.
Dissipative evolution and stability
For fixed mechanical and thermal controls, define the functional force
A η = − δ F δ η = − ∂ f ∂ η + ∂ i ( ∂ f ∂ ( ∂ i η ) ) . (23) A_\eta=-\frac{\delta\mathcal{F}}{\delta\eta}=-\frac{\partial f}{\partial\eta}+\partial_i\left(\frac{\partial f}{\partial(\partial_i\eta)}\right). \tag*{(23)} A η = − δ η δ F = − ∂ η ∂ f + ∂ i ( ∂ ( ∂ i η ) ∂ f ) . (23)
For example, for constant scalar κ \kappa κ and f = f l o c + κ ∣ ∇ η ∣ 2 / 2 f=f_{\mathrm{loc}}+\kappa|\nabla\boldsymbol{\eta}|^2/2 f = f loc + κ ∣∇ η ∣ 2 /2 , this reduces to A η = − ∂ η f l o c + κ ∇ 2 η \mathbf{A}_{\eta}=-\partial_{\eta}f_{\mathrm{loc}}+\kappa\nabla^2\boldsymbol{\eta} A η = − ∂ η f loc + κ ∇ 2 η . At fixed controls and mechanical equilibrium, the chain rule gives F ˙ = − ∫ Ω A η ⋅ η ˙ d V \dot{\mathcal{F}}=-\int_{\Omega}\mathbf{A}_{\eta}\cdot\dot{\boldsymbol{\eta}}\,\mathrm{d}V F ˙ = − ∫ Ω A η ⋅ η ˙ d V after the structural boundary term has vanished. This identifies the force–rate pair used to construct the dissipative law. Suppose the allowed rates lie in a linear constrained tangent space T η M T_{\eta}\mathcal{M} T η M , with orthogonal projection P η \mathbf{P}_{\eta} P η . For purely relaxational nonconserved kinetics consider
∂ t η = P η M η P η A η , M η = M η T ⪰ 0. (24) \partial_t\boldsymbol{\eta}=\mathbf{P}_{\eta}\mathbf{M}_{\eta}\mathbf{P}_{\eta}\mathbf{A}_{\eta},\qquad\mathbf{M}_{\eta}=\mathbf{M}_{\eta}^{\mathsf{T}}\succeq0. \tag*{(24)} ∂ t η = P η M η P η A η , M η = M η T ⪰ 0. (24)
For a positive definite mobility, this law can equivalently be obtained by minimizing the Rayleigh functional over admissible rates:
R [ v ] = ∫ Ω [ 1 2 v ⋅ M η − 1 v − A η ⋅ v ] d V , v ∈ T η M , M T = ( P η M η P η ) ∣ T η M . (25) \mathcal{R}[\mathbf{v}]=\int_{\Omega}\left[\frac{1}{2}\mathbf{v}\cdot\mathbf{M}_{\eta}^{-1}\mathbf{v}-\mathbf{A}_{\eta}\cdot\mathbf{v}\right]\mathrm{d}V,\qquad\mathbf{v}\in T_{\eta}\mathcal{M},\qquad\mathbf{M}_{T}=(\mathbf{P}_{\eta}\mathbf{M}_{\eta}\mathbf{P}_{\eta})|_{T_{\eta}\mathcal{M}}. \tag*{(25)} R [ v ] = ∫ Ω [ 2 1 v ⋅ M η − 1 v − A η ⋅ v ] d V , v ∈ T η M , M T = ( P η M η P η ) ∣ T η M . (25)
Stationarity yields v = M T P η A η \mathbf{v}=\mathbf{M}_{T}\mathbf{P}_{\eta}\mathbf{A}_{\eta} v = M T P η A η . The semidefinite case restricts rates to the range of M T \mathbf{M}_{T} M T . Using the orthogonality of P η \mathbf{P}_{\eta} P η in the chain rule then gives Eq. (26) . The mobility must transform covariantly under the parent group. For homogeneous thermal and mechanical controls with vanishing boundary structural power,
d F d t = − ∫ Ω ( P η A η ) ⋅ M η ( P η A η ) d V ≤ 0. (26) \frac{\mathrm{d}\mathcal{F}}{\mathrm{d}t}=-\int_{\Omega}(\mathbf{P}_{\eta}\mathbf{A}_{\eta})\cdot\mathbf{M}_{\eta}(\mathbf{P}_{\eta}\mathbf{A}_{\eta})\,\mathrm{d}V\leq0. \tag*{(26)} d t d F = − ∫ Ω ( P η A η ) ⋅ M η ( P η A η ) d V ≤ 0. (26)
For Q ∈ Sym 0 ( 3 ) \mathbf{Q}\in\operatorname{Sym}_{0}(3) Q ∈ Sym 0 ( 3 ) the projection is P 0 A = sym A − tr ( A ) I / 3 \mathbf{P}_{0}\mathbf{A}=\operatorname{sym}\mathbf{A}-\operatorname{tr}(\mathbf{A})\mathbf{I}/3 P 0 A = sym A − tr ( A ) I /3 . If external strain, stress, field, or temperature is driven, the total free energy can increase by external work; Eq. (26) must not be used to assert monotonic energy decrease for the driven problem.
A conserved scalar density n n n cannot obey an unconstrained Model-A-type law. With chemical potential μ n = δ F / δ n \mu_n=\delta\mathcal{F}/\delta n μ n = δ F / δ n , a standard isothermal choice is
∂ t n = ∇ ⋅ ( M n ∇ μ n ) , M n ≥ 0. (27) \partial_t n=\nabla\cdot\left(M_n\nabla\mu_n\right),\qquad M_n\geq0. \tag*{(27)} ∂ t n = ∇ ⋅ ( M n ∇ μ n ) , M n ≥ 0. (27)
under no-flux boundary conditions. The mass is preserved and
d F d t = − ∫ Ω M n ∣ ∇ μ n ∣ 2 d V ≤ 0. (28) \frac{\mathrm{d}\mathcal{F}}{\mathrm{d}t}=-\int_{\Omega}M_n|\nabla\mu_n|^2\,\mathrm{d}V\leq0. \tag*{(28)} d t d F = − ∫ Ω M n ∣∇ μ n ∣ 2 d V ≤ 0. (28)
A theory with several slow modes may require a block kinetic operator, reversible Poisson-bracket couplings, momentum transport, and thermal transport. Parent symmetry constrains these blocks but does not fix their magnitudes or the dynamic universality class. This follows the Model-A/Model-B distinction and its generalizations [19 ] .
For a constant mobility in an unconstrained nonconserved order-parameter space, a canonical stochastic closure is
∂ t η = − M η δ F δ η + ζ , (29) \partial_t\boldsymbol{\eta}=-\mathbf{M}_{\eta}\frac{\delta\mathcal{F}}{\delta\boldsymbol{\eta}}+\boldsymbol{\zeta}, \tag*{(29)} ∂ t η = − M η δ η δ F + ζ , (29)
⟨ ζ a ( X , t ) ⟩ = 0 , ⟨ ζ a ( X , t ) ζ b ( X ′ , t ′ ) ⟩ = 2 k B T ( M η ) a b δ ( X − X ′ ) δ ( t − t ′ ) . (30) \langle\zeta_a(\mathbf{X},t)\rangle=0,\qquad\langle\zeta_a(\mathbf{X},t)\zeta_b(\mathbf{X}',t')\rangle=2k_{\mathrm{B}}T(\mathbf{M}_{\eta})_{ab}\delta(\mathbf{X}-\mathbf{X}')\delta(t-t'). \tag*{(30)} ⟨ ζ a ( X , t )⟩ = 0 , ⟨ ζ a ( X , t ) ζ b ( X ′ , t ′ )⟩ = 2 k B T ( M η ) ab δ ( X − X ′ ) δ ( t − t ′ ) . (30)
The factor 2 k B T 2k_{\mathrm{B}}T 2 k B T is fixed by detailed balance. On a finite discretization with constant symmetric mobility, the probability density satisfies the Fokker–Planck equation
∂ t P = ∇ η ⋅ { M η [ ( ∇ η F ) P + k B T ∇ η P ] } . (31) \partial_t\mathcal{P}=\nabla_{\eta}\cdot\left\{\mathbf{M}_{\eta}\left[(\nabla_{\eta}\mathcal{F})\mathcal{P}+k_{\mathrm{B}}T\nabla_{\eta}\mathcal{P}\right]\right\}. \tag*{(31)} ∂ t P = ∇ η ⋅ { M η [ ( ∇ η F ) P + k B T ∇ η P ] } . (31)
For P e q ∝ exp [ − F / ( k B T ) ] \mathcal{P}_{\mathrm{eq}}\propto\exp[-\mathcal{F}/(k_{\mathrm{B}}T)] P eq ∝ exp [ − F / ( k B T )] , the quantity in square brackets vanishes identically. This establishes the stated equilibrium measure for additive noise, rather than simply appending a noise term to deterministic kinetics (Ohzeki, 2026). For state-dependent mobility, the stochastic calculus and corresponding drift must be specified to obtain the intended equilibrium measure. For a conserved scalar, noise must enter as a flux divergence:
∂ t n = ∇ ⋅ ( M n ∇ μ n + ξ ) , (32) \partial_t n=\nabla\cdot\left(M_n\nabla\mu_n+\boldsymbol{\xi}\right), \tag*{(32)} ∂ t n = ∇ ⋅ ( M n ∇ μ n + ξ ) , (32)
⟨ ξ i ( X , t ) ξ j ( X ′ , t ′ ) ⟩ = 2 k B T M n δ i j δ ( X − X ′ ) δ ( t − t ′ ) . (33) \langle\xi_i(\mathbf{X},t)\xi_j(\mathbf{X}',t')\rangle=2k_{\mathrm{B}}TM_n\delta_{ij}\delta(\mathbf{X}-\mathbf{X}')\delta(t-t'). \tag*{(33)} ⟨ ξ i ( X , t ) ξ j ( X ′ , t ′ )⟩ = 2 k B T M n δ ij δ ( X − X ′ ) δ ( t − t ′ ) . (33)
These expressions illustrate the required statistical extension. Additional hydrodynamic or inertial variables demand corresponding noise and reversible couplings.
Let η ∗ \eta_* η ∗ be a homogeneous stationary state and let H η \mathcal{H}_{\eta} H η denote the Hessian of its local potential on the admissible state space. With scalar mobility Γ \Gamma Γ and isotropic gradient coefficient κ \kappa κ , an eigenmode v a \mathbf{v}_a v a of H η \mathcal{H}_{\eta} H η with curvature h a h_a h a has Fourier growth exponent
δ η = v a e i q ⋅ X + ν a ( q ) t , ν a ( q ) = − Γ ( h a + κ ∣ q ∣ 2 ) . (34) \delta\eta=\mathbf{v}_a e^{i\mathbf{q}\cdot\mathbf{X}+\nu_a(\mathbf{q})t},\qquad\nu_a(\mathbf{q})=-\Gamma(h_a+\kappa\lvert\mathbf{q}\rvert^2). \tag*{(34)} δ η = v a e i q ⋅ X + ν a ( q ) t , ν a ( q ) = − Γ ( h a + κ ∣ q ∣ 2 ) . (34)
For h a > 0 h_a>0 h a > 0 , τ a ( q ) = [ Γ ( h a + κ q 2 ) ] − 1 \tau_a(\mathbf{q})=[\Gamma(h_a+\kappa q^2)]^{-1} τ a ( q ) = [ Γ ( h a + κ q 2 ) ] − 1 ; for h a < 0 h_a<0 h a < 0 , sufficiently long wavelengths are linearly unstable. An orientational Goldstone direction of an unbiased continuous orbit has zero homogeneous curvature, not the same positive curvature as an amplitude mode. For conserved dynamics, the analogous simple scalar linearization gives ν = − M n q 2 ( h n + κ q 2 ) \nu=-M_nq^2(h_n+\kappa q^2) ν = − M n q 2 ( h n + κ q 2 ) .
Indeed, the linearized nonconserved equation is ∂ t δ η = − Γ ( H η − κ ∇ 2 ) δ η \partial_t\delta\eta=-\Gamma(\mathcal{H}_{\eta}-\kappa\nabla^2)\delta\eta ∂ t δ η = − Γ ( H η − κ ∇ 2 ) δ η ; inserting the Fourier ansatz yields Eq. (34) . Here H η \mathcal{H}_{\eta} H η is evaluated under the specified mechanical constraint. If strain is relaxed, the appropriate Hessian is that of the mechanically reduced energy and may be nonlocal and direction dependent. In that case h a + κ q 2 h_a+\kappa q^2 h a + κ q 2 must be replaced by the eigenvalues of the complete Fourier Hessian.
For a deterministic equivariant ODE/PDE with uniqueness and exactly parent-symmetric initial state and controls, the solution retains that symmetry. For example, if η = 0 \eta=0 η = 0 is stationary for all imposed controls, then η ( 0 ) = 0 \eta(0)=0 η ( 0 ) = 0 implies η ( t ) = 0 \eta(t)=0 η ( t ) = 0 in a homogeneous polynomial gradient flow, even after that stationary state becomes linearly unstable. Parent symmetry by itself need not make the origin stationary when the representation contains invariant scalar components. Observed ordering requires fluctuations, a seed, boundaries, external bias, or a limiting procedure selecting a branch.
Critical scaling and fluctuations
Near a continuous equilibrium transition, define the reduced distance t = ( λ − λ c ) / λ 0 t=(\lambda-\lambda_c)/\lambda_0 t = ( λ − λ c ) / λ 0 , with a nonzero fixed scale λ 0 \lambda_0 λ 0 . If an ordered branch possesses a local order parameter, write the asymptotic laws
∥ η s p ∥ ∼ B ∣ t ∣ β , ξ ∼ ξ 0 ∣ t ∣ − ν , χ ∼ χ 0 ∣ t ∣ − γ , τ ∼ τ 0 ( ξ / ξ 0 ) z , (35) \lVert\eta_{\mathrm{sp}}\rVert\sim B\lvert t\rvert^\beta,\qquad\xi\sim\xi_0\lvert t\rvert^{-\nu},\qquad\chi\sim\chi_0\lvert t\rvert^{-\gamma},\qquad\tau\sim\tau_0(\xi/\xi_0)^z, \tag*{(35)} ∥ η sp ∥ ∼ B ∣ t ∣ β , ξ ∼ ξ 0 ∣ t ∣ − ν , χ ∼ χ 0 ∣ t ∣ − γ , τ ∼ τ 0 ( ξ / ξ 0 ) z , (35)
where the last relation pertains to the relevant slow dynamic sector. The exponents need not take their mean-field values. They depend on spatial dimension, range of interactions, symmetry and number of active components, conserved quantities, and other couplings. A symmetry stabilizer does not uniquely determine an equilibrium or dynamic universality class.
The distinction between symmetry and critical scaling is already visible in the Gaussian approximation. For one amplitude mode with curvature h > 0 h>0 h > 0 , write the quadratic energy and source as
F ( 2 ) = 1 2 ∫ d d q ( 2 π ) d ( h + κ q 2 ) ∣ φ ^ ( q ) ∣ 2 − ∫ j φ d V . (36) \mathcal{F}^{(2)}=\frac{1}{2}\int\frac{\mathrm{d}^d q}{(2\pi)^d}(h+\kappa q^2)\lvert\hat{\varphi}(\mathbf{q})\rvert^2-\int j\varphi\,\mathrm{d}V. \tag*{(36)} F ( 2 ) = 2 1 ∫ ( 2 π ) d d d q ( h + κ q 2 ) ∣ φ ^ ( q ) ∣ 2 − ∫ j φ d V . (36)
Minimizing at fixed j j j gives φ ^ = j ^ / ( h + κ q 2 ) \hat{\varphi}=\hat{j}/(h+\kappa q^2) φ ^ = j ^ / ( h + κ q 2 ) , so χ ( 0 ) = h − 1 \chi(0)=h^{-1} χ ( 0 ) = h − 1 and ξ = ( κ / h ) 1 / 2 \xi=(\kappa/h)^{1/2} ξ = ( κ / h ) 1/2 . The same pole in Eq. (34) gives τ ( q = 0 ) = ( Γ h ) − 1 = ξ 2 / ( Γ κ ) \tau(q=0)=(\Gamma h)^{-1}=\xi^2/(\Gamma\kappa) τ ( q = 0 ) = ( Γ h ) − 1 = ξ 2 / ( Γ κ ) , hence the Gaussian Model-A value z = 2 z=2 z = 2 . For the selected scalar branch of a φ 2 / 2 + c φ 4 / 4 a\varphi^2/2+c\varphi^4/4 a φ 2 /2 + c φ 4 /4 , φ ∗ 2 = − a / c \varphi_*^2=-a/c φ ∗ 2 = − a / c below onset; if a a a varies linearly with t t t , this additionally gives β = 1 / 2 \beta=1/2 β = 1/2 . These are approximation-specific results [18 , 19 , 41 ] .
An explicit route beyond mean field is to define the dimensionless Euclidean statistical action S = F / ( k B T ) S=\mathcal{F}/(k_{\mathrm{B}}T) S = F / ( k B T ) and introduce an infrared regulator R k R_k R k . Its scale-dependent effective average action obeys the standard functional renormalization-group equation
∂ k Γ k [ ϕ ] = 1 2 Tr [ ( Γ k ( 2 ) [ ϕ ] + R k ) − 1 ∂ k R k ] , (37) \partial_k\Gamma_k[\phi]=\frac{1}{2}\operatorname{Tr}\left[(\Gamma_k^{(2)}[\phi]+R_k)^{-1}\partial_kR_k\right], \tag*{(37)} ∂ k Γ k [ ϕ ] = 2 1 Tr [ ( Γ k ( 2 ) [ ϕ ] + R k ) − 1 ∂ k R k ] , (37)
for a bosonic/classical field setting with a suitable regularization [20 ] .
For clarity, introduce Δ S k = 1 2 ϕ ⋅ R k ϕ \Delta S_k=\frac{1}{2}\phi\cdot R_k\phi Δ S k = 2 1 ϕ ⋅ R k ϕ in the regulated partition function, let W k [ J ] = log Z k [ J ] W_k[J]=\log Z_k[J] W k [ J ] = log Z k [ J ] , and define Γ k [ Φ ] = J ⋅ Φ − W k [ J ] − Δ S k [ Φ ] \Gamma_k[\Phi]=J\cdot\Phi-W_k[J]-\Delta S_k[\Phi] Γ k [ Φ ] = J ⋅ Φ − W k [ J ] − Δ S k [ Φ ] , with Φ = δ W k / δ J \Phi=\delta W_k/\delta J Φ = δ W k / δ J . The connected covariance is W k ( 2 ) = ( Γ k ( 2 ) + R k ) − 1 W_k^{(2)}=(\Gamma_k^{(2)}+R_k)^{-1} W k ( 2 ) = ( Γ k ( 2 ) + R k ) − 1 . Differentiation at fixed Φ \Phi Φ cancels the mean-field part of ∂ k Δ S k \partial_k\Delta S_k ∂ k Δ S k and leaves ∂ k Γ k = 1 2 Tr ( W k ( 2 ) ∂ k R k ) \partial_k\Gamma_k=\frac{1}{2}\operatorname{Tr}(W_k^{(2)}\partial_kR_k) ∂ k Γ k = 2 1 Tr ( W k ( 2 ) ∂ k R k ) , which is Eq. (37) . This construction specifies which fluctuations have been integrated out. The group acts consistently at every coarse-graining scale if R k R_k R k preserves its symmetry. The running effective potential and kinetic coefficients can renormalize the phase diagram and critical exponents; Eq. (37) is a formal research route, not an assertion that a specified truncation is automatically controlled. A key question for the proposed symmetry-resolved framework is which invariant representations remain active at a stable fixed point and which operators are relevant, irrelevant, or dangerously irrelevant. A proposed prediction of a universal exponent must survive this analysis or a corresponding benchmark against known universality classes.
Constitutive symmetry and acoustic signatures
We now consider small-amplitude waves about a homogeneous equilibrium state. We first hold the internal state fixed during the probe and derive the directional selection rules for the elastic tangent. The final subsection allows the internal state to relax in response to the wave.
A parent-invariant constitutive map
At a stress-free, homogeneous base state let the strain energy have incremental tangent
C ( η ) = ∂ 2 W ( E , η ) ∂ E ∂ E ∣ E = E 0 ( η ) (38) \mathbb{C}(\boldsymbol{\eta}) = \left.\frac{\partial^{2} W(\mathbf{E},\boldsymbol{\eta})}{\partial\mathbf{E}\,\partial\mathbf{E}}\right|_{\mathbf{E}=\mathbf{E}_{0}(\boldsymbol{\eta})} \tag*{(38)} C ( η ) = ∂ E ∂ E ∂ 2 W ( E , η ) E = E 0 ( η ) (38)
Assume E 0 \mathbf{E}_0 E 0 transforms covariantly under the parent group. Then C \mathbb{C} C is an equivariant map. Define its stabilizer within the parent group by
G C ( η ) = { g ∈ G 0 : g ⋅ C ( η ) = C ( η ) } . G_{\mathbb{C}(\boldsymbol{\eta})}=\{g\in G_0:g\cdot\mathbb{C}(\boldsymbol{\eta})=\mathbb{C}(\boldsymbol{\eta})\}. G C ( η ) = { g ∈ G 0 : g ⋅ C ( η ) = C ( η )} .
The action on C \mathbb{C} C is the natural fourth-order tensor action. This definition distinguishes the inherited parent-group symmetry from any additional symmetry of the elastic tensor outside G 0 G_0 G 0 .
