Introduction
The Mpemba effect challenges the expectation that a preparation farther from equilibrium must take longer to relax. Its original formulation concerned hot and cold water [1 , 2 ] , but a controlled statistical-mechanical question can be stated without phase change: prepare two distributions and evolve them with the same final generator. The hotter preparation may initially have a larger distance to the common stationary state and later have a smaller one. Reversible Markov dynamics provide a precise setting for this comparison [3 ] .
The general spectral mechanism is already established. Relaxation rates belong to the final generator, whereas the amplitudes of its modes depend on the initial preparation. A reduced slow-mode overlap can produce anomalous ordering, and cancellation of the slowest active sector gives the strong Mpemba effect [3 , 4 ] . Anomalous relaxation of this kind has been predicted in granular fluids [5 , 6 ] , spin glasses [7 ] , Langevin dynamics [8 ] , and open and many-body quantum systems [9 , 10 , 11 ] , and observed in colloids [12 , 13 ] and trapped ions [14 , 15 ] . Inverse and precooling effects [13 , 16 , 17 ] and heating–cooling asymmetries [18 , 19 , 20 , 21 ] belong to the same broad family of relaxation speedups [22 ] .
Finite preparation accuracy raises a separate question. Exact removal of a slow component is an ideal condition: a residual amplitude restores that asymptotic decay rate. An ordinary finite-time crossing can nevertheless persist. Establishing this persistence requires a comparison over a specified time interval, rather than only an asymptotic exponent. Preparation-error sensitivity has also been studied in quantum Mpemba models [23 ] , and finite-time criteria occur in quantum imaginary-time evolution [24 ] . The present setting is a classical reversible semigroup with relative entropy as the specified distance.
Relative entropy is attractive because, for a Gibbs stationary state, it equals the excess nonequilibrium free energy divided by the bath temperature. Its derivative is minus the dimensionless entropy-production rate [25 , 26 ] , the central quantity of stochastic thermodynamics [27 , 28 ] , whose exact time dependence, together with the entropy-extraction rate, has been obtained for Brownian particles, motors, and heat engines [29 , 30 , 31 , 32 , 33 , 34 , 35 , 36 , 37 , 38 , 39 ] . Entropy dissipation and functional inequalities already constrain relaxation durations [40 , 41 , 42 , 43 ] . A global decay constant, however, does not retain the different mode weights of a particular pair of preparations. Conversely, a quadratic entropy approximation is useful near equilibrium but needs an explicit error bound before it can certify a crossing at finite amplitude.
We use the established bounded-likelihood-ratio comparison between relative entropy and chi-squared divergence [44 ] . Combining this static inequality with the Markov maximum principle provides exact time-dependent entropy envelopes. Their slow-sector form gives a transparent sufficient condition, denoted Γ > 0 \Gamma> 0 Γ > 0 , and an explicit time by which the initial ordering has reversed. Their all-mode form retains more of the spectrum, while a nonnegative remainder bound allows controlled use of only part of it. Degenerate slow eigenspaces are treated as a whole.
We then quantify robustness. A centered entropy-gradient estimate and probability contraction translate a prescribed preparation error into a uniform entropy error. Positive separation throughout an early interval, together with negative separation at a later endpoint, gives a robust first-meeting interval. Temperature and rate uncertainty require additional bounds, including the change of stationary reference when the generator changes. The proof identifies these requirements explicitly.
Our main results are as follows. (i) Theorem 1: for any finite reversible chain, a positive slow-sector margin Γ \Gamma Γ guarantees a crossing and gives an explicit upper time, without exact slow-mode cancellation (Figure 1 ). (ii) All-mode and partial-spectrum enclosures that tighten this time using more spectral information, with a controlled omitted tail. (iii) Theorem 2 and its corollaries: explicit tolerances for preparation, temperature, and rate errors, and a local sensitivity formula expressed through the normalized entropy-production rates. (iv) Tests on 2000 reversible networks with 4 to 64 states and on a reflecting diffusion, together with interval-arithmetic certificates, including a whole region of preparation temperatures (Figure 4 ).
Figure 1. How Theorem 1 works, for the three-state benchmark of Sec. VIII at ( T H , T C ) = ( 3 , 1.5 ) (T_H,T_C)=(3,1.5) ( T H , T C ) = ( 3 , 1.5 ) , which is not an exactly strong preparation. (a) Squared mode weights: the hot preparation has a much larger fast weight q H q_H q H but a slow weight w H w_H w H about 35 times smaller than w C w_C w C . (b) Relative entropies (lines) inside their exact finite-amplitude envelopes, Eq. (22) (shaded); the dot is the crossing t M ≃ 0.1702 t_M \simeq0.1702 t M ≃ 0.1702 and the dotted line the bound T u p ≃ 0.1849 T_{\mathrm{up}} \simeq0.1849 T up ≃ 0.1849 . (c) The entropy difference multiplied by e 2 λ s t e^{2\lambda_s t} e 2 λ s t (solid) and its upper bound from Eq. (26) (dashed). The bound becomes negative exactly at T u p T_{\mathrm{up}} T up and tends to − Γ -\Gamma − Γ ; the true difference changes sign earlier. A symmetric-logarithmic vertical scale resolves both signs.
Figure 4. Located crossing time of the three-state benchmark over the plane of preparation temperatures (double precision; gray: T H < T C T_H < T_C T H < T C or no crossing located). The blue contour is Γ = 0 \Gamma= 0 Γ = 0 , with Γ > 0 \Gamma> 0 Γ > 0 inside; the dashed line marks T ⋆ T_\star T ⋆ , where the hot slow weight vanishes. The black box is the rectangle of Eq. (67) , in which every pair is certified by interval arithmetic to have its first crossing in ( 0.161 , 0.179 ) (0.161, 0.179) ( 0.161 , 0.179 ) .
A companion manuscript develops the entropy-production-clock interpretation and a tightly certified three-state benchmark, and states the slow-sector comparison and preparation corollary in compact form [45 ] . Here we give a self-contained treatment of the general finite-state problem, the all-mode and partial-spectrum construction, parameter robustness, predefined network ensembles, and a physical diffusion model. The shared comparison ingredients are stated explicitly; the additional scope and numerical tests are the focus of this study.
The numerical work separates three kinds of evidence: evaluations in ordinary floating-point arithmetic, mesh and algorithm comparisons, and outward-rounded interval certificates. The network sample measures performance within a declared ensemble, not the prevalence of the effect among all Markov systems. Likewise, a finite-grid diffusion certificate is not a proof of a continuum limit. The ordering considered throughout is relative entropy; a simultaneous ordering of all monotones is a different question addressed by thermomajorization [46 ] .
Common dynamics, preparations, and thermodynamic distance
A fixed reversible Markov system
Consider N ≥ 2 N \ge2 N ≥ 2 states with transition rates k i j k_{ij} k ij for a jump from i i i to j j j . The probabilities are column vectors and obey
p ˙ i = ∑ j ≠ i ( k j i p j − k i j p i ) , p ˙ = W p . (1) \dot{p}_i=\sum_{j\ne i}(k_{ji}p_j-k_{ij}p_i),\qquad\dot{p}=Wp. \tag*{(1)} p ˙ i = j = i ∑ ( k j i p j − k ij p i ) , p ˙ = W p . (1)
Thus W i j = k j i W_{ij}=k_{ji} W ij = k j i for i ≠ j i\ne j i = j and W i i = − ∑ j ≠ i k i j W_{ii}=-\sum_{j\ne i}k_{ij} W ii = − ∑ j = i k ij ; every column of W W W sums to zero. We assume a finite, irreducible chain with a positive stationary distribution π \pi π and detailed balance,
W π = 0 , π i k i j = π j k j i . (2) W\pi=0,\qquad\pi_i k_{ij}=\pi_j k_{ji}. \tag*{(2)} W π = 0 , π i k ij = π j k j i . (2)
The graph may be sparse, provided it is connected. Both preparations evolve with precisely this same time-independent generator:
p α ( t ) = e W t p α ( 0 ) , α = H , C . (3) p^\alpha(t)=e^{Wt}p^\alpha(0),\qquad\alpha=H,C. \tag*{(3)} p α ( t ) = e W t p α ( 0 ) , α = H , C . (3)
The rates belong to the final dynamics; the preparations determine their mode amplitudes. This is the standard spectral setting for the Markovian Mpemba effect [3 , 4 , 47 , 48 ] .
The crossing bounds apply to arbitrary strictly positive initial distributions satisfying the entropy ordering specified below. A thermal realization has fixed energies E i E_i E i , bath temperature T b T_b T b , and
π i = e − E i / T b Z ( T b ) , p i α ( 0 ) = e − E i / T α Z ( T α ) , T H > T C > T b . (4) \pi_i=\frac{e^{-E_i/T_b}}{Z(T_b)},\qquad p_i^\alpha(0)=\frac{e^{-E_i/T_\alpha}}{Z(T_\alpha)},\qquad T_H>T_C>T_b. \tag*{(4)} π i = Z ( T b ) e − E i / T b , p i α ( 0 ) = Z ( T α ) e − E i / T α , T H > T C > T b . (4)
where Z ( T ) = ∑ i e − E i / T Z(T)=\sum_i e^{-E_i/T} Z ( T ) = ∑ i e − E i / T and k B = 1 k_{\mathrm{B}}=1 k B = 1 . Temperature labels refer to these initial equilibria; an evolving nonequilibrium distribution need not have a unique temperature.
Entropy, a first meeting, and the normalized decay rate
Use natural logarithms and the relative entropy
D α ( t ) = D [ p α ( t ) ∥ π ] = ∑ i p i α ( t ) ln p i α ( t ) π i . (5) D_\alpha(t)=D[p^\alpha(t)\Vert\pi]=\sum_i p_i^\alpha(t)\ln\frac{p_i^\alpha(t)}{\pi_i}. \tag*{(5)} D α ( t ) = D [ p α ( t ) ∥ π ] = i ∑ p i α ( t ) ln π i p i α ( t ) . (5)
For a Gibbs stationary state, define F [ p ] = ∑ i p i E i + T b ∑ i p i ln p i \mathcal{F}[p]=\sum_i p_iE_i+T_b\sum_i p_i\ln p_i F [ p ] = ∑ i p i E i + T b ∑ i p i ln p i . Then
F [ p ] − F [ π ] = T b D [ p ∥ π ] . (6) \mathcal{F}[p]-\mathcal{F}[\pi]=T_bD[p\Vert\pi]. \tag*{(6)} F [ p ] − F [ π ] = T b D [ p ∥ π ] . (6)
Relative entropy therefore measures excess free energy in units of T b T_b T b . For canonical cooling with nonconstant energies, D H ( 0 ) > D C ( 0 ) D_H(0)>D_C(0) D H ( 0 ) > D C ( 0 ) follows directly from T H > T C > T b T_H>T_C>T_b T H > T C > T b ; the derivative calculation is given in Appendix A.
With currents J i j = k i j p i − k j i p j J_{ij}=k_{ij}p_i-k_{ji}p_j J ij = k ij p i − k j i p j , the system entropy S = − ∑ i p i ln p i S=-\sum_i p_i\ln p_i S = − ∑ i p i ln p i obeys the balance
S ˙ = e ˙ p − h ˙ d , e ˙ p = 1 2 ∑ i , j J i j ln k i j p i k j i p j ≥ 0 , \dot{S}=\dot{e}_p-\dot{h}_d,\qquad\dot{e}_p=\frac{1}{2}\sum_{i,j}J_{ij}\ln\frac{k_{ij}p_i}{k_{ji}p_j}\ge0, S ˙ = e ˙ p − h ˙ d , e ˙ p = 2 1 i , j ∑ J ij ln k j i p j k ij p i ≥ 0 ,
h ˙ d = 1 2 ∑ i , j J i j ln k i j k j i . (7) \dot{h}_d=\frac{1}{2}\sum_{i,j}J_{ij}\ln\frac{k_{ij}}{k_{ji}}. \tag*{(7)} h ˙ d = 2 1 i , j ∑ J ij ln k j i k ij . (7)
where e ˙ p \dot{e}_p e ˙ p is the dimensionless entropy-production rate and h ˙ d \dot{h}_d h ˙ d the entropy-extraction (heat-dissipation) rate into the bath [25 , 26 , 27 , 28 ] . Sums over edges omit absent transitions. Under detailed balance ln ( k i j / k j i ) = ( E i − E j ) / T b \ln(k_{ij}/k_{ji})=(E_i-E_j)/T_b ln ( k ij / k j i ) = ( E i − E j ) / T b , so h ˙ d = − U ˙ / T b \dot{h}_d=-\dot{U}/T_b h ˙ d = − U ˙ / T b , and since D = − S + U / T b + ln Z ( T b ) D=-S+U/T_b+\ln Z(T_b) D = − S + U / T b + ln Z ( T b ) the balance yields the standard identity
− D ˙ α = e ˙ p α , r α = e ˙ p α D α = − d d t ln D α . (8) -\dot{D}_\alpha=\dot{e}_p^\alpha,\qquad r_\alpha=\frac{\dot{e}_p^\alpha}{D_\alpha}=-\frac{\mathrm{d}}{\mathrm{d}t}\ln D_\alpha. \tag*{(8)} − D ˙ α = e ˙ p α , r α = D α e ˙ p α = − d t d ln D α . (8)
Both initial entropies are assumed nonzero, and we impose
D H ( 0 ) > D C ( 0 ) > 0 , d ( t ) = D H ( t ) − D C ( t ) . (9) D_H(0)>D_C(0)>0,\qquad d(t)=D_H(t)-D_C(t). \tag*{(9)} D H ( 0 ) > D C ( 0 ) > 0 , d ( t ) = D H ( t ) − D C ( t ) . (9)
At any finite time a nonstationary preparation remains nonstationary, because e W t e^{Wt} e W t is invertible. Hence the rates in Eq. (8) remain defined.
Let
t M = inf { t > 0 : d ( t ) = 0 } , (10) t_M=\inf\{t>0:d(t)=0\}, \tag*{(10)} t M = inf { t > 0 : d ( t ) = 0 } , (10)
with t M = ∞ t_M=\infty t M = ∞ when the set is empty; t M t_M t M is the first-passage time of the deterministic difference d ( t ) d(t) d ( t ) through zero [49 , 50 , 51 , 52 ] . A meeting can be a tangency; a Mpemba reversal additionally requires d < 0 d<0 d < 0 afterward. Our sufficient theorem below guarantees a strict reversed ordering after a finite upper bound, but does not assume uniqueness of earlier meetings. Integrating Eq. (8) gives
Θ α ( t ) = ∫ 0 t r α ( s ) d s = ln D α ( 0 ) D α ( t ) , (11) \Theta_\alpha(t)=\int_0^t r_\alpha(s)\,\mathrm{d}s=\ln\frac{D_\alpha(0)}{D_\alpha(t)}, \tag*{(11)} Θ α ( t ) = ∫ 0 t r α ( s ) d s = ln D α ( t ) D α ( 0 ) , (11)
ln D H ( t ) D C ( t ) = ln D H ( 0 ) D C ( 0 ) − ∫ 0 t ( r H − r C ) d s . \ln\frac{D_H(t)}{D_C(t)}=\ln\frac{D_H(0)}{D_C(0)}-\int_0^t(r_H-r_C)\,\mathrm{d}s. ln D C ( t ) D H ( t ) = ln D C ( 0 ) D H ( 0 ) − ∫ 0 t ( r H − r C ) d s .
The dimensionless quantity Θ \Theta Θ measures accumulated fractional free-energy reduction. It is a progress variable constructed from the trajectory, not an additional physical time coordinate. Connections between entropy production and relaxation duration are well established [42 ] ; here the task is to bound the meeting of two specified preparations. A relative-entropy crossing need not coincide with an energy crossing or an ordering of every distance measure.
Relaxation modes and exact finite-amplitude entropy envelopes
The preparation weights
Because W π = 0 W\pi=0 W π = 0 , the deviation p α ( t ) − π p^\alpha(t)-\pi p α ( t ) − π obeys the same master equation and can be expanded in the eigenvectors of W W W ,
p α ( t ) = π + ∑ n = 2 N b n , α e − λ n t v n , W v n = − λ n v n . (12) p^\alpha(t)=\pi+\sum_{n=2}^{N}b_{n,\alpha}e^{-\lambda_n t}v_n,\qquad Wv_n=-\lambda_n v_n. \tag*{(12)} p α ( t ) = π + n = 2 ∑ N b n , α e − λ n t v n , W v n = − λ n v n . (12)
This is the whole mechanism in one line: the rates λ n \lambda_n λ n belong to the common generator, whereas the amplitudes b n , α b_{n,\alpha} b n , α record how strongly each preparation excites relaxation pattern n n n . Detailed balance makes the patterns orthogonal once each component is weighted by 1 / π i 1/\pi_i 1/ π i , so we normalize ∑ i v m , i v n , i / π i = δ m n \sum_i v_{m,i}v_{n,i}/\pi_i=\delta_{mn} ∑ i v m , i v n , i / π i = δ mn .
For the entropy it is convenient to use the relative deviation f i α = p i α / π i − 1 f_i^\alpha=p_i^\alpha/\pi_i-1 f i α = p i α / π i − 1 , the fractional excess or deficit of probability in state i i i , together with the equilibrium inner product ⟨ u , v ⟩ π = ∑ i π i u i v i \langle u,v\rangle_\pi=\sum_i\pi_i u_i v_i ⟨ u , v ⟩ π = ∑ i π i u i v i . Normalization gives ⟨ f α , 1 ⟩ π = 0 \langle f^\alpha,1\rangle_\pi=0 ⟨ f α , 1 ⟩ π = 0 , and f f f evolves with the likelihood-ratio generator
( L f ) i = ∑ j ≠ i k i j ( f j − f i ) , f ˙ = L f . (13) (Lf)_i=\sum_{j\ne i}k_{ij}(f_j-f_i),\qquad\dot{f}=Lf. \tag*{(13)} ( L f ) i = j = i ∑ k ij ( f j − f i ) , f ˙ = L f . (13)
which detailed balance makes self-adjoint in this inner product. Its eigenfunctions ϕ n , i = v n , i / π i \phi_{n,i}=v_{n,i}/\pi_i ϕ n , i = v n , i / π i are normalized as
L ϕ n = − λ n ϕ n , ⟨ ϕ m , ϕ n ⟩ π = δ m n , L\phi_n=-\lambda_n\phi_n,\qquad\langle\phi_m,\phi_n\rangle_\pi=\delta_{mn}, L ϕ n = − λ n ϕ n , ⟨ ϕ m , ϕ n ⟩ π = δ mn ,
ϕ 1 = 1 , 0 = λ 1 < λ 2 ≤ ⋯ ≤ λ N . (14) \phi_1=1,\qquad0=\lambda_1<\lambda_2\le\cdots\le\lambda_N. \tag*{(14)} ϕ 1 = 1 , 0 = λ 1 < λ 2 ≤ ⋯ ≤ λ N . (14)
The exact solution is
f i α ( t ) = ∑ n = 2 N b n , α e − λ n t ϕ n , i , f_i^\alpha(t)=\sum_{n=2}^{N}b_{n,\alpha}e^{-\lambda_n t}\phi_{n,i}, f i α ( t ) = n = 2 ∑ N b n , α e − λ n t ϕ n , i ,
b n , α = ⟨ f α ( 0 ) , ϕ n ⟩ π = ∑ i p i α ( 0 ) ϕ n , i . (15) b_{n,\alpha}=\langle f^\alpha(0),\phi_n\rangle_\pi=\sum_i p_i^\alpha(0)\phi_{n,i}. \tag*{(15)} b n , α = ⟨ f α ( 0 ) , ϕ n ⟩ π = i ∑ p i α ( 0 ) ϕ n , i . (15)
The last equality uses the zero stationary mean of every nonstationary eigenfunction. The weighted squared deviation is
Q α ( t ) = ∑ i π i ∣ f i α ( t ) ∣ 2 = ∑ n = 2 N b n , α 2 e − 2 λ n t = χ 2 ( p α ( t ) ∥ π ) . (16) Q_\alpha(t)=\sum_i\pi_i\lvert f_i^\alpha(t)\rvert^2=\sum_{n=2}^{N}b_{n,\alpha}^2e^{-2\lambda_n t}=\chi^2(p^\alpha(t)\mathbin{\|}\pi). \tag*{(16)} Q α ( t ) = i ∑ π i ∣ f i α ( t ) ∣ 2 = n = 2 ∑ N b n , α 2 e − 2 λ n t = χ 2 ( p α ( t ) ∥ π ) . (16)
This is an exact spectral identity. Relative entropy is only asymptotically Q / 2 Q/2 Q /2 near equilibrium; replacing it by that approximation at finite amplitude would leave an uncontrolled crossing-time error. Appendix A derives the orthogonality and the expansion
D α ( t ) = 1 2 Q α ( t ) [ 1 + O ( ∥ f α ( t ) ∥ ∞ ) ] . (17) D_\alpha(t)=\frac{1}{2}Q_\alpha(t)\left[1+O\left(\lVert f^\alpha(t)\rVert_\infty\right)\right]. \tag*{(17)} D α ( t ) = 2 1 Q α ( t ) [ 1 + O ( ∥ f α ( t ) ∥ ∞ ) ] . (17)
A scalar comparison that remains valid far from equilibrium
Define
g ( x ) = ( 1 + x ) ln ( 1 + x ) − x , x > − 1 , g(x)=(1+x)\ln(1+x)-x,\qquad x>-1, g ( x ) = ( 1 + x ) ln ( 1 + x ) − x , x > − 1 ,
F ( x ) = { g ( x ) / x 2 , x ≠ 0 , 1 / 2 , x = 0. (18) F(x)=
\begin{cases}
g(x)/x^2,&x\ne0,\\
1/2,&x=0.
