Summary and Synthesis

Module Summary: Your Dynamics Toolkit¶

This module has given you a complete toolkit for temporal evolution built on understanding fundamental trade-offs:

Part 1 connected Euler’s local approximation to its behavior on an oscillator:

• Forward Euler on an undamped oscillator creates systematic energy growth

• Local accuracy doesn’t guarantee global stability

• Oscillator phase errors accumulate with duration at fixed step

• Geometric structure matters more than we thought

Part 2 showed how Runge-Kutta methods achieve higher accuracy:

• Multiple sampling captures solution curvature

• RK4’s weights reduce to Simpson quadrature for a time-only right-hand side

• Adaptive timesteps handle varying timescales

• Higher order does not by itself guarantee energy conservation

Part 3 introduced symplectic integration’s profound insight:

• Preserve geometry while still measuring trajectory accuracy

• Leapfrog has favorable long-time energy behavior for resolved, smooth, bounded conservative trajectories

• Modified Hamiltonians explain near-conservation under stated assumptions; step stability must still be checked

• Time reversibility follows from the symmetric kick–drift–kick composition

Part 4 analyzed when methods become unstable:

• Stability regions diagnose amplification on the linear test equation

• Stiff equations can force explicit stability steps much smaller than accuracy requires

• Implicit methods trade computation for stability

• CFL conditions connect space and time discretization

These aren’t separate topics — they’re one framework showing how continuous physics becomes discrete computation while preserving what matters most.

Key Takeaways¶

• Local and global errors differ: Euler’s one-step defect is $O(h^2)$; its global error is $O(h)$ at fixed duration under appropriate assumptions.

• Geometric preservation complements accuracy; compare them for the observable and duration

• Near-energy conservation has conditions: smoothness, appropriate fixed steps, and bounded trajectories

• Classical RK4 has fourth-order global accuracy under its assumptions; the SHO example also exhibits small artificial damping.

• Linear stability tests diagnose mode amplification; they do not replace accuracy, nonlinear stability, or physical-admissibility checks.

• Stiff equations often favor implicit methods, although specialized explicit and exponential methods also exist

• Ordinary state-dependent step selection generally loses symplecticity; specialized constructions need a separate argument

• Modified Hamiltonians explain why symplectic methods work

• Hamiltonian flow preserves canonical phase-space volume; a dissipative physical model need not.

• Transfer the questions with care: identify the equations and observable before choosing the relevant diagnostics.

Investigation: a stable answer to the wrong question¶

The oscillator showed that a small energy error can coexist with a growing phase error. Now change the physics. Let $u$ be a population divided by its initial population and let $\tau$ be time divided by its decay timescale:

This is a dimensionless model of exponential loss. A physically meaningful solution is nonnegative and decreases monotonically. Conservation of $u$ would be wrong. The mathematical model itself removes population.

Predict from the update before looking at a trajectory¶

Euler gives

For each candidate step $h=0.25$, 1.5, and 3, predict three separate properties: boundedness, positivity, and accurate decay. Explain your prediction using $1-h$, rather than a memorized ranking of method names.

• $|1-h|<1$ means the discrete magnitude decreases.

• $0\le1-h\le1$ also preserves nonnegative, nonincreasing values.

• Neither statement establishes a particular accuracy over the requested duration.

import numpy as np
import matplotlib.pyplot as plt


def euler_decay_trace(h, end_time=6.0):
    """Fixed steps for the dimensionless decay example; endpoint must align."""
    if not np.isfinite(h) or not np.isfinite(end_time) or h <= 0 or end_time <= 0:
        raise ValueError("Use a finite positive step and end time")
    n = int(round(end_time/h))
    if n <= 0 or not np.isclose(n*h, end_time, rtol=0, atol=1e-12):
        raise ValueError("Use a positive step that divides this example's interval")
    times = h*np.arange(n+1)
    values = np.empty(n+1)
    values[0] = 1.0
    for i in range(n):
        values[i+1] = values[i] + h*(-values[i])
    return times, values