Equivariance constrains the constitutive symmetry inherited from the structural state. For a parent-equivariant constitutive map η ↦ C ( η ) \boldsymbol{\eta}\mapsto\mathbb{C}(\boldsymbol{\eta}) η ↦ C ( η ) ,
H η ⊆ G C ( η ) . (39) H_{\boldsymbol{\eta}}\subseteq G_{\mathbb{C}(\boldsymbol{\eta})}. \tag*{(39)} H η ⊆ G C ( η ) . (39)
If the map is injective on the relevant states, then H η = G C ( η ) H_{\boldsymbol{\eta}}=G_{\mathbb{C}(\boldsymbol{\eta})} H η = G C ( η ) , assuming the same group action and no independent structural variables.
The equivariance needed here follows by differentiating the energy invariance twice. For arbitrary symmetric increments H \mathbf{H} H and K \mathbf{K} K ,
( g H g T ) : C ( ρ ( g ) η ) : ( g K g T ) = H : C ( η ) : K . (40) (g\mathbf{H}g^{T}):\mathbb{C}(\rho(g)\boldsymbol{\eta}):(g\mathbf{K}g^{T})=\mathbf{H}:\mathbb{C}(\boldsymbol{\eta}):\mathbf{K}. \tag*{(40)} ( g H g T ) : C ( ρ ( g ) η ) : ( g K g T ) = H : C ( η ) : K . (40)
Thus C i j k l ( ρ ( g ) η ) = g i a g j b g k c g l d C a b c d ( η ) C_{ijkl}(\rho(g)\boldsymbol{\eta})=g_{ia}g_{jb}g_{kc}g_{ld}C_{abcd}(\boldsymbol{\eta}) C ij k l ( ρ ( g ) η ) = g ia g j b g k c g l d C ab c d ( η ) , provided the evaluation strain also transforms covariantly. The comparison is between stabilizers inside the same parent group. An elastic tensor can possess additional symmetries outside a finite structural parent [25 , 42 ] . If h ∈ H η h\in H_{\boldsymbol{\eta}} h ∈ H η , equivariance yields h ⋅ C ( η ) = C ( ρ ( h ) η ) = C ( η ) h\cdot\mathbb{C}(\boldsymbol{\eta})=\mathbb{C}(\rho(h)\boldsymbol{\eta})=\mathbb{C}(\boldsymbol{\eta}) h ⋅ C ( η ) = C ( ρ ( h ) η ) = C ( η ) . Conversely, if h h h preserves the tangent, equivariance implies C ( ρ ( h ) η ) = C ( η ) \mathbb{C}(\rho(h)\boldsymbol{\eta})=\mathbb{C}(\boldsymbol{\eta}) C ( ρ ( h ) η ) = C ( η ) . Injectivity gives ρ ( h ) η = η \rho(h)\boldsymbol{\eta}=\boldsymbol{\eta} ρ ( h ) η = η . Equality can fail when the constitutive map does not distinguish a state from one of its parent-group transforms, for example when the relevant coupling vanishes. Even when equality holds for the complete tangent, an incomplete measurement can fail to distinguish structural states; this is a separate loss of observational identifiability. The symmetry of a tangent evaluated under general finite deformation must also account for the deformation and prestress.
Christoffel/acoustic tensor and little-group selection
For a classical local elastic solid with density ρ 0 \rho_0 ρ 0 and a frozen, homogeneous tangent C i j k l C_{ijkl} C ij k l , evaluated under the specified thermal constraint, define
A i k ( n ^ ; η ) = C i j k l ( η ) n ^ j n ^ l , A p m = ρ 0 c m 2 p m , ω m = k c m . (41) A_{ik}(\hat{\mathbf{n}};\boldsymbol{\eta})=C_{ijkl}(\boldsymbol{\eta})\hat{n}_j\hat{n}_l,\qquad A_{pm}=\rho_0 c_m^2 p_m,\qquad\omega_m=k c_m. \tag*{(41)} A ik ( n ^ ; η ) = C ij k l ( η ) n ^ j n ^ l , A p m = ρ 0 c m 2 p m , ω m = k c m . (41)
The thermal constraint can be expressed explicitly for a homogeneous Helmholtz density f ( E , η , T ) f(\mathbf{E},\boldsymbol{\eta},T) f ( E , η , T ) per reference volume, with η \boldsymbol{\eta} η frozen. Let s = − ∂ T f \mathfrak{s}=-\partial_T f s = − ∂ T f be the entropy density, c E = T ( ∂ T s ) E , η > 0 c_E=T(\partial_T\mathfrak{s})_{\mathbf{E},\boldsymbol{\eta}}>0 c E = T ( ∂ T s ) E , η > 0 , and β t h = ( ∂ T σ ) E , η \boldsymbol{\beta}_{\mathrm{th}}=(\partial_T\boldsymbol{\sigma})_{\mathbf{E},\boldsymbol{\eta}} β th = ( ∂ T σ ) E , η . The Maxwell relation gives d s = ( c E / T ) d T − β t h : d E d\mathfrak{s}=(c_E/T)dT-\boldsymbol{\beta}_{\mathrm{th}}:\mathrm{d}\mathbf{E} d s = ( c E / T ) d T − β th : d E . Eliminating d T dT d T at constant entropy then yields
C S = C T + T c E β t h ⊗ β t h . (42) \mathbb{C}^{S}=\mathbb{C}^{T}+\frac{T}{c_E}\boldsymbol{\beta}_{\mathrm{th}}\otimes\boldsymbol{\beta}_{\mathrm{th}}. \tag*{(42)} C S = C T + c E T β th ⊗ β th . (42)
The superscripts T T T and S S S denote isothermal and isentropic frozen-state tangents under the same mechanical constraint. The tensor product acts as ( β t h ⊗ β t h ) : H = β t h ( β t h : H ) (\boldsymbol{\beta}_{\mathrm{th}}\otimes\boldsymbol{\beta}_{\mathrm{th}}):\mathbf{H}=\boldsymbol{\beta}_{\mathrm{th}}(\boldsymbol{\beta}_{\mathrm{th}}:\mathbf{H}) ( β th ⊗ β th ) : H = β th ( β th : H ) . Thermal exchange during a wave period determines which constraint approximates the measurement. Symmetry-consistent thermal coefficients preserve the inherited group and its protected doublets. Applying this conversion quantitatively requires thermal properties in addition to the elastic and structural data.
To obtain this eigenproblem, insert u = p exp [ i ( k n ^ ⋅ X − ω t ) ] \mathbf{u}=\mathbf{p}\exp[i(k\hat{\mathbf{n}}\cdot\mathbf{X}-\omega t)] u = p exp [ i ( k n ^ ⋅ X − ω t )] into ρ 0 u ¨ i = ∂ j ( C i j k l E k l ) \rho_0\ddot{u}_i=\partial_j(C_{ijkl}E_{kl}) ρ 0 u ¨ i = ∂ j ( C ij k l E k l ) . The minor symmetries of C \mathbf{C} C reduce the symmetric strain gradient to k 2 C i j k l n ^ j n ^ l p k k^2C_{ijkl}\hat{n}_j\hat{n}_l p_k k 2 C ij k l n ^ j n ^ l p k , giving k 2 A p = ρ 0 ω 2 p k^2\mathbf{A}\mathbf{p}=\rho_0\omega^2\mathbf{p} k 2 Ap = ρ 0 ω 2 p . Positive acoustic eigenvalues are required for propagating elastic waves in this local nondissipative approximation [43 ] . In pre-stressed or finite-strain settings, this is replaced by the appropriate incremental acoustic tensor. The relevant symmetry for a fixed wave normal is the little group
L n ^ ( η ) = { h ∈ H η : h n ^ = n ^ } . (43) L_{\hat{\mathbf{n}}}(\boldsymbol{\eta})=\{h\in H_{\boldsymbol{\eta}}:h\hat{\mathbf{n}}=\hat{\mathbf{n}}\}. \tag*{(43)} L n ^ ( η ) = { h ∈ H η : h n ^ = n ^ } . (43)
The acoustic operator commutes with the representation of the little group on polarization space. Irreducible representations of dimension greater than one can enforce modal degeneracy; lower symmetry may permit previously degenerate modes to separate. One must check the representation, measurement direction, and perturbation: symmetry breaking does not imply that every mode splits.
More explicitly, material covariance gives A ( h n ^ ; η ) = h A ( n ^ ; η ) h T \mathbf{A}(h\hat{\mathbf{n}};\boldsymbol{\eta})=h\mathbf{A}(\hat{\mathbf{n}};\boldsymbol{\eta})h^{T} A ( h n ^ ; η ) = h A ( n ^ ; η ) h T for h ∈ H η h\in H_{\boldsymbol{\eta}} h ∈ H η . At a fixed wave normal this becomes [ A , h ] = 0 [\mathbf{A},h]=0 [ A , h ] = 0 . On a single irreducible real polarization subspace, a commuting self-adjoint operator is scalar; equivalent copies may mix, but the dimension of the irreducible factor still enforces the corresponding degeneracy. For the uniaxial axial transverse plane, rotations about the axis therefore force a scalar transverse block. This is the operator statement behind the selection rule used in § 4 .
Degenerate perturbation theory and the first visible order
Suppose at η = 0 \boldsymbol{\eta}=0 η = 0 the acoustic operator has an isolated two-dimensional degenerate eigenvalue a 0 = ρ 0 c 0 2 a_0=\rho_0c_0^2 a 0 = ρ 0 c 0 2 for fixed n ^ \hat{\mathbf{n}} n ^ . Let P P P be its orthogonal eigenspace projector, and approach the reference state along a fixed normalized direction η = ϵ v \boldsymbol{\eta}=\epsilon\mathbf{v} η = ϵ v with ϵ ↓ 0 \epsilon\downarrow0 ϵ ↓ 0 . If the acoustic operator is analytic along this path,
A ( ϵ ) = A 0 + ϵ A 1 ( v ) + ϵ 2 A 2 ( v ) + ⋯ . (44) \mathbf{A}(\epsilon)=\mathbf{A}_0+\epsilon\mathbf{A}_1(\mathbf{v})+\epsilon^2\mathbf{A}_2(\mathbf{v})+\cdots. \tag*{(44)} A ( ϵ ) = A 0 + ϵ A 1 ( v ) + ϵ 2 A 2 ( v ) + ⋯ . (44)
At first order the two eigenvalue shifts are the eigenvalues of P A 1 P P\mathbf{A}_1P P A 1 P acting on the degenerate subspace. If it is proportional to the identity, the first-order splitting vanishes. Higher orders include both direct couplings and virtual mixing with other modes.
Let P ⊥ = I − P P_{\perp}=\mathbf{I}-P P ⊥ = I − P and define the reduced resolvent on the complement by
R 0 = P ⊥ [ ( a 0 I − A 0 ) ∣ range P ⊥ ] − 1 P ⊥ . (45) R_0=P_{\perp}\left[(a_0\mathbf{I}-\mathbf{A}_0)|_{\operatorname{range}P_{\perp}}\right]^{-1}P_{\perp}. \tag*{(45)} R 0 = P ⊥ [ ( a 0 I − A 0 ) ∣ range P ⊥ ] − 1 P ⊥ . (45)
The inverse is only on that complement, where spectral isolation makes it well defined. Eliminating the complementary polarization to the required order gives the effective doublet operator
A d b l = a 0 P + ϵ B 1 + ϵ 2 B 2 + O ( ϵ 3 ) , B 1 = P A 1 P , B 2 = P A 2 P + P A 1 R 0 A 1 P . (46) \mathbf{A}_{\mathrm{dbl}}=a_0P+\epsilon\mathbf{B}_1+\epsilon^2\mathbf{B}_2+O(\epsilon^3),\qquad\mathbf{B}_1=P\mathbf{A}_1P,\quad\mathbf{B}_2=P\mathbf{A}_2P+P\mathbf{A}_1R_0\mathbf{A}_1P. \tag*{(46)} A dbl = a 0 P + ϵ B 1 + ϵ 2 B 2 + O ( ϵ 3 ) , B 1 = P A 1 P , B 2 = P A 2 P + P A 1 R 0 A 1 P . (46)
The mixing term in B 2 \mathbf{B}_2 B 2 is the usual second-order correction from perturbation theory [44 ] . If B 1 \mathbf{B}_1 B 1 is scalar and B 2 \mathbf{B}_2 B 2 is not, then p = 2 p=2 p = 2 . In an orthonormal doublet basis the gap coefficient at that order is
D = [ ( B 2 , 11 − B 2 , 22 ) 2 + 4 B 2 , 12 2 ] 1 / 2 . (47) D=\left[(B_{2,11}-B_{2,22})^2+4B_{2,12}^2\right]^{1/2}. \tag*{(47)} D = [ ( B 2 , 11 − B 2 , 22 ) 2 + 4 B 2 , 12 2 ] 1/2 . (47)
Only the traceless part of the first nonscalar effective block produces splitting; scalar shifts move both modes together. Denote the first nonzero splitting order by p ∈ { 1 , 2 , … } p \in\{1,2,\ldots\} p ∈ { 1 , 2 , … } . Then, when that order exists,
∣ a + ( ε ) − a − ( ε ) ∣ = D ( n ^ , v ) ε p + o ( ε p ) , D > 0. (48) |a_{+}(\varepsilon)-a_{-}(\varepsilon)|=D(\hat{\mathbf{n}},\mathbf{v})\varepsilon^{p}+o(\varepsilon^{p}),\qquad D>0. \tag*{(48)} ∣ a + ( ε ) − a − ( ε ) ∣ = D ( n ^ , v ) ε p + o ( ε p ) , D > 0. (48)
Since ω = k a / ρ 0 \omega=k\sqrt{a/\rho_{0}} ω = k a / ρ 0 ,
Δ ω = k D ( n ^ , v ) 2 ρ 0 c 0 ε p + o ( k ε p ) . (49) \Delta\omega=\frac{kD(\hat{\mathbf{n}},\mathbf{v})}{2\rho_{0}c_{0}}\varepsilon^{p}+o(k\varepsilon^{p}). \tag*{(49)} Δ ω = 2 ρ 0 c 0 k D ( n ^ , v ) ε p + o ( k ε p ) . (49)
The conversion to a frequency gap can also be made without an independent Taylor expansion of each mode:
∣ ω + − ω − ∣ = k ∣ a + − a − ∣ ρ 0 ( c + + c − ) . (50) |\omega_{+}-\omega_{-}|=\frac{k|a_{+}-a_{-}|}{\rho_{0}(c_{+}+c_{-})}. \tag*{(50)} ∣ ω + − ω − ∣ = ρ 0 ( c + + c − ) k ∣ a + − a − ∣ . (50)
Since c ± → c 0 > 0 c_{\pm}\to c_{0}>0 c ± → c 0 > 0 , Eq. (49) follows. The positive nonzero reference speed is essential: the same exponent conversion need not hold for a mode softening to zero. A perturbation symmetry may forbid the leading matrix element. Accidental crossings can disappear under a generic perturbation, whereas protected degeneracies remain only when the relevant little group and its representation still enforce them. The value of p p p depends on the constitutive coupling, the selected state-space path, and the observation direction. Symmetry can forbid a coefficient, but an allowed coefficient can also vanish through parameter choice or cancellation; the stabilizer alone therefore does not fix p p p .
Conditional spectral critical-exponent relation
The spectral–critical reduction is restricted to a single selected ordered branch near a continuous transition, subject to the following conditions: (i) the measured system is sufficiently homogeneous or domain-selected for a nonzero η s p \eta_{\mathrm{sp}} η sp to characterize the probe volume; (ii) the selected branch lies on a fixed state-space ray, η s p = r ( t ) v \eta_{\mathrm{sp}}=r(t)\mathbf{v} η sp = r ( t ) v , with ∥ v ∥ = 1 \lVert\mathbf{v}\rVert=1 ∥ v ∥ = 1 and r ( t ) = B ∣ t ∣ β + o ( ∣ t ∣ β ) r(t)=B|t|^{\beta}+o(|t|^{\beta}) r ( t ) = B ∣ t ∣ β + o ( ∣ t ∣ β ) , B > 0 B>0 B > 0 ; (iii) the relevant spectral observable is analytic in the amplitude with first nonzero order p p p and nonzero coefficient; (iv) singular background elastic self-energies, critical attenuation, and probe-induced dynamic order-parameter relaxation do not supply a more relevant contribution to the measured gap; and (v) the experimental direction does not lie on a protected zero.
At fixed nonzero probe wave number k k k , under these assumptions and with a positive limiting reference speed c 0 c_{0} c 0 , the directional spectral gap obeys
Δ ω ( n ^ , t ) = A ( n ^ ) ∣ t ∣ p β + o ( ∣ t ∣ p β ) , A ( n ^ ) > 0. (51) \Delta\omega(\hat{\mathbf{n}},t)=\mathcal{A}(\hat{\mathbf{n}})|t|^{p\beta}+o(|t|^{p\beta}),\qquad\mathcal{A}(\hat{\mathbf{n}})>0. \tag*{(51)} Δ ω ( n ^ , t ) = A ( n ^ ) ∣ t ∣ pβ + o ( ∣ t ∣ pβ ) , A ( n ^ ) > 0. (51)
By hypothesis, the spectral gap has the asymptotic representation Δ ω = A ( n ^ ) ∥ η s p ∥ p + o ( ∥ η s p ∥ p ) \Delta\omega=\mathcal{A}(\hat{\mathbf{n}})\lVert\eta_{\mathrm{sp}}\rVert^{p}+o(\lVert\eta_{\mathrm{sp}}\rVert^{p}) Δ ω = A ( n ^ ) ∥ η sp ∥ p + o (∥ η sp ∥ p ) with A > 0 \mathcal{A}>0 A > 0 . Substitution of the ordered-branch scaling ∥ η s p ∥ = B ∣ t ∣ β + o ( ∣ t ∣ β ) \lVert\eta_{\mathrm{sp}}\rVert=B|t|^{\beta}+o(|t|^{\beta}) ∥ η sp ∥ = B ∣ t ∣ β + o ( ∣ t ∣ β ) gives Δ ω = A B p ∣ t ∣ p β + o ( ∣ t ∣ p β ) \Delta\omega=\mathcal{A}B^{p}|t|^{p\beta}+o(|t|^{p\beta}) Δ ω = A B p ∣ t ∣ pβ + o ( ∣ t ∣ pβ ) . Equation (51) is a conditional composition of two scaling laws, not a new identity obeyed by all critical spectra. A different RG scaling dimension for the probe operator, nonanalytic dynamic self-energies, macroscopic multidomain cancellation, a leading critical contribution unrelated to the order parameter, or a protected zero can invalidate it. The exponent p p p need not be universal and β \beta β alone is not fixed by the broken-symmetry subgroup.
For a curved branch, merely specifying η s p = B ∣ t ∣ β v + o ( ∣ t ∣ β ) \eta_{\mathrm{sp}}=B|t|^{\beta}\mathbf{v}+o(|t|^{\beta}) η sp = B ∣ t ∣ β v + o ( ∣ t ∣ β ) is not enough when p > 1 p>1 p > 1 : a subleading transverse component can activate a lower-degree observation channel and dominate the nominal ∣ t ∣ p β |t|^{p\beta} ∣ t ∣ pβ term. One must instead verify the same degree-p p p expansion along the actual branch, or show that all branch departures contribute o ( r p ) o(r^{p}) o ( r p ) to the gap. The distinction between an analytic constitutive readout and a fluctuation-renormalized critical observable also has to be checked (Schorgg and Schwabl, 1992).
For first-order transitions, an order-parameter jump may generate an observable jump,
[ Δ ω ] c o e x = Δ ω ( η + ) − Δ ω ( η − ) , (52) [\Delta\omega]_{\mathrm{coex}}=\Delta\omega(\eta_{+})-\Delta\omega(\eta_{-}), \tag*{(52)} [ Δ ω ] coex = Δ ω ( η + ) − Δ ω ( η − ) , (52)
but only if the selected measurement is sensitive to the difference. The jump should be evaluated at coexistence, not confused with a spinodal or kinetic switching threshold.
If the wave probe also drives the structural variable, linearized mechanics and internal-state dynamics generically take the form
\begin{pmatrix}
K_{uu}(\omega) & K_{u\eta}(\omega) \\
K_{\eta u}(\omega) & K_{\eta\eta}(\omega)
\end{pmatrix}
\begin{pmatrix}
\delta u \\
\delta\eta
\end{pmatrix}
= 0. \tag*{(53)}\end{pmatrix}
When the structural block is invertible, eliminating δ η \delta\eta δ η gives
K e f f ( ω ) = K u u ( ω ) − K u η ( ω ) K η η ( ω ) − 1 K η u ( ω ) . (54) K_{\mathrm{eff}}(\omega)=K_{uu}(\omega)-K_{u\eta}(\omega)K_{\eta\eta}(\omega)^{-1}K_{\eta u}(\omega). \tag*{(54)} K eff ( ω ) = K uu ( ω ) − K u η ( ω ) K η η ( ω ) − 1 K η u ( ω ) . (54)
Thus a slow structural relaxation can generate dispersion and attenuation in a dynamically relaxed tangent. The frozen-state wave approximation requires a comparison between probe and structural time scales; ω τ η ≫ 1 \omega\tau_{\eta}\gg1 ω τ η ≫ 1 is a common sufficient condition when coupling is present. If the linear cross block vanishes at the chosen stress-free base state, frozen and linear relaxed responses may coincide even without that inequality.