\end{cases} \tag*{(18)} F ( x ) = { g ( x ) / x 2 , 1/2 , x = 0 , x = 0. (18)
Since g ( 0 ) = g ′ ( 0 ) = 0 g(0)=g'(0)=0 g ( 0 ) = g ′ ( 0 ) = 0 and g ′ ′ ( x ) = 1 / ( 1 + x ) g''(x)=1/(1+x) g ′′ ( x ) = 1/ ( 1 + x ) , twice integrating gives
F ( x ) = ∫ 0 1 1 − u 1 + u x d u > 0 , F(x)=\int_0^1\frac{1-u}{1+ux}\,\mathrm{d}u>0, F ( x ) = ∫ 0 1 1 + ux 1 − u d u > 0 ,
F ′ ( x ) = − ∫ 0 1 u ( 1 − u ) ( 1 + u x ) 2 d u < 0. (19) F'(x)=-\int_0^1\frac{u(1-u)}{(1+ux)^2}\,\mathrm{d}u<0. \tag*{(19)} F ′ ( x ) = − ∫ 0 1 ( 1 + ux ) 2 u ( 1 − u ) d u < 0. (19)
Normalization cancels the linear term, so exactly D α = ∑ i π i g ( f i α ) = ∑ i π i ( f i α ) 2 F ( f i α ) D_\alpha= \sum_i \pi_i g(f_i^\alpha) = \sum_i \pi_i (f_i^\alpha)^2 F(f_i^\alpha) D α = ∑ i π i g ( f i α ) = ∑ i π i ( f i α ) 2 F ( f i α ) . For a reference time s s s , set
m α ( s ) = min i f i α ( s ) , M α ( s ) = max i f i α ( s ) , m_\alpha(s)=\min_i f_i^\alpha(s),\qquad M_\alpha(s)=\max_i f_i^\alpha(s), m α ( s ) = i min f i α ( s ) , M α ( s ) = i max f i α ( s ) ,
a α ( s ) = F ( M α ( s ) ) , c α ( s ) = F ( m α ( s ) ) . (20) a_\alpha(s)=F(M_\alpha(s)),\qquad c_\alpha(s)=F(m_\alpha(s)). \tag*{(20)} a α ( s ) = F ( M α ( s )) , c α ( s ) = F ( m α ( s )) . (20)
The likelihood-ratio dynamics maps every component at t ≥ s t\ge s t ≥ s to a convex combination of components at s s s :
f i ( t ) = ∑ j K i j ( t − s ) f j ( s ) , K i j ( u ) = [ e W u ] i j π j π i . (21) f_i(t)=\sum_j K_{ij}(t-s)f_j(s),\qquad K_{ij}(u)=[e^{Wu}]_{ij}\frac{\pi_j}{\pi_i}. \tag*{(21)} f i ( t ) = j ∑ K ij ( t − s ) f j ( s ) , K ij ( u ) = [ e W u ] ij π i π j . (21)
where K i j ≥ 0 K_{ij}\ge0 K ij ≥ 0 and ∑ j K i j = 1 \sum_j K_{ij}=1 ∑ j K ij = 1 . Therefore each later component remains in [ m α ( s ) , M α ( s ) ] [m_\alpha(s),M_\alpha(s)] [ m α ( s ) , M α ( s )] . Monotonicity of F F F gives
a α ( s ) Q α ( t ) ≤ D α ( t ) ≤ c α ( s ) Q α ( t ) , t ≥ s . (22) a_\alpha(s)Q_\alpha(t)\le D_\alpha(t)\le c_\alpha(s)Q_\alpha(t),\qquad t\ge s. \tag*{(22)} a α ( s ) Q α ( t ) ≤ D α ( t ) ≤ c α ( s ) Q α ( t ) , t ≥ s . (22)
The static bounded-likelihood-ratio comparison between KL and chi-squared divergences is established [44 ] . The dynamical construction retains the actual preparation weights and uses Markov contraction to keep the comparison valid over a time interval. There is no omitted Taylor remainder. For fixed s s s the factors remain fixed; updating the reference time tightens them as the likelihood-ratio range contracts toward zero.
Finite-time crossing bounds with an occupied hot slow sector
A sufficient condition for arbitrary finite state count
Let λ s = λ 2 \lambda_s=\lambda_2 λ s = λ 2 and let λ f > λ s \lambda_f>\lambda_s λ f > λ s be the next distinct positive rate. The explicit time formula in this subsection assumes that this next rate exists. Define
w α = ∑ n : λ n = λ s b n , α 2 , q H = ∑ n : λ n > λ s b n , H 2 , w_\alpha=\sum_{n:\lambda_n=\lambda_s} b_{n,\alpha}^2,\qquad q_H=\sum_{n:\lambda_n>\lambda_s} b_{n,H}^2, w α = n : λ n = λ s ∑ b n , α 2 , q H = n : λ n > λ s ∑ b n , H 2 ,
c H = F ( m H ( 0 ) ) , a C = F ( M C ( 0 ) ) , c_H=F(m_H(0)),\qquad a_C=F(M_C(0)), c H = F ( m H ( 0 )) , a C = F ( M C ( 0 )) ,
Γ = a C w C − c H w H . (23) \Gamma=a_Cw_C-c_Hw_H. \tag*{(23)} Γ = a C w C − c H w H . (23)
The slow weight includes the entire eigenspace, so it is unchanged by rotating the eigenbasis within a degeneracy. Physically, w H w_H w H and w C w_C w C measure the probability departure stored in the slowest sector; q H q_H q H measures all faster hot contributions.
Theorem 1 (slow-sector crossing bound). For the common finite irreducible reversible generator and positive preparations specified above, assume D H ( 0 ) > D C ( 0 ) D_H(0)>D_C(0) D H ( 0 ) > D C ( 0 ) and Γ > 0 \Gamma>0 Γ > 0 . Then a first meeting exists and
t M ≤ T u p = 1 2 ( λ f − λ s ) ln ( c H q H Γ ) . (24) t_M\le T_{\mathrm{up}}=\frac{1}{2(\lambda_f-\lambda_s)}\ln\left(\frac{c_Hq_H}{\Gamma}\right). \tag*{(24)} t M ≤ T up = 2 ( λ f − λ s ) 1 ln ( Γ c H q H ) . (24) Furthermore D H ( t ) < D C ( t ) D_H(t)<D_C(t) D H ( t ) < D C ( t ) for every t > T u p t>T_{\mathrm{up}} t > T up .
The proof is short enough to expose the physical content. Positivity of the spectral weights gives
Q H ( t ) ≤ w H e − 2 λ s t + q H e − 2 λ f t , Q_H(t)\le w_H e^{-2\lambda_s t}+q_H e^{-2\lambda_f t}, Q H ( t ) ≤ w H e − 2 λ s t + q H e − 2 λ f t ,
Q C ( t ) ≥ w C e − 2 λ s t . (25) Q_C(t)\ge w_C e^{-2\lambda_s t}. \tag*{(25)} Q C ( t ) ≥ w C e − 2 λ s t . (25)
Combining these with Eq. (22) at s = 0 s=0 s = 0 yields
d ( t ) ≤ e − 2 λ s t [ c H q H e − 2 ( λ f − λ s ) t − Γ ] . (26) d(t)\le e^{-2\lambda_s t}\left[c_Hq_H e^{-2(\lambda_f-\lambda_s)t}-\Gamma\right]. \tag*{(26)} d ( t ) ≤ e − 2 λ s t [ c H q H e − 2 ( λ f − λ s ) t − Γ ] . (26)
Initially d ( 0 ) > 0 d(0)>0 d ( 0 ) > 0 , so necessarily c H q H > Γ > 0 c_Hq_H>\Gamma>0 c H q H > Γ > 0 and Eq. (24) is positive and defined. Its time is exactly when the bracket becomes zero, after which the bracket is negative. Continuity proves the bound. Appendix B gives the complete decomposition and explains restarting the bound at a later reference time.
The condition permits w H > 0 w_H>0 w H > 0 ; exact cancellation is unnecessary. The initially larger hot distance can be dominated by fast contributions that disappear before the colder preparation’s slow contribution. Figure 1 shows this for a detuned three-state preparation with w H ≃ 1.81 × 10 − 5 > 0 w_H\simeq1.81\times10^{-5}>0 w H ≃ 1.81 × 1 0 − 5 > 0 , c H ≃ 6.41 × 10 − 4 c_H\simeq6.41\times10^{-4} c H ≃ 6.41 × 1 0 − 4 , and Γ ≃ 2.17 × 10 − 4 \Gamma\simeq2.17\times10^{-4} Γ ≃ 2.17 × 1 0 − 4 : the theorem guarantees a crossing by T u p ≃ 0.185 T_{\mathrm{up}}\simeq0.185 T up ≃ 0.185 , about 9% after the actual crossing t M ≃ 0.170 t_M\simeq0.170 t M ≃ 0.170 . The positive margin Γ \Gamma Γ quantifies the permitted hot slow weight under this comparison, while λ f − λ s \lambda_f-\lambda_s λ f − λ s controls how quickly the excess fast contribution becomes small. Strong asymptotic acceleration additionally requires w H = 0 w_H=0 w H = 0 . For w H > 0 w_H>0 w H > 0 and w C > 0 w_C>0 w C > 0 , both preparations ultimately have the same exponential decay rate and D H / D C → w H / w C D_H/D_C\to w_H/w_C D H / D C → w H / w C .
The theorem is sufficient and is not asserted to be uniformly sharp. It discards faster cold contributions and merges all faster hot rates. If Γ ≤ 0 \Gamma\le0 Γ ≤ 0 , it decides nothing about crossing. If there is only one distinct positive rate, this explicit formula does not apply; the entropy envelopes still do. The theorem does not establish uniqueness of meetings before T u p T_{\mathrm{up}} T up . The familiar asymptotic condition w H < w C w_H<w_C w H < w C , when w C > 0 w_C>0 w C > 0 , is weaker than the displayed finite-amplitude criterion; the latter supplies a finite upper time using the specified range factors. The ingredients are the established spectral mechanism [3 , 4 ] and divergence comparison [44 ] ; our claim concerns their preparation-dependent finite-time use.
All-mode bounds on the first meeting
Keeping all weights gives tighter comparisons than merging the fast sector. For any reference time s s s define
G − ( t ; s ) = a H ( s ) Q H ( t ) − c C ( s ) Q C ( t ) , G_-(t;s)=a_H(s)Q_H(t)-c_C(s)Q_C(t), G − ( t ; s ) = a H ( s ) Q H ( t ) − c C ( s ) Q C ( t ) ,
G + ( t ; s ) = c H ( s ) Q H ( t ) − a C ( s ) Q C ( t ) . (27) G_+(t;s)=c_H(s)Q_H(t)-a_C(s)Q_C(t). \tag*{(27)} G + ( t ; s ) = c H ( s ) Q H ( t ) − a C ( s ) Q C ( t ) . (27)
The entropy envelopes imply
G − ( t ; s ) ≤ d ( t ) ≤ G + ( t ; s ) , t ≥ s . (28) G_-(t;s)\le d(t)\le G_+(t;s),\qquad t\ge s. \tag*{(28)} G − ( t ; s ) ≤ d ( t ) ≤ G + ( t ; s ) , t ≥ s . (28)
If earlier analysis excludes meetings before s s s , and G − ( t ; s ) > 0 G_-(t;s)>0 G − ( t ; s ) > 0 throughout s ≤ t < L s\le t<L s ≤ t < L , then t M ≥ L t_M\ge L t M ≥ L . Any U ≥ s U \ge s U ≥ s with G + ( U ; s ) ≤ 0 G_+(U;s) \le0 G + ( U ; s ) ≤ 0 gives t M ≤ U t_M \le U t M ≤ U . A lower bound on the first meeting requires positivity throughout the earlier interval; a positive value at one endpoint does not exclude a previous meeting and recrossing.
For fixed s s s , these bounds are finite sums of exponentials. Subdivision and interval evaluation can establish their signs. A reference time can be advanced once the required early-time exclusion has been proved. The reduced range factors then improve the entropy comparison without changing the generator or initial amplitudes. Zeros of the bounding exponential sums do not directly count the zeros of the nonlinear entropy difference.
A controlled remainder when only slow modes are known
Computing a full spectrum may be unnecessary. Retain modes 2 , … , m 2,\ldots,m 2 , … , m and suppose every omitted rate is at least λ c u t > 0 \lambda_{\mathrm{cut}}>0 λ cut > 0 . Define
Q α r e t ( t ) = ∑ n = 2 m b n , α 2 e − 2 λ n t , Q_\alpha^{\mathrm{ret}}(t)=\sum_{n=2}^{m}b_{n,\alpha}^{2}e^{-2\lambda_n t}, Q α ret ( t ) = n = 2 ∑ m b n , α 2 e − 2 λ n t ,
R α ( 0 ) = Q α ( 0 ) − ∑ n = 2 m b n , α 2 . (29) R_\alpha(0)=Q_\alpha(0)-\sum_{n=2}^{m}b_{n,\alpha}^{2}. \tag*{(29)} R α ( 0 ) = Q α ( 0 ) − n = 2 ∑ m b n , α 2 . (29)
All omitted squared weights are nonnegative, so
0 ≤ Q α ( t ) − Q α r e t ( t ) ≤ R α ( 0 ) e − 2 λ c u t t . (30) 0\le Q_\alpha(t)-Q_\alpha^{\mathrm{ret}}(t)\le R_\alpha(0)e^{-2\lambda_{\mathrm{cut}}t}. \tag*{(30)} 0 ≤ Q α ( t ) − Q α ret ( t ) ≤ R α ( 0 ) e − 2 λ cut t . (30)
The lower retained norm and upper retained-plus-remainder norm can be substituted into Eq. (28) , with a hot upper norm and cold lower norm for an upper bound on d d d , and the reverse choices for a lower bound. Dropping the omitted modes without this remainder control does not provide a certificate.
Errors in the retained spectral data can be included explicitly. Suppose certified intervals enclose w ‾ n , α ≤ b n , α 2 ≤ w ‾ n , α \underline{w}_{n,\alpha}\le b_{n,\alpha}^{2}\le\overline{w}_{n,\alpha} w n , α ≤ b n , α 2 ≤ w n , α and λ ‾ n ≤ λ n ≤ λ ‾ n \underline{\lambda}_n\le\lambda_n\le\overline{\lambda}_n λ n ≤ λ n ≤ λ n , with nonnegative weight bounds and a certified R α ( 0 ) ≤ R ‾ α R_\alpha(0)\le\overline{R}_\alpha R α ( 0 ) ≤ R α . Then for t ≥ 0 t\ge0 t ≥ 0 ,
Q ‾ α ( t ) = ∑ n = 2 m w ‾ n , α e − 2 λ ‾ n t , \underline{Q}_\alpha(t)=\sum_{n=2}^{m}\underline{w}_{n,\alpha}e^{-2\overline{\lambda}_n t}, Q α ( t ) = n = 2 ∑ m w n , α e − 2 λ n t ,
Q ‾ α ( t ) = ∑ n = 2 m w ‾ n , α e − 2 λ ‾ n t + R ‾ α e − 2 λ c u t t , \overline{Q}_\alpha(t)=\sum_{n=2}^{m}\overline{w}_{n,\alpha}e^{-2\underline{\lambda}_n t}+\overline{R}_\alpha e^{-2\lambda_{\mathrm{cut}}t}, Q α ( t ) = n = 2 ∑ m w n , α e − 2 λ n t + R α e − 2 λ cut t ,
Q ‾ α ( t ) ≤ Q α ( t ) ≤ Q ‾ α ( t ) . (31) \underline{Q}_\alpha(t)\le Q_\alpha(t)\le\overline{Q}_\alpha(t). \tag*{(31)} Q α ( t ) ≤ Q α ( t ) ≤ Q α ( t ) . (31)
For example, an upper bound on Q α ( 0 ) Q_\alpha(0) Q α ( 0 ) minus the sum of retained weight lower bounds gives a remainder upper bound. The range factors, exponentials, and final sign comparisons must also be evaluated with controlled rounding. Closely spaced or degenerate eigenvalues require validated subspace weights or a justified orthonormal basis; a small floating-point eigenvector residual alone is not an enclosure.
Preparation tolerance and crossing-time sensitivity
An explicit probability-error budget
Keep the generator and stationary reference fixed. Let perturbed preparations satisfy
∥ p α ( 0 ) − p α ( 0 ) ∥ 1 ≤ ϵ α . (32) \lVert p^\alpha(0)-p^{\alpha}(0)\rVert_1\le\epsilon_\alpha. \tag*{(32)} ∥ p α ( 0 ) − p α ( 0 ) ∥ 1 ≤ ϵ α . (32)
Suppose all nominal and perturbed initial likelihood ratios lie in [ h α , − , h α , + ] [h_{\alpha,-},h_{\alpha,+}] [ h α , − , h α , + ] , with h α , − > 0 h_{\alpha,-}>0 h α , − > 0 . The stochastic semigroup contracts the probability 1-norm error, and its likelihood-ratio kernel preserves the specified range. A centered entropy-gradient estimate gives
∣ D ~ α ( t ) − D α ( t ) ∣ ≤ C α ϵ α , C α = 1 2 ln h α , + h α , − , \left|\widetilde{D}_{\alpha}(t)-D_{\alpha}(t)\right|\le C_{\alpha}\epsilon_{\alpha},\qquad C_{\alpha}=\frac{1}{2}\ln\frac{h_{\alpha,+}}{h_{\alpha,-}}, D α ( t ) − D α ( t ) ≤ C α ϵ α , C α = 2 1 ln h α , − h α , + ,
∣ d ~ ( t ) − d ( t ) ∣ ≤ δ D : = C H ϵ H + C C ϵ C . (33) \left|\widetilde{d}(t)-d(t)\right|\le\delta_D:=C_H\epsilon_H+C_C\epsilon_C. \tag*{(33)} d ( t ) − d ( t ) ≤ δ D := C H ϵ H + C C ϵ C . (33)
Appendix C derives the factor 1 / 2 1/2 1/2 by exploiting normalization, rather than using an uncentered entropy gradient. Theorem 2 (robust first-meeting interval). Suppose 0 ≤ L < U 0\le L<U 0 ≤ L < U and
d ( t ) ≥ η L > 0 ( 0 ≤ t ≤ L ) , d ( U ) ≤ − η U < 0. (34) d(t)\ge\eta_L>0\quad(0\le t\le L),\qquad d(U)\le-\eta_U<0. \tag*{(34)} d ( t ) ≥ η L > 0 ( 0 ≤ t ≤ L ) , d ( U ) ≤ − η U < 0. (34)
Every allowed perturbed pair has its first meeting in ( L , U ) (L,U) ( L , U ) if
δ D < min { η L , η U } . (35) \delta_D<\min\{\eta_L,\eta_U\}. \tag*{(35)} δ D < min { η L , η U } . (35)
Indeed Eq. (33) preserves the positive early-time gap and the negative endpoint gap. Continuity then places the first zero strictly between them. The theorem establishes persistence with a specified error budget; it does not require a perfectly canceled mode or prove that the perturbed trajectory crosses only once.
The range assumption is checkable. Since a normalized probability error with 1-norm at most ϵ \epsilon ϵ satisfies ∣ δ p i ∣ ≤ ϵ / 2 |\delta p_i|\le\epsilon/2 ∣ δ p i ∣ ≤ ϵ /2 , one may use
h − = min i p i ( 0 ) − ϵ / 2 π i , h + = max i p i ( 0 ) + ϵ / 2 π i . (36) h_-=\min_i\frac{p_i(0)-\epsilon/2}{\pi_i},\qquad
h_+=\max_i\frac{p_i(0)+\epsilon/2}{\pi_i}. \tag*{(36)} h − = i min π i p i ( 0 ) − ϵ /2 , h + = i max π i p i ( 0 ) + ϵ /2 . (36)
provided h − > 0 h_->0 h − > 0 . Direct bounds for the physical preparation family can be less conservative. The same distinction applies to the sign margins: they must be certified on the stated time set, rather than inferred from a few sampled points.
From temperature uncertainty to preparation uncertainty
For the Gibbs preparations of Eq. (4) ,
∂ p i ( T ) ∂ T = p i ( T ) [ E i − U ( T ) ] T 2 , U ( T ) = ∑ i p i ( T ) E i . (37) \frac{\partial p_i(T)}{\partial T}=\frac{p_i(T)[E_i-U(T)]}{T^2},\qquad U(T)=\sum_i p_i(T)E_i. \tag*{(37)} ∂ T ∂ p i ( T ) = T 2 p i ( T ) [ E i − U ( T )] , U ( T ) = i ∑ p i ( T ) E i . (37)
The mean absolute energy deviation is at most half the energy range. Consequently, for T , T ~ ≥ T min > 0 T,\widetilde{T}\ge T_{\min}>0 T , T ≥ T m i n > 0 ,
∥ p ( T ~ ) − p ( T ) ∥ 1 ≤ E max − E min 2 T min 2 ∣ T ~ − T ∣ . (38) \lVert p(\widetilde{T})-p(T)\rVert_1\le\frac{E_{\max}-E_{\min}}{2T_{\min}^2}|\widetilde{T}-T|. \tag*{(38)} ∥ p ( T ) − p ( T ) ∥ 1 ≤ 2 T m i n 2 E m a x − E m i n ∣ T − T ∣. (38)
The proof uses Cauchy–Schwarz and an elementary bounded-variable variance inequality, both shown in Appendix C. This converts a temperature interval into an explicit input for Theorem 2. Interval evaluation of the Gibbs probabilities can improve the numerical tolerance. A permitted temperature interval is a derived conclusion only when its likelihood-ratio and sign-margin conditions have also been verified.
