MOANSI 2026
Mathematics for Materials Modeling, Institute of Mathematics, EPFL
November 9, 2026
% Blackboard Bold % Calligraphic % Positive definite symmetric matrices
% Script % Fraktur % Other Symbols % Alternative definition, safer % % % General Physics/Math % Geometry and Neighborhoods % Hessian Related % Harmonic Oscillators % Simplified/Indexed Notation % Other Specific Notation $$
\[ \definecolor{gris_C}{RGB}{96,96,96} \definecolor{blanc_C}{RGB}{255,255,255} \definecolor{pistache_C}{RGB}{176,204,78} \definecolor{pistache2_C}{RGB}{150,171,91} \definecolor{jaune_C}{RGB}{253,235,125} \definecolor{jaune2_C}{RGB}{247,208,92} \defineclor{darkblue}{RGB}{} \definecolor{orange_C}{RGB}{244,157,84} \definecolor{orange2_C}{RGB}{239,119,87} \definecolor{bleu_C}{RGB}{126,151,206} \definecolor{bleu2_C}{RGB}{90,113,180} \definecolor{vert_C}{RGB}{96,180,103} \definecolor{vert2_C}{RGB}{68,141,96} \definecolor{pistache_light_C}{RGB}{235,241,212} \definecolor{jaune_light_C}{RGB}{254,250,224} \definecolor{orange_light_C}{RGB}{251,230,214} \definecolor{bleu_light_C}{RGB}{222,229,241} \definecolor{vert_light_C}{RGB}{216,236,218} \]
Let \(X\) be a continuous-time Markov process with values in \(\mathbb X\) (configuration space)
Fix a metastable state \(\mathcal{X}\), initial condition \(\eta_0^X = \mathrm{Law}(X_0)\in\mathcal{P}(\mathcal{X})\), and let \(\tau^X = \inf\{t>0:\,X_t\notin\mathcal{X}\}\) be the exit time from \(\mathcal{X}\)
Under broad conditions on \((X,\mathcal{X},\eta_0^X)\)1, time-marginals conditioned on remaining in \(\mathcal{X}\) converge to a unique quasi-stationary distribution (QSD)2 \(\nu^X\in\mathcal{P}(\mathcal{X})\) \[ \eta_t^X := \mathrm{Law}\left(X_t\,\middle|\,\tau^X>t\right)\xrightarrow[t\to\infty]{}\nu^X,\qquad \forall\,A\subset\mathcal{X},\quad \mathbb{P}_{\nu^X}\left(X_t\in A\,\middle|\,\tau^X>t\right) = \nu^X(A) \] the convergence on the left is Yaglom’s limit, the property on the right defines the QSD
Memoryless property of the QSD: starting from \(\nu^X\), \(\tau^X\) is an exponential random variable, independent from the exit point \[ \forall\,A\subset\partial\mathcal{X},\qquad \mathbb{P}_{\nu^X}\left(\tau^X>t,\,X_{\tau^X}\in A\right) = \mathrm{e}^{-\lambda^X t}\,\mathbb{P}_{\nu^X}\left(X_{\tau^X}\in A\right) \]
Accelerated MD algorithms (AMD, hyperdynamics1, parallel replica2, TAD3) exploit this to sample \((\tau^X,X_{\tau^X})\) efficiently once \(\eta_t^X \approx \nu^X\)
Requires i) defining states \(\mathcal{X}\), and ii) for each \(\mathcal{X}\), a decorrelation time \(t_0(\mathcal{X},\eta_0^X)\) (truncation parameter in Yaglom’s limit)