The relaxation correction can be derived directly from a quadratic local energy. Let e \mathbf{e} e be incremental strain coordinates and z \mathbf{z} z independent structural coordinates, and write
f ( 2 ) = 1 2 e T C ∞ e + e T B z + 1 2 z T H z , z ˙ = − M ( H z + B T e ) . (55) f^{(2)}=\frac{1}{2}\mathbf{e}^{T}C_{\infty}\mathbf{e}+\mathbf{e}^{T}B\mathbf{z}+\frac{1}{2}\mathbf{z}^{T}H\mathbf{z},\qquad
\dot{\mathbf{z}}=-M(H\mathbf{z}+B^{T}\mathbf{e}). \tag*{(55)} f ( 2 ) = 2 1 e T C ∞ e + e T B z + 2 1 z T H z , z ˙ = − M ( H z + B T e ) . (55)
Here H ≻ 0 H\succ0 H ≻ 0 and M ≻ 0 M\succ0 M ≻ 0 on the active structural subspace. With convention e − i ω t e^{-i\omega t} e − iω t , elimination gives
z = − ( H − i ω M − 1 ) − 1 B T e , C e f f ( ω ) = C ∞ − B ( H − i ω M − 1 ) − 1 B T , C e f f ( 0 ) = C ∞ − B H − 1 B T . (56) \mathbf{z}=-(H-i\omega M^{-1})^{-1}B^{T}\mathbf{e},\qquad
C_{\mathrm{eff}}(\omega)=C_{\infty}-B(H-i\omega M^{-1})^{-1}B^{T},\qquad
C_{\mathrm{eff}}(0)=C_{\infty}-BH^{-1}B^{T}. \tag*{(56)} z = − ( H − iω M − 1 ) − 1 B T e , C eff ( ω ) = C ∞ − B ( H − iω M − 1 ) − 1 B T , C eff ( 0 ) = C ∞ − B H − 1 B T . (56)
The static expression is the Schur complement of the full energy Hessian. For L ω = H − i ω M − 1 L_{\omega}=H-i\omega M^{-1} L ω = H − iω M − 1 and ω > 0 \omega>0 ω > 0 ,
− Im C e f f = ω B L ω − † M − 1 L ω − 1 B T ⪰ 0. (57) -\operatorname{Im}C_{\mathrm{eff}}=\omega B L_{\omega}^{-\dagger}M^{-1}L_{\omega}^{-1}B^{T}\succeq0. \tag*{(57)} − Im C eff = ω B L ω −† M − 1 L ω − 1 B T ⪰ 0. (57)
Thus eliminating a passive relaxing variable produces a passive complex modulus, with the stated sign convention [46 ] . For a single variable H = a H=a H = a , M = Γ η M=\Gamma_{\eta} M = Γ η , and coupling B = − g B=-g B = − g , Eq. (56) reduces to Eq. (105) . Stability of the complete quadratic energy requires C ∞ − B H − 1 B T ≻ 0 C_{\infty}-BH^{-1}B^{T}\succ0 C ∞ − B H − 1 B T ≻ 0 as well as H ≻ 0 H\succ0 H ≻ 0 ; symmetry alone supplies neither inequality.
Tensor and ferroelastic realizations
The two models below separate distinct sources of acoustic selectivity. For an isotropic parent, the mechanical couplings determine whether uniaxial ordering produces a linear or quadratic splitting. For a cubic parent, the little group of the propagation direction determines which transverse degeneracy is protected after tetragonal ordering.
Uniaxial ordering in a parent-isotropic solid
Using the tensor representation of § 2.1 , we recover the constitutive model of Ref. Zhu et al. [26 ] by specializing to G 0 = S O ( 3 ) G_{0}=SO(3) G 0 = S O ( 3 ) and η = Q ∈ Sym 0 ( 3 ) \boldsymbol{\eta}=\mathbf{Q}\in\operatorname{Sym}_{0}(3) η = Q ∈ Sym 0 ( 3 ) . At small strain, consider the following potential
Ψ ( E , Q , ∇ Q ) = λ L 2 ( tr E ) 2 + μ E : E + γ 1 ( tr E ) ( E : Q ) + γ 2 tr ( Q E 2 ) + a 2 tr Q 2 − b 3 tr Q 3 + c 4 ( tr Q 2 ) 2 + κ 2 ∣ ∇ Q ∣ 2 . (58) \Psi(\mathbf{E},\mathbf{Q},\nabla\mathbf{Q})=
\frac{\lambda_{L}}{2}(\operatorname{tr}\mathbf{E})^{2}
+\mu\mathbf{E}:\mathbf{E}
+\gamma_{1}(\operatorname{tr}\mathbf{E})(\mathbf{E}:\mathbf{Q})
+\gamma_{2}\operatorname{tr}(\mathbf{Q}\mathbf{E}^{2})
+\frac{a}{2}\operatorname{tr}\mathbf{Q}^{2}
-\frac{b}{3}\operatorname{tr}\mathbf{Q}^{3}
+\frac{c}{4}(\operatorname{tr}\mathbf{Q}^{2})^{2}
+\frac{\kappa}{2}\lvert\nabla\mathbf{Q}\rvert^{2}. \tag*{(58)} Ψ ( E , Q , ∇ Q ) = 2 λ L ( tr E ) 2 + μ E : E + γ 1 ( tr E ) ( E : Q ) + γ 2 tr ( Q E 2 ) + 2 a tr Q 2 − 3 b tr Q 3 + 4 c ( tr Q 2 ) 2 + 2 κ ∣ ∇ Q ∣ 2 . (58)
Here λ L \lambda_{L} λ L is a Lamé parameter (distinguished from the control parameter λ \lambda λ ), μ > 0 \mu>0 μ > 0 , 3 λ L + 2 μ > 0 3\lambda_{L}+2\mu>0 3 λ L + 2 μ > 0 , and γ 1 , γ 2 \gamma_{1},\gamma_{2} γ 1 , γ 2 are calibrated mechanical–structure couplings. The isotropic inequalities ensure a positive quadratic strain energy at Q = 0 \mathbf{Q}=0 Q = 0 . The coupled tangent must remain stable on the states probed. This model has zero stress at E = 0 \mathbf{E}=0 E = 0 for any frozen Q \mathbf{Q} Q , but the tangent depends on Q \mathbf{Q} Q . The first strain derivative, at fixed Q \mathbf{Q} Q , is
σ = λ L ( tr E ) I + 2 μ E + γ 1 [ ( E : Q ) I + ( tr E ) Q ] + γ 2 ( Q E + E Q ) . (59) \boldsymbol{\sigma}=\lambda_{L}(\operatorname{tr}\mathbf{E})\mathbf{I}+2\mu\mathbf{E}+\gamma_{1}\left[(\mathbf{E}:\mathbf{Q})\mathbf{I}+(\operatorname{tr}\mathbf{E})\mathbf{Q}\right]+\gamma_{2}(\mathbf{Q}\mathbf{E}+\mathbf{E}\mathbf{Q}). \tag*{(59)} σ = λ L ( tr E ) I + 2 μ E + γ 1 [ ( E : Q ) I + ( tr E ) Q ] + γ 2 ( QE + EQ ) . (59)
This follows from δ tr ( Q E 2 ) = ( Q E + E Q ) : δ E \delta\operatorname{tr}(\mathbf{Q}\mathbf{E}^{2})=(\mathbf{Q}\mathbf{E}+\mathbf{E}\mathbf{Q}):\delta\mathbf{E} δ tr ( Q E 2 ) = ( QE + EQ ) : δ E , without assuming that E \mathbf{E} E and Q \mathbf{Q} Q commute. Taking a further variation in a symmetric direction H \mathbf{H} H gives
C ( Q ) : H = λ L ( tr H ) I + 2 μ H + γ 1 [ ( H : Q ) I + ( tr H ) Q ] + γ 2 ( Q H + H Q ) . (60) \mathbb{C}(\mathbf{Q}):\mathbf{H}=\lambda_{L}(\operatorname{tr}\mathbf{H})\mathbf{I}+2\mu\mathbf{H}+\gamma_{1}\left[(\mathbf{H}:\mathbf{Q})\mathbf{I}+(\operatorname{tr}\mathbf{H})\mathbf{Q}\right]+\gamma_{2}(\mathbf{Q}\mathbf{H}+\mathbf{H}\mathbf{Q}). \tag*{(60)} C ( Q ) : H = λ L ( tr H ) I + 2 μ H + γ 1 [ ( H : Q ) I + ( tr H ) Q ] + γ 2 ( QH + HQ ) . (60)
The structural perturbation in components is
δ C i j k l = γ 1 ( δ i j Q k l + Q i j δ k l ) + γ 2 2 ( Q i k δ j l + Q i l δ j k + δ i k Q j l + δ i l Q j k ) . (61) \delta C_{ijkl}=\gamma_1(\delta_{ij}Q_{kl}+Q_{ij}\delta_{kl})+\frac{\gamma_2}{2}(Q_{ik}\delta_{jl}+Q_{il}\delta_{jk}+\delta_{ik}Q_{jl}+\delta_{il}Q_{jk}). \tag*{(61)} δ C ij k l = γ 1 ( δ ij Q k l + Q ij δ k l ) + 2 γ 2 ( Q ik δ j l + Q i l δ j k + δ ik Q j l + δ i l Q j k ) . (61)
A simple sufficient stability range follows by bounding the coupling energy for every symmetric H \mathbf{H} H :
W e l ( H , Q ) ≥ [ m 0 − ( 3 ∣ γ 1 ∣ + ∣ γ 2 ∣ ) ∥ Q ∥ ] ∥ H ∥ 2 , m 0 = min { μ , 3 λ L + 2 μ 2 } > 0. (62) W_{\mathrm{el}}(\mathbf{H},\mathbf{Q})\geq\left[m_0-(\sqrt{3}|\gamma_1|+|\gamma_2|)\lVert\mathbf{Q}\rVert\right]\lVert\mathbf{H}\rVert^2,\quad m_0=\min\left\{\mu,\frac{3\lambda_L+2\mu}{2}\right\}>0. \tag*{(62)} W el ( H , Q ) ≥ [ m 0 − ( 3 ∣ γ 1 ∣ + ∣ γ 2 ∣ ) ∥ Q ∥ ] ∥ H ∥ 2 , m 0 = min { μ , 2 3 λ L + 2 μ } > 0. (62)
Here ∣ tr H ∣ ≤ 3 ∥ H ∥ |\operatorname{tr}\mathbf{H}|\leq\sqrt{3}\lVert\mathbf{H}\rVert ∣ tr H ∣ ≤ 3 ∥ H ∥ and ∣ tr ( Q H 2 ) ∣ ≤ ∥ Q ∥ ∥ H ∥ 2 |\operatorname{tr}(\mathbf{Q}\mathbf{H}^2)|\leq\lVert\mathbf{Q}\rVert\lVert\mathbf{H}\rVert^2 ∣ tr ( Q H 2 ) ∣ ≤ ∥ Q ∥ ∥ H ∥ 2 . Positivity of the bracket is sufficient, although not necessary, for a positive strain Hessian and hence for strong ellipticity. The two conditions are not generally equivalent [47 ] . Equation (58) is used here as a local small-strain expansion on an admissible neighborhood with a stable elastic tangent. The linear-in-Q \mathbf{Q} Q elastic couplings do not ensure a globally bounded energy over unrestricted ( E , Q ) (\mathbf{E},\mathbf{Q}) ( E , Q ) . Applying the global minimization in Eq. (18) to this realization would therefore require a specified restricted admissible domain or a stabilizing constitutive completion. The local bound above does not provide that completion.
For constant isotropic structural mobility, the same potential gives the explicitly projected evolution
Q ˙ = − Γ Q [ ( a + c J 2 ) Q − b ( Q 2 − J 2 I / 3 ) − κ Q ∇ 2 Q + γ 1 ( tr E ) dev E + γ 2 dev ( E 2 ) ] . (63) \dot{\mathbf{Q}}=-\Gamma_Q\left[(a+cJ_2)\mathbf{Q}-b\left(\mathbf{Q}^2-J_2\mathbf{I}/3\right)-\kappa_Q\nabla^2\mathbf{Q}+\gamma_1(\operatorname{tr}\mathbf{E})\operatorname{dev}\mathbf{E}+\gamma_2\operatorname{dev}(\mathbf{E}^2)\right]. \tag*{(63)} Q ˙ = − Γ Q [ ( a + c J 2 ) Q − b ( Q 2 − J 2 I /3 ) − κ Q ∇ 2 Q + γ 1 ( tr E ) dev E + γ 2 dev ( E 2 ) ] . (63)
All terms are symmetric and traceless, so the evolution preserves the defining state space without an independent trace multiplier.
The structural part of Eq. (58) for a homogeneous, stress-free state is
f Q ( Q ) = a 2 J 2 − b 3 J 3 + c 4 J 2 2 , c > 0. (64) f_Q(\mathbf{Q})=\frac{a}{2}J_2-\frac{b}{3}J_3+\frac{c}{4}J_2^2,\qquad c>0. \tag*{(64)} f Q ( Q ) = 2 a J 2 − 3 b J 3 + 4 c J 2 2 , c > 0. (64)
For Q = s ( n ⊗ n − I / 3 ) \mathbf{Q}=s(\mathbf{n}\otimes\mathbf{n}-\mathbf{I}/3) Q = s ( n ⊗ n − I /3 ) , J 2 = 2 s 2 / 3 J_2=2s^2/3 J 2 = 2 s 2 /3 and J 3 = 2 s 3 / 9 J_3=2s^3/9 J 3 = 2 s 3 /9 , so
f Q ( s ) = a 3 s 2 − 2 b 27 s 3 + c 9 s 4 , f Q ′ ( s ) = 2 s 9 ( 3 a − b s + 2 c s 2 ) . (65) f_Q(s)=\frac{a}{3}s^2-\frac{2b}{27}s^3+\frac{c}{9}s^4,\qquad f'_Q(s)=\frac{2s}{9}(3a-bs+2cs^2). \tag*{(65)} f Q ( s ) = 3 a s 2 − 27 2 b s 3 + 9 c s 4 , f Q ′ ( s ) = 9 2 s ( 3 a − b s + 2 c s 2 ) . (65)
For b > 0 b>0 b > 0 , the uniaxial branch obeys
s ± = b ± b 2 − 24 a c 4 c , a t = b 2 27 c , s t = b 3 c . (66) s_{\pm}=\frac{b\pm\sqrt{b^2-24ac}}{4c},\qquad a_t=\frac{b^2}{27c},\qquad s_t=\frac{b}{3c}. \tag*{(66)} s ± = 4 c b ± b 2 − 24 a c , a t = 27 c b 2 , s t = 3 c b . (66)
To obtain coexistence, use stationarity a = ( b s − 2 c s 2 ) / 3 a=(bs-2cs^2)/3 a = ( b s − 2 c s 2 ) /3 to eliminate a a a from the energy:
f Q ( s ) ∣ f Q ′ = 0 = s 3 27 ( b − 3 c s ) , f Q ′ ′ ( s ) ∣ f Q ′ = 0 = 2 s 9 ( 4 c s − b ) . (67) \left.f_Q(s)\right|_{f'_Q=0}=\frac{s^3}{27}(b-3cs),\qquad\left.f''_Q(s)\right|_{f'_Q=0}=\frac{2s}{9}(4cs-b). \tag*{(67)} f Q ( s ) ∣ f Q ′ = 0 = 27 s 3 ( b − 3 cs ) , f Q ′′ ( s ) f Q ′ = 0 = 9 2 s ( 4 cs − b ) . (67)
Equality with f Q ( 0 ) = 0 f_Q(0)=0 f Q ( 0 ) = 0 gives s t = b / ( 3 c ) s_t=b/(3c) s t = b / ( 3 c ) and then a t = b 2 / ( 27 c ) a_t=b^2/(27c) a t = b 2 / ( 27 c ) . The nonzero pair appears at a = b 2 / ( 24 c ) a=b^2/(24c) a = b 2 / ( 24 c ) , while the parent curvature changes sign at a = 0 a=0 a = 0 . Neither threshold is the coexistence point. For b > 0 b>0 b > 0 , minimizing at fixed J 2 J_2 J 2 maximizes J 3 J_3 J 3 . The bound in Eq. (9) then selects a prolate uniaxial tensor, so this uniaxial reduction identifies the bulk minimum of the stated quartic model, rather than an arbitrary tensor restriction [48 ] . The values a t a_t a t and s t s_t s t in Eq. (66) are the coexistence control and order-parameter jump for the specified uniaxial branch, not universal numbers. For b = 0 b=0 b = 0 and a < 0 a<0 a < 0 , the quartic model fixes J 2 = − a / c J_2=-a/c J 2 = − a / c but does not select uniaxiality over biaxiality without another mechanism. Therefore assigning a uniaxial continuous branch in that case is an explicit selection assumption.
Let D i j = C i j k l ϵ k l D_{ij}=C_{ijkl}\epsilon_{kl} D ij = C ij k l ϵ k l and V i k = C i j k j V_{ik}=C_{ijkj} V ik = C ij k j denote the dilation and Voigt stiffness contractions. The isotropic parts are D 0 = ( 3 λ L + 2 μ ) I \mathbf{D}_0=(3\lambda_L+2\mu)\mathbf{I} D 0 = ( 3 λ L + 2 μ ) I and V 0 = ( λ L + 4 μ ) I \mathbf{V}_0=(\lambda_L+4\mu)\mathbf{I} V 0 = ( λ L + 4 μ ) I . Tracing the perturbation over ( k , l ) (k,l) ( k , l ) and ( j , l ) (j,l) ( j , l ) , respectively, uses tr Q = 0 \operatorname{tr}\mathbf{Q}=0 tr Q = 0 and removes all isotropic contributions after taking the deviator. This gives the two independent coefficients below. Equation (61) implies
dev D = ( 3 γ 1 + 2 γ 2 ) Q = : a D Q . (68) \operatorname{dev}\mathbf{D}=(3\gamma_1+2\gamma_2)\mathbf{Q}=:a_D\mathbf{Q}. \tag*{(68)} dev D = ( 3 γ 1 + 2 γ 2 ) Q =: a D Q . (68)
dev V = ( 2 γ 1 + 5 2 γ 2 ) Q = : a V Q . (69) \operatorname{dev}\mathbf{V}=\left(2\gamma_1+\frac{5}{2}\gamma_2\right)\mathbf{Q}=:a_V\mathbf{Q}. \tag*{(69)} dev V = ( 2 γ 1 + 2 5 γ 2 ) Q =: a V Q . (69)
The mapping ( γ 1 , γ 2 ) ↦ ( a D , a V ) (\gamma_1,\gamma_2) \mapsto(a_D,a_V) ( γ 1 , γ 2 ) ↦ ( a D , a V ) has determinant 7 / 2 7/2 7/2 . Hence, for any fixed ( γ 1 , γ 2 ) ≠ ( 0 , 0 ) (\gamma_1,\gamma_2) \ne(0,0) ( γ 1 , γ 2 ) = ( 0 , 0 ) ,
Q = a D dev D + a V dev V a D 2 + a V 2 (70) \mathbf{Q}=\frac{a_D\operatorname{dev}\mathbf{D}+a_V\operatorname{dev}\mathbf{V}}{a_D^2+a_V^2} \tag*{(70)} Q = a D 2 + a V 2 a D dev D + a V dev V (70)
and the calibrated complete tangent distinguishes Q \mathbf{Q} Q exactly. Consequently H Q = G C ( Q ) H_{\mathbf{Q}}=G_{\mathbb{C}}(\mathbf{Q}) H Q = G C ( Q ) as a proper rotational stabilizer in this restricted model. If both couplings vanish, the complete tangent is isotropic and cannot reveal a nonzero Q \mathbf{Q} Q . A particular incomplete set of wave speeds need not reconstruct the full tangent even when Eq. (70) holds.
Put Q = s ( N ⊗ N − I / 3 ) \mathbf{Q}=s(\mathbf{N}\otimes\mathbf{N}-\mathbf{I}/3) Q = s ( N ⊗ N − I /3 ) , and let ϑ \vartheta ϑ be the angle between wave normal n ^ \hat{\mathbf{n}} n ^ and the structural axis N \mathbf{N} N . For sufficiently small s s s , with a nonzero longitudinal–transverse separation in the isotropic reference, the two quasi-transverse modes obey
∣ c T 1 2 − c T 2 2 ∣ = ∣ γ 2 s ∣ 2 ρ 0 sin 2 ϑ + O ( s 2 ) . (71) \left|c_{T1}^{2}-c_{T2}^{2}\right|=\frac{|\gamma_2s|}{2\rho_0}\sin^2\vartheta+O(s^2). \tag*{(71)} c T 1 2 − c T 2 2 = 2 ρ 0 ∣ γ 2 s ∣ sin 2 ϑ + O ( s 2 ) . (71)
The γ 1 \gamma_1 γ 1 coupling drops out of the first-order transverse projection and affects the splitting through higher-order mixing. The axial zero at ϑ = 0 \vartheta=0 ϑ = 0 is protected by the uniaxial rotational little group. At fixed non-axial observation angle, if γ 2 ≠ 0 \gamma_2\ne0 γ 2 = 0 the gap has p = 1 p=1 p = 1 along the selected uniaxial path. If γ 2 = 0 \gamma_2=0 γ 2 = 0 and γ 1 ≠ 0 \gamma_1\ne0 γ 1 = 0 , the leading oblique splitting from the source model is instead
∣ c T 1 2 − c T 2 2 ∣ = γ 1 2 s 2 ρ 0 ( λ L + μ ) sin 2 ϑ cos 2 ϑ + O ( ∣ s ∣ 3 ) , (72) \left|c_{T1}^{2}-c_{T2}^{2}\right|=\frac{\gamma_1^2s^2}{\rho_0(\lambda_L+\mu)}\sin^2\vartheta\cos^2\vartheta+O(|s|^3), \tag*{(72)} c T 1 2 − c T 2 2 = ρ 0 ( λ L + μ ) γ 1 2 s 2 sin 2 ϑ cos 2 ϑ + O ( ∣ s ∣ 3 ) , (72)
assuming the positive-energy isotropic reference specified above. This gives p = 2 p=2 p = 2 at generic oblique angles. In this γ 1 \gamma_1 γ 1 -only model, both the axial and equatorial directions have zero transverse splitting: the axial zero is symmetry protected, whereas the equatorial zero follows from the selected coupling. Thus an angular zero must be interpreted together with the constitutive law, rather than attributed to structural symmetry alone.
