Samplepath Large Deviations: Theory and Numerical Tools

Summary: Rare but extreme events in complex systems can often efficiently be described by samplepath large deviations: In the limit of some smallness-parameter approaching zero (such as temperature for chemical reactions, inverse number of particles for thermodynamic limits, or inverse timescale separation for multiscale systems), probabilities and most likely pathways of occurence can are readily accessible. For large and strongly coupled stochastic systems, such as climate, atmosphere, or ocean, the corresponding computations pose a huge numerical challenge. These method borrow heavily from field theory, and represent the rare probability as a path integral, necessitating the computation of instantons and fluctuation determinants. In this project, we adress these challenges, including (1) how to compute the large deviation minimizer (instanton) for large systems, (2) how to compute next-order prefactor corrections, and (3) how to deal with heavy-tailed distributions

Relevant publications

  1. T. Schorlepp, T. Grafke, and R. Grauer, "Gel'fand-Yaglom type equations for calculating fluctuations around Instantons in stochastic systems", J. Phys. A: Math. Theor. 54 (2021), 235003 (link)

  2. T. Grafke, T. Schäfer, and E. Vanden-Eijnden, "Sharp Asymptotic Estimates for Expectations, Probabilities, and Mean First Passage Times in Stochastic Systems with Small Noise", ArXiv (2021) (link)

  3. M. Alqahtani, and T. Grafke, "Instantons for rare events in heavy-tailed distributions", J. Phys. A: Math. Theor. 54 (2021), 175001 (link)

  4. G. Ferré and T. Grafke, "Approximate Optimal Controls via Instanton Expansion for Low Temperature Free Energy Computation", SIAM Multiscale Model. Simul. 19(3) (2021), 1310 (link)

  5. T. Grafke, and E. Vanden-Eijnden, "Numerical computation of rare events via large deviation theory", Chaos 29 (2019), 063118 (link)

  6. T. Grafke, "String Method for Generalized Gradient Flows: Computation of Rare Events in Reversible Stochastic Processes", J. Stat. Mech. 2019/4 (2019), 043206 (link)

Multistable climate regimes

Earth's climate is a highly complex, non-equilibrium and chaotic stochastic system. In this project, we attempt to classify its chaotic attractors with methods from non-equilibrium statistical mechanics, large deviation theory and manifold learning. Concretely, due to the ice albedo feedback, the climate is known to exist in two locally stable states, the current (warm) climate, and a "snowball" state, where the globe is covered in ice. Some models even suggest additional metastable climate states, such as the slushball Earth. Transitions between these climate states, and their local stability, can in principle be analyzed in light of the non-equilibrium quasipotential, characterizing the expected transition times and most likely escape paths out of the current climate state.

Approximate Optimal Controls via Instanton Expansion for Low Temperature Free Energy Computation

G. Ferré and T. Grafke, SIAM Multiscale Model. Simul. 19(3) (2021) 1310–1332

Abstract

The computation of free energies is a common issue in statistical physics. A natural technique to compute such high- dimensional integrals is to resort to Monte Carlo simulations. However, these techniques generally suffer from a high variance in the low temperature regime, because the expectation is often dominated by high values corresponding to rare system trajectories. A standard way to reduce the variance of the estimator is to modify the drift of the dynamics with a control enhancing the probability of rare events, leading to so-called importance sampling estimators. In theory, the optimal control leads to a zero-variance estimator; it is, however, defined implicitly and computing it is of the same difficulty as the original problem. We propose here a general strategy to build approximate optimal controls in the small temperature limit for diffusion processes, with the first goal to reduce the variance of free energy Monte Carlo estimators. Our construction builds upon low noise asymptotics by expanding the optimal control around the instanton, which is the path describing most likely fluctuations at low temperature. This technique not only helps reducing variance, but it is also interesting as a theoretical tool since it differs from usual small temperature expansions (WKB ansatz). As a complementary consequence of our expansion, we provide a perturbative formula for computing the free energy in the small temperature regime, which refines the now standard Freidlin-Wentzell asymptotics. We compute this expansion explicitly for lower orders, and explain how our strategy can be extended to an arbitrary order of accuracy. We support our findings with illustrative numerical examples.


doi:10.1137/20M1385809

arXiv

Rogue Waves: Freaks of Nature Studied with Math and Lasers

The elusive waves, once thought to be myths, are explained by the same math that's found in a wide range of settings.

