Searcharxiv⌕ Search

arXiv subjects

Imre Boros

Publications and source records attributed to Imre Boros.

4 recordsLinked to original sources

Right shape, wrong time: phase errors of ODE solvers in the flame problem

The flame model $y'=y^2-y^3$, $y(0)=δ$, is a standard illustration of stiffness: for $δ=10^{-4}$ the stiff MATLAB solvers ode23s and ode15s need only 99 and 140 steps, against 3040 for ode45. We show that these cheap solutions have the right shape at the wrong time: they place the front 29 and 329 time units early, with maximal errors 0.95 and 1.00. Because all solutions are time translates of one master profile, the global error of every numerical solution with values in $[δ,1)$ is exactly a phase error, $y_k=y_δ(t_k-D_k)$. For fixed-step one-step methods of order $p$ we prove $D=h^p\int_δ^{y}g+O(h^{p+1})$ with a remainder bounded uniformly in $δ$, so that the delay grows at most like $h\ln(1/δ)$ for $p=1$ and stays bounded for $p\geq2$. For any computed solution the delay is exactly the sum of its defects with respect to the exact flow, converted into time; a heuristic model, confirmed by MATLAB runs, explains why it is created in the first few steps and why tightening the tolerances is an expensive cure. Integrating for $1/y$, which runs like a clock in the slow phase, with an absolute tolerance reduces the error by several orders of magnitude, at a median cost close to that of the standard formulation. The mechanism needs only that the equation is autonomous; its general form is the subject of a companion paper.

math.NA↗

Front-pinned spectral collocation for the flame problem and related ignition problems

We consider global Chebyshev collocation for the ignition problems $y'=y^q(1-y)$, $y(0)=δ$, $q\ge2$; for $q=2$ this is the flame problem, a standard stiff test equation. After rescaling to $[-1,1]$ the solution has an interior layer of width $O(μ)$, $μ=(q-1)δ^{q-1}$, at a $δ$-dependent position. Two obstacles arise: plain Chebyshev collocation needs $O(1/μ)$ nodes, and imposing the initial condition amplifies errors by a factor of order $1/μ$, which makes the collocation equations severely ill-conditioned. Our main contribution addresses the second one. Since all solutions are translates of one profile, the initial condition can be replaced by an equivalent interior condition. For an unknown $u=φ(y)$, imposing it where the speed $|φ'f_q|$ of $u$ along the orbit is maximal bounds the propagator of the linearized equation by one; for $u=\ln y$ this is $y(x^\ast)=(q-1)/q$. For the first obstacle we use the sinh map of Tee and Trefethen: the time function is explicit, the complex singularities of the profile lie exactly at $\pm(2k+1)πi$, and the map parameters follow in closed form, so that $O(\ln(1/μ))$ nodes suffice. In double precision the pinned scheme reaches the interpolation error of the exact solution in five or six Newton iterations for $μ$ down to $10^{-12}$, whereas the initial-condition form fails or converges to wrong discrete solutions, also in an integral (Birkhoff) reformulation. Computations in $256$-bit arithmetic confirm the predicted rate down to errors of $10^{-70}$, AAA approximation locates the singularities to within $2\%$, and the method carries over to an Arrhenius reaction rate, with all parameters computed numerically.

math.NA↗

New optimization algorithms for neural network training using operator splitting techniques

In the following paper we present a new type of optimization algorithms adapted for neural network training. These algorithms are based upon sequential operator splitting technique for some associated dynamical systems. Furthermore, we investigate through numerical simulations the empirical rate of convergence of these iterative schemes toward a local minimum of the loss function, with some suitable choices of the underlying hyper-parameters. We validate the convergence of these optimizers using the results of the accuracy and of the loss function on the MNIST, MNIST-Fashion and CIFAR 10 classification datasets.

cs.LG↗

Benchmark for numerical solutions of flow in heterogeneous groundwater formations

This article presents numerical investigations on accuracy and convergence properties of several numerical approaches for simulating steady state flows in heterogeneous aquifers. Finite difference, finite element, discontinuous Galerkin, spectral, and random walk methods are tested on one- and two-dimensional benchmark flow problems. Realizations of log-normal hydraulic conductivity fields are generated by Kraichnan algorithms in closed form as finite sums of random periodic modes, which allow direct code verification by comparisons with manufactured reference solutions. The quality of the methods is assessed for increasing number of random modes and for increasing variance of the log-hydraulic conductivity fields with Gaussian and exponential correlation. Experimental orders of convergence are calculated from successive refinements of the grid. The numerical methods are further validated by comparisons between statistical inferences obtained from Monte Carlo ensembles of numerical solutions and theoretical first-order perturbation results. It is found that while for Gaussian correlation of the log-conductivity field all the methods perform well, in the exponential case their accuracy deteriorates and, for large variance and number of modes, the benchmark problems are practically not tractable with reasonably large computing resources, for all the methods considered in this study.

cs.CE↗