If a continuous ordered branch has ∣ s ∣ ∼ ∣ t ∣ β |s|\sim|t|^\beta ∣ s ∣ ∼ ∣ t ∣ β and the regularity and observability assumptions underlying Eq. (51) hold, the observable exponent is β \beta β for the nonzero γ 2 \gamma_2 γ 2 channel and 2 β 2\beta 2 β for the leading γ 1 \gamma_1 γ 1 -only oblique channel. In the mean-field selected b = 0 b=0 b = 0 uniaxial branch, β = 1 / 2 \beta=1/2 β = 1/2 . This choice is not necessarily a generic full-tensor equilibrium phase and must be stabilized by the material’s selection mechanism. The cases above concern a specified coupling channel and observation angle. In particular, a vanishing coefficient caused by calibration or by an angular zero must be identified before assigning a fitted temperature exponent.
Cubic-to-tetragonal ferroelasticity
Consider a parent cubic crystal with point group G 0 = O h G_0=O_h G 0 = O h . A cubic-to-tetragonal distortion is represented at minimal order by a two-dimensional E g E_g E g multiplet
η = ( η 1 , η 2 ) , r 2 = η 1 2 + η 2 2 , η 1 = r cos ϕ , η 2 = r sin ϕ . (73) \boldsymbol{\eta}=(\eta_1,\eta_2),\qquad r^2=\eta_1^2+\eta_2^2,\qquad\eta_1=r\cos\phi,\ \eta_2=r\sin\phi. \tag*{(73)} η = ( η 1 , η 2 ) , r 2 = η 1 2 + η 2 2 , η 1 = r cos ϕ , η 2 = r sin ϕ . (73)
Fix the basis explicitly by assigning the multiplet the same transformation law as ( − e 3 , e 2 ) (-\mathbf{e}_3,\mathbf{e}_2) ( − e 3 , e 2 ) , with the strain components defined in Eq. (79) . Its diagonal tensor representative is
D E g ( η ) = diag ( − η 1 6 + η 2 2 , − η 1 6 − η 2 2 , 2 η 1 6 ) . (74) \mathbf{D}_{E_g}(\boldsymbol{\eta})=\operatorname{diag}\left(-\frac{\eta_1}{\sqrt{6}}+\frac{\eta_2}{\sqrt{2}},-\frac{\eta_1}{\sqrt{6}}-\frac{\eta_2}{\sqrt{2}},\frac{2\eta_1}{\sqrt{6}}\right). \tag*{(74)} D E g ( η ) = diag ( − 6 η 1 + 2 η 2 , − 6 η 1 − 2 η 2 , 6 2 η 1 ) . (74)
Cubic rotations permute its diagonal entries. Consequently tr D E g 2 = r 2 \operatorname{tr}\mathbf{D}_{E_g}^2=r^2 tr D E g 2 = r 2 and 6 tr D E g 3 = η 1 3 − 3 η 1 η 2 2 = r 3 cos 3 ϕ \sqrt{6}\operatorname{tr}\mathbf{D}_{E_g}^3=\eta_1^3-3\eta_1\eta_2^2=r^3\cos3\phi 6 tr D E g 3 = η 1 3 − 3 η 1 η 2 2 = r 3 cos 3 ϕ are invariants ([49 ] ). This fixes the angular convention used in the following polynomial; the unrotated ( e 2 , e 3 ) (\mathbf{e}_2,\mathbf{e}_3) ( e 2 , e 3 ) basis instead produces a sine form of the same cubic invariant. A minimal Landau density is
f E g ( r , ϕ ) = a 2 r 2 + b 3 r 3 cos 3 ϕ + c 4 r 4 , c > 0. (75) f_{E_g}(r,\phi)=\frac{a}{2}r^2+\frac{b}{3}r^3\cos3\phi+\frac{c}{4}r^4,\qquad c>0. \tag*{(75)} f E g ( r , ϕ ) = 2 a r 2 + 3 b r 3 cos 3 ϕ + 4 c r 4 , c > 0. (75)
For b > 0 b>0 b > 0 , the three preferred tetragonal variants have cos 3 ϕ = − 1 \cos3\phi=-1 cos 3 ϕ = − 1 . Their symmetry groups are conjugate tetragonal subgroups D 4 h ⊂ O h D_{4h}\subset O_h D 4 h ⊂ O h (up to extra couplings and specimen conditions); a generic nonzero E g E_g E g direction may have lower symmetry. The cubic parent-to-tetragonal transition can therefore be represented through a single invariant potential, without selecting three separate elastic laws. Aizu’s classification of ferroelastic species provides the established group-theoretic context [21 , 22 ] .
Along a preferred variant, Eq. (75) reduces to
f v a r ( r ) = a 2 r 2 − b 3 r 3 + c 4 r 4 , d f v a r d r = r ( a − b r + c r 2 ) . (76) f_{\mathrm{var}}(r)=\frac{a}{2}r^2-\frac{b}{3}r^3+\frac{c}{4}r^4,\qquad\frac{\mathrm{d}f_{\mathrm{var}}}{\mathrm{d}r}=r(a-br+cr^2). \tag*{(76)} f var ( r ) = 2 a r 2 − 3 b r 3 + 4 c r 4 , d r d f var = r ( a − b r + c r 2 ) . (76)
The stationary radii are
r = 0 , r ± = b ± b 2 − 4 a c 2 c . (77) r=0,\qquad r_{\pm}=\frac{b\pm\sqrt{b^2-4ac}}{2c}. \tag*{(77)} r = 0 , r ± = 2 c b ± b 2 − 4 a c . (77)
For b > 0 b>0 b > 0 the two equal-energy minima occur at
a t = 2 b 2 9 c , r t = 2 b 3 c . (78) a_t=\frac{2b^2}{9c},\qquad r_t=\frac{2b}{3c}. \tag*{(78)} a t = 9 c 2 b 2 , r t = 3 c 2 b . (78)
For r > 0 r>0 r > 0 , angular stationarity gives ∂ ϕ f E g = − b r 3 sin 3 ϕ = 0 \partial_{\phi}f_{E_g}=-br^3\sin3\phi=0 ∂ ϕ f E g = − b r 3 sin 3 ϕ = 0 and ∂ ϕ ϕ f E g = − 3 b r 3 cos 3 ϕ \partial_{\phi\phi}f_{E_g}=-3br^3\cos3\phi ∂ ϕϕ f E g = − 3 b r 3 cos 3 ϕ . Thus the preferred rays for b > 0 b>0 b > 0 have positive angular curvature. On those rays the radial curvature at a stationary point is f ′ ′ = r ( 2 c r − b ) f^{\prime\prime}=r(2cr-b) f ′′ = r ( 2 cr − b ) ; r + r_+ r + is locally stable radially and r − r_- r − is unstable whenever the roots are distinct and positive. At stationarity the energy is b r 3 / 6 − c r 4 / 4 br^3/6-cr^4/4 b r 3 /6 − c r 4 /4 , whose equality with zero yields Eq. (78) . Reversing the sign of b b b interchanges elongational and compressional preferred rays and replaces b b b by ∣ b ∣ \lvert b\rvert ∣ b ∣ in the preferred-ray radial polynomial. The nonzero stationary pair first appears at a = b 2 / ( 4 c ) a=b^2/(4c) a = b 2 / ( 4 c ) , whereas the cubic parent state loses quadratic local stability at a = 0 a=0 a = 0 . Coexistence, emergence of metastable stationary branches, and loss of metastability are distinct thresholds. At b = 0 b=0 b = 0 the quartic model has accidental continuous angular degeneracy at fixed r r r . Higher-order cubic invariants, elastic compatibility, and boundaries are needed to determine the selected variants. These are mean-field conclusions, subject to fluctuation and elastic renormalization.
Define symmetry-adapted small-strain combinations
e 1 = E x x + E y y + E z z 3 , e 2 = E x x − E y y 2 , e 3 = E x x + E y y − 2 E z z 6 . (79) e_1=\frac{E_{xx}+E_{yy}+E_{zz}}{\sqrt{3}},\qquad e_2=\frac{E_{xx}-E_{yy}}{\sqrt{2}},\qquad e_3=\frac{E_{xx}+E_{yy}-2E_{zz}}{\sqrt{6}}. \tag*{(79)} e 1 = 3 E xx + E y y + E z z , e 2 = 2 E xx − E y y , e 3 = 6 E xx + E y y − 2 E z z . (79)
The doublet ( e 2 , e 3 ) (e_2,e_3) ( e 2 , e 3 ) transforms as E g E_g E g , up to basis conventions. If η \boldsymbol{\eta} η denotes an internal optical/structural distortion distinct from the macroscopic strain, the leading coupling can be
W E g = K E 2 ( e 2 2 + e 3 2 ) − g ( η 2 e 2 − η 1 e 3 ) + W o t h e r ( E , η ) , K E > 0. (80) W_{E_g}=\frac{K_E}{2}(e_2^2+e_3^2)-g(\eta_2e_2-\eta_1e_3)+W_{\mathrm{other}}(\mathbf{E},\boldsymbol{\eta}),\qquad K_E>0. \tag*{(80)} W E g = 2 K E ( e 2 2 + e 3 2 ) − g ( η 2 e 2 − η 1 e 3 ) + W other ( E , η ) , K E > 0. (80)
This is the invariant scalar product of the two E g E_g E g doublets ( η 1 , η 2 ) (\eta_1,\eta_2) ( η 1 , η 2 ) and ( − e 3 , e 2 ) (-e_3,e_2) ( − e 3 , e 2 ) . If the strain is fully relaxed in a homogeneous stress-free idealization and W o t h e r W_{\mathrm{other}} W other does not couple these sectors,
( e 2 , e 3 ) = g K E ( η 2 , − η 1 ) , f E g r e l a x e d = f E g − g 2 2 K E r 2 . (81) (e_2,e_3)=\frac{g}{K_E}(\eta_2,-\eta_1),\qquad f_{E_g}^{\mathrm{relaxed}}=f_{E_g}-\frac{g^2}{2K_E}r^2. \tag*{(81)} ( e 2 , e 3 ) = K E g ( η 2 , − η 1 ) , f E g relaxed = f E g − 2 K E g 2 r 2 . (81)
Thus a e f f = a − g 2 / K E a_{\mathrm{eff}}=a-g^2/K_E a eff = a − g 2 / K E in this homogeneous stress-free idealization. This result follows by completing the square:
K E 2 [ ( e 2 − g η 2 K E ) 2 + ( e 3 + g η 1 K E ) 2 ] − g 2 2 K E r 2 . (82) \frac{K_E}{2}\left[\left(e_2-\frac{g\eta_2}{K_E}\right)^2+\left(e_3+\frac{g\eta_1}{K_E}\right)^2\right]-\frac{g^2}{2K_E}r^2. \tag*{(82)} 2 K E [ ( e 2 − K E g η 2 ) 2 + ( e 3 + K E g η 1 ) 2 ] − 2 K E g 2 r 2 . (82)
The bare and relaxed values of a a a must therefore be distinguished when comparing the branch thresholds. If a a a in Eq. (75) is already a stress-free fitted coefficient, the same relaxation correction must not be subtracted again. Finite-body compatibility, external constraints, domains, and nonlocal elastic interactions obstruct independent local strain minimization. If instead the strain is itself chosen as the primary order parameter, Eq. (80) must not introduce a redundant, independently evolving copy of the same degree of freedom.
A parent-invariant higher-order elastic energy can make the tangent phase dependent,
W e l a s t i c ( E , η ) = 1 2 E : C 0 : E + ∑ a = 1 2 η a B a ( E , E ) + O ( r 2 ∥ E ∥ 2 , ∥ E ∥ 3 ) , (83) W_{\mathrm{elastic}}(\mathbf{E},\boldsymbol{\eta}) = \frac{1}{2}\mathbf{E} \mathbin{:} \mathbb{C}_0 \mathbin{:} \mathbf{E} + \sum_{a=1}^{2} \eta_a \mathcal{B}_a(\mathbf{E},\mathbf{E}) + O\left(r^2\lVert\mathbf{E}\rVert^2,\lVert\mathbf{E}\rVert^3\right), \tag*{(83)} W elastic ( E , η ) = 2 1 E : C 0 : E + a = 1 ∑ 2 η a B a ( E , E ) + O ( r 2 ∥ E ∥ 2 , ∥ E ∥ 3 ) , (83)
where the bilinear elastic forms B a \mathcal{B}_a B a must transform as a covariant E g E_g E g doublet. This notation encodes allowed symmetry couplings but does not assume their coefficients or positive ellipticity a priori.
The distinction between the two couplings is explicit: the term linear in strain in Eq. (80) changes the equilibrium strain but has zero second strain derivative at fixed η \boldsymbol{\eta} η . It cannot by itself generate a phase-dependent frozen stiffness. For a concrete symmetry-allowed example of Eq. (83) , let d i d_i d i be the diagonal entries of D E \mathbf{D}_E D E and add
W a n = 2 γ E ( d x E y z 2 + d y E x z 2 + d z E x y 2 ) . (84) W_{\mathrm{an}} = 2\gamma_E\left(d_x E_{yz}^2 + d_y E_{xz}^2 + d_z E_{xy}^2\right). \tag*{(84)} W an = 2 γ E ( d x E y z 2 + d y E x z 2 + d z E x y 2 ) . (84)
This expression is invariant under cubic signed coordinate permutations. Since the tensorial shear energy is 2 C i j i j E i j 2 2C_{ijij}E_{ij}^2 2 C ij ij E ij 2 , it gives δ ( C 44 , C 55 , C 66 ) = γ E ( d x , d y , d z ) \delta(C_{44},C_{55},C_{66})=\gamma_E(d_x,d_y,d_z) δ ( C 44 , C 55 , C 66 ) = γ E ( d x , d y , d z ) . For a z z z -tetragonal ray η 2 = 0 \eta_2=0 η 2 = 0 , one obtains C 66 − C 44 = 3 / 2 γ E η 1 C_{66}-C_{44}=\sqrt{3}/2\gamma_E\eta_1 C 66 − C 44 = 3 /2 γ E η 1 . Thus a calibrated nonzero γ E \gamma_E γ E realizes p = 1 p=1 p = 1 explicitly; the abstract expansion in Eq. (87) also allows models in which that term is absent [50 ] .
For a tetragonal variant whose unique axis is z z z , consider wave propagation along x x x . Standard tetragonal stiffness notation gives
c T , y 2 = C 66 ρ 0 , c T , z 2 = C 44 ρ 0 , Δ c T 2 ∣ [ 100 ] = ∣ C 66 − C 44 ∣ ρ 0 . (85) c_{T,y}^2 = \frac{C_{66}}{\rho_0}, \qquad c_{T,z}^2 = \frac{C_{44}}{\rho_0}, \qquad\left.\Delta c_T^2\right|_{[100]} = \frac{\lvert C_{66}-C_{44}\rvert}{\rho_0}. \tag*{(85)} c T , y 2 = ρ 0 C 66 , c T , z 2 = ρ 0 C 44 , Δ c T 2 [ 100 ] = ρ 0 ∣ C 66 − C 44 ∣ . (85)
For the cubic parent C 66 = C 44 C_{66}=C_{44} C 66 = C 44 , so these two transverse branches are degenerate along [100]. In the tetragonal ordered state they generally split. Along the tetragonal [001] axis, by contrast, fourfold rotation protects the two transverse polarizations and
Δ c T 2 ∣ [ 001 ] = 0. (86) \left.\Delta c_T^2\right|_{[001]} = 0. \tag*{(86)} Δ c T 2 [ 001 ] = 0. (86)
The tetragonal acoustic matrix is diag ( C 11 , C 66 , C 44 ) \operatorname{diag}(C_{11},C_{66},C_{44}) diag ( C 11 , C 66 , C 44 ) along [100]. Along [001], it is diag ( C 44 , C 44 , C 33 ) \operatorname{diag}(C_{44},C_{44},C_{33}) diag ( C 44 , C 44 , C 33 ) . These diagonal matrices give both statements without identifying a general cubic propagation direction with a degenerate axis. The important claim is an angular selection rule, not a claim that every propagation direction is degenerate in a cubic crystal. At a small selected E g E_g E g amplitude r r r , suppose the calibrated constitutive map has
C 66 − C 44 = B E r p + o ( r p ) , B E ≠ 0. (87) C_{66}-C_{44}=B_Er^p+o(r^p), \qquad B_E\ne0. \tag*{(87)} C 66 − C 44 = B E r p + o ( r p ) , B E = 0. (87)
Then the [100] squared-speed gap has leading order ∣ B E ∣ r p / ρ 0 \lvert B_E\rvert r^p/\rho_0 ∣ B E ∣ r p / ρ 0 , whereas the [001] protected gap is zero. Near a continuous critical branch with r ∼ ∣ t ∣ β r\sim\lvert t\rvert^\beta r ∼ ∣ t ∣ β , the [100] gap follows ∣ t ∣ p β \lvert t\rvert^{p\beta} ∣ t ∣ pβ only under the hypotheses in § 3 . For a first-order branch, the signal can exhibit a finite coexistence discontinuity. Measuring both crystallographic directions distinguishes a predicted lifted degeneracy from a symmetry-protected one.
Numerical benchmarks
The benchmarks below test distinct parts of the framework using separately specified models, where the homogeneous strain marginal tests invariant free-energy reconstruction, the E g E_g E g lattice tests coexistence, the compatible strain model tests domain organization, and the fixed-cell bond network tests acoustic selection rules. The Ising calculations test prescribed linear and quadratic spectral readouts, with and without driving. These calculations provide complementary consistency checks.
Symmetry strata and free-energy reconstruction
In the basis of Eq. (79) , write the diagonal deviatoric strain as s B q s\mathbf{Bq} s Bq , where q = ( e 2 , e 3 ) / s \mathbf{q}=(e_2,e_3)/s q = ( e 2 , e 3 ) / s and
B = ( 1 / 2 1 / 6 − 1 / 2 1 / 6 0 − 2 / 6 ) . (88) \mathbf{B} =
\begin{pmatrix}
1/\sqrt{2} & 1/\sqrt{6} \\
-1/\sqrt{2} & 1/\sqrt{6} \\
0 & -2/\sqrt{6}
\end{pmatrix}. \tag*{(88)} B = 1/ 2 − 1/ 2 0 1/ 6 1/ 6 − 2/ 6 . (88)
For a proper-strain realization in the canonical basis of § 4.2 , the components are related by ( q 1 , q 2 ) = ( η 2 , − η 1 ) / s (q_1,q_2)=(\eta_2,-\eta_1)/s ( q 1 , q 2 ) = ( η 2 , − η 1 ) / s . This orthogonal basis change leaves the quadratic norm unchanged and converts the cosine cubic invariant into 3 q 1 q 2 2 − q 2 3 3q_1q_2^2-q_2^3 3 q 1 q 2 2 − q 2 3 . It must be applied before comparing angular positions across the two conventions. For the homogeneous lattice benchmark in Figure 2 , the deformation gradient is diag ( exp [ s ( B q ) j ] ) \operatorname{diag}(\exp[s(\mathbf{B}\mathbf{q})_j]) diag ( exp [ s ( Bq ) j ]) with s = 0.10 s=0.10 s = 0.10 . Thus the plotted order parameter is obtained from the logarithmic strain of the deformed face-centred cubic reference. This finite-deformation construction illustrates the symmetry classification of a homogeneous lattice; it is distinct from the infinitesimal-strain constitutive realizations in § 4 . A permutation of the three diagonal strains acts on the doublet as the cubic group projected onto E g E_g E g . The origin has stabilizer O h O_h O h . Two equal principal strains give D 4 h D_{4h} D 4 h on each of the three lines away from the origin. Three unequal principal strains give D 2 h D_{2h} D 2 h . The three elongational rays form one orbit. This classification concerns the diagonal strain state represented by the doublet.
Figure 2. Symmetry-stratified states in a cubic-to-tetragonal lattice benchmark. (a) Undeformed face-centred cubic reference and the three volume-preserving elongational variants. (b) Static bond energies along the indicated deformation paths. (c) WHAM free energy at t ∗ = 0.9881 t_* = 0.9881 t ∗ = 0.9881 with the equal-depth minima marked. (d) The origin, the tetragonal lines, and the generic directions have stabilizers O h O_h O h , D 4 h D_{4h} D 4 h , and D 2 h D_{2h} D 2 h , respectively. Blue points mark the elongational orbit. (e) Radial probability densities below, at, and above t ∗ t_* t ∗ . (f) Sixth-order invariant fit (lines) and an independent WHAM estimate along excluded directions (circles).