In most cases4, a quantitative version of Yaglom’s limit is verified \[ \forall\,t\geqslant 0,\qquad \mathrm{d}_{\rm TV}\left(\eta_t^X,\nu^X\right)\leqslant C(\eta_0^X)\,\mathrm{e}^{-\gamma^X t}, \] where \(\gamma^X>0\) is a spectral gap, and \(C(\eta_0^X)\) depends on the initial condition
| On-the-fly | \(\eta_0^X\)-dependent | Physical regime | Non-heuristic | |
|---|---|---|---|---|
| Physical heuristics, linearizations1 | ✗ | ✗ | solid-state/low \(T\) | ✗ |
| Semiclassical asymptotics2 | ✗ | ✗ | solid-state/low \(T\), reversible \(X\) | ✓ |
| Operator learning3 | ✗ | ✗ | reversible \(X\) | ✓ |
| MCMC diagnostics4 | ✓ | ✓ | adiabatic, nonreversible | ✗ |
| Coarse-graining (this work)5 | ✓ | ✓ | adiabatic, nonreversible | ✓ |
Dynamical setting: \(X\) is a jump process on a finite set \(\overline{\mathcal{X}} = \mathcal{X}\cup\partial^\mathcal{X}\), absorbed at the boundary \(\partial^\mathcal{X}\)
Fix a surjective collective variable \(\xi:\mathcal{X}\to\mathcal{Z}\), and define a surrogate process \(Z\) on \(\overline{\mathcal{Z}} = \mathcal{Z}\cup\partial^\mathcal{Z}\) (with \(\partial^\mathcal{Z}= \xi(\partial^\mathcal{X}\)))
Goal: uniform-in-time control on bias functionals \[ \underbrace{\mathrm{d}_{\rm TV}\left(\xi_\star\eta_t^X,\eta_t^Z\right)}_{\text{coarse-graining bias}},\qquad \underbrace{\left|\mathrm{d}_{\rm TV}\left(\eta_t^Z,\nu^Z\right)-\mathrm{d}_{\rm TV}\left(\eta_t^X,\nu^X\right)\right|}_{\text{diagnostic bias}} \] where \(\eta_t^Z = \mathrm{Law}\left(Z_t\,\middle|\,\tau^Z>t\right)\) and \(\nu^Z \in \mathcal{P}(\mathcal{Z})\) is the QSD of \(Z\) within \(\mathcal{Z}\)
Effective dynamics without absorption are well studied1: correspond to linear probability flows (linear Fokker-Planck equations)
The conditional laws \(\eta_t^X,\eta_t^Z\) evolve under non-linear probability flows
For \(Y\in\{X,Z\}\), define the killed semigroup and its Dirichlet generator \(\mathcal{L}^Y\) \[ \forall\,f\in\ell^{\infty}(\overline{\mathcal{Y}}):f(\partial^\mathcal{Y})=0,\quad Q_t^Y f(y) = \mathbb{E}_y\left[f(Y_t)\mathbf{1}_{\tau^Y>t}\right],\quad \partial_t Q_t^Y = \mathcal{L}^Y Q_t^Y \]
The conditioned time-marginals define a non-linear flow on \(\mathcal{P}(\mathcal{Y})\): \[ \eta_t^Y(f) = \mathbb{E}_{\eta_0^Y}\left[f(Y_t)\,\middle|\,\tau^Y>t\right] = \frac{\eta_0^Y\left(Q_t^Y f\right)}{\eta_0^Y\left(Q_t^Y\mathbf{1}_{\mathcal{Y}}\right)} \]
Perron–Frobenius1: unique principal eigentriplet \((\lambda^Y,\nu^Y,h^Y)\in(0,\infty)\times\mathcal{P}(\mathcal{Y})\times\ell^\infty(\mathcal{Y})\), \(\nu^Y,h^Y>0\), \(\nu^Y(h^Y)=1\), \[ \nu^Y Q_t^Y = \mathrm{e}^{-\lambda^Y t}\nu^Y,\qquad Q_t^Y h^Y = \mathrm{e}^{-\lambda^Y t}h^Y \] and \(\nu^Y\) is the QSD for \(Y\) in \(\mathcal{Y}\)
