This browser is older than PubPhys supports. Every page can be read, but signing in to contribute, voting and signing need Safari 15.4, Chrome or Edge 99, Firefox 97 or newer. Browsers
All problems
57 problems
Starting order drafted by AI. Votes change it. How ranking works
-
Predicting neural scaling-law exponents from properties of the data
Partially resolved 72For a deep network trained in the feature-learning regime on natural data, the test loss follows $L($$N$, $D$$) = L_{\infty} + A N^{-\beta_N} + B D^{-\beta_D}$ over several decades, with N parameters and D training examples or tokens. Derive $\beta_N, \beta_D$ and the compute-optimal allocation exponent from measurable statistics of the data distribution (for example the power-law decay exponent of the data covariance spectrum or of the target function's spectral decomposition). An answer is a theory that predicts measured exponents on a held-out dataset and architecture within quoted errors, or a proof that no data statistic of this kind fixes them.
-
For the 3D Edwards-Anderson Ising model H = -sum_<ij> J_ij s_i s_j with symmetric Gaussian or $\pm J$ couplings at $T < T_c$, determine in the thermodynamic limit whether the spin overlap distribution $P(q)$ keeps a continuous part between $-q_{\mathrm{EA}}$ and $q_{\mathrm{EA}}$ (replica symmetry breaking, RSB), whether system-spanning excitations cost $O(1)$ energy (RSB) or grow as $L^{\theta}$ with $\theta > 0$ (droplet picture), or whether the intermediate trivial-nontrivial (TNT) scenario holds. An answer is one of these classifications supported by finite-size analysis that controls the $L \to \infty$ limit, or by a proof.
-
For driven lattice gases and boundary-driven diffusive systems described by macroscopic fluctuation theory, determine whether intensive variables and a free-energy-like function, defined from isolated steady states (as in Sasa-Tasaki steady-state thermodynamics), predict beyond linear response the outcome of macroscopic operations: contact between systems, phase coexistence, and response to slow changes of parameters. Answer: yes with the construction and its domain, or a no-go result with counterexamples.
-
Finite predictability horizon in multiscale chaotic flows
Partially resolved 58For flows with energy spectrum $E(k) \sim k^{-\beta}$ over an inertial range, Lorenz predicted that for $\beta < 3$ the predictability time $T_p$ stays bounded as the initial error amplitude $\delta_0 \to 0$ at fixed error scale, since error moves upscale in eddy turnover times $\tau(k) \sim k^{((\beta-3)/2)}$. Determine whether this holds for 3D Navier-Stokes turbulence ($\beta = 5/3$) as the Reynolds number grows, and how $T_p$, in units of the large-eddy turnover time, depends on the dissipation range and on the scale of the initial error. Answer: yes or no, with $T_p$ and its dependences.
-
Explain, within a stated measure on initial conditions (Liouville measure, or a measure from a quantum theory of gravity), why the coarse-grained entropy of the observable universe at early times was far below its maximum (today approximately $10^{104}\,k_B$ against a black-hole bound of order $10^{123}\,k_B$), or show that the low-entropy initial state must be postulated (the Past Hypothesis). An answer is a mechanism that makes it typical, or an argument that none can.
-
For a planted distribution $P_n$ and a null distribution $Q_n$ on $\{0,1\}^M$ (for example random graphs), the degree-D advantage is the best correlation of a degree-D polynomial with the likelihood ratio $\mathrm{d}P_n/\mathrm{d}Q_n$. Identify an explicit, checkable condition on $(P_n, Q_n)$ such that a bounded advantage at $D \sim \operatorname{polylog}(n)$, together with the condition, implies that no polynomial-time algorithm distinguishes $P_n$ from $Q_n$ after independent noise; or prove that no condition covering standard planted problems (planted clique, sparse PCA, tensor PCA, community detection) can exist. An answer is a stated corrected conjecture with a proof, or a proof that the program fails.