Transition-rate uncertainty changes the reference as well
For two fixed irreducible generators W W W and W ~ \widetilde{W} W , put δ W = ∥ W ~ − W ∥ 1 → 1 \delta_W=\lVert\widetilde{W}-W\rVert_{1\to1} δ W = ∥ W − W ∥ 1 → 1 . Duhamel’s formula and stochastic contraction imply
∥ p ~ α ( t ) − p α ( t ) ∥ 1 ≤ ϵ α + t δ W . (39) \lVert\widetilde{p}^{\alpha}(t)-p^{\alpha}(t)\rVert_1\le\epsilon_{\alpha}+t\delta_W. \tag*{(39)} ∥ p α ( t ) − p α ( t ) ∥ 1 ≤ ϵ α + t δ W . (39)
Let C α C_{\alpha} C α now be a valid centered-gradient constant for both distributions compared with the nominal π \pi π over 0 ≤ t ≤ U 0\le t\le U 0 ≤ t ≤ U . If the perturbed stationary distribution is π ~ \widetilde{\pi} π , define δ π = max i ∣ ln ( π i / π ~ i ) ∣ \delta_{\pi}=\max_i|\ln(\pi_i/\widetilde{\pi}_i)| δ π = max i ∣ ln ( π i / π i ) ∣ . Then
∣ d ~ ( t ) − d ( t ) ∣ ≤ ∑ α = H , C C α ( ϵ α + U δ W ) + 2 δ π , |\widetilde{d}(t)-d(t)|\le\sum_{\alpha=H,C}C_{\alpha}(\epsilon_{\alpha}+U\delta_W)+2\delta_{\pi}, ∣ d ( t ) − d ( t ) ∣ ≤ α = H , C ∑ C α ( ϵ α + U δ W ) + 2 δ π ,
0 ≤ t ≤ U . (40) 0\le t\le U. \tag*{(40)} 0 ≤ t ≤ U . (40)
Here D ~ \widetilde{D} D is measured relative to π ~ \widetilde{\pi} π . If the right-hand side is below both margins in Eq. (34) , the first meeting remains in ( L , U ) (L,U) ( L , U ) . Appendix C3 supplies the proof and a way to bound the likelihood-ratio ranges on the finite time interval. Keeping the old stationary reference without the δ π \delta_{\pi} δ π correction would not establish the claimed uncertainty result.
Local sensitivity and the entropy-production advantage
For a smooth parameter θ \theta θ , consider a simple meeting root d ( t ∗ ( θ ) , θ ) = 0 d(t_*(\theta),\theta)=0 d ( t ∗ ( θ ) , θ ) = 0 . At the root write D ∗ = D H ( t ∗ ) = D C ( t ∗ ) > 0 D_*=D_H(t_*)=D_C(t_*)>0 D ∗ = D H ( t ∗ ) = D C ( t ∗ ) > 0 . The entropy balance gives
∂ t d ( t ∗ ) = − D ∗ [ r H ( t ∗ ) − r C ( t ∗ ) ] . (41) \partial_t d(t_*)=-D_*[r_H(t_*)-r_C(t_*)]. \tag*{(41)} ∂ t d ( t ∗ ) = − D ∗ [ r H ( t ∗ ) − r C ( t ∗ )] . (41)
When r H ( t ∗ ) ≠ r C ( t ∗ ) r_H(t_*)\ne r_C(t_*) r H ( t ∗ ) = r C ( t ∗ ) , implicit differentiation therefore yields
∂ t ∗ ∂ θ = ∂ θ d ( t ∗ , θ ) D ∗ [ r H ( t ∗ ) − r C ( t ∗ ) ] . (42) \frac{\partial t_*}{\partial\theta}=
\frac{\partial_\theta d(t_*,\theta)}
{D_*[r_H(t_*)-r_C(t_*)]}. \tag*{(42)} ∂ θ ∂ t ∗ = D ∗ [ r H ( t ∗ ) − r C ( t ∗ )] ∂ θ d ( t ∗ , θ ) . (42)
The partial derivative holds time fixed and includes any dependence of the stationary reference on θ \theta θ . For a fixed numerator, a small normalized-rate difference makes the meeting time sensitive. This tracks a local simple root; a first-meeting interpretation also needs an early-time exclusion. At a tangent meeting the denominator vanishes, and finite sign margins provide a separate route to robust conclusions.
Scope and a negative control
The spectral bound uses reversibility. The scalar entropy envelopes and the fixed-generator probability-error argument need only a fixed positive stationary distribution. For a nonreversible generator one must evaluate the full semigroup norm for Q Q Q instead of assuming orthogonal eigenmodes. Chi-squared divergence remains contractive even in that case; nonorthogonal modes do not imply transient growth of this divergence. Appendix D proves these statements and recalls why nonadiabatic entropy production replaces total entropy production in the fixed-reference clock [53 , 54 ] .
A two-state Gibbs cooling protocol is a useful negative control. Its excited-state probabilities remain ordered as p 1 H ( t ) > p 1 C ( t ) > π 1 p_1^H(t) > p_1^C(t) > \pi_1 p 1 H ( t ) > p 1 C ( t ) > π 1 at every finite time, and binary relative entropy increases with p 1 p_1 p 1 on that side of equilibrium. Hence it has no relative-entropy crossing. The explicit proof in Appendix D applies to T H > T C > T b T_H > T_C > T_b T H > T C > T b , rather than to arbitrary pairs of binary distributions.
Tests on reversible network ensembles
Sampling protocol and the quantities tested
We test the finite-time condition in Eq. (23) without adjusting a preparation temperature to cancel a slow mode. All network realizations use
( T H , T C , T b ) = ( 3 , 1.5 , 1 ) , k B = 1. (43) (T_H,T_C,T_b)=(3,1.5,1), \qquad k_B=1. \tag*{(43)} ( T H , T C , T b ) = ( 3 , 1.5 , 1 ) , k B = 1. (43)
For each N ∈ { 4 , 8 , 16 , 32 , 64 } N \in\{4,8,16,32,64\} N ∈ { 4 , 8 , 16 , 32 , 64 } and each of four graph families, we generate 100 realizations. The 2000 draws, including every unsuccessful draw, are retained in the supplied data.
The energy levels are a random permutation of
E j = 5 j N − 1 , j = 0 , … , N − 1. (44) E_j=\frac{5j}{N-1}, \qquad j=0,\ldots,N-1. \tag*{(44)} E j = N − 1 5 j , j = 0 , … , N − 1. (44)
Thus every size has the same energy range; the placement of those levels on the graph is randomized. Allowed undirected edges receive independent symmetric prefactors
γ i j = γ j i = e X i j , X i j ∼ Uniform [ − 4 , 2 ] , (45) \gamma_{ij}=\gamma_{ji}=e^{X_{ij}}, \qquad X_{ij}\sim\operatorname{Uniform}[-4,2], \tag*{(45)} γ ij = γ j i = e X ij , X ij ∼ Uniform [ − 4 , 2 ] , (45)
and the directed rates are
k i j = γ i j e − ( E j − E i ) / 2 . (46) k_{ij}=\gamma_{ij}e^{-(E_j-E_i)/2}. \tag*{(46)} k ij = γ ij e − ( E j − E i ) /2 . (46)
Detailed balance follows directly:
π i k i j = γ i j Z b e − ( E i + E j ) / 2 = π j k j i . (47) \pi_i k_{ij}=\frac{\gamma_{ij}}{Z_b}e^{-(E_i+E_j)/2}=\pi_j k_{ji}. \tag*{(47)} π i k ij = Z b γ ij e − ( E i + E j ) /2 = π j k j i . (47)
The graph families are complete graphs, rings, birth–death paths, and sparse connected graphs. A sparse graph contains a ring backbone, with each remaining undirected edge added independently with probability min ( 3 / N , 1 ) \min(3/N,1) min ( 3/ N , 1 ) . Each graph is therefore connected before any rate is drawn. Appendix E specifies the random-number streams, entropy evaluation, and crossing search.
The complete symmetric prefactor matrix is supplied as exact finite decimals in selected_network_exact_inputs.json; this file, Eq. (46) , and the energy vector define the reference model unambiguously. Its first rates and weights are
λ s ≃ 0.51374978 , λ f ≃ 2.27730802 , (48) \lambda_s \simeq0.51374978,\qquad\lambda_f \simeq2.27730802, \tag*{(48)} λ s ≃ 0.51374978 , λ f ≃ 2.27730802 , (48)
w H ≃ 4.1466755 × 10 − 6 , w C ≃ 4.5240625 × 10 − 4 , (49) w_H \simeq4.1466755\times10^{-6},\qquad w_C \simeq4.5240625\times10^{-4}, \tag*{(49)} w H ≃ 4.1466755 × 1 0 − 6 , w C ≃ 4.5240625 × 1 0 − 4 , (49)
q H ≃ 1.5733603 , Γ ≃ 1.2529349 × 10 − 4 . (50) q_H \simeq1.5733603,\qquad\Gamma\simeq1.2529349\times10^{-4}. \tag*{(50)} q H ≃ 1.5733603 , Γ ≃ 1.2529349 × 1 0 − 4 . (50)
For every draw we record the slow weights w H , w C w_H,w_C w H , w C , the sufficient margin Γ \Gamma Γ , the upper formula T + = T u p T_+=T_{\mathrm{up}} T + = T up in Eq. (24) , and the outcome of a finite-time entropy search. Times are compared in units τ = λ s t \tau=\lambda_s t τ = λ s t , which remove a uniform rescaling of all rates. The search uses a declared time grid followed by root refinement. We call its result a located crossing ; numerical root refinement alone does not establish that this is the first or only crossing.
Coverage and limitations of the sufficient condition
Table Table I reports both successes and failures. There are 86 located crossings and 45 positive-margin draws. The simple condition therefore covers 45 / 86 ≃ 52.3 % 45/86\simeq52.3\% 45/86 ≃ 52.3% of the crossings observed in this sample. Every located crossing also has w H < w C w_H<w_C w H < w C . No transient crossing with the opposite asymptotic ordering was detected by the declared search. The sampling and finite search do not exclude that possibility in another system.
Among the 45 positive-margin draws, the ratio of the upper formula to the located crossing ranges from 1.214 to 30.193, with median 2.154. The criterion is useful as a sufficient guarantee, but it can be loose. In particular, no dimension-independent sharpness claim follows from Figure 2 . The occurrence counts also depend on the chosen energy range, temperature pair, graph ensemble, and rate distribution. The absence of an observed crossing in the sampled four-state graphs does not imply a four-state impossibility result.
Figure 2. Reversible network ensembles at fixed preparation temperatures ( T H , T C , T b ) = ( 3 , 1.5 , 1 ) (T_H,T_C,T_b)=(3,1.5,1) ( T H , T C , T b ) = ( 3 , 1.5 , 1 ) . (a) Located crossings per 100 draws for four graph families; lines guide the eye. (b) Slow-sector upper bound T u p T_{\mathrm{up}} T up against the located crossing t M t_M t M for the 45 draws with Γ > 0 \Gamma>0 Γ > 0 , in units of λ s \lambda_s λ s ; the dashed diagonal marks equality, so every point lies on the conservative side. (c) Relative entropies of the selected eight-state complete graph. The dot marks the crossing; the dotted and dash-dotted lines are the slow-sector bound and the bound retaining two modes plus a controlled tail. (d) Squared weights of all seven nonstationary modes: the hot slow weight b 2 2 b_2^2 b 2 2 is small but nonzero. All curves are double-precision evaluations; the certified statements are listed in Table III .
Table III. First-meeting statements established by interval arithmetic for the exact inputs specified in the text and supplement. Each lower endpoint is supported by an exclusion on the whole preceding time interval. None of these rows asserts global uniqueness, certification of an entire sampled network ensemble, or a continuum limit.
An eight-state example and a partial-spectrum improvement
To fix a selection rule independently of the subsequent plots, we use the first positive-margin draw in the family-size-draw order specified in Appendix E. It is complete graph N = 8 N=8 N = 8 , draw 6. The energy vector is
E = 5 7 ( 5 , 2 , 3 , 0 , 7 , 6 , 1 , 4 ) . (51) E=\frac{5}{7}(5,2,3,0,7,6,1,4). \tag*{(51)} E = 7 5 ( 5 , 2 , 3 , 0 , 7 , 6 , 1 , 4 ) . (51)
The complete symmetric prefactor matrix is supplied as exact finite decimals in selected_network_exact_inputs.json; this file, Eq. (46) , and the energy vector define the reference model unambiguously. Its first rates and weights are
λ s ≃ 0.51374978 , λ f ≃ 2.27730802 , (52) \lambda_s\simeq0.51374978,\qquad\lambda_f\simeq2.27730802, \tag*{(52)} λ s ≃ 0.51374978 , λ f ≃ 2.27730802 , (52)
w H ≃ 4.1466755 × 10 − 6 , w C ≃ 4.5240625 × 10 − 4 , (53) w_H\simeq4.1466755\times10^{-6},\qquad w_C\simeq4.5240625\times10^{-4}, \tag*{(53)} w H ≃ 4.1466755 × 1 0 − 6 , w C ≃ 4.5240625 × 1 0 − 4 , (53)
q H ≃ 1.5733603 , Γ ≃ 1.2529349 × 10 − 4 . (54) q_H\simeq1.5733603,\qquad\Gamma\simeq1.2529349\times10^{-4}. \tag*{(54)} q H ≃ 1.5733603 , Γ ≃ 1.2529349 × 1 0 − 4 . (54)
Table 1. Located crossings / positive-Γ \Gamma Γ draws, out of 100 realizations in each cell, at the fixed temperatures in Eq. (43) . These are double-precision sample counts. A negative margin is inconclusive, and failure to detect a crossing in a finite search is not a no-crossing theorem.
The initial entropies are D H ( 0 ) ≃ 0.38363557 D_H(0)\simeq0.38363557 D H ( 0 ) ≃ 0.38363557 and D C ( 0 ) ≃ 0.06814960 D_C(0)\simeq0.06814960 D C ( 0 ) ≃ 0.06814960 . The located crossing and evaluated upper formula are
t c r o s s ≃ 1.2694268 , T + ≃ 2.5395000. (55) t_{\mathrm{cross}}\simeq1.2694268,\qquad T_+\simeq2.5395000. \tag*{(55)} t cross ≃ 1.2694268 , T + ≃ 2.5395000. (55)
An independent interval matrix exponential, together with a semigroup enclosure proving positivity throughout [ 0 , 1.26 ] [0,1.26] [ 0 , 1.26 ] , certifies the first crossing in ( 1.26 , 1.28 ) (1.26,1.28) ( 1.26 , 1.28 ) for the exact reference inputs. Appendix F supplies the interval-cover proof. This propagation proof is independent of a separate validated spectral calculation. The latter encloses the simple spectrum and its eigenvectors, proves w H > 0 w_H > 0 w H > 0 and Γ > 0 \Gamma> 0 Γ > 0 , and establishes the outward upper endpoint
T u p < 2.53950003286112. (56) T_{\mathrm{up}} < 2.53950003286112. \tag*{(56)} T up < 2.53950003286112. (56)
Neither calculation asserts global uniqueness of crossings. The partial-spectrum bound of Eq. (30) gives a useful improvement. Keep the first two nonstationary modes and denote their weights by b 2 , α 2 , b 3 , α 2 b_{2,\alpha}^2, b_{3,\alpha}^2 b 2 , α 2 , b 3 , α 2 . The next omitted rate is λ 4 ≃ 17.27305636 \lambda_4 \simeq17.27305636 λ 4 ≃ 17.27305636 . The hot norm is bounded above by
Q H ( t ) ≤ b 2 , H 2 e − 2 λ 2 t + b 3 , H 2 e − 2 λ 3 t + R H ( 0 ) e − 2 λ 4 t , (57) Q_H(t) \le b_{2,H}^2 e^{-2\lambda_2 t} + b_{3,H}^2 e^{-2\lambda_3 t} + R_H(0)e^{-2\lambda_4 t}, \tag*{(57)} Q H ( t ) ≤ b 2 , H 2 e − 2 λ 2 t + b 3 , H 2 e − 2 λ 3 t + R H ( 0 ) e − 2 λ 4 t , (57)
where R H ( 0 ) = Q H ( 0 ) − b 2 , H 2 − b 3 , H 2 R_H(0) = Q_H(0) - b_{2,H}^2 - b_{3,H}^2 R H ( 0 ) = Q H ( 0 ) − b 2 , H 2 − b 3 , H 2 . Dropping the nonnegative cold remainder supplies the corresponding lower bound for Q C ( t ) Q_C(t) Q C ( t ) . Inserting these into c H Q H − a C Q C c_H Q_H - a_C Q_C c H Q H − a C Q C gives an evaluated crossing upper bound of 1.5516211. The full-spectrum envelope gives the same value at the displayed precision. Thus a bound on the omitted contribution can retain most of the useful information in this example. Here the full spectrum was computed to evaluate and check the partial-information formula; we do not claim a measured computational speedup. A separate interval calculation validates both retained weights, their rates, and the omitted tail, using the exact lower cutoff λ c u t = 17.27305635 \lambda_{\mathrm{cut}} = 17.27305635 λ cut = 17.27305635 . At the exact time t = 1.55163 t = 1.55163 t = 1.55163 , the resulting upper envelope is strictly negative,
G + ( 1.55163 ; 0 ) < − 8.000112 × 10 − 10 < 0. (58) G_+(1.55163;0) < -8.000112 \times10^{-10} < 0. \tag*{(58)} G + ( 1.55163 ; 0 ) < − 8.000112 × 1 0 − 10 < 0. (58)
Thus t M < 1.55163 t_M < 1.55163 t M < 1.55163 is a rigorous partial-spectrum upper bound for the specified eight-state model. The decimal 1.5516211 above is a numerical root estimate for the envelope, not the endpoint used in this certificate.
A physical example: reflecting diffusion in an asymmetric potential
Dynamics, parameter search, and selection
Consider overdamped diffusion on the bounded interval [ − 3 , 3 ] [-3,3] [ − 3 , 3 ] ,
∂ t ρ ( x , t ) = ∂ x [ U ′ ( x ) ρ ( x , t ) + ∂ x ρ ( x , t ) ] , (59) \partial_t \rho(x,t) = \partial_x[U'(x)\rho(x,t) + \partial_x\rho(x,t)], \tag*{(59)} ∂ t ρ ( x , t ) = ∂ x [ U ′ ( x ) ρ ( x , t ) + ∂ x ρ ( x , t )] , (59)
with reflecting boundaries and units T b = γ = k B = 1 T_b = \gamma= k_B = 1 T b = γ = k B = 1 . Both preparations evolve with this same final potential and bath. Their initial densities are proportional to e − U ( x ) / T H e^{-U(x)/T_H} e − U ( x ) / T H and e − U ( x ) / T C e^{-U(x)/T_C} e − U ( x ) / T C , respectively. We use
U ( x ) = a x 2 2 − A e − ( x + 1 ) 2 / ( 2 s L 2 ) − B e − ( x − 1 ) 2 / ( 2 s R 2 ) . (60) U(x) = \frac{a x^2}{2} - A e^{-(x+1)^2/(2s_L^2)} - B e^{-(x-1)^2/(2s_R^2)}. \tag*{(60)} U ( x ) = 2 a x 2 − A e − ( x + 1 ) 2 / ( 2 s L 2 ) − B e − ( x − 1 ) 2 / ( 2 s R 2 ) . (60)
The finite search takes A ∈ { 3 , 4 , 5 , 6 } A \in\{3,4,5,6\} A ∈ { 3 , 4 , 5 , 6 } , B ∈ { 2 , 3 , 4 , 5 } B \in\{2,3,4,5\} B ∈ { 2 , 3 , 4 , 5 } with A > B A > B A > B , s L ∈ { 0.2 , 0.3 , 0.5 } s_L \in\{0.2,0.3,0.5\} s L ∈ { 0.2 , 0.3 , 0.5 } , s R ∈ { 0.4 , 0.7 , 1 } s_R \in\{0.4,0.7,1\} s R ∈ { 0.4 , 0.7 , 1 } , and a ∈ { 0 , 0.5 } a \in\{0,0.5\} a ∈ { 0 , 0.5 } , in that Cartesian-product order. At N = 128 N = 128 N = 128 this gives 180 potential choices. Seven show a located crossing and have w H < w C w_H < w_C w H < w C ; three satisfy Γ > 0 \Gamma> 0 Γ > 0 . We select the first positive-margin case:
( A , B , s L , s R , a ) = ( 3 , 2 , 0.2 , 0.7 , 0 ) . (61) (A,B,s_L,s_R,a) = (3,2,0.2,0.7,0). \tag*{(61)} ( A , B , s L , s R , a ) = ( 3 , 2 , 0.2 , 0.7 , 0 ) . (61)
The temperatures remain ( T H , T C , T b ) = ( 3 , 1.5 , 1 ) (T_H,T_C,T_b) = (3,1.5,1) ( T H , T C , T b ) = ( 3 , 1.5 , 1 ) throughout. The selected potential has a narrow deeper left well and a broader right well Figure 3 (a). Selection from a disclosed finite parameter search is different from a prevalence statement; neither temperature is found by solving a slow-mode zero.
Figure 3. Reflecting diffusion in the selected asymmetric potential. (a) Potential U ( x ) U(x) U ( x ) on [ − 3 , 3 ] [-3,3] [ − 3 , 3 ] (left axis) with the stationary density and the two initial Gibbs densities (right axis). (b) Relative entropies at N = 512 N=512 N = 512 ; the dot marks the crossing and the dotted line the slow-sector bound T u p T_{\mathrm{up}} T up . (c) Difference between the crossing on each grid and on the finest grid, with a ( Δ x ) 2 (\Delta x)^2 ( Δ x ) 2 guide. (d) Margin Γ \Gamma Γ at all 81 sampled temperature pairs (white points); the colored cells only visualize these point values.