The static energy sums over the three axial second-neighbour bonds counted per site, with reduced extensions d j = ( B q ) j d_j=(\mathbf{B}\mathbf{q})_j d j = ( Bq ) j ,
U ( q ) N E 0 = ∑ j = 1 3 ( − 0.375 d j 2 − 2.4 6 3 d j 3 + 1.2 d j 4 + 0.12 d j 6 ) . (89) \frac{U(\mathbf{q})}{N E_0}=\sum_{j=1}^{3}\left(-0.375d_j^2-\frac{2.4\sqrt{6}}{3}d_j^3+1.2d_j^4+0.12d_j^6\right). \tag*{(89)} N E 0 U ( q ) = j = 1 ∑ 3 ( − 0.375 d j 2 − 3 2.4 6 d j 3 + 1.2 d j 4 + 0.12 d j 6 ) . (89)
The displacements have polarization-dependent stiffness K j K_j K j on a periodic harmonic nearest-neighbour graph. For N = 512 N=512 N = 512 sites and t = k B T / E 0 t=k_{\mathrm{B}}T/E_0 t = k B T / E 0 , integration over the N − 1 N-1 N − 1 nonzero displacement modes per polarization gives the shape-dependent free energy
F t ( q ) N E 0 = U ( q ) N E 0 + t 2 ( 1 − 1 N ) ∑ j = 1 3 ln K j ( q ) K 0 + c o n s t . , K j K 0 = exp ( 1.15 ∣ q ∣ 2 3 ) ( 1 + 0.14 d j 2 ) . (90) \frac{F_t(\mathbf{q})}{NE_0}=\frac{U(\mathbf{q})}{NE_0}+\frac{t}{2}\left(1-\frac{1}{N}\right)\sum_{j=1}^{3}\ln\frac{K_j(\mathbf{q})}{K_0}+\mathrm{const.},\qquad\frac{K_j}{K_0}=\exp\left(\frac{1.15|\mathbf{q}|^2}{3}\right)(1+0.14d_j^2). \tag*{(90)} N E 0 F t ( q ) = N E 0 U ( q ) + 2 t ( 1 − N 1 ) j = 1 ∑ 3 ln K 0 K j ( q ) + const. , K 0 K j = exp ( 3 1.15∣ q ∣ 2 ) ( 1 + 0.14 d j 2 ) . (90)
The logarithmic term follows from a Gaussian integral. If ℓ α > 0 \ell_\alpha>0 ℓ α > 0 are the N − 1 N-1 N − 1 nonzero eigenvalues of the common periodic graph Laplacian, the mode energy is 1 2 K j ℓ α x j α 2 \frac{1}{2}K_j\ell_\alpha x_{j\alpha}^2 2 1 K j ℓ α x j α 2 . Hence
Z v i b ( q ) ∝ ∏ j = 1 3 ∏ α = 1 N − 1 ( β T K j ( q ) ℓ α ) − 1 / 2 , β T = ( k B T ) − 1 . (91) Z_{\mathrm{vib}}(\mathbf{q})\propto\prod_{j=1}^{3}\prod_{\alpha=1}^{N-1}(\beta_TK_j(\mathbf{q})\ell_\alpha)^{-1/2},\qquad\beta_T=(k_{\mathrm{B}}T)^{-1}. \tag*{(91)} Z vib ( q ) ∝ j = 1 ∏ 3 α = 1 ∏ N − 1 ( β T K j ( q ) ℓ α ) − 1/2 , β T = ( k B T ) − 1 . (91)
Taking − k B T log Z v i b -k_{\mathrm{B}}T\log Z_{\mathrm{vib}} − k B T log Z vib and dividing by N E 0 NE_0 N E 0 gives the coefficient t ( 1 − 1 / N ) / 2 t(1-1/N)/2 t ( 1 − 1/ N ) /2 in Eq. (90) . The translational zero modes are fixed, and graph and momentum factors independent of q \mathbf{q} q enter the constant. This simple product requires the specified factorization into K j ( q ) K_j(\mathbf{q}) K j ( q ) times a common graph operator ([51 ] ). The static bond energy includes the cubic invariant I 3 = ∑ j d j 3 = ( 3 q 1 2 q 2 − q 2 3 ) / 6 I_3=\sum_jd_j^3=(3q_1^2q_2-q_2^3)/\sqrt{6} I 3 = ∑ j d j 3 = ( 3 q 1 2 q 2 − q 2 3 ) / 6 , the basis-specific form of the angular term in Eq. (75) . Figure 2 b compares three deformations of this bond model.
Independent draws from 81 biased windows of the analytically integrated strain marginal were combined by weighted histogram analysis (WHAM) to obtain F t = − k B T ln P t + c o n s t . F_t=-k_{\mathrm{B}}T\ln P_t+\mathrm{const.} F t = − k B T ln P t + const. on a flat two-dimensional strain grid. The equal-depth temperature t ∗ = 0.9881 t_*=0.9881 t ∗ = 0.9881 gives one cubic and three tetragonal minima, with a barrier of 3.54 k B T ∗ 3.54k_{\mathrm{B}}T_* 3.54 k B T ∗ along a preferred variant ray. Equal depth refers to the minima of the two-dimensional map. The radial probability in Figure 2 e also includes the polar measure r d r d ϕ r\,\mathrm{d}r\,\mathrm{d}\phi r d r d ϕ , with r = ∣ q ∣ r=|\mathbf{q}| r = ∣ q ∣ , and the multiplicity of the tetragonal wells. An invariant fit using 1 1 1 , r 2 r^2 r 2 , I 3 I_3 I 3 , r 4 r^4 r 4 , r 2 I 3 r^2I_3 r 2 I 3 , r 6 r^6 r 6 , I 3 2 I_3^2 I 3 2 excludes the symmetry orbits of the two test directions and their angular neighbourhoods. Its root-mean-square error against a separate umbrella sample on those directions is 0.30 k B T ∗ 0.30k_{\mathrm{B}}T_* 0.30 k B T ∗ . This tests sampling, reconstruction, and angular prediction against the specified strain marginal. Because the reference marginal is analytically available, it is a controlled reconstruction benchmark, rather than an independent test of the material Hamiltonian.
Table 1 quantifies both the strain-marginal reconstruction and the invariant fit. We use the total-variation distance 1 2 ∑ ℓ ∣ P ℓ W H A M − P ℓ r e f ∣ \frac{1}{2}\sum_\ell|P_\ell^{\mathrm{WHAM}}-P_\ell^{\mathrm{ref}}| 2 1 ∑ ℓ ∣ P ℓ WHAM − P ℓ ref ∣ on the common discrete grid to compare normalized probabilities. The angular test uses an independently sampled reference and excludes the complete symmetry-related neighbourhoods of the test rays from fitting. The probability error and the energy-fit error measure complementary parts of this reconstruction.
Table 1. Sampling and free-energy reconstruction for the homogeneous lattice benchmark in Figure 2 . The reference marginal follows from the integrated harmonic model.
Coexistence and a smooth symmetry-stratum crossing
The finite-volume statistical test of § 2.2 uses a continuous E g E_g E g field q i = r i ( cos ϕ i , sin ϕ i ) \mathbf{q}_i = r_i(\cos\phi_i,\sin\phi_i) q i = r i ( cos ϕ i , sin ϕ i ) on a periodic cubic lattice. Here the doublet uses the rotated cosine convention of § 4.2 ; its angular origin is a preferred elongational ray for the negative b b b used below. It differs from the fixed ( e 2 , e 3 ) (e_2,e_3) ( e 2 , e 3 ) coordinates of § 5.1 . Its Hamiltonian is
H = ∑ i ( a 2 r i 2 + b 3 r i 3 cos 3 ϕ i + c 4 r i 4 ) + J 2 ∑ ⟨ i , j ⟩ ∣ q i − q j ∣ 2 . (92) H = \sum_i \left(\frac{a}{2}r_i^2 + \frac{b}{3}r_i^3\cos3\phi_i + \frac{c}{4}r_i^4\right) + \frac{J}{2}\sum_{\langle i,j\rangle} \lvert\mathbf{q}_i-\mathbf{q}_j\rvert^2. \tag*{(92)} H = i ∑ ( 2 a r i 2 + 3 b r i 3 cos 3 ϕ i + 4 c r i 4 ) + 2 J ⟨ i , j ⟩ ∑ ∣ q i − q j ∣ 2 . (92)
We use ( a , b , c , J ) = ( 19.10 , − 60 , 40 , 0.55 ) (a,b,c,J) = (19.10,-60,40,0.55) ( a , b , c , J ) = ( 19.10 , − 60 , 40 , 0.55 ) and sample by Metropolis updates with replica exchanges between umbrella windows in L − 3 ∑ i r i 2 L^{-3}\sum_i r_i^2 L − 3 ∑ i r i 2 and reweight to equal integrated basin weight at T L T_L T L , using 0.48 as the threshold in the window coordinate. The ordered weight includes all three variants. The reference sampling temperature is 0.374. Each of the three independent umbrella ladders discards 4500 sweeps and records 3200 measurements at intervals of three sweeps, with neighbouring window exchanges attempted each sweep. Temperature reweighting then locates T L T_L T L for each lattice size. Figure 3 a,Figure 3 b shows the distribution of the global average q ‾ = L − 3 ∑ i q i \overline{\mathbf{q}} = L^{-3}\sum_i \mathbf{q}_i q = L − 3 ∑ i q i and its amplitude, as well as the coexistence of the low- and high-amplitude basins. The energy-histogram barrier is
Figure 3. Thermodynamic coexistence and symmetry-stratum crossing. (a) Free-energy maps of the global E g E_g E g doublet for L = 6 L = 6 L = 6 at ordered-basin weights 0.9, 0.5, and 0.1. (b) Joint distribution of energy density and global amplitude at equal basin weight. (c) Energy histograms at each size’s equal-weight temperature. (d) Barrier from Eq. (93) against L 2 L^2 L 2 . Bars show the standard deviation across three umbrella runs. Dashed histograms and open markers denote limited mixing. The dashed line in (d) uses filled symbols with L ≥ 6 L \ge6 L ≥ 6 . (e) Mean amplitude and threshold-based basin populations. (f) Analytic field-driven stratum crossing in the harmonic model and exchange of the two restricted free-energy branches in the anharmonic model. The latter curves are centred on their common average.
Δ F L k B T L = ln P L ( E 1 ) P L ( E 2 ) P L ( E v ) (93) \frac{\Delta F_L}{k_{\mathrm{B}}T_L} = \ln\frac{\sqrt{P_L(E_1)P_L(E_2)}}{P_L(E_{\mathrm{v}})} \tag*{(93)} k B T L Δ F L = ln P L ( E v ) P L ( E 1 ) P L ( E 2 ) (93)
where E 1 E_1 E 1 , E 2 E_2 E 2 are the two peak positions and E v E_{\mathrm{v}} E v is the intervening valley. For L = 4 , 6 , 8 , 10 L = 4, 6, 8, 10 L = 4 , 6 , 8 , 10 , where repeated passages were confirmed in the umbrella or auxiliary fixed-bias runs, the peaks remain separated by 0.44–0.45 energy units per site and the barrier rises from 3.92 to 64.93 in units of k B T L k_{\mathrm{B}}T_L k B T L . These are finite-size first-order signatures. The L = 12 , 16 L = 12, 16 L = 12 , 16 data have limited mixing. For a periodic box with two planar interfaces of tension σ \sigma σ , the leading area contribution would be 2 σ L 2 2\sigma L^2 2 σ L 2 . The L 2 L^2 L 2 line is a finite-range guide. Figure 3 e labels samples with ∣ q ∣ < 0.18 |\mathbf{q}| < 0.18 ∣ q ∣ < 0.18 as cubic-like and the remaining samples within 10 ∘ 10^\circ 1 0 ∘ of a preferred ray as tetragonal-like. The response estimators available from the same histograms are c V = L 3 Var ( E / L 3 ) / ( k B T 2 ) c_V = L^3 \operatorname{Var}(E/L^3)/(k_{\mathrm{B}}T^2) c V = L 3 Var ( E / L 3 ) / ( k B T 2 ) and χ m = L 3 Var ( ∣ q ∣ ) / ( k B T ) \chi_m = L^3 \operatorname{Var}(|\mathbf{q}|)/(k_{\mathrm{B}}T) χ m = L 3 Var ( ∣ q ∣ ) / ( k B T ) . Here χ m \chi_m χ m measures fluctuations of the radial amplitude. The linear susceptibility to a vector field conjugate to q \mathbf{q} q is instead the component covariance L 3 [ ⟨ q α q β ⟩ − ⟨ q α ⟩ ⟨ q β ⟩ ] / ( k B T ) L^3[\langle q_\alpha q_\beta\rangle- \langle q_\alpha\rangle\langle q_\beta\rangle]/(k_{\mathrm{B}}T) L 3 [⟨ q α q β ⟩ − ⟨ q α ⟩ ⟨ q β ⟩] / ( k B T ) . The observable and its conjugate source must be specified when interpreting these fluctuation estimators as response functions.
The two paths in Figure 3 f isolate the distinction in § 2.2 . Path A uses a convex harmonic lattice with local energy ∣ q i ∣ 2 / 2 |\mathbf{q}_i|^2/2 ∣ q i ∣ 2 /2 and field term − h ⋅ q i -\mathbf{h}\cdot\mathbf{q}_i − h ⋅ q i , where h = ( 0.65 , h y ) \mathbf{h} = (0.65,h_y) h = ( 0.65 , h y ) . Shifting all q i \mathbf{q}_i q i by h \mathbf{h} h leaves the neighbour differences invariant. Consequently the free-energy density and the mean satisfy
f ( h y ) − f ( 0 ) = − h y 2 2 , ⟨ q ⟩ = h . (94) f(h_y) - f(0) = -\frac{h_y^2}{2}, \qquad\langle\mathbf{q}\rangle= \mathbf{h}. \tag*{(94)} f ( h y ) − f ( 0 ) = − 2 h y 2 , ⟨ q ⟩ = h . (94)
The mean passes through D 2 h D_{2h} D 2 h , D 4 h D_{4h} D 4 h , and D 2 h D_{2h} D 2 h as h y h_y h y crosses zero. The free energy is analytic at every size and in the thermodynamic limit. Path B compares the free energies restricted to the two basins of Eq. (92) . Their exchange identifies the equal-weight condition while the finite-volume partition function remains smooth.
Table 2 reports the finite-size estimates together with the passage diagnostics. The energy-density peak separation varies little over the simulated sizes, while the barrier grows substantially. The independent fixed-bias checks at L = 8 , 10 L = 8, 10 L = 8 , 10 support the corresponding umbrella estimates. At L = 12 L = 12 L = 12 the two checks give barriers of approximately 96.5 and 61.2, so that the small spread of the umbrella estimates understates the sampling ambiguity. The filled-symbol trend in Figure 3 d therefore uses the diagnostic-supported sizes L = 6 , 8 , 10 L = 6, 8, 10 L = 6 , 8 , 10 .
Table 2. Finite-size coexistence diagnostics for Figure 3 . The barrier central value uses the pooled umbrella reconstruction. Its uncertainty is the sample standard deviation across three independent umbrella ladders.
The peak gap is the separation of the energy-density maxima in model energy units. The L = 12 , 16 L = 12, 16 L = 12 , 16 rows are exploratory estimates with limited mixing. The present size range supports a finite-range area trend for the diagnostic-supported runs.
Compatible three-dimensional microstructure
To extend the proper-strain case in § 4.2 to spatially varying states, the periodic phase-field model takes q \mathbf{q} q to be the E g E_g E g part of the strain itself, expressed in units of s s s . The dimensionless energy and local density are
F [ q ] = ∫ Ω [ f ( q ) + κ 2 ∣ ∇ q ∣ 2 ] d V + F c o m p a t , f ( q ) = a 2 ∣ q ∣ 2 + b 3 ( 3 q 1 2 q 2 − q 2 3 ) + c 4 ∣ q ∣ 4 , (95) F[\mathbf{q}] = \int_{\Omega}\left[f(\mathbf{q}) + \frac{\kappa}{2}|\nabla\mathbf{q}|^2\right]\,\mathrm{d}V + F_{\mathrm{compat}}, \qquad f(\mathbf{q}) = \frac{a}{2}|\mathbf{q}|^2 + \frac{b}{3}(3q_1^2q_2 - q_2^3) + \frac{c}{4}|\mathbf{q}|^4, \tag*{(95)} F [ q ] = ∫ Ω [ f ( q ) + 2 κ ∣∇ q ∣ 2 ] d V + F compat , f ( q ) = 2 a ∣ q ∣ 2 + 3 b ( 3 q 1 2 q 2 − q 2 3 ) + 4 c ∣ q ∣ 4 , (95)
with ( a , b , c ) = ( − 0.50 , − 1.50 , 2 ) (a,b,c) = (-0.50,-1.50,2) ( a , b , c ) = ( − 0.50 , − 1.50 , 2 ) . At k ≠ 0 \mathbf{k} \ne0 k = 0 , let n = k / ∣ k ∣ \mathbf{n} = \mathbf{k}/|\mathbf{k}| n = k /∣ k ∣ and ε ^ / s = sym ( n ⊗ v ) \hat{\varepsilon}/s = \operatorname{sym}(\mathbf{n}\otimes\mathbf{v}) ε ^ / s = sym ( n ⊗ v ) . This is the Fourier representation of the Saint-Venant condition for compatible linear strain. Set
A α i = B i α n i , Q = μ I + ( λ + μ ) n ⊗ n , G = A Q − 1 A T , P = G G + . (96) A_{\alpha i}=B_{i\alpha}n_i,\quad\mathbf{Q}=\mu\mathbf{I}+(\lambda+\mu)\mathbf{n}\otimes\mathbf{n},\quad\mathbf{G}=\mathbf{A}\mathbf{Q}^{-1}\mathbf{A}^{\mathsf{T}},\quad\mathbf{P}=\mathbf{G}\mathbf{G}^{+}. \tag*{(96)} A α i = B i α n i , Q = μ I + ( λ + μ ) n ⊗ n , G = A Q − 1 A T , P = G G + . (96)
Minimizing the auxiliary isotropic elastic energy subject to A v = q ^ \mathbf{A}\mathbf{v}=\hat{\mathbf{q}} Av = q ^ gives v = Q − 1 A T G + q ^ \mathbf{v}=\mathbf{Q}^{-1}\mathbf{A}^{\mathsf{T}}\mathbf{G}^{+}\hat{\mathbf{q}} v = Q − 1 A T G + q ^ . Removing the auxiliary E g E_g E g penalty, already represented in the local potential, leaves the compatibility kernel
U ( n ) = G + − 2 μ P , q ^ ∈ range A , F c o m p a t = ∣ Ω ∣ 2 ∑ k ≠ 0 q ^ † U ( n ) q ^ . (97) \mathbf{U}(\mathbf{n})=\mathbf{G}^{+}-2\mu\mathbf{P},\qquad\hat{\mathbf{q}}\in\operatorname{range}\mathbf{A},\qquad F_{\mathrm{compat}}=\frac{|\Omega|}{2}\sum_{\mathbf{k}\ne0}\hat{\mathbf{q}}^{\dagger}\mathbf{U}(\mathbf{n})\hat{\mathbf{q}}. \tag*{(97)} U ( n ) = G + − 2 μ P , q ^ ∈ range A , F compat = 2 ∣Ω∣ k = 0 ∑ q ^ † U ( n ) q ^ . (97)
where the Fourier coefficients are normalized by ∣ Ω ∣ − 1 |\Omega|^{-1} ∣Ω ∣ − 1 . The pseudoinverse handles the rank-one modes along cubic axes. The homogeneous mode is clamped in the matched comparison. This construction follows the general strain-compatibility approach for ferroelastics [28 ] .
The elimination can be checked with a multiplier Λ \mathbf{\Lambda} Λ . For the auxiliary energy 1 2 v † Q v \frac{1}{2}\mathbf{v}^{\dagger}\mathbf{Q}\mathbf{v} 2 1 v † Qv and constraint A v = q ^ \mathbf{A}\mathbf{v}=\hat{\mathbf{q}} Av = q ^ , stationarity gives Q v = A T Λ \mathbf{Q}\mathbf{v}=\mathbf{A}^{\mathsf{T}}\mathbf{\Lambda} Qv = A T Λ and G Λ = q ^ \mathbf{G}\mathbf{\Lambda}=\hat{\mathbf{q}} GΛ = q ^ . The minimum energy is therefore 1 2 q ^ † G + q ^ \frac{1}{2}\hat{\mathbf{q}}^{\dagger}\mathbf{G}^{+}\hat{\mathbf{q}} 2 1 q ^ † G + q ^ . Orthogonality of the two diagonal strain basis tensors makes their auxiliary energy μ ∣ q ^ ∣ 2 \mu|\hat{\mathbf{q}}|^2 μ ∣ q ^ ∣ 2 ; subtracting it gives Eq. (97) . This remaining energy is nonnegative on range A \mathbf{A} A . Components outside that range are incompatible and must be excluded, rather than assigned a finite energy by a pseudoinverse. The projector P \mathbf{P} P in the evolution enforces this restriction [52 ] .