-
For finite-state continuous-time Markov jump processes whose rates obey local detailed balance, $k(x\to y)/k(y\to x) = \exp(s(x,y))$ with s the entropy flux to the environment, determine whether a variational functional built from entropy production and dynamical activity (frenesy, the time-symmetric part of the path weight) has an extremum that gives the stationary distribution to all orders in the driving, as minimum entropy production does to linear order. The Landauer blowtorch example shows stationary occupations depend on kinetic details beyond the entropy flux. Answer: yes with the functional and its domain, or a no-go theorem.
-
In the Erdos-Renyi graph $G(n, d/n)$ with large fixed average degree $d$, the largest independent set has size about $2 (\ln d / d)n$, while greedy and known polynomial-time algorithms reach about $(\ln d / d)n$. Does any polynomial-time algorithm output, with high probability, an independent set of size $(1 + \epsilon)(\ln d / d)n$ for a fixed $\epsilon > 0$ and all sufficiently large $d$? The answer is yes (an algorithm with proof) or no (a proof of hardness for all polynomial-time algorithms under a stated assumption).
-
For a chaotic system with known equations (Kolmogorov flow, high-dimensional Lorenz-96) and initial uncertainty $\delta_0$, determine the maximal lead time at which events of an observable above its 99.9th percentile can be forecast with given precision and recall, in units of the inverse maximal Lyapunov exponent, and its relation to finite-time Lyapunov exponents along precursor trajectories. An answer is a scaling law for lead time versus $\delta_0$ and event size.
-
Predictive theory of collisionless violent relaxation
Partially resolved 54For Vlasov dynamics of long-range systems (Hamiltonian mean-field model, 1D self-gravitating sheets, 3D gravitating N-body systems as $N\to\infty$), determine the conditions on initial data (virial ratio, waterbag levels) under which Lynden-Bell's maximum-entropy prediction of the quasi-stationary state is accurate, and give a predictive theory of the core-halo states produced when mixing is incomplete.
-
Test error of interpolating two-layer networks in the feature-learning regime
Partially resolved 52Two-layer network $f(x) = \sum_{j=1..m} a_j \sigma(w_j . x)$ trained by gradient descent from small initialization to zero training loss on n samples $y = f_{*}(x) + \mathrm{noise}$ of variance $\sigma_n^2$, $x \sim \mathrm{N}(0, I_d)$ with $I_d$ the d x d identity, in the limit $n, d \to \infty$ with $\psi = n/d$ fixed and $m >> n$. For targets $f_{*}(x) = g(U x)$ with $U$ an r x d matrix and r fixed, compute the limiting test mean-squared error on noisy labels $E(\psi, \sigma_n^2, g)$ at fixed $\psi$, and classify the interpolation by the $\psi \to \infty$ limit of $E$: benign if $E \to \sigma_n^2$, tempered if $E \to c\,\sigma_n^2$ with $1 < c < \infty$, catastrophic if $E$ diverges. The answer is an asymptotic formula, proved or confirmed numerically to high precision.
-
For the 3D Ising spin glass in a uniform or random field $h$, determine whether a thermodynamic phase transition occurs at nonzero $h_{\mathrm{AT}}(T)$ for $T < T_c$, signaled by a diverging replicon susceptibility and replicon correlation length. Answer: yes or no, and the line $h_{\mathrm{AT}}(T)$ if yes.
-
Determine whether physically realizable driven dissipative systems exist whose steady state is critical (avalanche distributions with a cutoff diverging with system size) for an open set of control parameters, as opposed to requiring the double limit of drive rate $h \to 0$ and bulk dissipation $\epsilon \to 0$ that maps self-organized criticality onto a tuned absorbing-state transition. Answer: yes with a model or experiment, or no with a general argument.
-
For directional random surface gravity waves in deep water described by the Zakharov equation or the 2D nonlinear Schrodinger equation, determine how the exceedance probability of crest-to-trough height $H > 2 H_s$ depends on the Benjamin-Feir index BFI (steepness divided by spectral bandwidth) and on directional spreading, and whether it exceeds second-order (bound-wave) statistics in realistic directional seas. An answer is a quantitative law tested against field and tank data.