A reversible discretization and mesh refinement
Let x i = − 3 + ( i + 1 / 2 ) Δ x x_i = -3 + (i+1/2)\Delta x x i = − 3 + ( i + 1/2 ) Δ x , i = 0 , … , N − 1 i = 0,\ldots,N-1 i = 0 , … , N − 1 , with Δ x = 6 / N \Delta x = 6/N Δ x = 6/ N . Only nearest neighbors communicate, with
k i , i + 1 = ( Δ x ) − 2 e − [ U ( x i + 1 ) − U ( x i ) ] / 2 , k_{i,i+1} = (\Delta x)^{-2}e^{-[U(x_{i+1})-U(x_i)]/2}, k i , i + 1 = ( Δ x ) − 2 e − [ U ( x i + 1 ) − U ( x i )] /2 ,
k i + 1 , i = ( Δ x ) − 2 e + [ U ( x i + 1 ) − U ( x i ) ] / 2 . (62) k_{i+1,i} = (\Delta x)^{-2}e^{+[U(x_{i+1})-U(x_i)]/2}. \tag*{(62)} k i + 1 , i = ( Δ x ) − 2 e + [ U ( x i + 1 ) − U ( x i )] /2 . (62)
No jumps leave the interval. The finite model obeys detailed balance with the midpoint Gibbs weights. Appendix Appendix § 2.1.4 derives the flux consistency with Eq. (59) . The finite-state theorems apply to every such discretization.
Table II shows that the located crossing and upper formula stabilize under refinement. The difference between the N = 512 N = 512 N = 512 and N = 1024 N = 1024 N = 1024 located crossings is 1.6005 × 10 − 4 1.6005 \times10^{-4} 1.6005 × 1 0 − 4 , approximately 0.0052% of the finer value. This is a numerical convergence check; it is not an interval bound on the continuum crossing time. In particular, no extrapolated continuum value is used as a certified result.
At N = 512 N = 512 N = 512 , the computed slow weights are w H ≃ 1.1371813 × 10 − 5 w_H \simeq1.1371813 \times10^{-5} w H ≃ 1.1371813 × 1 0 − 5 and w C ≃ 9.8849729 × 10 − 4 w_C \simeq9.8849729 \times10^{-4} w C ≃ 9.8849729 × 1 0 − 4 . The hot contribution is small and nonzero at numerical precision. For the exact 64-cell discretization, independent interval propagation gives
D H ( 3 ) − D C ( 3 ) > 0 , D H ( 3.2 ) − D C ( 3.2 ) < 0. (63) D_H(3) - D_C(3) > 0,\qquad D_H(3.2) - D_C(3.2) < 0. \tag*{(63)} D H ( 3 ) − D C ( 3 ) > 0 , D H ( 3.2 ) − D C ( 3.2 ) < 0. (63)
An additional semigroup enclosure proves D H ( t ) − D C ( t ) > 0 D_H(t) - D_C(t) > 0 D H ( t ) − D C ( t ) > 0 throughout 0 ≤ t ≤ 3 0 \le t \le3 0 ≤ t ≤ 3 . Together these results certify the first crossing in ( 3 , 3.2 ) (3,3.2) ( 3 , 3.2 ) for that finite model. The interval-cover proof is given in Appendix Appendix F. A separate validated eigensystem also establishes w H > 0 w_H > 0 w H > 0 , Γ > 0 \Gamma> 0 Γ > 0 , and T u p < 4.0214384794728 T_{\mathrm{up}} < 4.0214384794728 T up < 4.0214384794728 for this grid. Global uniqueness and the continuum limit are not certified.
Preparation variation and a negative control
A separate 9 × 9 9 \times9 9 × 9 pointwise survey at N = 128 N = 128 N = 128 takes T H ∈ [ 2.8 , 3.2 ] T_H \in[2.8,3.2] T H ∈ [ 2.8 , 3.2 ] and T C ∈ [ 1.4 , 1.6 ] T_C \in[1.4,1.6] T C ∈ [ 1.4 , 1.6 ] , with equally spaced values on each axis. Every one of the 81 pairs has a positive computed margin, between 3.5897 × 10 − 4 3.5897 \times10^{-4} 3.5897 × 1 0 − 4 and 4.1289 × 10 − 4 4.1289 \times10^{-4} 4.1289 × 1 0 − 4 , and a located crossing between 3.01944 and 3.21891. This supports robustness of the observed diffusion example over the sampled preparations. It does not certify the full continuous temperature rectangle. The separate three-state rectangle in Appendix F does carry that stronger interval guarantee.
Table 2. Mesh-refinement evidence for the selected reflecting diffusion. Each row is a double-precision evaluation for its finite grid. The certificate in Appendix F applies separately to N = 64 N=64 N = 64 .
For an analytical negative control, take a two-state Gibbs system with energies ( 0 , 2 ) (0,2) ( 0 , 2 ) , symmetric prefactor one, and the same temperatures. Its excited-state occupation satisfies
p 1 α ( t ) = π 1 + [ p 1 α ( 0 ) − π 1 ] e − λ t . (64) p_{1}^{\alpha}(t)=\pi_{1}+[p_{1}^{\alpha}(0)-\pi_{1}]e^{-\lambda t}. \tag*{(64)} p 1 α ( t ) = π 1 + [ p 1 α ( 0 ) − π 1 ] e − λ t . (64)
Since T H > T C > T b T_H>T_C>T_b T H > T C > T b , one has p 1 H ( t ) > p 1 C ( t ) > π 1 p_{1}^{H}(t)>p_{1}^{C}(t)>\pi_{1} p 1 H ( t ) > p 1 C ( t ) > π 1 at every finite time. The derivative of binary relative entropy is
d D ( p ∥ π 1 ) d p = ln p ( 1 − π 1 ) π 1 ( 1 − p ) > 0 when p > π 1 . (65) \frac{dD(p\Vert\pi_{1})}{dp}=\ln\frac{p(1-\pi_{1})}{\pi_{1}(1-p)}>0 \quad\text{when }p>\pi_{1}. \tag*{(65)} d p d D ( p ∥ π 1 ) = ln π 1 ( 1 − p ) p ( 1 − π 1 ) > 0 when p > π 1 . (65)
Consequently D H ( t ) > D C ( t ) D_H(t)>D_C(t) D H ( t ) > D C ( t ) and there is no crossing. The numerical control agrees with this exact conclusion. The theorem’s next-distinct-rate assumption also explains why its two-sector formula is not applicable to this one-mode case.
Certified crossings and a continuous preparation region
The general bounds are analytical inequalities. Their evaluation becomes a numerical certificate only when the input errors and sign comparisons are also enclosed. Table III summarizes the claims verified independently with Arb ball arithmetic; the corresponding proof algorithms are given in Appendix F. The selected network and diffusion certificates use direct semigroup propagation for their first-meeting intervals. Separate validated eigenvalue and eigenvector enclosures establish their positive slow-sector margins and upper formulas. The eight-state partial-spectrum endpoint t M < 1.55163 t_M < 1.55163 t M < 1.55163 is also validated, including its omitted-tail cutoff.
To demonstrate a full region of thermal preparations, consider the exact three-state benchmark
E = ( 0 , 2 / 5 , 3 ) , ( γ 01 , γ 02 , γ 12 ) = ( 1 , 23 / 5 , 3 / 25 ) , (66) E = (0, 2/5, 3), \qquad(\gamma_{01},\gamma_{02},\gamma_{12}) = (1, 23/5, 3/25), \tag*{(66)} E = ( 0 , 2/5 , 3 ) , ( γ 01 , γ 02 , γ 12 ) = ( 1 , 23/5 , 3/25 ) , (66)
with symmetric prefactors, T b = 1 T_b = 1 T b = 1 , and the rate law in Eq. (46) . Both preparations always experience this same fixed generator. At ( T H , T C ) = ( 3 , 3 / 2 ) (T_H,T_C) = (3, 3/2) ( T H , T C ) = ( 3 , 3/2 ) , the hot slow weight is strictly positive and the first meeting lies in ( 0.17022 , 0.17024 ) (0.17022, 0.17024) ( 0.17022 , 0.17024 ) . This reference point is not an exactly strong preparation.
More substantially, every independent temperature pair in the closed rectangle
2.9 ≤ T H ≤ 3.1 , 1.45 ≤ T C ≤ 1.55 (67) 2.9 \le T_H \le3.1, \qquad1.45 \le T_C \le1.55 \tag*{(67)} 2.9 ≤ T H ≤ 3.1 , 1.45 ≤ T C ≤ 1.55 (67)
has its first meeting in
0.161 < t M < 0.179. (68) 0.161 < t_M < 0.179. \tag*{(68)} 0.161 < t M < 0.179. (68)
Throughout this entire rectangle, interval evaluation gives w H > 3.2401 × 10 − 6 w_H > 3.2401 \times10^{-6} w H > 3.2401 × 1 0 − 6 and Γ > 1.8624 × 10 − 4 \Gamma> 1.8624 \times10^{-4} Γ > 1.8624 × 1 0 − 4 . Thus no hot preparation in the certified region cancels its slow mode.
Figure 4 places this rectangle in the full plane of preparation temperatures. In double-precision evaluations on a 177 × 136 177 \times136 177 × 136 grid, crossings are located only in a lobe around the strong temperature T ⋆ ≃ 2.823 T b T_\star\simeq2.823 T_b T ⋆ ≃ 2.823 T b , where the hot slow weight is small. The margin is positive in the interior of that lobe, covering about 75% of the grid points with a located crossing, and every grid point with Γ > 0 \Gamma> 0 Γ > 0 has a located crossing, as Theorem 1 requires. The certified rectangle lies well inside the Γ > 0 \Gamma> 0 Γ > 0 region; crossings therefore do not depend on tuning T H T_H T H to T ⋆ T_\star T ⋆ .
The interior of this rectangle is an open set of preparation pairs with the same crossing property. The statement includes intermediate temperatures and the boundary; it is not inferred from a point grid.
The certificate covers the hot and cold temperature axes by adjacent rational intervals, encloses all Gibbs probabilities and modal coefficients over each interval, and then applies entropy envelopes to their combined ranges. A bootstrap excludes every meeting up to 0.161, while a uniform negative upper comparison at 0.179 supplies the other endpoint. The construction keeps the uncertainty correlations within each preparation enclosed and allows the two temperatures to vary independently. It does not prove that this is the largest permitted temperature region or that all trajectories have only one meeting.
An arbitrary probability perturbation has a different uncertainty. At the central pair ( 3 , 3 / 2 ) (3, 3/2) ( 3 , 3/2 ) , independent normalized initial errors ∥ δ p H ∥ 1 , ∥ δ p C ∥ 1 ≤ 10 − 6 \lVert\delta p_H\rVert_1,\lVert\delta p_C\rVert_1 \le10^{-6} ∥ δ p H ∥ 1 , ∥ δ p C ∥ 1 ≤ 1 0 − 6 preserve a first meeting in ( 0.169 , 0.172 ) (0.169, 0.172) ( 0.169 , 0.172 ) . For all such errors the likelihood ratios remain in [ 0.7 , 6 ] [0.7,6] [ 0.7 , 6 ] , giving C H = C C = 1 2 ln ( 6 / 0.7 ) C_H=C_C=\frac{1}{2}\ln(6/0.7) C H = C C = 2 1 ln ( 6/0.7 ) . The resulting total entropy-error budget is less than 2.149 × 10 − 6 2.149 \times10^{-6} 2.149 × 1 0 − 6 , below the separately certified sign margins. This is a direct application of Theorem 2. The larger thermal rectangle is obtained by enclosing its specific Gibbs family rather than treating every direction in probability space as an allowed error.
Physical interpretation, scope, and limitations
The crossing condition compares how the two preparations distribute their initial departures from equilibrium across slow and fast relaxation patterns. A hot state may have a larger total departure but a smaller entropy-weighted slow contribution. The excess fast contribution then decays sufficiently quickly for the ordering to reverse. The positive margin Γ \Gamma Γ supplies a quantitative allowance for a residual hot slow weight; the separation λ f − λ s \lambda_f-\lambda_s λ f − λ s sets the time needed for the remaining fast contribution to become smaller than that margin.
This interpretation distinguishes ordinary crossing from strong asymptotic acceleration. If both slow weights are nonzero, both relative entropies eventually decay with the same exponent, even though their amplitudes and threshold times differ. The strong effect requires cancellation of the entire slowest active eigenspace. Finite-time sign margins are therefore useful when exact cancellation is unavailable or unnecessary for the operational comparison.
Generality of the theorem does not mean uniform sharpness. Its assumptions refer to any finite size, but its constants depend on the initial likelihood-ratio ranges and spectral separation. Small stationary probabilities or a weak separation can make the bound loose. The network study displays this directly: among the 45 positive-margin cases, the ratio of the simple upper formula to the located crossing ranges from about 1.21 to 30.19, with median 2.15. Retaining more spectral information can improve a particular bound substantially, as the eight-state example demonstrates, but no dimension-independent accuracy guarantee follows.
The partial-spectrum construction is an information bound. To use it as a numerical certificate, one needs enclosures for the retained weights and rates, the total initial squared deviation, and a valid lower bound on the omitted spectrum. The demonstration here computes the full spectrum once and then evaluates bounds with fewer retained modes. For the selected eight-state model, interval arithmetic validates this construction and the endpoint t M < 1.55163 t_M<1.55163 t M < 1.55163 . This does not establish the computational complexity or speedup of a large matrix eigensolver. Certified iterative spectral methods and sharper tail estimates would be useful subsequent developments.
The diffusion example connects the finite-state argument to an explicit Brownian relaxation model. The grid generator preserves detailed balance and probability exactly at the analytical level, and the observed crossing stabilizes under refinement. The interval certificate applies to a specified finite grid. A continuum crossing theorem with a numerical error enclosure would additionally require a discretization estimate for both distributions or their entropy difference. On an unbounded domain, hotter Gibbs preparations can also have unbounded likelihood ratios, requiring tail control beyond the present finite-range comparison.
Several extensions have distinct mathematical requirements. The scalar entropy comparison remains valid for a fixed nonreversible stationary generator, and the chi-squared divergence remains contractive, but the orthogonal sum of modal squares generally does not survive. The exact nonadiabatic entropy balance still holds at a fixed stationary reference [53 ] ; it does not justify replacing reversible decay rates by arbitrary complex eigenvalues. Time-dependent generators and changing reference states also introduce additional terms. Finally, an entropy crossing alone need not imply a crossing of mean energy, temperature, or every monotone distance [46 , 55 ] . The present analysis therefore makes no statement about which water sample freezes first.
The results identify a concrete combination of established ingredients and derived applications. The spectral Mpemba mechanism, stochastic contraction, entropy dissipation, and the static divergence comparison are prior results. Their use here yields finite-amplitude crossing enclosures, a controlled omitted-spectrum contribution, quantitative preparation tolerances, and reproducible tests with explicitly separated certification levels. No claim of priority for all finite-time Mpemba criteria, or of a universal optimization law, is needed for these statements.
Conclusions
A finite-time relative-entropy crossing can be guaranteed without exact hot slow-mode cancellation. For a reversible finite-state system, a positive slow-sector margin gives an explicit upper meeting-time bound, with the complete slow eigenspace included in any degeneracy. Exact entropy envelopes retain additional modes, and a bounded omitted tail makes partial spectral information usable without silently discarding its error. Probability contraction and an entropy-gradient estimate convert sign margins into preparation tolerances.
The 2000-network study and the reflecting diffusion example show how these results operate beyond a specially canceled three-state preparation. The simple condition detects 45 of 86 numerically located network crossings in the declared sample and can be conservative. Independent interval arithmetic verifies selected spectral bounds, first crossings, a partial-spectrum improvement, and a continuous temperature rectangle in an exactly specified benchmark. Together these results supply a general finite-state framework with explicit hypotheses and model-dependent numerical quality, while separating ordinary crossings from strong asymptotic acceleration.
Data and code availability
The Supplemental Material contains all 2000 network draws, all 180 diffusion-search outcomes, the mesh-refinement and temperature-survey data, exact inputs for the certified examples, the numerical driver (run_generalization.py), the figure script, and the interval-arithmetic program (certify_generalization.py), with a README giving commands, software versions, and the scope of each certificate. Rounded values in tables and figures do not replace the exact model definitions used by the certificates.
Entropy balance, canonical ordering, and reversible modes
Thermodynamic meaning and canonical initial ordering
For π i = e − E i / T b / Z ( T b ) \pi_i=e^{-E_i/T_b}/Z(T_b) π i = e − E i / T b / Z ( T b ) ,
T b D ( p ∥ π ) = T b ∑ i p i ln p i + ∑ i p i E i + T b ln Z ( T b ) (69) T_bD(p\Vert\pi)=T_b\sum_i p_i\ln p_i+\sum_i p_iE_i+T_b\ln Z(T_b) \tag*{(69)} T b D ( p ∥ π ) = T b i ∑ p i ln p i + i ∑ p i E i + T b ln Z ( T b ) (69)
= F [ p ] − F [ π ] , (70) =F[p]-F[\pi], \tag*{(70)} = F [ p ] − F [ π ] , (70)
because F [ π ] = − T b ln Z ( T b ) F[\pi]=-T_b\ln Z(T_b) F [ π ] = − T b ln Z ( T b ) . This identifies the dimensionless entropy distance with excess free energy when the reference is Gibbsian.
For an initial Gibbs state at inverse temperature β 0 = 1 / T 0 \beta_0=1/T_0 β 0 = 1/ T 0 , write U ( β 0 ) = ∑ i E i e − β 0 E i / Z ( β 0 ) U(\beta_0)=\sum_i E_i e^{-\beta_0E_i}/Z(\beta_0) U ( β 0 ) = ∑ i E i e − β 0 E i / Z ( β 0 ) and β b = 1 / T b \beta_b=1/T_b β b = 1/ T b . Direct substitution gives
D 0 ( β 0 ) = ( β b − β 0 ) U ( β 0 ) + ln Z ( β b ) Z ( β 0 ) . (71) D_0(\beta_0)=(\beta_b-\beta_0)U(\beta_0)+\ln\frac{Z(\beta_b)}{Z(\beta_0)}. \tag*{(71)} D 0 ( β 0 ) = ( β b − β 0 ) U ( β 0 ) + ln Z ( β 0 ) Z ( β b ) . (71)
Using ∂ β 0 ln Z = − U \partial_{\beta_0}\ln Z=-U ∂ β 0 ln Z = − U and ∂ β 0 U = − Var T 0 ( E ) \partial_{\beta_0}U=-\operatorname{Var}_{T_0}(E) ∂ β 0 U = − Var T 0 ( E ) , differentiation cancels the two terms proportional to U U U and yields
∂ D 0 ∂ T 0 = β b − β 0 T 0 2 Var T 0 ( E ) > 0 ( T 0 > T b ) . (72) \frac{\partial D_0}{\partial T_0}=\frac{\beta_b-\beta_0}{T_0^2}\operatorname{Var}_{T_0}(E)>0\qquad(T_0>T_b). \tag*{(72)} ∂ T 0 ∂ D 0 = T 0 2 β b − β 0 Var T 0 ( E ) > 0 ( T 0 > T b ) . (72)
The variance is strictly positive for nonconstant energies. Thus the canonical cooling protocol starts with D H ( 0 ) > D C ( 0 ) > 0 D_H(0)>D_C(0)>0 D H ( 0 ) > D C ( 0 ) > 0 .
The entropy-production identity
Differentiate D = ∑ i p i ln ( p i / π i ) D=\sum_i p_i\ln(p_i/\pi_i) D = ∑ i p i ln ( p i / π i ) with fixed π \pi π . Normalization removes the constant derivative term, giving
D ˙ = ∑ i p ˙ i ln p i π i . (73) \dot D=\sum_i\dot p_i\ln\frac{p_i}{\pi_i}. \tag*{(73)} D ˙ = i ∑ p ˙ i ln π i p i . (73)
Since p ˙ i = − ∑ j J i j \dot p_i=-\sum_jJ_{ij} p ˙ i = − ∑ j J ij and J i j = − J j i J_{ij}=-J_{ji} J ij = − J j i ,
D ˙ = − 1 2 ∑ i , j J i j [ ln p i π i − ln p j π j ] . (74) \dot D=-\frac{1}{2}\sum_{i,j}J_{ij}\left[\ln\frac{p_i}{\pi_i}-\ln\frac{p_j}{\pi_j}\right]. \tag*{(74)} D ˙ = − 2 1 i , j ∑ J ij [ ln π i p i − ln π j p j ] . (74)
Detailed balance identifies the logarithmic difference with ln ( k i j p i / k j i p j ) \ln(k_{ij}p_i/k_{ji}p_j) ln ( k ij p i / k j i p j ) . Therefore D ˙ = − e ˙ p \dot D=-\dot e_p D ˙ = − e ˙ p . Each paired contribution to e ˙ p \dot e_p e ˙ p has the form ( a − b ) ln ( a / b ) ≥ 0 (a-b)\ln(a/b)\ge0 ( a − b ) ln ( a / b ) ≥ 0 . Irreducibility gives equality only at p = π p=\pi p = π . Integrating − d ln D / d t = e ˙ p / D -\mathrm{d}\ln D/\mathrm{d}t=\dot e_p/D − d ln D / d t = e ˙ p / D proves Eq. (11) .