We evolved the two components by projected semi-implicit gradient flow with λ = μ = 4 \lambda=\mu=4 λ = μ = 4 and gradient coefficient κ = 0.85 \kappa=0.85 κ = 0.85 . The primary comparison uses 240 steps of size 0.78, unit structural mobility, and stabilization coefficient 2. The gradient and compatibility terms are implicit, and the local potential derivative is explicit with the matching stabilization term. Both grids resolve the same physical box and initial field. The mean doublet is clamped to zero, and the continuous-derivative Nyquist planes are set to zero in both comparisons. On the 128 3 128^3 12 8 3 grid, 83.8% of candidate compatible walls have normals within 15 ∘ 15^\circ 1 5 ∘ of their ideal pair-specific { 110 } \{110\} { 110 } directions, compared with 10.8% in the local control initialized with the same field. The 256 3 256^3 25 6 3 compatible result is 82.8%. Both compatible runs give a normalized Fourier compatibility residual below 1.6 × 10 − 7 1.6 \times10^{-7} 1.6 × 1 0 − 7 . Let v j \mathbf{v}_j v j be unit vectors along the three variant rays. The distance from the E g E_g E g projection to the closure of the tetragonal stratum and the frozen local [100] squared-speed gap used in Figure 4 are
Figure 4. Three-dimensional microstructure under nonlocal elastic compatibility. (a) Cutaway coloured by the nearest tetragonal variant. (b) Pairwise walls and junction region in a selected subvolume. (c) Finite-tolerance stabilizer classes of the E g E_g E g projection. (d) Distance to the closure of the tetragonal stratum. (e) Frozen local squared-speed splitting along [100] from Eq. (98) . (f) Matched mid-plane cuts and transverse plane-wave spectral measures with and without compatibility. The spectra use the same frozen constitutive law for both microstructures.
d 4 h / s = min j ∣ q − ( v j ⋅ q ) v j ∣ , Δ c T 2 ( [ 100 ] ) = γ ρ 0 ∣ d z − d y ∣ , d i = ( B q ) i . (98) d_{4h/s} = \min_j \left\lvert\mathbf{q} - (\mathbf{v}_j \cdot\mathbf{q})\mathbf{v}_j \right\rvert, \qquad\Delta c_{\mathrm{T}}^2([100]) = \frac{\gamma}{\rho_0} \lvert d_z - d_y \rvert, \qquad d_i = (\mathbf{B}\mathbf{q})_i. \tag*{(98)} d 4 h / s = j min ∣ q − ( v j ⋅ q ) v j ∣ , Δ c T 2 ([ 100 ]) = ρ 0 γ ∣ d z − d y ∣ , d i = ( Bq ) i . (98)
The generic D 2 h D_{2h} D 2 h stratum is dense in the doublet plane, so its distance is identically zero. The speed gap uses the illustrative frozen constitutive map ( C y z y z , C x z x z , C x y x y ) = μ 0 + γ ( d x , d y , d z ) (C_{yzyz}, C_{xzxz}, C_{xyxy}) = \mu_0 + \gamma(d_x, d_y, d_z) ( C y z y z , C x z x z , C x y x y ) = μ 0 + γ ( d x , d y , d z ) , with μ 0 = 4 \mu_0 = 4 μ 0 = 4 , γ = 2.3 \gamma= 2.3 γ = 2.3 , and ρ 0 = 1 \rho_0 = 1 ρ 0 = 1 . The classes in Figure 4 c apply tolerances ∣ q ∣ < 0.16 \lvert\mathbf{q} \rvert< 0.16 ∣ q ∣ < 0.16 and d 4 h / s < 0.075 d_{4h/s} < 0.075 d 4 h / s < 0.075 in this projected state space.
For the spectrum in Figure 4 f, the same frozen law defines the spatial wave operator K = ρ 0 − 1 D ∗ C ( x ) D \mathbf{K} = \rho_0^{-1}\mathcal{D}^{*}\mathbb{C}(\mathbf{x})\mathcal{D} K = ρ 0 − 1 D ∗ C ( x ) D , where D \mathcal{D} D is the periodic discrete symmetric gradient. The plotted density is the spectral measure of K \sqrt{\mathbf{K}} K projected on the transverse plane-wave inputs. At the selected [100] wavevector on the 64 3 64^3 6 4 3 wave grid, its normalized width is 0.0995 with compatibility and 0.1404 in the local control. These widths describe the mode distribution in a finite periodic body.
Table 3 compares the compatible and local evolutions under the same clamped mean strain, physical box of side 64, and smooth initial field. Refining the compatible grid changes the selected-wall fraction by about one percentage point, with very small field discrepancy after comparison on the same 64 3 64^3 6 4 3 grid. The frozen wave calculation uses the identical strain-dependent stiffness map for both microstructures. Its normalized spectral width is reduced by approximately 29.2% in the compatible case. This comparison connects domain arrangement to the plane-wave spectral measure under the chosen frozen constitutive map. Here the wave stiffness is prescribed as an observation law, separately from the auxiliary elastic energy used to construct the compatibility kernel. The comparison therefore tests this specified microstructure-to-spectrum map. It does not establish that the wave operator is the Hessian of the microstructure-generating functional.
Table 3. Compatibility, discretization and frozen-wave checks for Figure 4 . All cases use the same physical domain. Percentages refer to the candidate pairwise walls.
The residual is ∥ P ⊥ E ^ P ⊥ ∥ / ∥ E ^ ∥ \lVert P_{\perp}\hat{E}P_{\perp}\rVert/\lVert\hat{E}\rVert ∥ P ⊥ E ^ P ⊥ ∥ / ∥ E ^ ∥ , evaluated over nonzero Fourier modes with P ⊥ = I − n ⊗ n P_{\perp} = I - \mathbf{n} \otimes\mathbf{n} P ⊥ = I − n ⊗ n . Spectral widths are standard deviations of the projected frequencies before visualization smoothing, divided by the homogeneous reference frequency. An em dash marks an unevaluated quantity.
Polarization-resolved finite-temperature lattice spectra
We test the directional rules in Eqs. (85) and (86) with classical molecular dynamics on a monatomic face-centred cubic bond network. The homogeneous cell is assigned to the cubic or z z z -tetragonal branch of Figure 2 . Each trajectory evolves at fixed cell shape. The bond energy is k b δ b 2 / 2 + α δ b 4 / 4 k_b\delta_b^2/2 + \alpha\delta_b^4/4 k b δ b 2 /2 + α δ b 4 /4 with α = 1 \alpha= 1 α = 1 , δ b = n b ⋅ ( u j − u i ) \delta_b = n_b \cdot(\mathbf{u}_j - \mathbf{u}_i) δ b = n b ⋅ ( u j − u i ) , and n b = R b / ∣ R b ∣ \mathbf{n}_b = \mathbf{R}_b/\lvert\mathbf{R}_b\rvert n b = R b / ∣ R b ∣ along the deformed reference bond. The six reference-diagonal bonds have k b = 4.5 + 2.10 d b k_b = 4.5 + 2.10d_b k b = 4.5 + 2.10 d b , where d b d_b d b is the reduced strain normal to the bond plane. The three axial bonds have k b = 4.5 k_b = 4.5 k b = 4.5 . The harmonic frequency-squared matrix in unit mass is
D 0 ( k ) = 2 ∑ b k b [ 1 − cos ( k ⋅ R b ) ] n b ⊗ n b , A 0 ( n ) = lim k → 0 D 0 ( k n ) / k 2 . (99) \mathbf{D}_0(\mathbf{k}) = 2 \sum_b k_b[1-\cos(\mathbf{k}\cdot\mathbf{R}_b)]\,\mathbf{n}_b \otimes\mathbf{n}_b, \qquad\mathbf{A}_0(\mathbf{n}) = \lim_{k\to0}\mathbf{D}_0(k\mathbf{n})/k^2. \tag*{(99)} D 0 ( k ) = 2 b ∑ k b [ 1 − cos ( k ⋅ R b )] n b ⊗ n b , A 0 ( n ) = k → 0 lim D 0 ( k n ) / k 2 . (99)
The sum uses one bond per undirected direction per site. Dashed dispersion curves use D 0 \mathbf{D}_0 D 0 and the directional maps use A 0 \mathbf{A}_0 A 0 . The standard simulation uses 2048 atoms in an 8 3 8^3 8 3 conventional-cell box. Each branch is equilibrated before a microcanonical production run. The original temperature series uses 4200 BAOAB Langevin steps with unit friction, followed by velocity-Verlet production at Δ t = 0.012 \Delta t = 0.012 Δ t = 0.012 with 8192 frames spaced by 0.072 model time units. The updated T = 0.967 T = 0.967 T = 0.967 point uses 24 independent states equilibrated for 33334 steps at Δ t = 0.003 \Delta t = 0.003 Δ t = 0.003 , with unit friction. Each production record then has 32768 frames spaced by 0.072, for duration 2359.296. After removing centre-of-mass motion, the same equilibrated state is copied to production at Δ t = 0.012 , 0.006 , 0.003 \Delta t = 0.012, 0.006, 0.003 Δ t = 0.012 , 0.006 , 0.003 . The main near-transition point uses the finest step. Mass and k B k_{\mathrm{B}} k B are one. The common path in the reduced reciprocal coordinates of the conventional cell is X x X_x X x –Γ \Gamma Γ –X z X_z X z –W z W_z W z –L L L –Γ \Gamma Γ –K K K .
Table 4 tests the two directional selection rules over the sampled temperatures and records the principal resolution checks at T = 0.967 T = 0.967 T = 0.967 . The tetragonal states give a positive [100] peak difference, while the [001] estimates fluctuate near zero. Above the assigned branch-selection temperature, both axial differences fluctuate near zero. At T = 0.967 T = 0.967 T = 0.967 the ratio of the [100] gap to the linewidth sum is approximately 0.41. A matched-state and matched-duration calculation now compares 24 paired replicas at three integration steps. The [100] gap difference between Δ t = 0.006 \Delta t = 0.006 Δ t = 0.006 and 0.003 is −0.0030, with a paired 95% bootstrap interval [−0.0093, 0.0036]. This interval lies within the preset absolute tolerance [−0.01, 0.01] in model angular-frequency units. The near-transition estimate therefore satisfies this finite-precision step criterion. The earlier six-replica, short-record comparison is superseded by this controlled calculation.
Table 4. Polarization-resolved MD peak differences and numerical controls for Figure 5 , at reduced wavevector 0.5. The main data use 2048 atoms. Frequencies are in model angular-frequency units and quoted errors are one trajectory-jackknife standard error.
The signed gaps are ω y − ω z \omega_y - \omega_z ω y − ω z along [100] and ω x − ω y \omega_x - \omega_y ω x − ω y along [001]. The tabulated FWHM values refer to the Lorentzian before window convolution. The original temperature points and enlarged-box check retain their original protocol. The refined T = 0.967 T = 0.967 T = 0.967 data have spacing 0.00266 and a single-mode Hann reference FWHM of 0.00384. The matched-state controls all use duration 2359.296 and the same sampling interval, with trajectory-jackknife errors. The half refined record has duration 1179.648. An em dash marks an inapplicable quantity.
The MD observable is a polarization-resolved velocity power spectrum (Thomas et al., [29 ] ). For a unit polarization e \mathbf{e} e , reference sites R i \mathbf{R}_i R i , and a Hann taper w ( t ) w(t) w ( t ) , its positive-frequency contribution is proportional to
S e ( k , ω ) ∝ ⟨ ∣ ∫ w ( t ) e i ω t ∑ i e − i k ⋅ R i e ⋅ u ˙ i ( t ) d t ∣ 2 ⟩ . (100) S_{\mathbf{e}}(\mathbf{k},\omega) \propto\left\langle\left| \int w(t)e^{i\omega t}\sum_i e^{-i\mathbf{k}\cdot\mathbf{R}_i}\mathbf{e}\cdot\dot{\mathbf{u}}_i(t)\,\mathrm{d}t \right|^2 \right\rangle. \tag*{(100)} S e ( k , ω ) ∝ ⟨ ∫ w ( t ) e iω t i ∑ e − i k ⋅ R i e ⋅ u ˙ i ( t ) d t 2 ⟩ . (100)
The estimate combines both temporal frequency signs and averages independent trajectories. The original record spacing in frequency is 0.01065. At T = 0.967 T = 0.967 T = 0.967 the refined records give 0.00266, with a single-mode Hann reference width of 0.00384. For peak initialization, smoothing uses the larger of two bins and 0.02 model angular-frequency units, so that a longer record retains a comparable physical smoothing scale. This smoothing initializes the fit to the unsmoothed power. Peak positions and linewidths are fitted independently to the polarizations using Lorentzians convolved with the window response.
On the cubic branch the axial little group is C 4 v C_{4v} C 4 v for both directions. On the z z z -tetragonal branch, the [001] little group retains its two-dimensional transverse representation. Along [100] it reduces to C 2 v C_{2v} C 2 v , whose transverse polarizations belong to distinct representations. These symmetry constraints also apply to the ensemble-averaged velocity correlations within each branch. Figure 5 summarizes the resulting polarization-resolved spectral signatures. The cubic and z z z -tetragonal dispersions in Figure 5 a,Figure 5 b distinguish the symmetry-protected transverse doublet from the branch-dependent splitting along the common reciprocal-space path. The polarization-resolved [100] profiles in Figure 5 c resolve the two fitted peak centres, whose signed difference is compared with harmonic references in Figure 5 d. The directional squared-speed gap in Figure 5 e provides the complementary long-wavelength symmetry diagnostic. Finally, Figure 5 f compares the finite-temperature [100] peak separation with the combined linewidth and the window-resolution reference, distinguishing a nonzero polarization-resolved splitting from two peaks that would necessarily be resolved in a polarization-summed spectrum. At T = 0.967 T = 0.967 T = 0.967 and reduced wavevector 0.5, the signed MD peak difference along [100] is 0.121 ± \pm ± 0.002, whereas the [001] difference is − 0.003 ± 0.003 -0.003 \pm0.003 − 0.003 ± 0.003 . The uncertainty is a one-standard-error trajectory jackknife. Along [100] the sum of fitted linewidths is 0.296 ± \pm ± 0.007. Thus the split channels overlap in the polarization-summed spectrum. The bare finite-wavevector splitting at the same state is 0.444. An enlarged box of 6912 atoms gives 0.100 ± \pm ± 0.024$ for
Figure 5. Protected and lifted transverse degeneracies in a finite-temperature lattice model. (a,b) Velocity power on a common reduced reciprocal-space path for cubic and z z z -tetragonal cells. Dashed curves are the harmonic eigenfrequencies. (c) Transverse power along [100], centred on the fitted peak midpoint at each wavevector. Circles indicate the two fits at reduced wavevector 0.5, horizontally offset for visibility. (d) Signed MD peak differences and the harmonic finite-wavevector and long-wave references. (e) Directional squared-speed gap of the two quasi-transverse acoustic branches. (f) Absolute [100] gap and sum of fitted linewidths, with the summed record-specific Hann-window width dotted. The T = 0.967 T = 0.967 T = 0.967 point and tetragonal spectra use the refined trajectories described in Table 4 . The right-hand profiles show independent polarizations at T = 0.967 T = 0.967 T = 0.967 with fitted centres marked at the baselines.
the [100] MD gap at the same reduced wavevector. These comparisons separate the symmetry constraint, the undamped reference, and the anharmonic fitted peak positions.
Independent static exponents and held-out spectral scaling
To test Eq. (51) with non-mean-field fluctuations, we sampled the periodic nearest-neighbour Ising model H = − J ∑ ⟨ i , j ⟩ s i s j \mathcal{H}=-J\sum_{\langle i,j\rangle}s_i s_j H = − J ∑ ⟨ i , j ⟩ s i s j in three dimensions using Wolff updates on sizes L = 8 , 12 , 16 , 24 , 32 , 48 L=8,12,16,24,32,48 L = 8 , 12 , 16 , 24 , 32 , 48 with J = k B = 1 J=k_{\mathrm{B}}=1 J = k B = 1 . Half of the independent chains were used to determine the static parameters, using eight of the sixteen chains at each size. The main chains are sampled at T 0 = 4.5115 T_0=4.5115 T 0 = 4.5115 with one stored observation every two Wolff cluster flips. Single-histogram reweighting supplies the nearby temperatures used in the fits. With m = L − 3 ∑ i s i m=L^{-3}\sum_i s_i m = L − 3 ∑ i s i , the Binder cumulant and second-moment length are
U 4 = 1 − ⟨ m 4 ⟩ 3 ⟨ m 2 ⟩ 2 , ξ L = 1 2 sin ( π / L ) [ S ( 0 ) S ( k min ) − 1 ] 1 / 2 , S ( 0 ) = L 3 ⟨ m 2 ⟩ . (101) U_4=1-\frac{\langle m^4\rangle}{3\langle m^2\rangle^2},\qquad\xi_L=\frac{1}{2\sin(\pi/L)}\left[\frac{S(0)}{S(k_{\min})}-1\right]^{1/2},\qquad S(0)=L^3\langle m^2\rangle. \tag*{(101)} U 4 = 1 − 3 ⟨ m 2 ⟩ 2 ⟨ m 4 ⟩ , ξ L = 2 sin ( π / L ) 1 [ S ( k m i n ) S ( 0 ) − 1 ] 1/2 , S ( 0 ) = L 3 ⟨ m 2 ⟩ . (101)
The slopes of U 4 U_4 U 4 and ξ L / L \xi_L/L ξ L / L near the common crossing scale as L 1 / ν L^{1/\nu} L 1/ ν , up to corrections. Together with ⟨ ∣ m ∣ ⟩ T c ∝ L − β / ν \langle|m|\rangle_{T_c}\propto L^{-\beta/\nu} ⟨ ∣ m ∣ ⟩ T c ∝ L − β / ν , the fit on L ≥ 12 L\geq12 L ≥ 12 gives T c = 4.51190 ± 0.00054 T_c=4.51190\pm0.00054 T c = 4.51190 ± 0.00054 , ν = 0.626 ± 0.041 \nu=0.626\pm0.041 ν = 0.626 ± 0.041 , and β = 0.322 ± 0.023 \beta=0.322\pm0.023 β = 0.322 ± 0.023 . The fits include a leading correction proportional to L − 0.83 L^{-0.83} L − 0.83 . The uncertainties use whole-chain resampling. Figure 6 c compares different lower size cutoffs and established Ising reference values [30 ] .
Figure 6. Critical universality and symmetry-selected spectral scaling. (a) Magnetization collapse using parameters from the static training chains. (b) Binder cumulant and correlation-length ratio, with the common fitted critical temperature dotted. (c) Static exponents across lower size cutoffs. The filled symbols and bars mark the primary estimates and bootstrap standard errors. (d) Eigenbranches of the direct and mediated probes. (e) Their transverse gaps at fixed k k k versus magnetization amplitude. (f) Critical size dependence and spectral collapse on held-out chains. The dashed guides on the left use the static estimate of β / ν \beta/\nu β / ν multiplied by p = 1 p=1 p = 1 or 2 2 2 .
The two auxiliary probes in Figure 6 d are prescribed algebraic readouts of the sampled magnetization and share a transverse doublet at m = 0 m=0 m = 0 . Their dimensionless stiffness matrices are
D A k 2 = ( v T 2 g m g m v T 2 ) , D B k 2 = ( v T 2 0 g m 0 v T 2 0 g m 0 v L 2 ) , v L > v T . (102) \frac{\mathbf{D}_{A}}{k^2}=\begin{pmatrix}v_T^2&gm\\gm&v_T^2\end{pmatrix},\qquad\frac{\mathbf{D}_{B}}{k^2}=\begin{pmatrix}v_T^2&0&gm\\0&v_T^2&0\\gm&0&v_L^2\end{pmatrix},\qquad v_L>v_T. \tag*{(102)} k 2 D A = ( v T 2 g m g m v T 2 ) , k 2 D B = v T 2 0 g m 0 v T 2 0 g m 0 v L 2 , v L > v T . (102)
The numerical probes use v T = 1.50 v_T=1.50 v T = 1.50 , v L = 2.20 v_L=2.20 v L = 2.20 , and g = 1.05 g=1.05 g = 1.05 . Their positivity conditions are ∣ g m ∣ < v T 2 |gm|<v_T^2 ∣ g m ∣ < v T 2 for A and g 2 m 2 < v T 2 v L 2 g^2m^2<v_T^2v_L^2 g 2 m 2 < v T 2 v L 2 for B. These conditions hold throughout the sampled Ising range ∣ m ∣ ≤ 1 |m|\leq1 ∣ m ∣ ≤ 1 , so both readouts have real positive frequencies at every sampled magnetization.