Spectral gap \(\gamma^Y>0\): with \(P^Y f = f - \nu^Y(f)h^Y\), \[ \left\|Q_t^Y P^Y\right\|\leqslant C\,\mathrm{e}^{-(\lambda^Y+\gamma^Y)t}\qquad\Longrightarrow\qquad \mathrm{d}_{\rm TV}\left(\eta_t^Y,\nu^Y\right)\leqslant C(\eta_0^Y)\,\mathrm{e}^{-\gamma^Y t} \]
Pushforward of measures and pullback of observables: \[ \forall\,(\mu,f)\in\mathcal{P}(\mathcal{X})\times \ell^\infty(\mathcal{Z}),\qquad \xi_\star\mu(f) = \mu(\xi^\star f),\qquad \xi^\star f = f\circ\xi \]
We fix a family of fiber measures \(\mathcal F = (\pi_z)_{z\in\mathcal{Z}}\), where \(\pi_z\in\mathcal{P}\left(\xi^{-1}\{z\}\right)\), and define \[ \Pi^Z f(z) = \pi_z(f),\qquad \Pi^X = \xi^\star\Pi^Z \]
\(\Pi^X\) is a projector onto fiber-wise constant observables, defined via averaging under \(\mathcal F\)
Backward error measures à la1
For \(Y\in\{X,Z\}\) and \(0\leqslant t\leqslant T\), define “backward error” probability flows \[ \psi_{t,T}^Y(g) = \frac{\eta_0^Y\left(Q_t^Y\Pi^Y Q_{T-t}^X g\right)}{\eta_0^Y\left(Q_t^Y\Pi^Y Q_{T-t}^X\mathbf{1}_{\mathcal{X}}\right)} \] Coarse-grained evolution on \([0,t]\), exact evolution on \([t,T]\), normalized
Remark: boundary values
If \(\xi_\star\eta_0^X = \eta_0^Z\), notice that \[ \psi_{0,T}^X(\xi^\star f) = \psi_{0,T}^Z(\xi^\star f) \] and that \[ \psi_{T,T}^X(\xi^\star f) = \xi_\star\eta_T^X(f),\qquad \psi_{T,T}^Z(\xi^\star f) = \eta_T^Z(f) \]
Consequence (coarse-graining bias formula2)
\[ \begin{aligned} \xi_\star\eta_T^X(f)-\eta_T^Z(f) &= \int_0^T\partial_t\left[\psi_{t,T}^X-\psi_{t,T}^Z\right](\xi^\star f)\,\mathrm{d}t\\ &= \int_0^T\frac{\eta_t^X\left(A^X Q_{T-t}^X\left(\xi^\star f-\psi_{t,T}^X(\xi^\star f)\mathbf{1}_{\mathcal{X}}\right)\right)}{\eta_t^X\left(\Pi^X Q_{T-t}^X\mathbf{1}_{\mathcal{X}}\right)}\,\mathrm{d}t\\ &\quad-\int_0^T\frac{\eta_t^Z\left(A^Z Q_{T-t}^X\left(\xi^\star f-\psi_{t,T}^Z(\xi^\star f)\mathbf{1}_{\mathcal{X}}\right)\right)}{\eta_t^Z\left(\Pi^Z Q_{T-t}^X\mathbf{1}_{\mathcal{X}}\right)}\,\mathrm{d}t \end{aligned} \]
Defect operators
\[ A^X =\underbrace{ \mathcal{L}^X\Pi^X-\Pi^X\mathcal{L}^X}_{\text{closure defect}},\qquad A^Z = \underbrace{\mathcal{L}^Z\Pi^Z-\Pi^Z\mathcal{L}^X}_{\text{intertwining defect}} \]
Definition (canonical model1)
Jump rates of \(Z\) are QSD-conditional averages on fibers of \(\xi\) (analogy with Gibbs-conditional average coefficients of effective diffusions2)
Proposition (properties of the canonical model3)
The following algebraic relations hold \[ \begin{aligned} \nu^Z &= \xi_\star\nu^X\,\text{(exact QSD pushforward)}, & \lambda^Z &= \lambda^X\,\text{(exact rate)}\\ A^Z &= \Pi^Z A^X \,\text{(closure $\implies$ intertwining)}&& \end{aligned} \] and for \(Y\in\{X,Z\}\), \[ \nu^Y(\Pi^Y f) = \nu^X(f), \qquad A^Y = P^Y A^Y\,\text{(fast decay of defects)}. \]
Theorem1