Self-adjointness and orthogonal amplitudes
Put Π = diag ( π ) \Pi=\operatorname{diag}(\pi) Π = diag ( π ) . Dividing the master equation by π i \pi_i π i and using detailed balance gives
f ˙ i = ∑ j ≠ i k i j ( f j − f i ) , W = Π L Π − 1 . (75) \dot f_i=\sum_{j\ne i}k_{ij}(f_j-f_i),\qquad W=\Pi L\Pi^{-1}. \tag*{(75)} f ˙ i = j = i ∑ k ij ( f j − f i ) , W = Π L Π − 1 . (75)
For real vectors u , v u,v u , v ,
⟨ u , L v ⟩ π = ∑ i , j π i k i j u i ( v j − v i ) (76) \langle u,Lv\rangle_\pi=\sum_{i,j}\pi_i k_{ij}u_i(v_j-v_i) \tag*{(76)} ⟨ u , Lv ⟩ π = i , j ∑ π i k ij u i ( v j − v i ) (76)
= − 1 2 ∑ i , j π i k i j ( u i − u j ) ( v i − v j ) (77) =-\frac{1}{2}\sum_{i,j}\pi_i k_{ij}(u_i-u_j)(v_i-v_j) \tag*{(77)} = − 2 1 i , j ∑ π i k ij ( u i − u j ) ( v i − v j ) (77)
= ⟨ L u , v ⟩ π . (78) =\langle Lu,v\rangle_\pi. \tag*{(78)} = ⟨ Lu , v ⟩ π . (78)
The second line follows by interchanging i , j i,j i , j in one copy of the sum and averaging, with π i k i j = π j k j i \pi_i k_{ij}=\pi_j k_{ji} π i k ij = π j k j i . Thus L L L is self-adjoint and nonpositive. Its constant null eigenfunction is ϕ 1 = 1 \phi_1=1 ϕ 1 = 1 ; irreducibility means that null space is one-dimensional. The finite-dimensional spectral theorem gives the remaining orthonormal eigenfunctions in Eq. (14) . Since ⟨ f ( 0 ) , 1 ⟩ π = 0 \langle f(0),1\rangle_\pi=0 ⟨ f ( 0 ) , 1 ⟩ π = 0 , no stationary coefficient appears. Taking inner products of the mode expansion with ϕ n \phi_n ϕ n gives Eq. (15) .
Squaring that expansion and averaging makes the cross terms vanish:
Q α ( t ) = ∑ m , n ≥ 2 b m , α b n , α e − ( λ m + λ n ) t ⟨ ϕ m , ϕ n ⟩ π (79) Q_\alpha(t)=\sum_{m,n\ge2}b_{m,\alpha}b_{n,\alpha}e^{-(\lambda_m+\lambda_n)t}\langle\phi_m,\phi_n\rangle_\pi \tag*{(79)} Q α ( t ) = m , n ≥ 2 ∑ b m , α b n , α e − ( λ m + λ n ) t ⟨ ϕ m , ϕ n ⟩ π (79)
= ∑ n ≥ 2 b n , α 2 e − 2 λ n t . (80) =\sum_{n\ge2}b_{n,\alpha}^2e^{-2\lambda_nt}. \tag*{(80)} = n ≥ 2 ∑ b n , α 2 e − 2 λ n t . (80)
For numerical diagonalization, the similar matrix Π − 1 / 2 W Π 1 / 2 \Pi^{-1/2}W\Pi^{1/2} Π − 1/2 W Π 1/2 is real symmetric. An orthonormal eigenvector z n z_n z n of this matrix gives ϕ n , i = z n , i / π i \phi_{n,i}=z_{n,i}/\sqrt{\pi_i} ϕ n , i = z n , i / π i .
The quadratic limit and the asymptotic weight criterion
For ∣ x ∣ < 1 |x|<1 ∣ x ∣ < 1 , the scalar entropy function has the convergent expansion
g ( x ) = x 2 2 − x 3 6 + x 4 12 − ⋯ . (81) g(x)=\frac{x^2}{2}-\frac{x^3}{6}+\frac{x^4}{12}-\cdots. \tag*{(81)} g ( x ) = 2 x 2 − 6 x 3 + 12 x 4 − ⋯ . (81)
If η = ∥ f ∥ ∞ < 1 \eta=\lVert f\rVert_{\infty}<1 η = ∥ f ∥ ∞ < 1 , its remainder after the quadratic term is bounded componentwise by a constant times η f i 2 \eta f_i^2 η f i 2 , with a constant bounded on each fixed range η ≤ η 0 < 1 \eta\le\eta_0<1 η ≤ η 0 < 1 . Summation gives D = Q / 2 + O ( η Q ) D=Q/2+O(\eta Q) D = Q /2 + O ( η Q ) and proves Eq. (17) .
If the cold preparation has w C > 0 w_C>0 w C > 0 , then Q C ( t ) = w C e − 2 λ s t + o ( e − 2 λ s t ) Q_C(t)=w_Ce^{-2\lambda_s t}+o(e^{-2\lambda_s t}) Q C ( t ) = w C e − 2 λ s t + o ( e − 2 λ s t ) . For the hot preparation the corresponding leading coefficient is w H w_H w H , with a vanishing leading term if w H = 0 w_H=0 w H = 0 . Both relative deviations tend to zero, so
lim t → ∞ D H ( t ) D C ( t ) = w H w C . (82) \lim_{t\to\infty}\frac{D_H(t)}{D_C(t)}=\frac{w_H}{w_C}. \tag*{(82)} t → ∞ lim D C ( t ) D H ( t ) = w C w H . (82)
Thus w H < w C w_H<w_C w H < w C and the opposite initial ordering imply an eventual meeting. This asymptotic statement alone supplies no finite upper time. Since m H ( 0 ) ≤ 0 ≤ M C ( 0 ) m_H(0)\le0\le M_C(0) m H ( 0 ) ≤ 0 ≤ M C ( 0 ) and F ( 0 ) = 1 / 2 F(0)=1/2 F ( 0 ) = 1/2 , one has c H ≥ 1 / 2 c_H\ge1/2 c H ≥ 1/2 and a C ≤ 1 / 2 a_C\le1/2 a C ≤ 1/2 . Hence Γ > 0 \Gamma>0 Γ > 0 implies w H / w C < a C / c H ≤ 1 w_H/w_C<a_C/c_H\le1 w H / w C < a C / c H ≤ 1 , explaining why Theorem 1 is a stronger sufficient condition.
Finite-amplitude comparison and the crossing theorems
The scalar bound and Markov range contraction
Taylor’s integral formula, or two direct integrations of g ′ ′ ( x ) = 1 / ( 1 + x ) g''(x)=1/(1+x) g ′′ ( x ) = 1/ ( 1 + x ) , gives
g ( x ) = ∫ 0 x x − y 1 + y d y = x 2 ∫ 0 1 1 − u 1 + u x d u . (83) g(x)=\int_0^x\frac{x-y}{1+y}\,\mathrm{d}y=x^2\int_0^1\frac{1-u}{1+ux}\,\mathrm{d}u. \tag*{(83)} g ( x ) = ∫ 0 x 1 + y x − y d y = x 2 ∫ 0 1 1 + ux 1 − u d u . (83)
For x > − 1 x>-1 x > − 1 the denominator is positive for 0 ≤ u ≤ 1 0\le u\le1 0 ≤ u ≤ 1 . This proves positivity of F F F and permits differentiation under the integral to obtain Eq. (19) . Also ∑ i π i f i = 0 \sum_i\pi_i f_i=0 ∑ i π i f i = 0 , so
D ( p ∥ π ) = ∑ i π i [ ( 1 + f i ) ln ( 1 + f i ) − f i ] = ∑ i π i f i 2 F ( f i ) . (84) D(p\mathbin{\|}\pi)=\sum_i\pi_i[(1+f_i)\ln(1+f_i)-f_i]=\sum_i\pi_i f_i^2F(f_i). \tag*{(84)} D ( p ∥ π ) = i ∑ π i [( 1 + f i ) ln ( 1 + f i ) − f i ] = i ∑ π i f i 2 F ( f i ) . (84)
Let P ( u ) = e W u P(u)=e^{Wu} P ( u ) = e W u . Its entries are nonnegative and its columns sum to one. Stationarity gives P ( u ) π = π P(u)\pi=\pi P ( u ) π = π , so the ratio kernel K i j ( u ) = P i j ( u ) π j / π i K_{ij}(u)=P_{ij}(u)\pi_j/\pi_i K ij ( u ) = P ij ( u ) π j / π i has row sums one. Therefore
m α ( s ) ≤ ∑ j K i j ( t − s ) f j α ( s ) ≤ M α ( s ) , t ≥ s . (85) m_\alpha(s)\le\sum_jK_{ij}(t-s)f_j^\alpha(s)\le M_\alpha(s),\qquad t\ge s. \tag*{(85)} m α ( s ) ≤ j ∑ K ij ( t − s ) f j α ( s ) ≤ M α ( s ) , t ≥ s . (85)
Because F F F decreases, every term in the exact entropy sum lies between π i ( f i α ) 2 F ( M α ( s ) ) \pi_i(f_i^\alpha)^2F(M_\alpha(s)) π i ( f i α ) 2 F ( M α ( s )) and π i ( f i α ) 2 F ( m α ( s ) ) \pi_i(f_i^\alpha)^2F(m_\alpha(s)) π i ( f i α ) 2 F ( m α ( s )) . Summing proves Eq. (22) without a small-amplitude assumption.
A whole slow eigenspace and the explicit upper time
Let S = { n ≥ 2 : λ n = λ s } S=\{n\ge2:\lambda_n=\lambda_s\} S = { n ≥ 2 : λ n = λ s } and F = { n ≥ 2 : λ n > λ s } \mathcal{F}=\{n\ge2:\lambda_n>\lambda_s\} F = { n ≥ 2 : λ n > λ s } . Under Theorem 1, F \mathcal{F} F is nonempty and λ f \lambda_f λ f is its smallest rate. The weights are w α = ∑ n ∈ S b n , α 2 w_\alpha=\sum_{n\in S}b_{n,\alpha}^2 w α = ∑ n ∈ S b n , α 2 and q H = ∑ n ∈ F b n , H 2 q_H=\sum_{n\in\mathcal{F}}b_{n,H}^2 q H = ∑ n ∈ F b n , H 2 . An orthogonal rotation inside S S S preserves this sum of squares, so a degenerate sector is handled without selecting a special basis vector.
Since every spectral weight is nonnegative,
Q H ( t ) = w H e − 2 λ s t + ∑ n ∈ F b n , H 2 e − 2 λ n t (86) Q_H(t)=w_He^{-2\lambda_s t}+\sum_{n\in\mathcal{F}}b_{n,H}^2e^{-2\lambda_n t} \tag*{(86)} Q H ( t ) = w H e − 2 λ s t + n ∈ F ∑ b n , H 2 e − 2 λ n t (86)
≤ w H e − 2 λ s t + q H e − 2 λ f t , (87) \le w_He^{-2\lambda_s t}+q_He^{-2\lambda_f t}, \tag*{(87)} ≤ w H e − 2 λ s t + q H e − 2 λ f t , (87)
and dropping the faster cold terms gives Q C ( t ) ≥ w C e − 2 λ s t Q_C(t)\ge w_Ce^{-2\lambda_s t} Q C ( t ) ≥ w C e − 2 λ s t . Inserting these two inequalities into d ( t ) ≤ c H Q H ( t ) − a C Q C ( t ) d(t)\le c_HQ_H(t)-a_CQ_C(t) d ( t ) ≤ c H Q H ( t ) − a C Q C ( t ) proves Eq. (26) .
At zero time d ( 0 ) > 0 d(0)>0 d ( 0 ) > 0 , while that upper expression equals c H q H − Γ c_Hq_H-\Gamma c H q H − Γ . Thus c H q H > Γ > 0 c_Hq_H>\Gamma>0 c H q H > Γ > 0 ; in particular q H > 0 q_H>0 q H > 0 and w C > 0 w_C>0 w C > 0 . Solving
c H q H e − 2 ( λ f − λ s ) t = Γ (88) c_Hq_He^{-2(\lambda_f-\lambda_s)t}=\Gamma \tag*{(88)} c H q H e − 2 ( λ f − λ s ) t = Γ (88)
gives the positive time in Eq. (24) . The comparison is nonpositive at that time and strictly negative afterward. The continuous function d d d starts positive, so it has a first zero in ( 0 , T u p ] (0,T_{\mathrm{up}}] ( 0 , T up ] and a strict reversed ordering for every t > T u p t>T_{\mathrm{up}} t > T up . This proves Theorem 1 and does not exclude an earlier agency or additional earlier crossings.
If d ( s ) > 0 d(s)>0 d ( s ) > 0 , the proof can be restarted using the distributions at s s s . The weights must be those of f α ( s ) f^\alpha(s) f α ( s ) , and the exponential variable is t − s t-s t − s . A positive restarted margin gives an absolute upper time s + T u p ( s ) s+T_{\mathrm{up}}^{(s)} s + T up ( s ) . This also bounds the original first meeting even if there was an earlier meeting and recrossing. If d ( s ) ≤ 0 d(s)\le0 d ( s ) ≤ 0 , continuity already gives t M ≤ s t_M\le s t M ≤ s ; a negative restarted logarithm must not be used to move the bound into the past of its reference time.
All-mode signs and the omitted spectral tail
For fixed s s s , Eq. (22) bounds each entropy separately. Subtracting the upper cold bound from the lower hot bound gives G − ( t ; s ) ≤ d ( t ) G_-(t;s)\le d(t) G − ( t ; s ) ≤ d ( t ) . Subtracting the lower cold bound from the upper hot bound gives d ( t ) ≤ G + ( t ; s ) d(t)\le G_+(t;s) d ( t ) ≤ G + ( t ; s ) . This proves Eq. (28) . Positivity of the lower bound on an entire earlier interval excludes zeros there. Nonpositivity of the upper bound at a later endpoint guarantees that at least one zero has already occurred, since the initial difference was positive. Neither inference requires monotonicity of the bounding exponential sums.
For omitted modes,
R α ( t ) = ∑ n = m + 1 N b n , α 2 e − 2 λ n t . (89) R_\alpha(t)=\sum_{n=m+1}^{N}b_{n,\alpha}^2e^{-2\lambda_n t}. \tag*{(89)} R α ( t ) = n = m + 1 ∑ N b n , α 2 e − 2 λ n t . (89)
Every term is nonnegative and, by the rate lower bound, is no larger than b n , α 2 e − 2 λ c u t t b_{n,\alpha}^2e^{-2\lambda_{\mathrm{cut}}t} b n , α 2 e − 2 λ cut t . Summation proves Eq. (30) . If uncertain spectral inputs are enclosed as in Eq. (31) , monotonicity of e − 2 λ t e^{-2\lambda t} e − 2 λ t for t ≥ 0 t\ge0 t ≥ 0 gives
w ‾ n , α e − 2 λ ‾ n t ≤ b n , α 2 e − 2 λ n t ≤ w ‾ n , α e − 2 λ ‾ n t . (90) \underline{w}_{n,\alpha}e^{-2\overline{\lambda}_n t}\le b_{n,\alpha}^2e^{-2\lambda_n t}\le\overline{w}_{n,\alpha}e^{-2\underline{\lambda}_n t}. \tag*{(90)} w n , α e − 2 λ n t ≤ b n , α 2 e − 2 λ n t ≤ w n , α e − 2 λ n t . (90)
Adding these retained terms and the omitted upper tail proves the stated norm bounds. A fully explicit entropy-difference enclosure is therefore
d ( t ) ≥ a H ( s ) Q H − ( t ) − c C ( s ) Q C + ( t ) , d ( t ) ≤ c H ( s ) Q H + ( t ) − a C ( s ) Q C − ( t ) , t ≥ s . (91) \begin{aligned}
d(t) &\ge a_H(s)Q_H^{-}(t)-c_C(s)Q_C^{+}(t),\\
d(t) &\le c_H(s)Q_H^{+}(t)-a_C(s)Q_C^{-}(t),\qquad t\ge s. \tag*{(91)}
\end{aligned} d ( t ) d ( t ) ≥ a H ( s ) Q H − ( t ) − c C ( s ) Q C + ( t ) , ≤ c H ( s ) Q H + ( t ) − a C ( s ) Q C − ( t ) , t ≥ s . (91)
If the factors themselves are intervals, their nonnegative upper or lower endpoints must be chosen according to these inequality directions. The formulas require valid interval inputs, not merely approximate spectral estimates.
Proofs of preparation and parameter tolerance
Probability contraction and the centered entropy gradient
For the common stochastic semigroup P ( t ) = e W t P(t)=e^{Wt} P ( t ) = e W t ,
∥ p ~ ( t ) − p ( t ) ∥ 1 = ∑ i ∣ ∑ j P i j ( t ) [ p ~ j ( 0 ) − p j ( 0 ) ] ∣ (92) \lVert\widetilde{p}(t)-p(t)\rVert_1
=\sum_i\left|\sum_j P_{ij}(t)[\widetilde{p}_j(0)-p_j(0)]\right| \tag*{(92)} ∥ p ( t ) − p ( t ) ∥ 1 = i ∑ j ∑ P ij ( t ) [ p j ( 0 ) − p j ( 0 )] (92)
≤ ∑ i , j P i j ( t ) ∣ p ~ j ( 0 ) − p j ( 0 ) ∣ (93) \le\sum_{i,j}P_{ij}(t)|\widetilde{p}_j(0)-p_j(0)| \tag*{(93)} ≤ i , j ∑ P ij ( t ) ∣ p j ( 0 ) − p j ( 0 ) ∣ (93)
= ∥ p ~ ( 0 ) − p ( 0 ) ∥ 1 . (94) =\lVert\widetilde{p}(0)-p(0)\rVert_1. \tag*{(94)} = ∥ p ( 0 ) − p ( 0 ) ∥ 1 . (94)
The last equality uses the unit column sums. The ratio kernel separately preserves the common initial likelihood-ratio range for both trajectories.
At a fixed time put δ p = p ~ − p \delta p=\widetilde{p}-p δ p = p − p and p u = p + u δ p p_u=p+u\delta p p u = p + u δ p , 0 ≤ u ≤ 1 0\le u\le1 0 ≤ u ≤ 1 . Every point on this segment is a normalized probability distribution with ratios in [ h − , h + ] [h_-,h_+] [ h − , h + ] . Differentiating the entropy along the segment and integrating gives
D ( p ~ ∥ π ) − D ( p ∥ π ) = ∫ 0 1 ∑ i δ p i [ ln p u , i π i + 1 ] d u . (95) D(\widetilde{p}\mathbin\Vert\pi)-D(p\mathbin\Vert\pi)
=\int_0^1\sum_i\delta p_i\left[\ln\frac{p_{u,i}}{\pi_i}+1\right]\,du. \tag*{(95)} D ( p ∥ π ) − D ( p ∥ π ) = ∫ 0 1 i ∑ δ p i [ ln π i p u , i + 1 ] d u . (95)
Since ∑ i δ p i = 0 \sum_i\delta p_i=0 ∑ i δ p i = 0 , subtract the constant 1 + 1 2 ( ln h − + ln h + ) 1+\frac{1}{2}(\ln h_-+\ln h_+) 1 + 2 1 ( ln h − + ln h + ) inside the brackets without changing the integral. The remaining factor has magnitude at most C = 1 2 ln ( h + / h − ) C=\frac{1}{2}\ln(h_+/h_-) C = 2 1 ln ( h + / h − ) . Therefore
∣ D ( p ~ ∥ π ) − D ( p ∥ π ) ∣ ≤ C ∥ p ~ − p ∥ 1 . (96) \left|D(\widetilde{p}\mathbin\Vert\pi)-D(p\mathbin\Vert\pi)\right|
\le C\lVert\widetilde{p}-p\rVert_1. \tag*{(96)} ∣ D ( p ∥ π ) − D ( p ∥ π ) ∣ ≤ C ∥ p − p ∥ 1 . (96)
Combining this with probability contraction proves each entropy-error bound in Eq. (33) ; the triangle inequality gives the bound on d ~ − d \widetilde{d}-d d − d .
Under Theorem 2’s error budget, for every 0 ≤ t ≤ L 0\le t\le L 0 ≤ t ≤ L one has d ~ ( t ) ≥ η L − δ D > 0 \widetilde{d}(t)\ge\eta_L-\delta_D>0 d ( t ) ≥ η L − δ D > 0 , while d ~ ( U ) ≤ − η U + δ D < 0 \widetilde{d}(U)\le-\eta_U+\delta_D<0 d ( U ) ≤ − η U + δ D < 0 . Hence no earlier zero exists and continuity places a first zero in ( L , U ) (L,U) ( L , U ) . This proves Theorem 2.
To justify the component error used in Eq. (36) , decompose a normalized difference into positive and negative parts. Their total masses are equal because ∑ i δ p i = 0 \sum_i\delta p_i=0 ∑ i δ p i = 0 , and their sum is ∥ δ p ∥ 1 ≤ ϵ \lVert\delta p\rVert_1\le\epsilon ∥ δ p ∥ 1 ≤ ϵ . Each mass is therefore at most ϵ / 2 \epsilon/2 ϵ /2 , implying ∣ δ p i ∣ ≤ ϵ / 2 |\delta p_i|\le\epsilon/2 ∣ δ p i ∣ ≤ ϵ /2 for every state. The stated ratio endpoints contain every allowed preparation if their lower endpoint is positive.