-
Construct a coarse-grained entropy functional $S_{\mathrm{grav}}$ for classical self-gravitating systems or perturbed Friedmann spacetimes (for example built from the Weyl tensor, the tidal part of spacetime curvature, following the Weyl curvature hypothesis) that is nondecreasing under gravitational clustering and reduces to the Bekenstein-Hawking entropy $A/(4\ell_{P}^{2})$ for black holes. An answer is a definition with a monotonicity proof for a stated class of dynamics, or a no-go result.
-
For 1D anharmonic chains conserving energy, momentum and stretch (e.g., FPUT with asymmetric or symmetric potentials), determine the exponent delta in the conductivity $\kappa(L) \sim L^\delta$. Nonlinear fluctuating hydrodynamics predicts $\delta = 1/3$ for generic asymmetric potentials (FPUT-alpha-beta: KPZ sound modes and a Levy 5/3 heat mode); for the symmetric FPUT-beta chain numerics and mode-coupling theory support $\delta = 2/5$. An answer is the exponent for each class, confirmed at sizes where finite-size drift is controlled.
-
When single optimal paths dominate extreme events
Partially resolved 46For stochastically forced systems (Burgers, 2D and 3D Navier-Stokes with Gaussian forcing of strength $\epsilon$) and for deterministic chaotic systems, determine the observables $O$ and thresholds $z$ for which $P(O > z)$ is given asymptotically ($z \to \infty \text{ or } \epsilon \to 0$) by a single minimizer of the Freidlin-Wentzell action, including the prefactor, and characterize failures due to fluctuations around the instanton or several competing paths. An answer is a classification with verified prefactors.
-
Optimal estimate of total dissipation from partial observation
Partially resolved 46Given a stationary Markov process of which only a coarse-grained observable is recorded (a subset of transitions, or lumped states), determine the largest lower bound on the total entropy production rate $\sigma$ computable from the observed statistics (waiting-time distributions, current fluctuations, time-irreversibility of observed trajectories), and the conditions under which it equals $\sigma$. An answer is an estimator with a proof that no other estimator using the same data does better.
-
For the disordered discrete nonlinear Schrodinger equation i dpsi_n/dt = eps_n psi_n + beta |psi_n|^2 psi_n - (psi_(n+1) + psi_(n-1)) with i.i.d. random eps_n, and the analogous Klein-Gordon chain, determine whether the second moment $m_2(t)$ of an initially localized packet diverges as $t \to \infty$ (simulations show $m_2 \sim t^{1/3}$ over many decades) or saturates once the packet density falls below a chaos threshold. Answer: yes or no, with the asymptotic law.
-
For the nearest-neighbor Ising model on the simple cubic lattice, find a closed-form expression for the free energy per site, or for the universal exponents $\nu = 0.629971(4)$ and $\eta = 0.0362978(20)$ (conformal bootstrap values), or show that none exists within a stated class of closed forms. Computing Ising partition functions on general nonplanar graphs with arbitrary couplings is NP-hard (Istrail 2000), but this does not exclude an exact solution for the uniform cubic lattice; none is known. The open part is whether the free energy or the universal exponents have a closed form.
-
Sample complexity of gradient training for multi-index targets
Partially resolved 42Inputs $x \sim N(0, I_d)$ and labels $y = g(W x)$ with $W$ an $r$ x d matrix of rank r fixed as $d \to \infty$ and $g$ a fixed link function. For a standard two-layer network trained by online or multi-pass SGD on the square loss without data preprocessing, determine the exponent $\gamma(g)$ in $n \sim d^\gamma$ required to reach test error $o(1)$, and determine whether it equals the benchmark for statistical-query and low-degree algorithms, $n \sim d^{\max(1, k_*/2)}$, with $k_*$ the generative exponent for single-index targets and the generative leap exponent for multi-index targets (Damian, Lee and Bruna, arXiv:2506.05500). The answer is $\gamma$ as a function of the Hermite structure of $g$, with proof.