Let \(\mathcal F\), \(Z\) be the canonical choice, \(\xi_\star\eta_0^X = \eta_0^Z\), and assume an adiabatic condition: for some \(\zeta>2(\lambda^X+\gamma^X)\), \[ \forall\,t\geqslant 0,\qquad\left\|\mathrm{e}^{t(I-\Pi^X)\mathcal{L}^X(I-\Pi^X)}(I-\Pi^X)\right\|\leqslant C_\zeta\,\mathrm{e}^{-\zeta t} \] Then there exists \(C>0\), depending on \(\eta_0^X\) but not on \(T\), such that for all \(T>0\), \[ \mathrm{d}_{\rm TV}\left(\xi_\star\eta_T^X,\eta_T^Z\right)\leqslant C\frac{\|A^X\|}{\zeta}\left(\mathrm{e}^{-\gamma^X T}+T\mathrm{e}^{-(\gamma^Z\wedge\gamma^X)T}\right) \] and \[ \left|\mathrm{d}_{\rm TV}\left(\eta_T^Z,\nu^Z\right)-\mathrm{d}_{\rm TV}\left(\eta_T^X,\nu^X\right)\right|\leqslant C\left(\mathrm{e}^{-(\zeta-\lambda^X)T}+\frac{\|A^X\|}{\zeta}(1+T)\,\mathrm{e}^{-(\gamma^Z\wedge\gamma^X)T}\right) \]
Quasimode estimates2
Under the same assumptions, \[ \|h^X-\Pi^X h^X\|\leqslant C \frac{\|A^X\|}{\zeta},\qquad \|\Pi^Z h^X - h^Z\| \leqslant C \frac{\|A^X\|^2}{\zeta}. \]
Hence \(\xi^\star h^Z\) is a quasimode for \(h^X\) in the adiabatic regime \(\|A^X\|/\zeta \ll 1\)
Two main ingredients
Use the properties of \(Z\) to obtain improved spectral decay estimates for \(\partial_t \psi_{t,T}^Y(\xi^\star f)\) (e.g. \(P^Y A^Y = A^Y\), \(A^Z=\Pi^Z A^X\) lead to many cancellations)
Refine these estimates using a splitting of the semigroup onto slow and fast variables \[ \partial_t (I-\Pi^X)Q_t^X = (I-\Pi^X)\mathcal{L}^X(I-\Pi^X)Q_t^X + (I-\Pi^X)\mathcal{L}^X\Pi^X Q_t^X, \] using a Duhamel-Dyson formula to exploit the adiabatic condition
Fix a partition \(\mathcal{X}= \bigsqcup_{m=1}^M\mathcal{X}_m\) into fibers of \(\xi = \sum_m m\mathbf{1}_{\mathcal{X}_m}\), and consider \[ \mathcal{L}_\varepsilon^X = \mathcal{L}_{\rm mac} - K + \varepsilon^{-1}\mathcal{L}_{\rm mic},\qquad \mathcal{L}_{\rm mic} = \bigoplus_{m=1}^M A_m \] with \(\mathcal{L}_{\rm mac},A_m\) irreducible generators on \(\mathcal{X}\) and \(\mathcal{X}_m\) respectively, \(K\geqslant 0\) is the diagonal killing operator
Theorem 1
There exist \(C,\varepsilon_0>0\) such that for all \(\varepsilon\in(0,\varepsilon_0)\), \[ \sup_{t\geqslant 0}\,\mathrm{d}_{\rm TV}\left(\eta_{t,\varepsilon}^Z,\xi_\star\eta_{t,\varepsilon}^X\right)\leqslant C\varepsilon,\qquad \sup_{t\geqslant 2C\varepsilon|\log\varepsilon|}\left|\mathrm{d}_{\rm TV}\left(\eta_{t,\varepsilon}^Z,\nu_\varepsilon^Z\right)-\mathrm{d}_{\rm TV}\left(\eta_{t,\varepsilon}^X,\nu_\varepsilon^X\right)\right|\leqslant C\varepsilon \]
Proof via spectral perturbation theory2 as \(\varepsilon\to 0\), to control constants from the general theorem uniformly in \(\varepsilon\), with in particular \[ \|A_\varepsilon^X\| \leqslant C,\qquad \zeta_\varepsilon\geqslant\frac{1}{C\varepsilon} \]