Gibbs temperature sensitivity
For p i ( T ) = e − E i / T / Z ( T ) p_i(T)=e^{-E_i/T}/Z(T) p i ( T ) = e − E i / T / Z ( T ) ,
∂ ln p i ∂ T = E i T 2 − ∂ ln Z ∂ T = E i − U ( T ) T 2 . (97) \frac{\partial\ln p_i}{\partial T}
=\frac{E_i}{T^2}-\frac{\partial\ln Z}{\partial T}
=\frac{E_i-U(T)}{T^2}. \tag*{(97)} ∂ T ∂ ln p i = T 2 E i − ∂ T ∂ ln Z = T 2 E i − U ( T ) . (97)
Multiplication by p i ( T ) p_i(T) p i ( T ) proves Eq. (37) and summation of absolute values gives
∥ ∂ p ( T ) ∂ T ∥ 1 = ⟨ ∣ E − U ∣ ⟩ T T 2 ≤ Var T ( E ) T 2 . (98) \left\lVert\frac{\partial p(T)}{\partial T}\right\rVert_1
=\frac{\langle|E-U|\rangle_T}{T^2}
\le\frac{\sqrt{\operatorname{Var}_T(E)}}{T^2}. \tag*{(98)} ∂ T ∂ p ( T ) 1 = T 2 ⟨ ∣ E − U ∣ ⟩ T ≤ T 2 Var T ( E ) . (98)
The last step is Cauchy–Schwarz. To bound the variance, let a = E min a=E_{\min} a = E m i n , b = E max b=E_{\max} b = E m a x and use ( E − a ) ( b − E ) ≥ 0 (E-a)(b-E)\ge0 ( E − a ) ( b − E ) ≥ 0 . Taking expectations gives ⟨ E 2 ⟩ ≤ ( a + b ) U − a b \langle E^2\rangle\le(a+b)U-ab ⟨ E 2 ⟩ ≤ ( a + b ) U − ab , hence
Var ( E ) ≤ ( U − a ) ( b − U ) ≤ ( b − a ) 2 4 . (99) \operatorname{Var}(E)\le(U-a)(b-U)\le\frac{(b-a)^2}{4}. \tag*{(99)} Var ( E ) ≤ ( U − a ) ( b − U ) ≤ 4 ( b − a ) 2 . (99)
Thus the norm of the temperature derivative is at most ( E max − E min ) / ( 2 T 2 ) (E_{\max}-E_{\min})/(2T^2) ( E m a x − E m i n ) / ( 2 T 2 ) . Integrating along the interval between T T T and T ~ \widetilde{T} T , whose temperatures are all at least T min > 0 T_{\min}>0 T m i n > 0 , gives Eq. (38) .
A change of rates and stationary reference
Let Δ W = W ~ − W \Delta W=\widetilde{W}-W Δ W = W − W . Differentiating e W ~ ( t − s ) e W s e^{\widetilde{W}(t-s)}e^{Ws} e W ( t − s ) e W s with respect to s s s and integrating yields
e W ~ t − e W t = ∫ 0 t e W ~ ( t − s ) Δ W e W s d s . (100) e^{\widetilde{W}t}-e^{Wt}
=\int_0^t e^{\widetilde{W}(t-s)}\Delta W e^{Ws}\,ds. \tag*{(100)} e W t − e W t = ∫ 0 t e W ( t − s ) Δ W e W s d s . (100)
Each stochastic semigroup has induced 1-norm equal to one. Splitting the probability difference into its initial-state and generator parts,
p ~ ( t ) − p ( t ) = e W ~ t [ p ~ ( 0 ) − p ( 0 ) ] (101) \widetilde{p}(t)-p(t)
=e^{\widetilde{W}t}[\widetilde{p}(0)-p(0)] \tag*{(101)} p ( t ) − p ( t ) = e W t [ p ( 0 ) − p ( 0 )] (101)
+ ( e W ~ t − e W t ) p ( 0 ) , (102) +\left(e^{\widetilde{W}t}-e^{Wt}\right)p(0), \tag*{(102)} + ( e W t − e W t ) p ( 0 ) , (102)
and using ∥ p ( 0 ) ∥ 1 = 1 \lVert p(0)\rVert_1=1 ∥ p ( 0 ) ∥ 1 = 1 proves Eq. (39) .
The centered-gradient bound can now compare p ~ ( t ) \widetilde{p}(t) p ( t ) and p ( t ) p(t) p ( t ) relative to the same nominal π \pi π . Its ratio interval must be valid for 0 ≤ t ≤ U 0\le t\le U 0 ≤ t ≤ U , since the perturbed generator need not preserve a range relative to the old reference. A conservative construction is available from Eq. (39) . Put e α = ϵ α + U δ W e_\alpha=\epsilon_\alpha+U\delta W e α = ϵ α + U δ W and π min = min i π i \pi_{\min}=\min_i\pi_i π m i n = min i π i . The nominal range is preserved, and every component difference is at most e α / 2 e_\alpha/2 e α /2 . Thus
h α , − = 1 + m α ( 0 ) − e α 2 π min , (103) h_{\alpha,-}=1+m_\alpha(0)-\frac{e_\alpha}{2\pi_{\min}}, \tag*{(103)} h α , − = 1 + m α ( 0 ) − 2 π m i n e α , (103)
h α , + = 1 + M α ( 0 ) + e α 2 π min , h_{\alpha,+}=1+M_\alpha(0)+\frac{e_\alpha}{2\pi_{\min}}, h α , + = 1 + M α ( 0 ) + 2 π m i n e α ,
encloses both distributions’ ratios to π \pi π whenever h α , − > 0 h_{\alpha,-}>0 h α , − > 0 . More detailed component bounds can improve it.
The stationary-reference correction is exact:
D ( p ~ ∥ π ~ ) − D ( p ~ ∥ π ) = ∑ i p ~ i ln π i π ~ i . (104) D(\tilde p\Vert\tilde\pi)-D(\tilde p\Vert\pi)=\sum_i \tilde p_i\ln\frac{\pi_i}{\tilde\pi_i}. \tag*{(104)} D ( p ~ ∥ π ~ ) − D ( p ~ ∥ π ) = i ∑ p ~ i ln π ~ i π i . (104)
Its absolute value is at most δ π = max i ∣ ln ( π i / π ~ i ) ∣ \delta_\pi=\max_i|\ln(\pi_i/\tilde\pi_i)| δ π = max i ∣ ln ( π i / π ~ i ) ∣ . Combining this with the distribution error for each preparation and adding the two bounds proves Eq. (40). The perturbed stationary distribution must itself be computed or bounded for the specified rate family. The resulting statement is finite-time: the linear-in-U U U Duhamel estimate need not remain useful on arbitrarily long intervals.
Differentiating a transverse meeting
For a simple root d ( t ∗ ( θ ) , θ ) = 0 d(t_*(\theta),\theta)=0 d ( t ∗ ( θ ) , θ ) = 0 , the chain rule gives
0 = ∂ t d ∂ θ t ∗ + ∂ θ d . (105) 0=\partial_t d\,\partial_\theta t_*+\partial_\theta d. \tag*{(105)} 0 = ∂ t d ∂ θ t ∗ + ∂ θ d . (105)
At the root the two positive entropies equal D ∗ D_* D ∗ , and the entropy balance gives
∂ t d = − e ˙ p H + e ˙ p C = − D ∗ ( r H − r C ) . (106) \partial_t d=-\dot e_p^{H}+\dot e_p^{C}=-D_*(r_H-r_C). \tag*{(106)} ∂ t d = − e ˙ p H + e ˙ p C = − D ∗ ( r H − r C ) . (106)
Substitution proves Eq. (42) . The nonzero derivative condition is essential to the local implicit-function argument. It establishes neither global uniqueness nor absence of earlier roots; those questions require separate sign information.
A negative control and the scope beyond detailed balance
Two-state canonical cooling cannot reverse entropy order
Let E 1 > E 0 E_1>E_0 E 1 > E 0 and denote the excited-state probability by u ( T ) u(T) u ( T ) . For a Gibbs preparation,
u ( T ) = 1 1 + e ( E 1 − E 0 ) / T . (107) u(T)=\frac{1}{1+e^{(E_1-E_0)/T}}. \tag*{(107)} u ( T ) = 1 + e ( E 1 − E 0 ) / T 1 . (107)
which is strictly increasing in T T T . Write u b = u ( T b ) u_b=u(T_b) u b = u ( T b ) . The two-state master equation has one nonzero relaxation rate λ 2 = k 01 + k 10 \lambda_2=k_{01}+k_{10} λ 2 = k 01 + k 10 and exact solution
u α ( t ) = u b + [ u ( T α ) − u b ] e − λ 2 t . (108) u_\alpha(t)=u_b+[u(T_\alpha)-u_b]e^{-\lambda_2t}. \tag*{(108)} u α ( t ) = u b + [ u ( T α ) − u b ] e − λ 2 t . (108)
Thus T H > T C > T b T_H>T_C>T_b T H > T C > T b implies u H ( t ) > u C ( t ) > u b u_H(t)>u_C(t)>u_b u H ( t ) > u C ( t ) > u b for every finite time. Binary relative entropy is
D ( u ∥ u b ) = u ln u u b + ( 1 − u ) ln 1 − u 1 − u b , (109) D(u\Vert u_b)=u\ln\frac{u}{u_b}+(1-u)\ln\frac{1-u}{1-u_b}, \tag*{(109)} D ( u ∥ u b ) = u ln u b u + ( 1 − u ) ln 1 − u b 1 − u , (109)
and its derivative is
d D d u = ln u ( 1 − u b ) u b ( 1 − u ) > 0 ( u > u b ) . (110) \frac{\mathrm dD}{\mathrm du}=\ln\frac{u(1-u_b)}{u_b(1-u)}>0\qquad(u>u_b). \tag*{(110)} d u d D = ln u b ( 1 − u ) u ( 1 − u b ) > 0 ( u > u b ) . (110)
Consequently D H ( t ) > D C ( t ) D_H(t)>D_C(t) D H ( t ) > D C ( t ) at every finite time. This is a negative control for the stated canonical cooling protocol, not a claim about arbitrary binary preparations on opposite sides of equilibrium.
What survives without reversibility
For any fixed irreducible Markov generator with positive stationary distribution π \pi π , the ratio kernel in Eq. (21) still has nonnegative entries and unit row sums. Therefore the scalar entropy envelopes and likelihood-ratio range contraction remain valid. The exact squared deviation can be written as
Q α ( t ) = ∥ Π − 1 / 2 e W t [ p α ( 0 ) − π ] ∥ 2 2 . (111) Q_\alpha(t)=\left\|\Pi^{-1/2}e^{Wt}[p^\alpha(0)-\pi]\right\|_2^2. \tag*{(111)} Q α ( t ) = Π − 1/2 e W t [ p α ( 0 ) − π ] 2 2 . (111)
Without detailed balance it cannot generally be replaced by a sum of independent squared eigenmode amplitudes, because the modes can be nonorthogonal or complex. Nevertheless Q Q Q is nonincreasing. Probability conservation gives ∑ i π i K i j = π j \sum_i\pi_iK_{ij}=\pi_j ∑ i π i K ij = π j , and Jensen’s inequality yields
∑ i π i ( ∑ j K i j f j ) 2 ≤ ∑ i , j π i K i j f j 2 (112) \sum_i\pi_i\left(\sum_jK_{ij}f_j\right)^2\leq\sum_{i,j}\pi_iK_{ij}f_j^2 \tag*{(112)} i ∑ π i ( j ∑ K ij f j ) 2 ≤ i , j ∑ π i K ij f j 2 (112)
= ∑ j π j f j 2 . (113) =\sum_j\pi_jf_j^2. \tag*{(113)} = j ∑ π j f j 2 . (113)
This proves chi-squared contraction without reversibility. Nonorthogonal spectral representations do not imply transient amplification of this stationary-weighted divergence.
For completeness, set h i = p i / π i h_i=p_i/\pi_i h i = p i / π i . Differentiating relative entropy with a fixed stationary reference gives
− D ˙ = ∑ i , j π i k i j h i ln h i h j . (114) -\dot D=\sum_{i,j}\pi_i k_{ij}h_i\ln\frac{h_i}{h_j}. \tag*{(114)} − D ˙ = i , j ∑ π i k ij h i ln h j h i . (114)
Stationarity cancels ∑ i , j π i k i j ( h i − h j ) \sum_{i,j}\pi_i k_{ij}(h_i-h_j) ∑ i , j π i k ij ( h i − h j ) , so
− D ˙ = ∑ i , j π i k i j [ h i ln h i h j − h i + h j ] ≥ 0. (115) -\dot D=\sum_{i,j}\pi_i k_{ij}\left[h_i\ln\frac{h_i}{h_j}-h_i+h_j\right]\geq0. \tag*{(115)} − D ˙ = i , j ∑ π i k ij [ h i ln h j h i − h i + h j ] ≥ 0. (115)
This is the nonadiabatic entropy-production rate. With bidirectional allowed transitions, finite total entropy production can also contain a housekeeping contribution [53 ] . A time-dependent reference adds − ∑ i p ˙ i ∂ t ln π i -\sum_i\dot p_i\partial_t\ln\pi_i − ∑ i p ˙ i ∂ t ln π i to D ˙ \dot D D ˙ , so the fixed-reference clock must not be applied unchanged to that situation.
Finally, every finite detailed-balance discretization of a diffusion model falls within the finite-state results. A continuum theorem would also need the required operator and likelihood-ratio bounds in the continuum setting. On an unbounded domain a hotter Gibbs distribution can have an unbounded ratio to the final Gibbs density, making the proposed global lower entropy factor trivial. Mesh refinement supplies numerical evidence of convergence; it does not by itself certify a continuum crossing interval.
Reproducible numerical methods and consistency checks
Generator assembly and random streams
The generator acts on column probability vectors. From jump rates k i j k_{ij} k ij for i → j i \to j i → j , the off-diagonal entry is W j i = k i j W_{ji}=k_{ij} W j i = k ij and the diagonal entry is W i i = − ∑ j ≠ i k i j W_{ii}=-\sum_{j\ne i}k_{ij} W ii = − ∑ j = i k ij . Thus each column sums to zero. For a reversible model, diagonalize the symmetric positive semidefinite matrix
S = − Π − 1 / 2 W Π 1 / 2 , Π = diag ( π 1 , … , π N ) . (116) S=-\Pi^{-1/2}W\Pi^{1/2}, \qquad\Pi=\operatorname{diag}(\pi_1,\ldots,\pi_N). \tag*{(116)} S = − Π − 1/2 W Π 1/2 , Π = diag ( π 1 , … , π N ) . (116)
If u n u_n u n is a Euclidean-normalized eigenvector of S S S , then ϕ n = Π − 1 / 2 u n \phi_n=\Pi^{-1/2}u_n ϕ n = Π − 1/2 u n is normalized in the weighted inner product used in the theory. The nonstationary coefficient can be computed without subtracting large entropy values:
b n , α = ∑ i ϕ n , i [ p i ( α ) ( 0 ) − π i ] . (117) b_{n,\alpha}=\sum_i\phi_{n,i}\left[p_i^{(\alpha)}(0)-\pi_i\right]. \tag*{(117)} b n , α = i ∑ ϕ n , i [ p i ( α ) ( 0 ) − π i ] . (117)
The stationary eigenvector is omitted analytically. For the rates in Eq. (46) , the off-diagonal entries of S S S are simply − γ i j -\gamma_{ij} − γ ij , avoiding a numerical symmetrization mismatch. Random continuous prefactors give nondegenerate computed slow rates in the reported samples; the theorem itself groups a degenerate slow eigenspace.
The network iteration order is complete, sparse connected, ring, birth–death; within each family the sizes are 4, 8, 16, 32, 64, and within each size the draw index runs from 0 to 99. Every case uses a fresh NumPy default_rng initialized by
SeedSequence ( 20261024 , spawn_key = ( f , N , d ) ) , (118) \operatorname{SeedSequence}(20261024,\newline
\operatorname{spawn\_key}=(f,N,d)), \tag*{(118)} SeedSequence ( 20261024 , spawn_key = ( f , N , d )) , (118)
where f f f is the zero-based family index and d d d the zero-based draw index. The energy permutation is drawn first. For a sparse graph, edge Bernoulli variables are drawn next over upper-triangular pairs in NumPy order, and the ring backbone is then imposed by a logical union. Finally the log-prefactors are drawn on allowed pairs in the same order. Specifying this order makes every realization independently reproducible. The energy permutation and spawn key are retained for every draw.
The selected network’s interval calculation uses the rational energy grid in Eq. (51) and the supplied decimal prefactors as exact inputs. Ordinary numerical calculations use their floating-point representations. The rigorous input file, rather than an assumed exact meaning of a pseudorandom exponential, defines the reference model for that certificate.
Stable entropy and crossing search
For nonstationary modal propagation,
f i α ( t ) = ∑ n = 2 N b n , α ϕ n , i e − λ n t , (119) f_i^\alpha(t)=\sum_{n=2}^{N}b_{n,\alpha}\phi_{n,i}e^{-\lambda_n t}, \tag*{(119)} f i α ( t ) = n = 2 ∑ N b n , α ϕ n , i e − λ n t , (119)
we evaluate
D α ( t ) = ∑ i π i h ( f i α ( t ) ) , h ( x ) = ( 1 + x ) ln ( 1 + x ) − x . (120) D_\alpha(t)=\sum_i\pi_i h(f_i^\alpha(t)), \qquad h(x)=(1+x)\ln(1+x)-x. \tag*{(120)} D α ( t ) = i ∑ π i h ( f i α ( t )) , h ( x ) = ( 1 + x ) ln ( 1 + x ) − x . (120)
The subtracted linear term sums to zero exactly in the mathematical model. For ∣ x ∣ < 10 − 3 |x|<10^{-3} ∣ x ∣ < 1 0 − 3 , direct evaluation can lose relative accuracy, so the implementation instead sums
h ( x ) = ∑ n = 2 18 ( − 1 ) n x n n ( n − 1 ) + R 19 ( x ) . (121) h(x)=\sum_{n=2}^{18}\frac{(-1)^n x^n}{n(n-1)}+\mathcal{R}_{19}(x). \tag*{(121)} h ( x ) = n = 2 ∑ 18 n ( n − 1 ) ( − 1 ) n x n + R 19 ( x ) . (121)
The absolute remainder obeys
∣ R 19 ( x ) ∣ ≤ ∣ x ∣ 19 19 ⋅ 18 ( 1 − ∣ x ∣ ) , ∣ x ∣ < 1. (122) |\mathcal{R}_{19}(x)|\le\frac{|x|^{19}}{19\cdot18(1-|x|)}, \qquad|x|<1. \tag*{(122)} ∣ R 19 ( x ) ∣ ≤ 19 ⋅ 18 ( 1 − ∣ x ∣ ) ∣ x ∣ 19 , ∣ x ∣ < 1. (122)
The value of F ( x ) = h ( x ) / x 2 F(x)=h(x)/x^2 F ( x ) = h ( x ) / x 2 is evaluated using the corresponding series near zero, with F ( 0 ) = 1 / 2 F(0)=1/2 F ( 0 ) = 1/2 . All initial likelihood ratios are strictly positive. No small-deviation approximation replaces the entropy curves.
For each numerical model the time search uses τ = λ s t = 0 \tau=\lambda_s t=0 τ = λ s t = 0 and 500 logarithmically spaced points from 10 − 6 10^{-6} 1 0 − 6 to 64. When available, the slow-sector upper endpoint is included if it is at most τ = 250 \tau=250 τ = 250 . In the reported network study all added endpoints are below 64. To locate an interchange, the code searches for adjacent points where ln ( D H / D C ) \ln(D_H/D_C) ln ( D H / D C ) changes from positive to nonpositive, and refines that bracket with Brent’s method. The absolute time tolerance is 10 − 13 10^{-13} 1 0 − 13 and the relative tolerance 2 × 10 − 14 2\times10^{-14} 2 × 1 0 − 14 ; these are solver settings, not rigorous root-error bounds. The finite time grid cannot exclude an earlier narrow crossing and recrossing. Accordingly, reported roots are called located crossings unless an additional interval proof establishes their first-crossing property.
The smallest absolute computed Γ \Gamma Γ across the network draws is 8.76 × 10 − 9 8.76\times10^{-9} 8.76 × 1 0 − 9 , and the smallest absolute computed difference ∣ w H − w C ∣ |w_H-w_C| ∣ w H − w C ∣ is 8.97 × 10 − 8 8.97\times10^{-8} 8.97 × 1 0 − 8 . Every nonstationary rate is positive at working precision; the smallest is 2.5039 × 10 − 5 2.5039\times10^{-5} 2.5039 × 1 0 − 5 . Every sampled slow rate is simple at working precision. More specifically,
min λ f − λ s λ f ≃ 0.0075115 , (123) \min\frac{\lambda_f-\lambda_s}{\lambda_f}\simeq0.0075115, \tag*{(123)} min λ f λ f − λ s ≃ 0.0075115 , (123)
min λ f − λ s λ max ≃ 3.0244 × 10 − 7 . (124) \min\frac{\lambda_f-\lambda_s}{\lambda_{\max}}\simeq3.0244\times10^{-7}. \tag*{(124)} min λ m a x λ f − λ s ≃ 3.0244 × 1 0 − 7 . (124)
The driver rejects a slow gap smaller than 10 − 10 λ max 10^{-10}\lambda_{\max} 1 0 − 10 λ m a x , because it uses one slow eigenvector; a degenerate model must instead group its whole slow eigenspace as required by the theorem. The largest absolute computed stationary eigenvalue is 9.27 × 10 − 14 9.27\times10^{-14} 9.27 × 1 0 − 14 . This quantity is a stationary-eigenvalue diagnostic, not a residual norm for the entire eigensystem. These checks describe the numerical calculation; they do not substitute for rigorous eigenvector or sign-error enclosures for all 2000 draws.