Link

Stability of Atmospheric Jets

Atmospheric flows on rotating planets are prone to form jets that are surprisingly stable in time. This behaviour can be predicted by simulating the planet's atmosphere numerically. In certain parameter regimes, the underlying dynamical system has multiple locally stable solutions, corresponding to atmosphere configurations with different numbers of jets. This project explores the mechanisms by which random turbulent fluctuations in the atmosphere drive the system to transition between these fixed points, effectively creating or destroying atmospheric jets in the process.

Rogue Waves and Large Deviations

Summary: A statistical theory of rogue waves is proposed and tested against experimental data collected in a long water tank where random waves with different degrees of nonlinearity are mechanically generated and free to propagate along the flume. Strong evidence is given that the rogue waves observed in the tank are hydrodynamic instantons, that is, saddle point configurations of the action associated with the stochastic model of the wave system. As shown here, these hydrodynamic instantons are complex spatio-temporal wave field configurations which can be defined using the mathematical framework of Large Deviation Theory and calculated via tailored numerical methods. These results indicate that the instantons describe equally well rogue waves that originate from a simple linear superposition mechanism (in weakly nonlinear conditions) or from a nonlinear focusing one (in strongly nonlinear conditions), paving the way for the development of a unified explanation to rogue wave formation.

Experimental Setup

The experimental data were recorded in the 270m long wave flume at Marintek (Norway), schematically represented in the schematic. At one end of the tank a plane-wave generator perturbs the water surface with a predefined random signal, here taken to be the JONSWAP spectrum, modelling a random sea state, with enhancement factor \(\gamma\). These perturbations create long-crested wave trains that propagate along the tank toward the opposite end, where they eventually break on a smooth beach that suppresses most of the reflections. The water surface \(\eta(x,t)\) is measured by probes placed at different distances from the wave maker (\(x\)-coordinate). The signal at the wave maker \(\eta(x=0,t) \equiv \eta_0(t)\) is prepared according to the stationary random-phase statistics with deterministic spectral amplitudes \(C(\omega_j)\). Experimental data were collected for three different regimes: quasi-linear (\(\gamma=1\), \(H_s=0.11\) m), intermediate (\(\gamma=3.3\), \(H_s=0.13\) m), and highly nonlinear (\(\gamma=6\), \(H_s=0.15\) m). Note that these three regimes have comparable significant wave heights \(H_s\), but the difference in their enhancement factors \(\gamma\) has significant dynamical consequences.

Extreme-event filtering: Extracting rogue waves from experimental data

To characterize the dynamics leading to extreme events of the water surface, we adopt the following procedure: at a fixed location \(x=L\) along the flume, we select small observation windows around all temporal maxima of \(\eta\) that exceed a threshold \(z\). The choice of the threshold \(z\) is meant to select extreme events with a similar probability for all sets: the values of \(z=H_s=4\sigma\) for the quasi-linear set, \(z=1.1\,H_s=4.4\sigma\) for the intermediate set and \(z=1.2\,H_s=4.8\sigma\) for the highly-nonlinear set, where the maximum of the surface elevation exceeds the threshold at the \(45\) m probe, \(\eta(x=45\) m\(,t)\ge z\). We track the wave packet backward in space and look at its shape at earlier points in the channel. This allows us to build a collection of extreme events and monitor their precursors.

Theoretical description of rogue waves via instantons of NLSE

To avoid solving fully nonlinear water wave equations that are complicated from both theoretical and computational viewpoints, it is customary to use simplified models such as the Nonlinear Schr\"odinger equation (NLSE). If we exclude very nonlinear initial data, it is known that NLSE captures the statistical properties of one dimensional wave propagation to a good degree of accuracy up to a certain time and it can be improved upon by using higher order envelope equations. Because of their simplicity, NLSE and extensions thereof have been successfully used to explain basics mechanisms such as the modulational instability in water waves. With the aim of capturing leading order effects, rather than describing the full wave dynamics, here we restrict ourselves to the NLSE as a prototype model for describing the nonlinear and dispersive waves in the wave flume.