fig, axes = plt.subplots(1, 2, figsize=(9, 4), layout="constrained")
fine_times = np.linspace(0, 6, 301)
for ax in axes:
    ax.plot(fine_times, np.exp(-fine_times), "k--", label="exact decay")
    ax.axhline(0, color="gray", lw=0.7)
    ax.set(xlabel="dimensionless time", ylabel="population / initial population")
for h, ax in ((0.25, axes[0]), (1.5, axes[1]), (3.0, axes[1])):
    ts, us = euler_decay_trace(h)
    ax.plot(ts, us, "o-", ms=4, label=f"Euler h={h}")
    print(f"h={h}: multiplier={1-h:+.2f}, minimum={us.min():+.4f}, "
          f"final absolute error={abs(us[-1]-np.exp(-6)):.3e}")
axes[0].set_title("Positive and decaying; still approximate")
axes[1].set_title("Alternating sign: stable or unstable?")
for ax in axes:
    ax.legend()
plt.show()

h=0.25: multiplier=+0.75, minimum=+0.0010, final absolute error=1.475e-03
h=1.5: multiplier=-0.50, minimum=-0.5000, final absolute error=6.002e-02
h=3.0: multiplier=-2.00, minimum=-2.0000, final absolute error=3.998e+00

The panels use different vertical ranges so both the resolved trajectory and the larger failures are readable. This is a comparison of mechanisms, not a visual claim that equal distances on the two panels mean equal errors.

Does an implicit step settle the question?¶

Backward Euler gives $u_{n+1}=u_n/(1+h)$. For every positive $h$, it preserves positive decay in this example. At $h=3$, however, its one-step factor is $1/4$, whereas the exact factor is $e^{-3}\approx0.0498$. A stable step can retain much more population than the model predicts. Whether that discrepancy is acceptable depends on the desired observable and accuracy, not the word “implicit.”

This example separates model behavior, numerical stability, and numerical accuracy. We need evidence for all three when they matter to the scientific claim.

Transfer: which conclusion survives the evidence?¶

Consider these reports without choosing a new method immediately:

Choose one row and design the smallest additional comparison that could contradict the report’s implied claim. State what each possible outcome would mean. For example, compare oscillator phase at a fixed physical time under refinement; do not merely extend an energy plot and call it a different test.

Looking Forward¶

The next scientific application should determine the diagnostic. Orbital motion needs phase and trajectory checks as well as conservation. A cooling or reacting system needs the correct dissipative behavior and resolved slow evolution. A Hamiltonian Monte Carlo proposal also needs the sampler’s own acceptance and reversibility arguments; an accurate ODE trajectory alone is not a sampling validation.

These are applications of the numerical ideas, not a revised assignment map. Project 2 remains quadrature; Project 3 introduces root finding and the Star class. The assignment briefs specify deliverables.

Before transferring a solver, record the equations, units, initial state, duration, observable, comparison, and acceptance criterion. Then ask which failure from this module could survive your planned checks.

Quick Reference Tables¶

Integration Methods Comparison¶

The energy and area entries below refer to the linear oscillator, except for the decay behavior explicitly identified for backward Euler. They are not universal properties of every system a method could integrate.

Timestep selection: restrictions and measurements¶

Stability regions summary¶

For the scalar test equation, the entries give the portion of each axis with $|R(z)|\le1$. Boundary points may fail to damp a decaying mode. The leapfrog row instead describes the two-state oscillator update; its endpoints are excluded because repeated eigenvalues can allow growth.

Diagnose before choosing a repair¶

Glossary¶

A-stable: A method whose scalar linear stability region contains the entire left half-plane. This removes that linear stability step restriction; it does not guarantee accuracy or rapid damping of stiff modes.

Adaptive timestep: Adjusting the step using an error estimate. Ordinary state-dependent adaptation generally does not preserve a fixed-step method’s symplectic structure.