Both obey D ( − m ) = S D ( m ) S − 1 \mathbf{D}(-m)=\mathbf{S}\mathbf{D}(m)\mathbf{S}^{-1} D ( − m ) = SD ( m ) S − 1 , with S = diag ( 1 , − 1 ) \mathbf{S}=\operatorname{diag}(1,-1) S = diag ( 1 , − 1 ) and diag ( 1 , 1 , − 1 ) \operatorname{diag}(1,1,-1) diag ( 1 , 1 , − 1 ) , respectively. The equal transverse reference stiffness is imposed in both probes; spin-inversion symmetry alone does not enforce this degeneracy. In B, the same parity of the transverse states rules out terms odd in m m m in the transverse block. Diagonalization gives
Δ ω A k = ∣ g ∣ v T ∣ m ∣ + O ( ∣ m ∣ 3 ) , Δ ω B k = g 2 m 2 2 v T ( v L 2 − v T 2 ) + O ( m 4 ) . (103) \frac{\Delta\omega_A}{k}=\frac{|g|}{v_T}|m|+O(|m|^3),\qquad\frac{\Delta\omega_B}{k}=\frac{g^2m^2}{2v_T(v_L^2-v_T^2)}+O(m^4). \tag*{(103)} k Δ ω A = v T ∣ g ∣ ∣ m ∣ + O ( ∣ m ∣ 3 ) , k Δ ω B = 2 v T ( v L 2 − v T 2 ) g 2 m 2 + O ( m 4 ) . (103)
For A the two squared frequencies are k 2 ( v T 2 ± g m ) k^2(v_T^2\pm gm) k 2 ( v T 2 ± g m ) . For B, the uncoupled transverse eigenvalue is v T 2 v_T^2 v T 2 , and the lower eigenvalue of the coupled block is
λ − ( m ) = v T 2 + v L 2 − ( v L 2 − v T 2 ) 2 + 4 g 2 m 2 2 = v T 2 − g 2 m 2 v L 2 − v T 2 + O ( m 4 ) . (104) \lambda_-(m)=\frac{v_T^2+v_L^2-\sqrt{(v_L^2-v_T^2)^2+4g^2m^2}}{2}=v_T^2-\frac{g^2m^2}{v_L^2-v_T^2}+O(m^4). \tag*{(104)} λ − ( m ) = 2 v T 2 + v L 2 − ( v L 2 − v T 2 ) 2 + 4 g 2 m 2 = v T 2 − v L 2 − v T 2 g 2 m 2 + O ( m 4 ) . (104)
Taking the square roots gives Eq. (103) within the positive-stiffness range. If the critical magnetization density obeys P L ( m ) = L β / ν P ∗ ( m L β / ν ) P_L(m)=L^{\beta/\nu}\mathcal{P}_*(mL^{\beta/\nu}) P L ( m ) = L β / ν P ∗ ( m L β / ν ) up to scaling corrections, substitution of x = m L β / ν x=mL^{\beta/\nu} x = m L β / ν into the moment integral gives ⟨ ∣ m ∣ p ⟩ = L − p β / ν ∫ ∣ x ∣ p P ∗ ( x ) d x \langle|m|^p\rangle=L^{-p\beta/\nu}\int|x|^p\mathcal{P}_*(x)\,\mathrm{d}x ⟨ ∣ m ∣ p ⟩ = L − pβ / ν ∫ ∣ x ∣ p P ∗ ( x ) d x . This is the moment-level origin of the two gap exponents; it does not replace the average gap by the gap at the averaged signed magnetization (Xu et al., 2019). For the held-out chains, we average the positive gap of each configuration at fixed k k k . Finite-size scaling then predicts ⟨ Δ ω ⟩ / k ∝ L − β / ν \langle\Delta\omega\rangle/k\propto L^{-\beta/\nu} ⟨ Δ ω ⟩ / k ∝ L − β / ν at T c T_c T c , with a collapse coordinate ( T − T c ) L 1 / ν / T c (T-T_c)L^{1/\nu}/T_c ( T − T c ) L 1/ ν / T c . The two independently fitted size exponents are 0.492 ± 0.065 0.492\pm0.065 0.492 ± 0.065 and 1.012 ± 0.101 1.012\pm0.101 1.012 ± 0.101 , consistent with the static predictions 0.514 0.514 0.514 and 1.028 1.028 1.028 . For the second channel, the measured ⟨ m 2 ⟩ / ⟨ ∣ m ∣ ⟩ 2 \langle m^2\rangle/\langle|m|\rangle^2 ⟨ m 2 ⟩ / ⟨ ∣ m ∣ ⟩ 2 is 1.22 1.22 1.22 –1.24 1.24 1.24 , so taking the mean before evaluating the probe would change the leading amplitude. The test concerns controlled eigenfrequency gaps evaluated on fluctuating configurations. As an ordered-side check, independent L = 32 L=32 L = 32 runs over 4.29 ≤ T ≤ 4.45 4.29\leq T\leq4.45 4.29 ≤ T ≤ 4.45 give effective temperature exponents of 0.304 0.304 0.304 for ∣ m ∣ |m| ∣ m ∣ and 0.308 0.308 0.308 and 0.587 0.587 0.587 for the two gaps. These finite-window slopes complement the critical size scaling.
Table 5 makes the independence of the static calibration and spectral test explicit. The static chains supply β / ν \beta/\nu β / ν and its uncertainty, while the other chains supply the two spectral size exponents. The differences between the measured size exponents and p β / ν p\beta/\nu pβ / ν are evaluated within the joint whole-chain bootstrap, including the uncertainty of T c T_c T c . Both differences are compatible with zero at the quoted precision. This comparison tests the moment-level prediction in Eq. (103) for two specified readouts.
Table 5. Independent static calibration and held-out spectral scaling in Figure 6 . The primary fits use L ≥ 12 L \ge12 L ≥ 12 , a leading L − 0.83 L^{-0.83} L − 0.83 correction, and 500 whole-chain bootstrap resamples. Errors are bootstrap standard errors.
At criticality, ⟨ m 2 ⟩ / ⟨ ∣ m ∣ ⟩ 2 \langle m^2\rangle/\langle|m|\rangle^2 ⟨ m 2 ⟩ / ⟨ ∣ m ∣ ⟩ 2 spans 1.22–1.24 over the sampled sizes. Consequently, replacing the second moment by the square of the mean amplitude would lower the leading quadratic-response amplitude by approximately 18–19%. Retaining the configuration-level gap preserves this fluctuation contribution.
For the gap rows, the static column is the prediction p β / ν p\beta/\nu pβ / ν , and the difference is the held-out estimate minus this prediction. Covariance is retained when computing the difference error, and differences are evaluated before rounding. Eight independent chains per size supply each of the static and held-out groups. An em dash marks a quantity outside that group’s assigned role.
Relaxing probes and nonequilibrium critical dynamics
For a scalar mechanical coordinate coupled to one relaxing internal variable, Eq. (54) takes the form
C e f f ( ω ) = C ∞ − g 2 a − i ω / Γ η , τ η = ( a Γ η ) − 1 , a > 0 , a C ∞ > g 2 . (105) C_{\mathrm{eff}}(\omega) = C_{\infty} - \frac{g^2}{a-\mathrm{i}\omega/\Gamma_{\eta}}, \qquad\tau_{\eta} = (a\Gamma_{\eta})^{-1}, \qquad a > 0,\quad aC_{\infty} > g^2. \tag*{(105)} C eff ( ω ) = C ∞ − a − i ω / Γ η g 2 , τ η = ( a Γ η ) − 1 , a > 0 , a C ∞ > g 2 . (105)
Here the time convention is e − i ω t e^{-\mathrm{i}\omega t} e − i ω t , so the loss modulus is − Im C e f f -\operatorname{Im} C_{\mathrm{eff}} − Im C eff . Figure 7 a,Figure 7 b use C ∞ = 4 C_{\infty} = 4 C ∞ = 4 , a = 1 a = 1 a = 1 , g = 1.35 g = 1.35 g = 1.35 , and mechanical damping γ m = 0.12 \gamma_m = 0.12 γ m = 0.12 . At low frequency the modulus tends to C ∞ − g 2 / a C_{\infty} - g^2/a C ∞ − g 2 / a and at high frequency to C ∞ C_{\infty} C ∞ . For unit mass and wave number with mechanical damping γ m \gamma_m γ m , the response H ( ω ) = [ C e f f ( ω ) − ω 2 − i γ m ω ] − 1 H(\omega) = [C_{\mathrm{eff}}(\omega) - \omega^2 - \mathrm{i}\gamma_m\omega]^{-1} H ( ω ) = [ C eff ( ω ) − ω 2 − i γ m ω ] − 1 agrees with the Fourier transform of the integrated coupled impulse response in Figure 7 b. The frequency changes the effective response while the symmetry of the underlying model remains fixed.
Figure 7. Nonequilibrium critical dynamics and limits of the frozen response. (a) Storage and loss moduli for three internal relaxation times. (b) Frozen, relaxed and coupled spectra, with numerical impulse results overlaid. (c) Oppositely signed coarse-grained domains during a three-dimensional Ising quench. (d) Crossover length at T c T_c T c and normalized frozen-probe gaps for adaptive and fixed spatial windows. The dashed length guide is the Model-A Kibble–Zurek reference. (e) Short-time spectra of the driven probe velocities. White dashed curves show the frozen reference. (f) Exponent deviation for the mediated probe under relaxation. The solid contour is the prescribed 0.06 exponent tolerance. The dashed contour equates the smallest storage-modulus gap proxy across the fitted rate interval with the prescribed linewidth sum.
The kinetic Ising calculations use local nonconserved heat-bath updates with T ( t ) = T c ( 1 − t / τ Q ) T(t) = T_c(1 - t/\tau_Q) T ( t ) = T c ( 1 − t / τ Q ) . For this driving schedule the fixed numerical value is T c = 4.511524 T_c = 4.511524 T c = 4.511524 , which lies inside the uncertainty interval of the independent static fit. The seven rates are τ Q = 48 , 96 , 192 , 384 , 768 , 1536 , 3072 \tau_Q = 48, 96, 192, 384, 768, 1536, 3072 τ Q = 48 , 96 , 192 , 384 , 768 , 1536 , 3072 sweeps. One sweep updates every spin once, providing the time unit for this local dynamics. We averaged 20 independent quenches at each of seven rates on 48 3 48^3 4 8 3 sites and checked two rates on 64 3 64^3 6 4 3 sites. At t = 0 t = 0 t = 0 we define ξ × \xi_\times ξ × by the first crossing of 35% of the initial nonzero-shell value of r 1 + η I C ( r ) r^{1+\eta_I}C(r) r 1 + η I C ( r ) , where C ( r ) C(r) C ( r ) is the shell-averaged spin correlation and η I = 0.03627 \eta_I = 0.03627 η I = 0.03627 is the Ising anomalous dimension [30 ] . For τ Q ≥ 192 \tau_Q \ge192 τ Q ≥ 192 sweeps the fitted exponent of ξ × ∼ τ Q α ^ ξ \xi_\times\sim\tau_Q^{\hat{\alpha}_{\xi}} ξ × ∼ τ Q α ^ ξ is 0.255 ± 0.019 0.255 \pm0.019 0.255 ± 0.019 . Changing the radial threshold yields 0.237 0.237 0.237 –0.263 0.263 0.263 . Using the static estimates above and the literature Model-A z = 2.0245 z = 2.0245 z = 2.0245 gives the Kibble–Zurek reference ν / ( 1 + z ν ) = 0.276 \nu/(1+z\nu) = 0.276 ν / ( 1 + z ν ) = 0.276 [31 , 32 ] .
For a linear ramp, let ϵ T = ( T − T c ) / T c = − t / τ Q \epsilon_T = (T - T_c)/T_c = -t/\tau_Q ϵ T = ( T − T c ) / T c = − t / τ Q . Equating the equilibrium relaxation time τ 0 ∣ ϵ T ∣ − z ν \tau_0|\epsilon_T|^{-z\nu} τ 0 ∣ ϵ T ∣ − z ν to the remaining driving time ∣ ϵ T / ϵ ˙ T ∣ = τ Q ∣ ϵ T ∣ |\epsilon_T/\dot{\epsilon}_T| = \tau_Q|\epsilon_T| ∣ ϵ T / ϵ ˙ T ∣ = τ Q ∣ ϵ T ∣ gives
∣ ϵ ^ T ∣ ∼ ( τ 0 τ Q ) 1 / ( 1 + z ν ) , ξ ^ ∼ τ Q ν / ( 1 + z ν ) , r ^ ∼ τ Q − β / ( 1 + z ν ) . (106) |\hat{\epsilon}_T| \sim\left(\frac{\tau_0}{\tau_Q}\right)^{1/(1+z\nu)}, \qquad\hat{\xi} \sim\tau_Q^{\nu/(1+z\nu)}, \qquad\hat{r} \sim\tau_Q^{-\beta/(1+z\nu)}. \tag*{(106)} ∣ ϵ ^ T ∣ ∼ ( τ Q τ 0 ) 1/ ( 1 + z ν ) , ξ ^ ∼ τ Q ν / ( 1 + z ν ) , r ^ ∼ τ Q − β / ( 1 + z ν ) . (106)
A degree-p p p frozen readout then has the conditional rate exponent p β / ( 1 + z ν ) p\beta/(1+z\nu) pβ / ( 1 + z ν ) when its spatial coarse-graining length remains a fixed multiple of ξ ^ \hat{\xi} ξ ^ and the dynamic scaling regime is reached. A fixed microscopic window probes a different moment and need not obey this same rate law [54 ] .
The same two frozen probes as in Eq. (102) are applied to locally averaged magnetization. For a Gaussian width of 0.55 ξ ^ 0.55\hat{\xi} 0.55 ξ ^ in lattice units, ⟨ Δ ω ⟩ ∼ τ Q − α \langle\Delta\omega\rangle\sim\tau_Q^{-\alpha} ⟨ Δ ω ⟩ ∼ τ Q − α gives α = 0.123 ± 0.010 \alpha= 0.123 \pm0.010 α = 0.123 ± 0.010 and 0.242 ± 0.022 0.242 \pm0.022 0.242 ± 0.022 for p = 1 , 2 p = 1, 2 p = 1 , 2 . The conditional estimates p β / ( 1 + z ν ) p\beta/(1+z\nu) pβ / ( 1 + z ν ) are 0.142 0.142 0.142 and 0.284 0.284 0.284 . A fixed Gaussian width of 3 lattice units gives − 0.204 ± 0.010 -0.204 \pm0.010 − 0.204 ± 0.010 for the second probe. The sign change reflects the different coarse-grained observable. For the transient spectra in Figure 7 e, the sampled block magnetization m B ( t ) m_B(t) m B ( t ) drives two damped oscillators with independent white-noise forcing with instantaneous squared frequencies 4 [ 2.25 ± 1.05 m B ( t ) ] 4[2.25 \pm1.05m_B(t)] 4 [ 2.25 ± 1.05 m B ( t )] . The plotted power uses a finite-time Hann-window transform of the velocities from the driven dynamics over 40 sweeps. The probe is a one-way readout of the spin dynamics.
For the two-channel parameter scan in Figure 7 f, the stiffer channel relaxes by λ δ C / ( 1 − i ω τ η ) \lambda\delta C/(1-i\omega\tau_\eta) λ δ C / ( 1 − iω τ η ) , where δ C > 0 \delta C > 0 δ C > 0 is the frozen stiffness contrast and 0 ≤ λ ≤ 1 0 \le\lambda\le1 0 ≤ λ ≤ 1 . The storage-modulus contrast is therefore
Δ C ′ ( ω ) = δ C [ 1 − λ 1 + ( ω τ η ) 2 ] , τ η ∝ ξ z η . (107) \Delta C'(\omega) = \delta C \left[1-\frac{\lambda}{1+(\omega\tau_\eta)^2}\right], \qquad\tau_\eta\propto\xi^{z_\eta}. \tag*{(107)} Δ C ′ ( ω ) = δ C [ 1 − 1 + ( ω τ η ) 2 λ ] , τ η ∝ ξ z η . (107)
Taking the square-root difference of the two storage moduli defines a gap proxy. This proxy is not, in general, the separation of the peaks or complex poles of the full dissipative response. The scan compares its fitted quench-rate exponent with 2 β / ( 1 + z ν ) 2\beta/(1+z\nu) 2 β / ( 1 + z ν ) and its magnitude with a prescribed linewidth sum. The contours mark an exponent-tolerance criterion and a prescribed detectability criterion for this storage-modulus proxy. Establishing resolved peaks in a material requires evaluation of the complete dissipative response and the measurement window.
Table 6 compares the measured rate exponents with the conditional Kibble–Zurek predictions using the static calibration and fixed literature value z = 2.0245 z = 2.0245 z = 2.0245 . The adaptive-window gap exponents lie below the central reference values over this finite rate interval. The fixed-window exponents have the opposite sign, showing that the spatial averaging scale changes the measured rate dependence.
Table 6. Nonequilibrium rate exponents in Figure 7 . All fits use τ Q ≥ 192 \tau_Q \ge192 τ Q ≥ 192 sweeps in the primary 48 3 48^3 4 8 3 runs, with 20 independent quenches per rate. Measured errors are trajectory-bootstrap standard errors.
Gap exponents use ⟨ Δ ω ⟩ ∝ τ Q − α \langle\Delta\omega\rangle\propto\tau_Q^{-\alpha} ⟨ Δ ω ⟩ ∝ τ Q − α . The reference uncertainties propagate the static bootstrap samples with z z z fixed. A fixed microscopic window has a distinct coarse-graining protocol. The em dashes mark cases outside the stated reference scaling hypothesis.
Changing the correlation threshold from 35% to 25% or 50% gives length exponents between 0.237 and 0.263. This spread quantifies an observable-definition sensitivity in addition to the bootstrap error. Independently, the numerical impulse response reproduces the scalar transfer spectrum with relative error 1.44 × 10 − 5 1.44 \times10^{-5} 1.44 × 1 0 − 5 , checking the mechanical elimination used in Figure 7 a,Figure 7 b.
Discussion
Experimental identification of the structural state
Let y = O ( η , ν ) \mathbf{y} = \mathcal{O}(\boldsymbol{\eta},\boldsymbol{\nu}) y = O ( η , ν ) collect measurable wave speeds, modal polarization projectors, stiffness contractions, thermographic observables, or diffraction-sensitive structural data. The vector ν \boldsymbol{\nu} ν denotes nuisance parameters such as density, prestress, calibration coefficients, and spatial texture. Near a regular state,
δ y = J η δ η + J ν δ ν , J η = ∂ O ∂ η . (108) \delta\mathbf{y} = J_{\eta}\delta\boldsymbol{\eta} + J_{\nu}\delta\boldsymbol{\nu}, \qquad J_{\eta} = \frac{\partial\mathcal{O}}{\partial\boldsymbol{\eta}}. \tag*{(108)} δ y = J η δ η + J ν δ ν , J η = ∂ η ∂ O . (108)
If nuisance parameters are independently fixed, full column rank of J η J_{\eta} J η is sufficient for first-order local distinguishability, though not global injectivity. If nuisance parameters vary, let Σ \Sigma Σ be the noise covariance and define
J ~ η = Σ − 1 / 2 J η , J ~ ν = Σ − 1 / 2 J ν , P ν = J ~ ν J ~ ν + . (109) \widetilde{J}_{\eta} = \Sigma^{-1/2}J_{\eta}, \qquad\widetilde{J}_{\nu} = \Sigma^{-1/2}J_{\nu}, \qquad P_{\nu} = \widetilde{J}_{\nu}\widetilde{J}_{\nu}^{+}. \tag*{(109)} J η = Σ − 1/2 J η , J ν = Σ − 1/2 J ν , P ν = J ν J ν + . (109)
The nuisance-projected information is
I η ∣ ν = J ~ η T ( I − P ν ) J ~ η . (110) \mathcal{I}_{\eta\mid\nu} = \widetilde{J}_{\eta}^{T}(\mathbf{I} - P_{\nu})\widetilde{J}_{\eta}. \tag*{(110)} I η ∣ ν = J η T ( I − P ν ) J η . (110)
To derive the projection, consider a Gaussian measurement model with parameter-independent Σ ≻ 0 \Sigma\succ0 Σ ≻ 0 . For a trial structural increment, nuisance optimization minimizes ∥ J ~ η δ η + J ~ ν δ ν ∥ 2 \left\lVert\widetilde{J}_{\eta}\delta\boldsymbol{\eta} + \widetilde{J}_{\nu}\delta\boldsymbol{\nu}\right\rVert^2 J η δ η + J ν δ ν 2 . The fitted nuisance increment removes the component in range J ~ ν \widetilde{J}_{\nu} J ν , leaving ( I − P ν ) J ~ η δ η (\mathbf{I} - P_{\nu})\widetilde{J}_{\eta}\delta\boldsymbol{\eta} ( I − P ν ) J η δ η . The squared residual is the quadratic form in Eq. (110) . If the nuisance block has full column rank, the same result is the ordinary Schur complement
I η ∣ ν = I η η − I η ν I ν ν − 1 I ν η , I a b = J a T Σ − 1 J b . (111) \mathcal{I}_{\eta\mid\nu} = \mathcal{I}_{\eta\eta} - \mathcal{I}_{\eta\nu}\mathcal{I}_{\nu\nu}^{-1}\mathcal{I}_{\nu\eta}, \qquad\mathcal{I}_{ab} = J_a^T\Sigma^{-1}J_b. \tag*{(111)} I η ∣ ν = I η η − I η ν I ν ν − 1 I ν η , I ab = J a T Σ − 1 J b . (111)
The projector form remains valid with redundant nuisance columns. Its interpretation as Fisher information uses the stated Gaussian noise model; more general likelihoods require their own information metric. Known state constraints and unobservable orientation gauges must be removed before applying a full-rank test (Gupta and Chouhan, [55 ] ). Positive definiteness is a regular local identifiability condition; small eigenvalues imply noise amplification. At exact modal degeneracy the polarization basis is not unique. Use invariant projectors and subspace distances instead of arbitrary eigenvector labels.
A practical calibration for cubic-to-tetragonal transformation is to track ( e 2 , e 3 ) (e_2, e_3) ( e 2 , e 3 ) or diffraction-derived variants concurrently with C 44 C_{44} C 44 , C 66 C_{66} C 66 , and the angular shear gap. At a first-order transition, compare coexistence predictions with measured hysteresis, not merely a fitted discontinuity in a normalized phase fraction. At a continuous transition, test the spectral exponent against structurally measured β \beta β , rather than extracting both p p p and β \beta β from one gap curve.
Assessment against NiTi measurements
Figure 8 assesses three aspects of the state-variable description using NiTi data. The single-crystal data compare an effective elastic degeneracy with a calorimetric transformation marker. Two separate wire analyses assess the predictive value of phase-fraction and loading information. These comparisons test the value of additional state information within the stated analysis models; they do not directly test the cubic-to-tetragonal constitutive realization or the continuous critical relation in Eq. (51) .
We use the cooling and heating branches of the single-crystal resonant-ultrasound data of Bodnárová et al. [33 ] . For a cubic effective stiffness, the [100] transverse subspace is a doublet with squared speed C 44 / ρ 0 C_{44}/\rho_0 C 44 / ρ 0 . For propagation along [110], the Christoffel eigenvectors polarized along [001] and [ 1 1 ˉ 0 ] [1\bar{1}0] [ 1 1 ˉ 0 ] have eigenvalues C 44 C_{44} C 44 and C ′ C' C ′ , respectively, in units of stiffness. Thus the signed velocity difference and Zener anisotropy are
δ c [ 110 ] = C 44 − C ′ ρ 0 , C ′ = C 11 − C 12 2 , A = C 44 C ′ . (112) \delta c_{[110]} = \frac{\sqrt{C_{44}}-\sqrt{C'}}{\sqrt{\rho_0}}, \qquad C' = \frac{C_{11}-C_{12}}{2}, \qquad A = \frac{C_{44}}{C'}. \tag*{(112)} δ c [ 110 ] = ρ 0 C 44 − C ′ , C ′ = 2 C 11 − C 12 , A = C ′ C 44 . (112)
The equality C 44 = C ′ C_{44}=C' C 44 = C ′ is a tuned degeneracy relative to the cubic class. At this equality the entire fitted cubic tensor has the isotropic Lamé form with λ = C 12 \lambda=C_{12} λ = C 12 and μ = C 44 \mu=C_{44} μ = C 44 . The stabilizer of this effective tangent can therefore exceed the symmetry identified by a structural probe.