In the limit of deep-water, small-steepness, and narrow-band properties, the evolution of the system is described, to leading order in nonlinearity and dispersion, by the one-dimensional NLSE:

$$ \frac{\partial \psi}{\partial x} + 2\frac{k_0}{\omega_0} \frac{\partial \psi}{\partial t} + {i} \frac{k_0}{\omega_0^2} \frac{\partial^2 \psi}{\partial t^2}+ 2 i {k_0^3} |\psi|^2\psi = 0\,. $$

The NLSE describes the change of the complex envelope \(\psi\equiv\psi(x,t)\) that relates to the surface elevation via the Stokes series truncated at second order:

$$ \eta = |\psi|\cos(\theta) + \tfrac12 k_0 |\psi|^2\cos(2\theta) + O(k_0^2 |\psi|^3)\,, $$

where \(\theta=k_0 x - \omega_0 t + \beta\) and \(\beta\) is the phase of \(\psi\). In this expression the second order term can be neglected when the field amplitude \(|\psi|\) is small — this is the case near the wave maker at \(x=0\), where we will specify initial conditions for the NLSE. However, this second order correction is important when \(|\psi|\) becomes large, i.e. when rogue waves develop.

The NLSE is written as an evolution equation in space (rather than in time) in order to facilitate the comparison with experimental data which are taken along the spatial extend of the flume. Consistent with the wave generator located at \(x=0\), we specify \(\psi(x=0,t)=\psi_0(t)\) as initial condition for the NLSE, which we take to be a Gaussian random field with a covariance whose Fourier transform is related to the JONSWAP spectrum.

Large Deviation Theory and Instanton Calculus

Our analytical and computational descriptions of rare events rely on instanton theory. Developed originally in the context of quantum field theory, at its core lies the realization that the evolution of any stochastic system, be it quantum and classical, reduces to a well-defined (semi-classical) limit in the presence of a small parameter. Concretely, the simultaneous evaluation of all possible realizations of the system subject to a given constraint results in a (classical or path-) integral whose integrand contains an action functional \(S(\psi)\). The dominating realization can then be obtained by approximating the integral by its saddle point approximation, using the solution to \(\delta S(\psi^*)/\delta \psi=0\). This critical point \(\psi^*\) of the action functional is called the instanton, and it yields the maximum likelihood realization of the event. This conclusion can also be justified mathematically within Large Deviation Theory.

Specifically, we are interested in the probability

$$ P_L(z) \equiv \mathbb{P}(\eta(L,0)\ge z) $$

i.e. the probability of the surface elevation at position \(L\) at an arbitrary time \(t=0\) exceeding a threshold \(z\). This probability can in principle be obtained by integrating the distribution of the initial conditions over the set

$$ \Lambda(z)=\{\psi_0: \eta(L,0))\ge z\},\label{eq:13} $$

i.e. the set of all initial conditions \(\psi_0\) at the wave maker \(x=0\) that exceed the threshold \(z\) further down the flume at \(x=L\). Since the initial field \(\psi_0(t)\) is Gaussian, the probability \(P_L(z)\) can therefore be formally written as the path integral

$$ P_L(z) = Z^{-1} \int_{\Lambda(z)} \exp(-\tfrac12 \|\psi_0\|^2_C)\,D[\psi_0]\,, $$

where \(Z\) is a normalization constant and the \(C\)-norm is induced by the energy spectrum (in this case, JONSWAP) of the initial condition. The set \(\Lambda(z)\) has a very complicated shape in general, that depends non-trivially on the nonlinear dynamics of the NLSE since it involves the field at \(x=L>0\) down the flume rather than \(x=0\). One way around this difficulty is to estimate the integral via Laplace's method. This strategy is the essence of Large deviation theory (LDT), or, equivalently, instanton calculus, and it is justified for large \(z\), when the probability of the set \(\Lambda(z)\) is dominated by a single \(\psi_0\) contributing most to the integral. The optimal condition leads to the constrained minimization problem

$$ \tfrac12\min_{\psi_0\in \Lambda(z)}\, \| \psi_0\|^2_C\equiv I_L(z)\,, $$

and gives the large deviation estimate for \(P_L(z)\) of,

$$ P_L(z) \asymp \exp\left(-I_L (z) \right)\,,$$

where the symbol \(\asymp\) means asymptotic logarithmic equivalence, i.e. the ratio of the logarithms of the two sides tends to 1 as \(z\to\infty\), or, in other words, the exponential portion of both sides scales in the same way with \(z\). Intuitively, this estimate says that, in the limit of extremely strong (and unlikely) waves, their probability is dominated by their least unlikely realization, the instanton.