Diffusion discretization and its flux limit
Let P i ( t ) P_i(t) P i ( t ) denote the probability in cell i i i and write P i = ρ i Δ x P_i=\rho_i\Delta x P i = ρ i Δ x . Define the rightward probability flux at the interface by
J i + 1 / 2 = k i , i + 1 P i − k i + 1 , i P i + 1 . (125) J_{i+1/2}=k_{i,i+1}P_i-k_{i+1,i}P_{i+1}. \tag*{(125)} J i + 1/2 = k i , i + 1 P i − k i + 1 , i P i + 1 . (125)
The master equation is then
ρ ˙ i = − J i + 1 / 2 − J i − 1 / 2 Δ x . (126) \dot{\rho}_i=-\frac{J_{i+1/2}-J_{i-1/2}}{\Delta x}. \tag*{(126)} ρ ˙ i = − Δ x J i + 1/2 − J i − 1/2 . (126)
Using Eq. (62) , the interface flux becomes
J i + 1 / 2 = e − Δ U i / 2 ρ i − e Δ U i / 2 ρ i + 1 Δ x , Δ U i = U i + 1 − U i . (127) J_{i+1/2}=\frac{e^{-\Delta U_i/2}\rho_i-e^{\Delta U_i/2}\rho_{i+1}}{\Delta x},\qquad\Delta U_i=U_{i+1}-U_i. \tag*{(127)} J i + 1/2 = Δ x e − Δ U i /2 ρ i − e Δ U i /2 ρ i + 1 , Δ U i = U i + 1 − U i . (127)
Expand about the midpoint x i + 1 / 2 x_{i+1/2} x i + 1/2 . For smooth U U U and ρ \rho ρ ,
Δ U i = Δ x U ′ ( x i + 1 / 2 ) + O ( Δ x 3 ) , (128) \Delta U_i=\Delta x\,U'(x_{i+1/2})+O(\Delta x^3), \tag*{(128)} Δ U i = Δ x U ′ ( x i + 1/2 ) + O ( Δ x 3 ) , (128)
ρ i = ρ − 1 2 Δ x ρ ′ + O ( Δ x 2 ) , (129) \rho_i=\rho-\frac{1}{2}\Delta x\,\rho'+O(\Delta x^2), \tag*{(129)} ρ i = ρ − 2 1 Δ x ρ ′ + O ( Δ x 2 ) , (129)
ρ i + 1 = ρ + 1 2 Δ x ρ ′ + O ( Δ x 2 ) . (130) \rho_{i+1}=\rho+\frac{1}{2}\Delta x\,\rho'+O(\Delta x^2). \tag*{(130)} ρ i + 1 = ρ + 2 1 Δ x ρ ′ + O ( Δ x 2 ) . (130)
The even terms cancel between the two products in Eq. (127) . Thus
J i + 1 / 2 = − ρ ′ − U ′ ρ + O ( Δ x 2 ) . (131) J_{i+1/2}=-\rho'-U'\rho+O(\Delta x^2). \tag*{(131)} J i + 1/2 = − ρ ′ − U ′ ρ + O ( Δ x 2 ) . (131)
Taking the discrete divergence gives the stated Fokker–Planck equation in the smooth-grid limit. At the two exterior interfaces, the flux is set to zero by omitting external jumps. Also, when P i ∝ e − U i P_i\propto e^{-U_i} P i ∝ e − U i , every interface flux vanishes exactly; equilibrium is preserved by the discrete model.
The discretized stationary and initial Gibbs distributions are normalized on equal-width cells, so the factor Δ x \Delta x Δ x cancels. The symmetric matrix is tridiagonal, with off-diagonal entries − ( Δ x ) − 2 -(\Delta x)^{-2} − ( Δ x ) − 2 and diagonal entries equal to the total escape rate. A symmetric tridiagonal eigensolver is used for the mesh sequence N = 32 , 64 , 128 , 256 , 512 , 1024 N=32,64,128,256,512,1024 N = 32 , 64 , 128 , 256 , 512 , 1024 . The parameter search uses N = 128 N=128 N = 128 throughout, before the selected potential is refined. The 81-point temperature survey also uses N = 128 N=128 N = 128 throughout. Search results, selection order, and all mesh values are stored, including failures.
The formal interior flux consistency and the measured mesh stabilization do not alone prove a uniform error bound for the continuum first crossing. Such a proof would additionally control the evolution and entropy errors, boundary treatment, and the transversality of the continuum crossing. The present interval claims are therefore made for specified finite grids only.
Independent propagation and reproducibility
For the selected eight-state network, modal and dense Padé matrix-exponential propagation agree on entropy-difference signs at t = 1.26 , 1.28 , 1.56 , 2.54 t=1.26,1.28,1.56,2.54 t = 1.26 , 1.28 , 1.56 , 2.54 ; the largest absolute difference is 4.42 × 10 − 18 4.42\times10^{-18} 4.42 × 1 0 − 18 . For the N = 64 N=64 N = 64 diffusion grid, the corresponding checks at t = 3.08 , 3.11 , 4.03 t=3.08,3.11,4.03 t = 3.08 , 3.11 , 4.03 differ by at most 5.88 × 10 − 16 5.88\times10^{-16} 5.88 × 1 0 − 16 . These checks use two different floating-point propagation methods. Appendix F gives the distinct interval arithmetic calculation, its exact inputs, and the assertions it proves.
The supplied scripts use Python 3.12, NumPy 2.3.5, SciPy 1.17.0, and Matplotlib 3.10.8. The interval program separately requires python-flint 0.9.0. The main numerical driver regenerates all draw-level tables, search results, mesh results, selected-model checks, and the two-state negative control. The plotting driver reads these saved results and produces the two vector figures. No unsuccessful realization is deleted from the supplied data.
Interval certification of the model statements
This appendix specifies the exact models, the interval algorithms, and the distinction between a verified first-crossing enclosure and a floating-point estimate. All the calculations described here are self-contained in certify-generalization.py; they do not require data or code from an earlier manuscript. Table IV summarizes the certified statements. No global uniqueness assertion is made for these examples.
Table IV. Scope of the interval certificates. Temperatures are in units with T b = 1 T_b = 1 T b = 1 . An interval for t M t_M t M refers to the first meeting of relative entropies, not a global uniqueness claim. The diffusion row concerns the specified finite discretization.
Arithmetic and acceptance rules
The certificate uses Python 3.12.14, python-flint 0.9.0, and FLINT 3.6.0. Arb represents a real quantity by a midpoint and an outward error radius. The main run uses 384-bit midpoint precision; a separate 320-bit run passes the same proof obligations. Increasing precision is a consistency check. The enclosure property of the arithmetic, together with the analytic bounds below, is the certification argument. The documented scalar and matrix interfaces provide interval elementary functions and matrix exponentials.2
Every physical input is supplied as an integer, exact fraction, or finite decimal string. Decimal strings are first converted to exact rational numbers. No binary floating-point number is used as a model parameter in the certificate. Functions such as the potential and the jump rates are then evaluated in ball arithmetic. A sign condition is accepted only if the complete resulting interval is strictly on the required side of zero. A check whose sign is unresolved stops the calculation. No small eigenvalue, coefficient, probability, or imaginary part is replaced by zero to obtain a successful test.
For an interval X = [ x − , x + ] X = [x^-, x^+] X = [ x − , x + ] , the scalar function
g ( x ) = ( 1 + x ) ln ( 1 + x ) − x (132) g(x) = (1 + x)\ln(1 + x) - x \tag*{(132)} g ( x ) = ( 1 + x ) ln ( 1 + x ) − x (132)
decreases on ( − 1 , 0 ) (-1, 0) ( − 1 , 0 ) and increases on ( 0 , ∞ ) (0, \infty) ( 0 , ∞ ) , because g ′ ( x ) = ln ( 1 + x ) g'(x) = \ln(1+x) g ′ ( x ) = ln ( 1 + x ) . Its interval range is therefore enclosed by endpoint evaluation, including its exact minimum g ( 0 ) = 0 g(0) = 0 g ( 0 ) = 0 whenever 0 ∈ X 0 \in X 0 ∈ X . Similarly, F ( x ) = g ( x ) / x 2 F(x) = g(x)/x^2 F ( x ) = g ( x ) / x 2 , with F ( 0 ) = 1 / 2 F(0) = 1/2 F ( 0 ) = 1/2 , is decreasing. Its enclosure is obtained by evaluating the endpoints in reverse order. These monotonicity rules reduce cancellation and interval overestimation; they do not truncate a series or discard a remainder.
If intervals X i = [ x i − , x i + ] X_i = [x_i^-, x_i^+] X i = [ x i − , x i + ] enclose several components, then
min i x i ∈ [ min i x i − , min i x i + ] (133) \min_i x_i \in[\min_i x_i^-, \min_i x_i^+] \tag*{(133)} i min x i ∈ [ i min x i − , i min x i + ] (133)
max i x i ∈ [ max i x i − , max i x i + ] . (134) \max_i x_i \in[\max_i x_i^-, \max_i x_i^+]. \tag*{(134)} i max x i ∈ [ i max x i − , i max x i + ] . (134)
The code uses these enclosure formulas, so it does not assume that the index of an extremizing component is constant over a temperature cell. For output, Arb’s integer midpoint–radius decimal representation is converted to explicit outward decimal endpoints. All shorter numerical inequalities printed below are widened from those endpoints.
Exact three-state model and its two-dimensional spectrum
The three states are labeled 0 , 1 , 2 0,1,2 0 , 1 , 2 , with exact energies and symmetric kinetic prefactors
E = ( 0 , 2 / 5 , 3 ) , ( γ 01 , γ 02 , γ 12 ) = ( 1 , 23 / 5 , 3 / 25 ) . (135) E = (0, 2/5, 3),
\qquad
(\gamma_{01}, \gamma_{02}, \gamma_{12}) = (1, 23/5, 3/25). \tag*{(135)} E = ( 0 , 2/5 , 3 ) , ( γ 01 , γ 02 , γ 12 ) = ( 1 , 23/5 , 3/25 ) . (135)
At the fixed final bath temperature T b = 1 T_b = 1 T b = 1 , define
k i j = γ i j e ( E i − E j ) / 2 ( i ≠ j ) , π i = e − E i ∑ j e − E j . (136) k_{ij} = \gamma_{ij}e^{(E_i-E_j)/2}
\quad(i \ne j),
\qquad
\pi_i = \frac{e^{-E_i}}{\sum_j e^{-E_j}}. \tag*{(136)} k ij = γ ij e ( E i − E j ) /2 ( i = j ) , π i = ∑ j e − E j e − E i . (136)
The backward generator is L i j = k i j L_{ij} = k_{ij} L ij = k ij for i ≠ j i \ne j i = j and L i i = − ∑ j ≠ i k i j L_{ii} = -\sum_{j\ne i} k_{ij} L ii = − ∑ j = i k ij . Probability columns evolve with W = L T W = L^T W = L T . Every preparation at temperature T T T is
p i ( T ) = e − E i / T ∑ j e − E j / T . (137) p_i(T) = \frac{e^{-E_i/T}}{\sum_j e^{-E_j/T}}. \tag*{(137)} p i ( T ) = ∑ j e − E j / T e − E i / T . (137)
The final rates are independent of the preparation temperature.
The two nonstationary modes can be constructed analytically, without a floating-point eigensolver. Set q i = π i q_i = \sqrt{\pi_i} q i = π i and
S = − diag ( q ) L diag ( q ) − 1 . (138) S = -\operatorname{diag}(q)L\operatorname{diag}(q)^{-1}. \tag*{(138)} S = − diag ( q ) L diag ( q ) − 1 . (138)
Detailed balance makes the exact matrix S S S real symmetric, and S q = 0 Sq = 0 S q = 0 . With d 0 = π 0 + π 1 d_0 = \sqrt{\pi_0+\pi_1} d 0 = π 0 + π 1 , the vectors
u = ( q 1 , − q 0 , 0 ) d 0 , v = ( q 0 q 2 d 0 , q 1 q 2 d 0 , − d 0 ) (139) u = \frac{(q_1,-q_0,0)}{d_0},
\qquad
v = \left(\frac{q_0q_2}{d_0},\frac{q_1q_2}{d_0},-d_0\right) \tag*{(139)} u = d 0 ( q 1 , − q 0 , 0 ) , v = ( d 0 q 0 q 2 , d 0 q 1 q 2 , − d 0 ) (139)
are orthonormal and perpendicular to q q q . Define α = u T S u \alpha= u^T S u α = u T S u , β = u T S v \beta= u^T S v β = u T S v , and δ = v T S v \delta= v^T S v δ = v T S v . The two positive relaxation rates are exactly
λ s , f = α + δ ∓ ( α − δ ) 2 + 4 β 2 2 . (140) \lambda_{s,f} = \frac{\alpha+\delta\mp\sqrt{(\alpha-\delta)^2+4\beta^2}}{2}. \tag*{(140)} λ s , f = 2 α + δ ∓ ( α − δ ) 2 + 4 β 2 . (140)
The certificate verifies 0 < λ s < λ f 0 < \lambda_s < \lambda_f 0 < λ s < λ f and β ≠ 0 \beta\ne0 β = 0 . A normalized backward eigenfunction for either rate is
ϕ n , i = β u i + ( λ n − α ) v i q i β 2 + ( λ n − α ) 2 , n = s , f . (141) \phi_{n,i} = \frac{\beta u_i + (\lambda_n-\alpha)v_i}{q_i\sqrt{\beta^2+(\lambda_n-\alpha)^2}},
\qquad n=s,f. \tag*{(141)} ϕ n , i = q i β 2 + ( λ n − α ) 2 β u i + ( λ n − α ) v i , n = s , f . (141)
Thus ∑ i π i ϕ n , i = 0 \sum_i \pi_i\phi_{n,i}=0 ∑ i π i ϕ n , i = 0 , ∑ i π i ϕ n , i 2 = 1 \sum_i \pi_i\phi_{n,i}^2=1 ∑ i π i ϕ n , i 2 = 1 , and L ϕ n = − λ n ϕ n L\phi_n=-\lambda_n\phi_n L ϕ n = − λ n ϕ n . These identities follow from the two-dimensional construction. Residual intervals containing zero are additional implementation checks, rather than substitutes for that construction.
For each preparation, calculate
b n ( T ) = ∑ i p i ( T ) ϕ n , i , b_n(T) = \sum_i p_i(T)\phi_{n,i}, b n ( T ) = i ∑ p i ( T ) ϕ n , i ,
f i ( t ; T ) = ∑ n = s , f b n ( T ) ϕ n , i e − λ n t . (142) f_i(t;T) = \sum_{n=s,f} b_n(T)\phi_{n,i}e^{-\lambda_n t}. \tag*{(142)} f i ( t ; T ) = n = s , f ∑ b n ( T ) ϕ n , i e − λ n t . (142)
The relative entropy is evaluated as D ( t ; T ) = ∑ i π i g ( f i ( t ; T ) ) D(t;T) = \sum_i \pi_i g(f_i(t;T)) D ( t ; T ) = ∑ i π i g ( f i ( t ; T )) . This expression is exact and has no linear contribution that must cancel numerically.
The fixed detuned preparation and arbitrary preparation errors
For T H = 3 T_H = 3 T H = 3 and T C = 3 / 2 T_C = 3/2 T C = 3/2 , direct interval evaluation gives
1.81217 < 10 5 w H < 1.81218 , 1.81217 < 10^{5}w_H < 1.81218, 1.81217 < 1 0 5 w H < 1.81218 ,
2.1699246260254279 < 10 4 Γ , 2.1699246260254279 < 10^{4}\Gamma, 2.1699246260254279 < 1 0 4 Γ ,
10 4 Γ < 2.1699246260254281 , 10^{4}\Gamma< 2.1699246260254281, 1 0 4 Γ < 2.1699246260254281 ,
0.18492278483565589 < T u p , 0.18492278483565589 < T_{\mathrm{up}}, 0.18492278483565589 < T up ,
T u p < 0.18492278483565590. (143) T_{\mathrm{up}} < 0.18492278483565590. \tag*{(143)} T up < 0.18492278483565590. (143)
Here w H = b s ( T H ) 2 w_H=b_s(T_H)^2 w H = b s ( T H ) 2 , w C = b s ( T C ) 2 w_C=b_s(T_C)^2 w C = b s ( T C ) 2 , q H = b f ( T H ) 2 q_H=b_f(T_H)^2 q H = b f ( T H ) 2 , and Γ = F ( M C ( 0 ) ) w C − F ( m H ( 0 ) ) w H \Gamma=F(M_C(0))w_C-F(m_H(0))w_H Γ = F ( M C ( 0 )) w C − F ( m H ( 0 )) w H . The hot slow-mode weight is strictly nonzero.
Write d ( t ) = D H ( t ) − D C ( t ) d(t)=D_H(t)-D_C(t) d ( t ) = D H ( t ) − D C ( t ) . Equation (F11) is evaluated over 1000 adjacent, closed, rational time intervals covering [ 0 , 0.169 ] [0,0.169] [ 0 , 0.169 ] . Every interval gives d > 0 d>0 d > 0 ; the smallest certified lower margin exceeds 7.4054 × 10 − 6 7.4054\times10^{-6} 7.4054 × 1 0 − 6 . To refine the crossing location, use
d D d t = ∑ i π i ln ( 1 + f i ) f ˙ i , \frac{dD}{dt}=\sum_i \pi_i\ln(1+f_i)\dot{f}_i, d t d D = i ∑ π i ln ( 1 + f i ) f ˙ i ,
f ˙ i = − ∑ n = s , f λ n b n ϕ n , i e − λ n t . (144) \dot{f}_i=-\sum_{n=s,f}\lambda_n b_n\phi_{n,i}e^{-\lambda_n t}. \tag*{(144)} f ˙ i = − n = s , f ∑ λ n b n ϕ n , i e − λ n t . (144)
Thirty adjacent intervals cover [ 0.169 , 0.172 ] [0.169,0.172] [ 0.169 , 0.172 ] and verify d ′ < 0 d'<0 d ′ < 0 on every cell. A separate evaluation of e W t p H ( 0 ) e^{W t}p_H(0) e W t p H ( 0 ) and e W t p C ( 0 ) e^{W t}p_C(0) e W t p C ( 0 ) then gives
d ( 0.17022 ) > 3.9872 × 10 − 8 , d(0.17022)>3.9872\times10^{-8}, d ( 0.17022 ) > 3.9872 × 1 0 − 8 ,
d ( 0.17024 ) < − 8.1501 × 10 − 8 . (145) d(0.17024)<-8.1501\times10^{-8}. \tag*{(145)} d ( 0.17024 ) < − 8.1501 × 1 0 − 8 . (145)
The early positivity excludes any earlier meeting, and the local derivative check excludes an earlier root within the final bracket. Consequently 0.17022 < t M < 0.17024 0.17022<t_M<0.17024 0.17022 < t M < 0.17024 . The derivative claim is local; it is not extended to all later times.
The same exact nominal preparations permit a conservative error certificate. Let p α ( 0 ) p^\alpha(0) p α ( 0 ) be any normalized probability vector satisfying ∥ p α ( 0 ) − p α ( 0 ) ∥ 1 ≤ 10 − 6 \lVert p^\alpha(0)-p^\alpha(0)\rVert_1\le10^{-6} ∥ p α ( 0 ) − p α ( 0 ) ∥ 1 ≤ 1 0 − 6 independently for α = H , C \alpha=H,C α = H , C . The code verifies componentwise that
0.7 < p i α ( 0 ) − 10 − 6 π i , 0.7<\frac{p_i^\alpha(0)-10^{-6}}{\pi_i}, 0.7 < π i p i α ( 0 ) − 1 0 − 6 ,
p i α ( 0 ) + 10 − 6 π i < 6. (146) \frac{p_i^\alpha(0)+10^{-6}}{\pi_i}<6. \tag*{(146)} π i p i α ( 0 ) + 1 0 − 6 < 6. (146)
The original and perturbed likelihood ratios therefore remain in [ 0.7 , 6 ] [0.7,6] [ 0.7 , 6 ] at all later times by the Markov maximum principle. Centering the entropy gradient on this range gives the Lipschitz constant C = 1 2 ln ( 6 / 0.7 ) < 1.074218 C=\frac{1}{2}\ln(6/0.7)<1.074218 C = 2 1 ln ( 6/0.7 ) < 1.074218 . The combined error in d d d is below 2 C × 10 − 6 < 2.148436 × 10 − 6 2C\times10^{-6}<2.148436\times10^{-6} 2 C × 1 0 − 6 < 2.148436 × 1 0 − 6 . This is less than the verified positive margin on [ 0 , 0.169 ] [0,0.169] [ 0 , 0.169 ] and the magnitude of d ( 0.172 ) < − 1.03253 × 10 − 5 d(0.172)<-1.03253\times10^{-5} d ( 0.172 ) < − 1.03253 × 1 0 − 5 . Every permitted pair consequently has its first meeting in ( 0.169 , 0.172 ) (0.169,0.172) ( 0.169 , 0.172 ) . The tolerance is sufficient; no maximal tolerance is claimed.
Certification of an entire temperature rectangle
Now allow both preparations to vary throughout the closed rectangle
T H ∈ [ 2.9 , 3.1 ] , T C ∈ [ 1.45 , 1.55 ] , T b = 1. (147) T_H\in[2.9,3.1],\qquad T_C\in[1.45,1.55],\qquad T_b=1. \tag*{(147)} T H ∈ [ 2.9 , 3.1 ] , T C ∈ [ 1.45 , 1.55 ] , T b = 1. (147)
This is a continuum of initial conditions. The calculation covers the hot interval with the 200 closed cells
I j H = [ 29 10 + j 1000 , 29 10 + j + 1 1000 ] , j = 0 , … , 199. (148) I_j^H=\left[\frac{29}{10}+\frac{j}{1000},\frac{29}{10}+\frac{j+1}{1000}\right],\qquad j=0,\ldots,199. \tag*{(148)} I j H = [ 10 29 + 1000 j , 10 29 + 1000 j + 1 ] , j = 0 , … , 199. (148)
and the cold interval with the 100 closed cells
I j C = [ 29 20 + j 1000 , 29 20 + j + 1 1000 ] , j = 0 , … , 99. (149) I_j^C=\left[\frac{29}{20}+\frac{j}{1000},\frac{29}{20}+\frac{j+1}{1000}\right],\qquad j=0,\ldots,99. \tag*{(149)} I j C = [ 20 29 + 1000 j , 20 29 + 1000 j + 1 ] , j = 0 , … , 99. (149)
All endpoints are exact rationals. The first and last endpoints, the common width, and equality of every pair of adjacent endpoints are asserted explicitly. Thus there are no gaps, including at the boundary of the rectangle.