Interpolation of C 44 − C ′ C_{44}-C' C 44 − C ′ gives crossings at 234.99 and 252.02 K on cooling and heating, respectively (Figure 8 a). Interpolating A − 1 A-1 A − 1 gives 234.75 and 251.63 K. The difference is a consequence of interpolating two different nonlinear functions of the sampled moduli. At an exactly resolved temperature with C ′ > 0 C'>0 C ′ > 0 , the conditions C 44 − C ′ = 0 C_{44}-C'=0 C 44 − C ′ = 0 and A − 1 = 0 A-1=0 A − 1 = 0 are equivalent. The measured brackets are 231.20–236.15 K and 250.19–255.25 K. The crossings lie within the independent, same-specimen calorimetric intervals of 202–249 K on cooling and 234–257 K on heating (Figure 8 b). The inferred shear speeds use the fitted elastic constants, and the longitudinal pulse-echo data contribute to their elastic inversion. The internal friction in Figure 8 c comes from the resonances. The calorimetry independently locates the thermal transformation interval, while identifying local variant symmetry and phase fraction requires structural data.
Figure 8. Effective elastic degeneracy and state-conditioned response in NiTi. (a) Signed [110] transverse speed difference inferred from the two fitted shear constants, with interpolated zeros shown as open circles. (b) Same-specimen DSC anomalies normalized separately on cooling and heating. Short bars indicate the reported transformation intervals and dashed lines indicate the zero crossings in (a). (c) Internal friction obtained from the single-crystal resonance measurements. (d) Diffraction-derived martensite fraction from a matching wire experiment. Coloured markers give the fractions used at the withheld temperatures. (e) Withheld wire moduli and predictions using the pure-phase calibration. Shading spans the scalar Voigt and Reuss predictions based on diffraction. (f) Wire-modulus prediction error by withheld stress level for three regression baselines.
Table 7 places the two elastic crossing definitions alongside the independently measured calorimetric markers. The cooling–heating separation is 17.03 K for the modulus-difference interpolation and 16.88 K for the ratio interpolation. Their shared branch dependence and their locations inside the calorimetric transformation intervals support the role of thermal history in the effective response.
Table 7. Elastic degeneracy and calorimetric markers from the single-crystal NiTi archive used in Figure 8 a–Figure 8 c ([33 ] ). Temperatures are in kelvin.
Both elastic zeros follow interpolation of archived measurements. The modulus-difference crossing is used for the markers in Figure 8 a,Figure 8 b. The DSC peaks use the branch-normalized anomaly with the stated baseline treatment. Brackets indicate adjacent measured temperatures enclosing the crossing.
For any single-valued prediction f ( T ) f(T) f ( T ) of A A A , pointwise completion of the square in the errors on the cooling and heating branches gives the uniformly temperature-weighted bound
RMSE p a i r ( f ) ≥ [ ∫ T 1 T 2 [ A c o o l ( T ) − A h e a t ( T ) ] 2 d T 4 ( T 2 − T 1 ) ] 1 / 2 = 0.0981. (113) \operatorname{RMSE}_{\mathrm{pair}}(f) \ge\left[\frac{\int_{T_1}^{T_2}[A_{\mathrm{cool}}(T)-A_{\mathrm{heat}}(T)]^2\,\mathrm{d}T}{4(T_2-T_1)}\right]^{1/2}=0.0981. \tag*{(113)} RMSE pair ( f ) ≥ [ 4 ( T 2 − T 1 ) ∫ T 1 T 2 [ A cool ( T ) − A heat ( T ) ] 2 d T ] 1/2 = 0.0981. (113)
Here T 1 = 165.085 T_1=165.085 T 1 = 165.085 K and T 2 = 294.9872 T_2=294.9872 T 2 = 294.9872 K bound the common interval of the two piecewise-linear interpolants. Equality holds for the midpoint of the branches.
Explicitly, set A c = A c o o l A_c=A_{\mathrm{cool}} A c = A cool , A h = A h e a t A_h=A_{\mathrm{heat}} A h = A heat and A ‾ = ( A c + A h ) / 2 \overline{A}=(A_c+A_h)/2 A = ( A c + A h ) /2 . Pointwise,
( f − A c ) 2 + ( f − A h ) 2 2 = ( f − A ‾ ) 2 + ( A c − A h ) 2 4 . (114) \frac{(f-A_c)^2+(f-A_h)^2}{2}=(f-\overline{A})^2+\frac{(A_c-A_h)^2}{4}. \tag*{(114)} 2 ( f − A c ) 2 + ( f − A h ) 2 = ( f − A ) 2 + 4 ( A c − A h ) 2 . (114)
Integrating over the common interval and dropping the first, nonnegative term proves Eq. (113) for the specified equal-branch, uniform-temperature weighting. This bound quantifies the predictive information carried by thermal history for these interpolated data. Its numerical value is conditional on the specified interpolation and temperature weighting. It is a deterministic bound for the two measured paths, with measurement uncertainty outside the present bound calculation.
For the wire experiment of Sedlák et al. ([34 ] ), we fit linear martensite and austenite endmember Young’s moduli E M ( T ) E_M(T) E M ( T ) and E A ( T ) E_A(T) E A ( T ) to six reference-phase points each. The martensite calibration uses the six heating points at or below 100 °C, and the austenite calibration uses the six cooling points at or above 100 °C. The phase fractions are linearly interpolated to the withheld measurement temperatures. Ten mixed-state heating measurements are withheld. The martensite fraction φ ( T ) \varphi(T) φ ( T ) comes from diffraction on a matching wire under a matched thermomechanical protocol ([35 ] ). The scalar mixture closures are
E V = φ E M + ( 1 − φ ) E A , E R = [ φ E M + ( 1 − φ ) E A ] − 1 , E H = ( E V + E R ) / 2. (115) E_V = \varphi E_M + (1-\varphi)E_A,\qquad E_R = \left[\frac{\varphi}{E_M}+\frac{(1-\varphi)}{E_A}\right]^{-1},\qquad E_H = (E_V+E_R)/2. \tag*{(115)} E V = φ E M + ( 1 − φ ) E A , E R = [ E M φ + E A ( 1 − φ ) ] − 1 , E H = ( E V + E R ) /2. (115)
These scalar closures follow from distinct loading idealizations. Under a common axial strain, averaging phase stresses gives E V E_V E V ; under a common axial stress, averaging phase strains gives E R E_R E R . The Hill expression is their arithmetic mean ([56 ] ). For textured, constrained, or anisotropic transforming wires, these scalar relations are illustrative approximations: rigorous elastic homogenization uses tensor stiffness and compliance together with the appropriate boundary conditions. With the same endmember calibration and Hill closure, the withheld RMSE is 11.45 GPa with diffraction and 20.56 GPa with a fraction that decreases linearly from one at 100 °C to zero at 320 °C (Figure 8 d,Figure 8 e). The diffraction-conditioned Voigt and Reuss RMSE values are 7.72 and 30.02 GPa. Extrapolating the endmember stiffnesses with a positive exponential instead of a linear law changes the ranking of the closures.
In a separate martensite-only test, we withhold one entire stress group at a time across five groups spanning 400–800 MPa with 40 total observations. The pooled RMSE of the three-parameter temperature quadratic is 7.59 GPa. The three-parameter temperature–stress plane gives 2.12 GPa and the six-parameter quadratic surface gives 1.19 GPa (Figure 8 f). The comparison at equal parameter count quantifies the predictive information supplied by stress for these conditions. The scalar closures and regressions are classical baselines.
Table 8 summarizes two complementary withheld-data tests. At fixed endmember calibration and Hill closure, replacing the prescribed temperature-ramp fraction by diffraction-derived fraction reduces the mixed-state RMSE by 44.3% for linear endmember continuation. The improvement persists with positive exponential continuation of the martensite endmember, while the ranking of the mixture closures changes. For the stress test, adding stress to a model with three coefficients reduces the pooled RMSE by 72.1%. These comparisons quantify the value of structural and loading information within specified classical models.
Table 8. Withheld NiTi wire-modulus prediction errors for the data in Figure 8 d–Figure 8 f ([34 , 35 ] ). RMSE values are in gigapascals.
Mixed-state tests share six martensite and six austenite calibration points and the same ten mixed-state test temperatures, spanning 116.65–259.91 °C. The exponential sensitivity check fits the logarithm of the martensite calibration moduli versus temperature. The austenite continuation is linear in both blocks. The diffraction data come from a matching wire experiment. The stress test withholds complete groups at 400, 500, 600, 700, 800 MPa, preserving all observations at a given stress in the same fold. Its errors pool all withheld predictions, including the two endpoint extrapolation folds. The input column in the regression block lists the basis functions, with σ \sigma σ denoting tensile stress.
The strongest closed-form results concern the frozen, homogeneous constitutive models and their direction-dependent acoustic doublets. For comparison with a physical transforming solid, one must additionally specify the thermal constraint of the measured tangent, calibrate the structural couplings, resolve domain averaging, and compare the probe period with internal relaxation times. The numerical readout models and NiTi reanalyses illuminate these requirements without completing all of them for a single material.
A finite-dimensional local order parameter and its stabilizer cannot completely classify topologically ordered quantum phases or every defect-mediated transition. The 2D XY/BKT mechanism, for example, requires vortex physics and long-distance correlation behavior ([23 , 24 ] ). When the physical system admits meaningful topological or nonlocal descriptors, a larger phase-state record may be written
S p h a s e = ( [ H η ] , T t o p , F t h e r m , D s l o w , O m e a s ) , (116) \mathcal{S}_{\mathrm{phase}} = ([H_{\eta}], \mathcal{T}_{\mathrm{top}}, \mathcal{F}_{\mathrm{therm}}, \mathcal{D}_{\mathrm{slow}}, \mathcal{O}_{\mathrm{meas}}), \tag*{(116)} S phase = ([ H η ] , T top , F therm , D slow , O meas ) , (116)
where T t o p \mathcal{T}_{\mathrm{top}} T top is a system-specific topological or nonlocal descriptor, not a universally existing scalar invariant. This expression is a classification program rather than a proven one-to-one parametrization of all phases.
No single internal variable (scalar, vector, or tensor) is guaranteed to identify all phases. Neither the group-subgroup relation nor the form of a local Landau polynomial alone fixes microscopic energy scales, the RG fixed point, elastic constants, nucleation rates, or dissipation. The spectrum additionally depends on the full constitutive operator. As a result, the most defensible target is a common covariance/thermodynamics/observation architecture, with explicit hypotheses under which particular cross-material laws are valid.
Conclusions
We have formulated a symmetry-stratified description of structural phase transitions that links an internal-state representation to an invariant free energy, incremental elasticity, and directional acoustic observables. The framework distinguishes orbit motion, changes of structural symmetry, and thermodynamic transitions.
The central mechanical results are the inheritance of the structural stabilizer by a parent-equivariant frozen tangent and the identification of the first nonzero acoustic splitting order by degenerate perturbation theory. For a selected continuous branch, fixed wave number, and positive reference wave speed, these results imply Δ ω ∼ ∣ t ∣ p β \Delta\omega\sim|t|^{p\beta} Δ ω ∼ ∣ t ∣ pβ only when the stated regularity and observability conditions hold. The tensor model provides an explicit reconstruction from calibrated stiffness contractions and distinguishes linear and quadratic splitting channels. The cubic-to-tetragonal realization contrasts a lifted [100] doublet with a protected [001] doublet.
The numerical benchmarks test complementary parts of this construction. They address invariant free-energy reconstruction, finite-size coexistence, compatible domain patterns, and polarization-resolved spectra at fixed cell shape. The Ising-based probes verify moment-level scaling for prescribed readouts and demonstrate sensitivity to coarse-graining and relaxation.
The NiTi data illustrate the distinction between effective elastic degeneracy and structural symmetry. They quantify thermal-history dependence, improved phase-fraction predictions within the specified Hill comparisons, and the value of stress in the reported regression baselines. A material-specific test should combine independently resolved structural states with directional, polarization-resolved spectra and a calibrated tangent under matched loading and thermal conditions. This defines a concrete route from the present symmetry constraints to experimentally testable constitutive predictions.
Declaration of Competing Interest
The authors declare that they have no known competing financial interests or personal relationships that could have appeared to influence the work reported in this paper.
Acknowledgments
This work was supported by the Adolf Martens Fellowship (Grant No. BAM-AMF-2025-1).
Data availability
Data will be made available on request.
CRediT authorship contribution statement
Pengfei Zhu: Methodology, Investigation, Visualization, Writing – original draft, Writing – review & editing. Leiming Du: Investigation, Writing – review & editing.
Acoustic splitting in the structural-tensor model
For completeness, the following calculation makes explicit why one mechanical coupling can be spectrally silent at first order. Let m = n ^ \mathbf{m}=\hat{\mathbf{n}} m = n ^ and q m = m ⋅ Q m q_m=\mathbf{m}\cdot\mathbf{Q}\mathbf{m} q m = m ⋅ Qm . Contracting Eq. (61) with m j m l m_jm_l m j m l yields
δ A = γ 1 [ m ⊗ ( Q m ) + ( Q m ) ⊗ m ] + γ 2 2 [ Q + q m I + ( Q m ) ⊗ m + m ⊗ ( Q m ) ] . (117) \delta\mathbf{A}=\gamma_1\left[\mathbf{m}\otimes(\mathbf{Q}\mathbf{m})+(\mathbf{Q}\mathbf{m})\otimes\mathbf{m}\right]+\frac{\gamma_2}{2}\left[\mathbf{Q}+q_m\mathbf{I}+(\mathbf{Q}\mathbf{m})\otimes\mathbf{m}+\mathbf{m}\otimes(\mathbf{Q}\mathbf{m})\right]. \tag*{(117)} δ A = γ 1 [ m ⊗ ( Qm ) + ( Qm ) ⊗ m ] + 2 γ 2 [ Q + q m I + ( Qm ) ⊗ m + m ⊗ ( Qm ) ] . (117)
Let P T = I − m ⊗ m \mathbf{P}_T=\mathbf{I}-\mathbf{m}\otimes\mathbf{m} P T = I − m ⊗ m be the unperturbed shear projector. Since P T m = 0 \mathbf{P}_T\mathbf{m}=0 P T m = 0 ,
P T δ A P T = γ 2 2 [ P T Q P T + q m P T ] . (118) \mathbf{P}_T\delta\mathbf{A}\mathbf{P}_T=\frac{\gamma_2}{2}\left[\mathbf{P}_T\mathbf{Q}\mathbf{P}_T+q_m\mathbf{P}_T\right]. \tag*{(118)} P T δ A P T = 2 γ 2 [ P T Q P T + q m P T ] . (118)
Thus the γ 1 \gamma_1 γ 1 term is exactly absent from the first-order shear block. For uniaxial Q = s ( N ⊗ N − I / 3 ) \mathbf{Q}=s(\mathbf{N}\otimes\mathbf{N}-\mathbf{I}/3) Q = s ( N ⊗ N − I /3 ) , the orthonormal transverse basis
e 1 = P T N sin ϑ , e 2 = m × N sin ϑ ( sin ϑ ≠ 0 ) (119) \mathbf{e}_1=\frac{\mathbf{P}_T\mathbf{N}}{\sin\vartheta},\qquad\mathbf{e}_2=\frac{\mathbf{m}\times\mathbf{N}}{\sin\vartheta}\qquad(\sin\vartheta\ne0) \tag*{(119)} e 1 = sin ϑ P T N , e 2 = sin ϑ m × N ( sin ϑ = 0 ) (119)
diagonalizes the block:
( P T δ A P T ) { e 1 , e 2 } = γ 2 s 2 ( 1 / 3 0 0 cos 2 ϑ − 2 / 3 ) (120) (P_T\delta A P_T)_{\{\mathbf{e}_1,\mathbf{e}_2\}}=\frac{\gamma_2 s}{2}\begin{pmatrix} 1/3 & 0 \\ 0 & \cos^2\vartheta-2/3 \end{pmatrix} \tag*{(120)} ( P T δ A P T ) { e 1 , e 2 } = 2 γ 2 s ( 1/3 0 0 cos 2 ϑ − 2/3 ) (120)
The difference of the two shear corrections is γ 2 s sin 2 ϑ / 2 \gamma_2s\sin^2\vartheta/2 γ 2 s sin 2 ϑ /2 , proving Eq. (71) after division by density. At ϑ = 0 \vartheta=0 ϑ = 0 the chosen basis is undefined but the symmetric eigenvalue pair remains protected by axial rotational symmetry.
For γ 2 = 0 \gamma_2=0 γ 2 = 0 , the leading shear splitting originates instead from second-order coupling between the in-plane transverse mode and the longitudinal mode. The unperturbed longitudinal and transverse acoustic eigenvalues are a L = λ L + 2 μ a_L=\lambda_L+2\mu a L = λ L + 2 μ and a T = μ a_T=\mu a T = μ , respectively. In the basis ( e 1 , e 2 , m ) (\mathbf{e}_1,\mathbf{e}_2,\mathbf{m}) ( e 1 , e 2 , m ) , the only transverse–longitudinal matrix element is
e 1 ⋅ δ A m = γ 1 e 1 ⋅ Q m = γ 1 s sin ϑ cos ϑ , e 2 ⋅ δ A m = 0. (121) \mathbf{e}_1\cdot\delta A\mathbf{m}=\gamma_1\mathbf{e}_1\cdot Q\mathbf{m}=\gamma_1s\sin\vartheta\cos\vartheta,\qquad\mathbf{e}_2\cdot\delta A\mathbf{m}=0. \tag*{(121)} e 1 ⋅ δ A m = γ 1 e 1 ⋅ Q m = γ 1 s sin ϑ cos ϑ , e 2 ⋅ δ A m = 0. (121)
Since the source model is linear in Q Q Q at the tangent level, there is no direct A 2 A_2 A 2 term. The reduced resolvent is R 0 = − m ⊗ m / ( λ L + μ ) R_0=-\mathbf{m}\otimes\mathbf{m}/(\lambda_L+\mu) R 0 = − m ⊗ m / ( λ L + μ ) , so
δ a T 1 ( 2 ) = − γ 1 2 s 2 sin 2 ϑ cos 2 ϑ λ L + μ , δ a T 2 ( 2 ) = 0. (122) \delta a_{T1}^{(2)}=-\frac{\gamma_1^2s^2\sin^2\vartheta\cos^2\vartheta}{\lambda_L+\mu},\qquad\delta a_{T2}^{(2)}=0. \tag*{(122)} δ a T 1 ( 2 ) = − λ L + μ γ 1 2 s 2 sin 2 ϑ cos 2 ϑ , δ a T 2 ( 2 ) = 0. (122)
Taking the magnitude of their difference and dividing by density gives Eq. (72) ; the positive-energy isotropic reference used here has λ L + μ > 0 \lambda_L+\mu>0 λ L + μ > 0 . The mechanism is the same virtual mixing term in Eq. (46) (Chernyshev et al., [44 ] ). Its angular zeros differ from the generic first-order γ 2 \gamma_2 γ 2 channel, illustrating that p p p contains actual constitutive information.
Free-energy comparison for the two-component cubic invariant
The cubic-to-tetragonal preferred-variant polynomial is f ( r ) = a r 2 / 2 − b r 3 / 3 + c r 4 / 4 f(r)=ar^2/2-br^3/3+cr^4/4 f ( r ) = a r 2 /2 − b r 3 /3 + c r 4 /4 . Along a nonzero stationary branch, a = b r − c r 2 a=br-cr^2 a = b r − c r 2 . Inserting this identity into f ( r ) f(r) f ( r ) gives
f ( r ) ∣ f ′ = 0 = b 6 r 3 − c 4 r 4 . (123) \left.f(r)\right|_{f'=0}=\frac{b}{6}r^3-\frac{c}{4}r^4. \tag*{(123)} f ( r ) ∣ f ′ = 0 = 6 b r 3 − 4 c r 4 . (123)
Equality with the parent branch f ( 0 ) = 0 f(0)=0 f ( 0 ) = 0 requires r t = 2 b / ( 3 c ) r_t=2b/(3c) r t = 2 b / ( 3 c ) . Inserting into a = b r − c r 2 a=br-cr^2 a = b r − c r 2 produces a t = 2 b 2 / ( 9 c ) a_t=2b^2/(9c) a t = 2 b 2 / ( 9 c ) . The nonzero roots exist for a ≤ b 2 / ( 4 c ) a\le b^2/(4c) a ≤ b 2 / ( 4 c ) , while the parent-state curvature is f ′ ′ ( 0 ) = a f''(0)=a f ′′ ( 0 ) = a . In this elementary homogeneous setting a = b 2 / ( 4 c ) a=b^2/(4c) a = b 2 / ( 4 c ) is an ordered-branch saddle-node threshold, a = 0 a=0 a = 0 is the parent-state spinodal, and a = a t a=a_t a = a t is equilibrium coexistence.
Mathematical notation
Table 9 lists the principal symbols used in the derivations, numerical benchmarks, and experimental analysis. The last column identifies the relevant context when a symbol has several meanings. The benchmarks in Figure 2 –Figure 7 use dimensionless model units. The NiTi temperatures and moduli in Figure 8 retain their experimental units.
Table 9. Principal mathematical symbols and their meanings.
Operators and statistical conventions
caption:
:::