Now, the stochastic sampling problem is replaced by a deterministic optimization problem, which we solve numerically. The trajectory initiated from the minimizer \(\psi_0^*\) of the action will be referred to as the instanton trajectory, and in the following we compare it to trajectories obtained from the experiment.

Experimental Rogue Waves compared to Instantons

In the figure above, numerically computed instanton realizations are depicted in solid black, while experimental averages and their standard deviation are in color. We compare the evolution of rogue waves observed in the experiment and averaged over many realizations to that of the instanton, both constrained at \(x=45\) m. In all cases the instanton tracks the dynamics of the averaged wave very closely during the whole evolution. Moreover, in the focusing region the standard deviation around the mean is small, especially toward the end of the evolution. This observation in itself is a statement that indeed all of the rogue waves such that \(\eta(L,0)\ge z\) resemble the instanton plus small random fluctuations. The instanton approximation shows excellent agreement not only across different degrees of nonlinearity (and therefore substantially different physical mechanisms), but also captures the behavior of precursors earlier along the channel.

It is worth stressing that the instanton approach captures both the linear and the fully nonlinear cases, unlike previous theories that could describe each of these regimes individually but not both. To make that point, in the next two sections we compare the predictions of our approach to those of the quasi-determinism and semi-classical theories that hold in the dispersive and nonlinear regimes, respectively.

Conclusion

Here we have proposed a unifying framework based on Large Deviation Theory and Instanton Calculus that is capable to describe with the same accuracy the shape of rogue waves that result either from a linear superposition or a nonlinear focusing mechanism. In the limit of large nonlinearity, the instantons closely resemble the Peregrine soliton to describe extreme events, but with the added bonus that our framework predicts their likelihood; in the limit of linear waves, the instanton reduces to the autocorrelation function. A smooth transition between the two limiting regimes is also observed, and these predictions are fully supported by experiments performed in a large wave tank with different degrees of nonlinearity. These results were obtained for one dimensional propagation, but there are no obstacles to apply the approach to two horizontal dimensions, which may finally explain the origin and shape of rogue waves in different setups, including the ocean.

Relevant publications

  1. G. Dematteis, T. Grafke, M. Onorato, and E. Vanden-Eijnden, "Experimental Evidence of Hydrodynamic Instantons: The Universal Route to Rogue Waves", Phys. Rev. X 9 (2019), 041057

  2. G. Dematteis, T. Grafke, and E. Vanden-Eijnden, "Extreme event quantification in dynamical systems with random components", J. Uncertainty Quantification 7 (2019), 1029

  3. G. Dematteis, T. Grafke, and E. Vanden-Eijnden, "Rogue Waves and Large Deviations in Deep Sea", PNAS 115 (2018), 855-860

  4. T. Grafke, R. Grauer, and T. Schäfer, "The Instanton Method and its Numerical Implementation in Fluid Mechanics", J. Phys. A: Math. Theor. 48 (2015), 333001

New Model Predicts Sudden Rogue Waves

Unified theory describes formation of huge, mysterious waves.

Link

Monsterwellen auf der Spur

Seit Jahren streiten Wissenschaftler darüber, wie die rätselhaften Giganten des Meeres entstehen, die schon etlichen Menschen das Leben gekostet haben. Ein neuer Ansatz aus der Wahrscheinlichkeitstheorie könnte Monsterwellen nun vorhersagen – unabhängig davon, was sie verursacht.

Link

The Grand Unified Theory of Rogue Waves

Rogue waves — enigmatic giants of the sea — were thought to be caused by two different mechanisms. But a new idea that borrows from the hinterlands of probability theory has the potential to predict them all.

Article, Podcast

A Unifying Framework for Describing Rogue Waves

A theory for rogue waves based on instantons—a mathematical concept developed in quantum chromodynamics—has been successfully tested in controlled laboratory experiments.

Link