For each temperature cell, interval evaluation of the Gibbs formula and Eq. (F11) encloses every probability and modal coefficient throughout that cell. Taking an interval hull over all hot cells and, separately, all cold cells gives bounds valid for every pair in the product rectangle. This independent enclosure of the two preparations can be conservative, but it cannot omit a pair. In particular, the certificate gives uniformly
w H > 3.24019 × 10 − 6 , w_H>3.24019\times10^{-6}, w H > 3.24019 × 1 0 − 6 ,
Γ > 1.8624069 × 10 − 4 , \Gamma>1.8624069\times10^{-4}, Γ > 1.8624069 × 1 0 − 4 ,
T u p < 0.190298. (150) T_{\mathrm{up}}<0.190298. \tag*{(150)} T up < 0.190298. (150)
No hot preparation in this rectangle has an exactly canceled slow mode.
The all-mode envelopes improve the time interval. For compactness in this subsection only, write H = b s ( T H ) 2 H=b_s(T_H)^2 H = b s ( T H ) 2 , A = b f ( T H ) 2 A=b_f(T_H)^2 A = b f ( T H ) 2 , C = b s ( T C ) 2 C=b_s(T_C)^2 C = b s ( T C ) 2 , B = b f ( T C ) 2 B=b_f(T_C)^2 B = b f ( T C ) 2 , and Δ = λ f − λ s \Delta=\lambda_f-\lambda_s Δ = λ f − λ s . At a certified reference s s s , let α α = F ( M α ( s ) ) \alpha_\alpha=F(M_\alpha(s)) α α = F ( M α ( s )) and c α = F ( m α ( s ) ) c_\alpha=F(m_\alpha(s)) c α = F ( m α ( s )) . For every t ≥ s t\ge s t ≥ s the exact lower comparison is
e 2 λ s t d ( t ) ≥ ( a H H − c C C ) + ( a H A − c C B ) e − 2 Δ t . (151) e^{2\lambda_s t}d(t)\ge(a_HH-c_CC)+(a_HA-c_CB)e^{-2\Delta t}. \tag*{(151)} e 2 λ s t d ( t ) ≥ ( a H H − c C C ) + ( a H A − c C B ) e − 2Δ t . (151)
The coefficient a H A − c C B a_HA-c_CB a H A − c C B is strictly positive throughout the whole temperature rectangle at each reference used below. For each fixed pair of temperatures, the right-hand side of Eq. (F20) is therefore decreasing with t t t . If its interval enclosure is positive at t = L t=L t = L , positivity is proved for every t ∈ [ s , L ] t\in[s,L] t ∈ [ s , L ] , not merely at L L L . Table V gives the four successive reference intervals. Induction covers all times from zero through 0.161.
Table V. Uniform temperature-rectangle certificate. The last two columns are conservative lower bounds on the fast coefficient in Eq. (151) and on its right-hand side at t = L t=L t = L . Every entry applies to the entire rectangle (F16).
The upper comparison at the last reference is
e 2 λ s t d ( t ) ≤ ( c H H − a C C ) + ( c H A − a C B ) e − 2 Δ t . (152) e^{2\lambda_s t}d(t)\le(c_HH-a_CC)+(c_HA-a_CB)e^{-2\Delta t}. \tag*{(152)} e 2 λ s t d ( t ) ≤ ( c H H − a C C ) + ( c H A − a C B ) e − 2Δ t . (152)
Its interval upper endpoint at t = 0.179 t=0.179 t = 0.179 is below − 3.12237 × 10 − 5 -3.12237\times10^{-5} − 3.12237 × 1 0 − 5 . Hence d ( 0.179 ) < 0 d(0.179)<0 d ( 0.179 ) < 0 for every pair in the rectangle. Combining this sign with the absence of an earlier meeting proves
0.161 < t M ( T H , T C ) < 0.179 (153) 0.161 < t_M(T_H,T_C) < 0.179 \tag*{(153)} 0.161 < t M ( T H , T C ) < 0.179 (153)
throughout the closed rectangle. In particular, its interior provides an open set of preparations with a certified crossing.
A general certificate excluding earlier crossings
For the larger examples we avoid a spectral assumption in the first-crossing computation. The following bound turns interval propagation at cell endpoints into a statement about every intervening time. Let P ( u ) = e W u P(u)=e^{Wu} P ( u ) = e W u and let z z z be any signed vector. Positivity and unit column sums imply
∥ P ( u ) z ∥ 1 ≤ ∑ i , j P i j ( u ) ∣ z j ∣ = ∑ j ∣ z j ∣ = ∥ z ∥ 1 . (154) \lVert P(u)z\rVert_1 \le\sum_{i,j}P_{ij}(u)|z_j|=\sum_j|z_j|=\lVert z\rVert_1. \tag*{(154)} ∥ P ( u ) z ∥ 1 ≤ i , j ∑ P ij ( u ) ∣ z j ∣ = j ∑ ∣ z j ∣ = ∥ z ∥ 1 . (154)
At a left endpoint t j t_j t j , define v α = ∥ W p α ( t j ) ∥ 1 v_\alpha=\lVert Wp^\alpha(t_j)\rVert_1 v α = ∥ W p α ( t j ) ∥ 1 . Because W W W commutes with its semigroup,
∥ p ˙ α ( t j + u ) ∥ 1 = ∥ P ( u ) W p α ( t j ) ∥ 1 ≤ v α . (155) \lVert\dot{p}^{\alpha}(t_j+u)\rVert_1=\lVert P(u)Wp^\alpha(t_j)\rVert_1\le v_\alpha. \tag*{(155)} ∥ p ˙ α ( t j + u ) ∥ 1 = ∥ P ( u ) W p α ( t j ) ∥ 1 ≤ v α . (155)
Integration for 0 ≤ u ≤ δ t 0\le u\le\delta t 0 ≤ u ≤ δ t gives ∥ p α ( t j + u ) − p α ( t j ) ∥ 1 ≤ u v α \lVert p^\alpha(t_j+u)-p^\alpha(t_j)\rVert_1\le uv_\alpha ∥ p α ( t j + u ) − p α ( t j ) ∥ 1 ≤ u v α . The difference has zero total mass, so the absolute displacement of one component is at most half its L 1 L^1 L 1 norm. Consequently
p i α ( t j ) − δ t v α / 2 π i ≤ p i α ( t j + u ) π i ≤ p i α ( t j ) + δ t v α / 2 π i . (156) \frac{p_i^\alpha(t_j)-\delta t\,v_\alpha/2}{\pi_i}
\le\frac{p_i^\alpha(t_j+u)}{\pi_i}
\le\frac{p_i^\alpha(t_j)+\delta t\,v_\alpha/2}{\pi_i}. \tag*{(156)} π i p i α ( t j ) − δ t v α /2 ≤ π i p i α ( t j + u ) ≤ π i p i α ( t j ) + δ t v α /2 . (156)
The minimum lower endpoint and maximum upper endpoint in Eq. (156) define h α , − h_{\alpha,-} h α , − and h α , + h_{\alpha,+} h α , + . The code requires h α , − > 0 h_{\alpha,-}>0 h α , − > 0 .
To bound the entropy change, join the two probability vectors by a straight segment. Its likelihood ratios stay in this same range. The gradient of D D D has components ln ( p i / π i ) + 1 \ln(p_i/\pi_i)+1 ln ( p i / π i ) + 1 . Subtracting the midpoint of the gradient range does not change its contraction with a zero-mass probability difference. Therefore, with
C α = 1 2 ln h α , + h α , − , (157) C_\alpha=\frac{1}{2}\ln\frac{h_{\alpha,+}}{h_{\alpha,-}}, \tag*{(157)} C α = 2 1 ln h α , − h α , + , (157)
integration along that segment gives
∣ D α ( t j + u ) − D α ( t j ) ∣ ≤ C α ∥ p α ( t j + u ) − p α ( t j ) ∥ 1 ≤ δ t C α v α . (158) \begin{aligned}
|D_\alpha(t_j+u)-D_\alpha(t_j)|
&\le C_\alpha\lVert p^\alpha(t_j+u)-p^\alpha(t_j)\rVert_1\\
&\le\delta t\,C_\alpha v_\alpha. \tag*{(158)}
\end{aligned} ∣ D α ( t j + u ) − D α ( t j ) ∣ ≤ C α ∥ p α ( t j + u ) − p α ( t j ) ∥ 1 ≤ δ t C α v α . (158)
Thus the computable cell margin
M j = d ( t j ) − δ t ( C H v H + C C v C ) (159) \mathcal{M}_j=d(t_j)-\delta t(C_Hv_H+C_Cv_C) \tag*{(159)} M j = d ( t j ) − δ t ( C H v H + C C v C ) (159)
is a lower bound for d ( t ) d(t) d ( t ) throughout [ t j , t j + δ t ] [t_j,t_j+\delta t] [ t j , t j + δ t ] . All quantities on the right are interval evaluated; a strictly positive lower endpoint certifies the entire cell.
For each uniform partition the program computes the interval matrix P ( δ t ) P(\delta t) P ( δ t ) once, then propagates both interval probability columns by repeated multiplication. Dependency between components may widen an enclosure but does not remove the exact trajectory. Exact rational partition endpoints establish complete time coverage. The exact target probability vectors remain normalized by the generator construction; the proof does not assume that the component sum of a finite-width interval vector equals the point interval [ 1 , 1 ] [1,1] [ 1 , 1 ] .
Exact larger-model inputs and certified spectral margins
For the selected eight-state network, the exact energies are
E = 1 7 ( 25 , 10 , 15 , 0 , 35 , 30 , 5 , 20 ) . (160) E=\frac{1}{7}(25,10,15,0,35,30,5,20). \tag*{(160)} E = 7 1 ( 25 , 10 , 15 , 0 , 35 , 30 , 5 , 20 ) . (160)
Table VII supplies all 28 independent prefactors; γ j i = γ i j \gamma_{ji}=\gamma_{ij} γ j i = γ ij and γ i i = 0 \gamma_{ii}=0 γ ii = 0 . The printed finite decimals define exact rational prefactors in the certificate. They are the recorded prefactors of the selected numerical sample. The certificate uses the exact rational energy grid rather than the binary approximations used in exploratory diagonalization. These definitions fix the model without requiring access to a random-number generator or its internal state. Rates are again k i j = γ i j exp [ ( E i − E j ) / 2 ] k_{ij}=\gamma_{ij}\exp[(E_i-E_j)/2] k ij = γ ij exp [( E i − E j ) /2 ] , with T b = 1 T_b=1 T b = 1 and preparations at T H = 3 T_H=3 T H = 3 , T C = 3 / 2 T_C=3/2 T C = 3/2 .
Table VII. Exact finite-decimal prefactors of the certified eight-state network. State labels run from 0 to 7; reverse-edge prefactors are equal.
For the diffusion certificate, take N = 64 N=64 N = 64 , δ x = 6 / N \delta x=6/N δ x = 6/ N , and x i = − 3 + ( i + 1 / 2 ) δ x x_i=-3+(i+1/2)\delta x x i = − 3 + ( i + 1/2 ) δ x , i = 0 , … , N − 1 i=0,\ldots,N-1 i = 0 , … , N − 1 . Define the potential exactly by
U ( x ) = − 3 exp [ − ( x + 1 ) 2 2 ( 1 / 5 ) 2 ] − 2 exp [ − ( x − 1 ) 2 2 ( 7 / 10 ) 2 ] . (161) U(x)=-3\exp\left[-\frac{(x+1)^2}{2(1/5)^2}\right]-2\exp\left[-\frac{(x-1)^2}{2(7/10)^2}\right]. \tag*{(161)} U ( x ) = − 3 exp [ − 2 ( 1/5 ) 2 ( x + 1 ) 2 ] − 2 exp [ − 2 ( 7/10 ) 2 ( x − 1 ) 2 ] . (161)
Only adjacent cells communicate, with
k i , i + 1 = δ x − 2 exp [ − U ( x i + 1 ) − U ( x i ) 2 ] , (162) k_{i,i+1}=\delta x^{-2}\exp\left[-\frac{U(x_{i+1})-U(x_i)}{2}\right], \tag*{(162)} k i , i + 1 = δ x − 2 exp [ − 2 U ( x i + 1 ) − U ( x i ) ] , (162)
and the reverse rate obtained by exchanging the indices. No jump leaves the domain. Every coordinate, width, and potential parameter is rational before the exponentials are evaluated with Arb. The stationary probabilities are proportional to e − U ( x i ) e^{-U(x_i)} e − U ( x i ) ; the preparation weights are proportional to e − U ( x i ) / T e^{-U(x_i)/T} e − U ( x i ) / T at T = 3 , 3 / 2 T=3,3/2 T = 3 , 3/2 .
For both larger fixed models the general theorem’s inputs are also certified. Form the real symmetric matrix S = − diag ( π ) L diag ( π ) − 1 S=-\operatorname{diag}(\sqrt{\pi})L\operatorname{diag}(\sqrt{\pi})^{-1} S = − diag ( π ) L diag ( π ) − 1 . The program calls the validated Rump eigensolver with eigenvectors requested. It requires a successful isolation of all eigenvalues and checks that the real-part intervals are strictly ordered, with an interval containing zero followed by strictly positive intervals. The approximate-only eigensolver option is never used. The validated routine encloses actual eigenpairs; a small residual alone would not supply this guarantee. The corresponding software specification is given in the documented acb_mat.eig interface.3
The exact matrix is real symmetric, so its eigenvalues are real. The real part of an enclosed complex eigenvector is a real eigenvector for the same exact real eigenvalue. Its Euclidean norm is certified strictly positive before division. Normalizing that vector and dividing its components by π i \sqrt{\pi_i} π i gives the weighted-normalized slow eigenfunction. The spectral weights w H , w C w_H,w_C w H , w C then follow from the Gibbs projections. The remaining hot weight is computed exactly as
q H = ∑ i π i ( p i H ( 0 ) π i − 1 ) 2 − w H . (163) q_H=\sum_i\pi_i\left(\frac{p_i^H(0)}{\pi_i}-1\right)^2-w_H. \tag*{(163)} q H = i ∑ π i ( π i p i H ( 0 ) − 1 ) 2 − w H . (163)
The isolated spectrum verifies that the slow eigenspace is one-dimensional in these two particular models. The general theorem itself allows degeneracy when the whole slow subspace is used.
Direct, separately computed matrix exponentials give
d 8 ( 1.26 ) > 2.0557824 × 10 − 6 , (164) d_8(1.26)>2.0557824\times10^{-6}, \tag*{(164)} d 8 ( 1.26 ) > 2.0557824 × 1 0 − 6 , (164)
d 8 ( 1.28 ) < − 2.1815267 × 10 − 6 , (165) d_8(1.28)<-2.1815267\times10^{-6}, \tag*{(165)} d 8 ( 1.28 ) < − 2.1815267 × 1 0 − 6 , (165)
d 64 ( 3 ) > 1.3766163 × 10 − 5 , (166) d_{64}(3)>1.3766163\times10^{-5}, \tag*{(166)} d 64 ( 3 ) > 1.3766163 × 1 0 − 5 , (166)
d 64 ( 3.2 ) < − 1.2647043 × 10 − 5 . (167) d_{64}(3.2)<-1.2647043\times10^{-5}. \tag*{(167)} d 64 ( 3.2 ) < − 1.2647043 × 1 0 − 5 . (167)
Combined with Table VI , these signs establish 1.26 < t M ( 8 ) < 1.28 1.26<t_M^{(8)}<1.28 1.26 < t M ( 8 ) < 1.28 and 3 < t M ( 64 ) < 3.2 3<t_M^{(64)}<3.2 3 < t M ( 64 ) < 3.2 . The first-crossing calculation therefore has a propagation-based verification independent of the spectral-margin calculation.
Table VI. Absence of an earlier crossing for the larger fixed models. The margin is a lower bound for every time in each cell, including the cell boundaries. A negative sign at the separate right endpoint then gives the stated first-crossing interval.
A certified upper endpoint using two retained modes
The selected eight-state model also permits a direct certificate for the partial-spectrum construction. In the conventional ordering 0 = λ 1 < λ 2 < λ 3 < λ 4 < ⋯ 0=\lambda_1<\lambda_2<\lambda_3<\lambda_4<\cdots 0 = λ 1 < λ 2 < λ 3 < λ 4 < ⋯ , the validated eigensolver gives
17.27305635 < λ 4 < 17.27305637. (168) 17.27305635<\lambda_4<17.27305637. \tag*{(168)} 17.27305635 < λ 4 < 17.27305637. (168)
Choose the exact rational lower cutoff λ 4 = 17.27305635 \lambda_4=17.27305635 λ 4 = 17.27305635 , and retain only the two coefficients b 2 , α b_{2,\alpha} b 2 , α and b 3 , α b_{3,\alpha} b 3 , α in each preparation. Their certified eigenfunctions are normalized as above. Define the remaining hot norm
R H ( 0 ) = χ 2 ( p H ( 0 ) ∥ π ) − b 2 , H 2 − b 3 , H 2 > 0. (169) R_H(0)=\chi^2(p_H(0)\Vert\pi)-b_{2,H}^2-b_{3,H}^2>0. \tag*{(169)} R H ( 0 ) = χ 2 ( p H ( 0 ) ∥ π ) − b 2 , H 2 − b 3 , H 2 > 0. (169)
Orthogonality and the cutoff imply
Q H ( t ) ≤ Q ‾ H ( t ) , Q_H(t)\leq\overline{Q}_H(t), Q H ( t ) ≤ Q H ( t ) ,
Q ‾ H ( t ) = b 2 , H 2 e − 2 λ 2 t + b 3 , H 2 e − 2 λ 3 t + R H ( 0 ) e − 2 λ 4 t . (170) \overline{Q}_H(t)=b_{2,H}^2e^{-2\lambda_2t}+b_{3,H}^2e^{-2\lambda_3t}
+R_H(0)e^{-2\lambda_4t}. \tag*{(170)} Q H ( t ) = b 2 , H 2 e − 2 λ 2 t + b 3 , H 2 e − 2 λ 3 t + R H ( 0 ) e − 2 λ 4 t . (170)
Table VIII. Conservative outward intervals for the spectral sufficient condition in the two larger fixed models. The full eigenvalue and eigenvector-derived enclosures are in the machine-readable certificate.
Q C ( t ) ≥ Q ‾ C ( t ) , Q_C(t) \ge\overline{Q}_C(t), Q C ( t ) ≥ Q C ( t ) ,
Q ‾ C ( t ) = b 2 , C 2 e − 2 λ 2 t + b 3 , C 2 e − 2 λ 3 t . (171) \overline{Q}_C(t) = b_{2,C}^{2}e^{-2\lambda_2t} + b_{3,C}^{2}e^{-2\lambda_3t}. \tag*{(171)} Q C ( t ) = b 2 , C 2 e − 2 λ 2 t + b 3 , C 2 e − 2 λ 3 t . (171)
In particular, the omitted hot contribution is bounded, rather than set to zero. At the exact time t = 155163 / 100000 = 1.55163 t = 155163/100000 = 1.55163 t = 155163/100000 = 1.55163 , the interval calculation gives
c H Q ‾ H ( t ) − a C Q ‾ C ( t ) ∈ ( − 8.000113 , − 8.000112 ) × 10 − 10 . (172) c_H\overline{Q}_H(t) - a_C\overline{Q}_C(t)
\in(-8.000113,-8.000112) \times10^{-10}. \tag*{(172)} c H Q H ( t ) − a C Q C ( t ) ∈ ( − 8.000113 , − 8.000112 ) × 1 0 − 10 . (172)
The right-hand side is an upper bound on d ( t ) d(t) d ( t ) , so continuity and the verified initial ordering imply the certified upper endpoint t M < 1.55163 t_M < 1.55163 t M < 1.55163 . This sign test certifies an endpoint, not the numerical root of the envelope or uniqueness of the entropy crossing. It also does not establish a computational speedup: the full eigensystem was computed to validate the cutoff and retained projections in this demonstration.
Reproduction and limits of the certificates
Install the pinned dependency in validation/requirements.txt, then execute certify_generalization.py. The options --bits 384 and --bits 320 reproduce the two recorded runs. Python’s optimization flag -O must not be used because the assertion statements are explicit proof obligations. The output records the software versions, source and exact-input hashes, spectral intervals, parameter-cell counts, reference-stage coefficients, time-cell counts, smallest positive margins, and endpoint signs. An unresolved interval or failed assumption terminates the program.
The continuous temperature-rectangle result and the stated fixed-model inequalities are certified. The finite network-ensemble classifications, the numerical error and coverage summaries across that ensemble, and the other diffusion mesh sizes are ordinary floating-point results unless explicitly identified otherwise. In particular, successful mesh refinement is not a rigorous continuum-error bound. Neither the network sampling frequencies nor these selected model certificates imply that Mpemba crossings are common under every sampling law, or that the analytic bounds are uniformly sharp. The algorithms combine standard entropy comparisons, contraction, validated eigenpairs, and interval propagation; no priority claim for those individual ingredients is made.