-
Universal scaling of the FPUT equipartition time
Partially resolved 42For alpha- and beta-FPUT chains with $N$ particles at energy density $\epsilon$ and effective nonlinearity g ($g \sim \alpha^2 \epsilon \text{ or } \beta \epsilon$), determine whether the equipartition time obeys $T_{\mathrm{eq}} \sim g^{-2}$ for $N > N*(g)$ in the thermodynamic limit, as wave-turbulence theory of resonant wave interactions predicts, and determine $N*(g)$ and the small-N regime where exact resonances are absent and the exponent differs. An answer is the exponent with its range of validity.
-
Upper critical dimension of the KPZ class
Contested 42Determine whether there is a finite $d_u$ such that for $d \ge d_u$ the strong-coupling KPZ fixed point has $\alpha = 0$ and $z = 2$ (mode-coupling and some functional RG treatments suggest $d_u = 4$), or whether $\alpha > 0$ for every finite $d$. Answer: yes with $d_u$, or no.
-
Zeroth law for two driven systems in weak contact
Contested 40For two driven lattice gases (e.g., driven exclusion processes or the Katz-Lebowitz-Spohn model) exchanging particles through a weak contact, determine whether an intensive variable defined from each isolated steady state, such as the chemical potential of Sasa-Tasaki steady-state thermodynamics, predicts the joint stationary densities to leading order in the contact rate, independent of the contact dynamics. Answer: yes with the construction, or a class of counterexamples.
-
Why directed percolation is rarely seen in experiments
Partially resolved 40Identify which conditions of the Janssen-Grassberger conjecture (a single absorbing state, short-range interactions, no extra conservation laws or quenched disorder) fail in candidate experiments (catalytic surface reactions, cyclically sheared suspensions, granular flows, transitional shear flows), and quantify how each deviation shifts the measured exponents. An answer is a predictive criterion for when DP exponents are observed.
-
Minimal dissipation per coherent oscillation of a stochastic clock
Partially resolved 40For a finite Markov jump network with stationary entropy production rate $\sigma$ and second eigenvalue $-\lambda_R + i \lambda_I$ of the rate matrix, the conjecture $\sigma \ge \lambda_I^{2}/\lambda_R$ (arXiv:2112.01607) fails (counterexample arXiv:2609.34352), while a version weighted by an eigenvector-uniformity factor is proven (arXiv:2606.05498). Find the tightest bound on entropy produced per oscillation period in terms of the coherence of an observable correlation function $C(t)$ (number of oscillations before decay), and a model that saturates it.
-
For KPZ growth in $d = 2$ with flat, curved (radial) and stationary initial conditions, write $h(x,t)=v t+(\Gamma t)^{\beta}\chi$ and identify the limiting distribution of $\chi$ in each geometry, together with its spatial and temporal correlation functions, and determine whether it has a solvable structure such as a Fredholm determinant or determinantal point process. An answer is an explicit distribution matching simulations.
-
Mechanism of memory and rejuvenation in aging spin glasses
Partially resolved 38In temperature-cycling protocols on spin-glass samples and in large simulations of the $3\mathrm{D}$ Edwards-Anderson model, identify the mechanism (growing correlation lengths, temperature chaos, hierarchical states) that yields both rejuvenation of the out-of-phase susceptibility $\chi''(\omega,t)$ after a step $\Delta T$ and memory of earlier aging plateaus. An answer is a theory that predicts $\chi''(\omega,t)$ quantitatively as a function of $\Delta T$ and waiting times.
-
For the Manna model and conserved lattice gases (fixed-energy sandpiles) in $d = 1, 2, 3$, determine whether the absorbing transition belongs to a class distinct from DP, and whether it coincides with depinning of a disordered elastic interface (quenched Edwards-Wilkinson class) as field-theory mappings suggest. An answer is a classification with exponent values in each $d$.
-
For the KPZ equation $dh/dt = \nu \mathrm{lap}\ h + (\lambda/2)(\mathrm{grad}\ h)^2 + \eta(x,t)$ with Gaussian white noise $\eta$ in $d = 2$ spatial dimensions, determine the roughness exponent $\alpha$ (with $z = 2 - \alpha$ from Galilean invariance) exactly or to $10^{-4}$, and decide whether it is a simple rational number. Large simulations give $\alpha$ approximately 0.387 to 0.39; the conjecture $\alpha = 2/5$ is disfavored.