Butcher tableau: Compact notation organizing the coefficients (aᵢⱼ, bᵢ, cᵢ) that define a Runge-Kutta method.

CFL condition: The numerical domain of dependence must contain the physical one. A common explicit bound is $h\le C_{\rm CFL}\Delta x/c_{\max}$; the coefficient depends on the space/time discretization and dimension, and the bound alone is not a sufficient stability proof.

Conjugate momenta: In Hamiltonian mechanics, the momenta p conjugate to generalized coordinates q, related by p = $\partial$L/$\partial$q̇.

Energy drift: Systematic change in total energy that should be conserved, indicating non-physical behavior of integrator.

Euler’s method: A first-order update using the initial slope. Its suitability depends on the equations, step length, duration, and required accuracy; its oscillator instability is one specific failure.

Global error: Total accumulated error over entire integration interval, typically O(h^p) for method of order p.

Hamiltonian: Function H(q,p,t) representing total energy of system. H = T + V for kinetic T and potential V.

Implicit method: Integration scheme where unknown appears on both sides, requiring an algebraic solution each step. Its stability properties depend on the particular method.

Initial Value Problem (IVP): Finding function x(t) given dx/dt = f(x,t) and x(t₀) = x₀. Fundamental problem of dynamics.

Leapfrog/Verlet: Symplectic integrator that staggers position and momentum updates to preserve symplectic structure in exact arithmetic for the separable Hamiltonian setting described here.

Liouville’s theorem: Phase space volume is preserved under Hamiltonian flow. Symplectic maps preserve canonical volume in exact arithmetic; floating-point implementations add rounding error.

Local truncation error: Error introduced in single step assuming perfect starting values, typically $O(h^{p+1})$ for order-$p$ method.

Modified Hamiltonian: A formal nearby Hamiltonian describing a symplectic map order by order. Suitable truncations explain long-time near-conservation under stated assumptions; the infinite series need not converge.

Order of accuracy: Power $p$ in global error scaling $O(h^p)$. Higher order doesn’t guarantee better long-term behavior.

Phase space: Space of all possible states $(q,p)$ of Hamiltonian system. For an autonomous Hamiltonian, trajectories remain on constant-energy surfaces.

Poisson bracket: $\{f,g\} = \partial f/\partial q \cdot \partial g/\partial p - \partial f/\partial p \cdot \partial g/\partial q$. Fundamental operation in Hamiltonian mechanics.

Predictor-corrector: Two-stage integration where initial estimate (predictor) is refined using additional information (corrector).

RK4 (Runge-Kutta 4): Classical fourth-order method using weighted average of four derivative evaluations whose update reduces to Simpson quadrature for a time-only right-hand side.

Stability function: $R(z)$ relating consecutive values $y_{n+1} = R(h\lambda)y_n$ for test equation $y' = \lambda y$.

Stability region: Set of complex values $z = h\lambda$ where $|R(z)| \leq 1$, ensuring bounded numerical solutions.

Stiff equation: An ODE for which an explicit method’s stability requirement forces much smaller steps than accuracy requires over the interval of interest. Multiple separated modes are one common mechanism, not a requirement.

Symplectic integrator: Numerical method preserving symplectic structure (including canonical phase-space volume, but not necessarily time reversibility) of Hamiltonian systems.

Symplectic structure: Geometric structure of phase space preserved by Hamiltonian flow, characterized by 2-form $\omega = dq \wedge dp$.

Time reversibility: Property where integrating forward then backward returns exactly to start. Holds for symmetric methods in exact arithmetic; symplecticity alone is insufficient.

Yoshida coefficients: Special timestep ratios $w_0$, $w_1$ involving $2^{1/3}$ that achieve fourth-order accuracy while maintaining symplecticity.

A useful numerical conclusion states the observable, the measured error or comparison, and the conditions under which the evidence applies. The methods in this module give you ways to build that conclusion and ways to challenge it.