An interactive tour of my dissertation, Variational Embedded Solitons, and Traveling Wavetrains Generated by Generalized Hopf Bifurcations, in some NLPDE Systems. Every simulation on this page runs the actual equations from the thesis, live in your browser.
Most waves spread out and fade. Drop a pebble in a pond and the ripples flatten as they go. But some nonlinear systems carry solitary waves, or solitons: compact pulses that travel for long distances without changing shape. The first half of the dissertation builds formulas for these pulses in two models of elastic materials. It includes a rare kind called the embedded soliton, which survives in conditions where it shouldn't.
The second half asks a different question. When does a system that sits perfectly still, like a laser, a population of animals, or a chemical reaction, suddenly start to oscillate? That tipping point is called a Hopf bifurcation. I studied its unusual, "degenerate" versions. There, the usual rulebook stops applying, and the system can pulse, wobble on a torus, turn chaotic, or blow up.
Guess the shape of a wave, then let the calculus of variations tune the guess. It works for ordinary solitons and for embedded ones.
A semiconductor laser and a predator–prey model with memory, both pushed to a highly degenerate tipping point.
Two oscillations are born at once, and they interfere into periodic and quasiperiodic waves of predators and prey moving across a landscape.
Chapter 3 studies a model of waves moving through a microstructured solid, a material with internal structure such as a composite, a granular medium, or a crystal lattice. Two effects compete in the model. Dispersion pulls a pulse apart, because short wiggles travel at different speeds than long ones. Nonlinearity steepens it, because the tall parts move differently from the short parts. When the two balance exactly, you get a solitary wave.
Below, the green pulse is an exact traveling solution of the full nonlinear equation. The orange curve starts with the same shape but obeys only the linear part of the equation. Watch it fall apart. Then change the speed: in this model, faster solitons are taller and narrower.
φ is the macroscopic deformation of the material. Looking for waves of the form φ(x − ct) and integrating twice gives the ODE φ″ = qφ + (r/2)φ², where q = (c² − b)/(δ(βc² − γ)) and r = −μ/(δ(βc² − γ)). Its exact soliton is φ = −(3q/r) sech²(√q·z/2). The demo uses b = δ = β = μ = 1 and γ = 1/2. The orange curve is evolved exactly in Fourier space using the linear dispersion relation ω² (1 + δβk²) = k²(b + δγk²).
Far from its peak, a solitary wave is tiny, so its tail follows the linearized equation. That gives a quick test. For the generalized Pochhammer–Chree equation of Chapter 2, which models longitudinal waves in an elastic rod, the tail behaves like eλz with λ = ±√(1 − a₁).
If the soliton moves faster than every ordinary ripple the rod can carry, λ is real and the tail decays cleanly. That is a regular soliton. If it moves within the band of ripple speeds, λ turns imaginary. Some ripple then travels at exactly the soliton's speed, the two resonate, and the pulse normally leaks energy into an oscillating tail. A soliton whose tail amplitude still comes out exactly zero is an embedded soliton: it is "embedded" in the continuous spectrum of linear waves. These are rare and delicate, and constructing them variationally was the novel part of Part II.
Speed of linear ripples vs. wavenumber k. Dashed line: soliton speed c.
What that means for the wave's tail.
These equations are not integrable, so there is usually no exact formula for the pulse. The Rayleigh–Ritz variational method works around that. Every solution of the wave equation makes a quantity called the action stationary. So you pick a family of trial shapes, here a Gaussian bump A·e−z²/ρ² with height A and width ρ, and compute the action for every member. Then you find where the action is flat. That point is the best Gaussian.
Drag the white dot around the action landscape to try different Gaussians. Watch the residual, which measures how badly the guess violates the real equation. Press Snap to jump to the stationary point, the saddle where the contours cross. Because the thesis solves for A and ρ in closed form, the residual can be computed exactly. That is unusual for variational solutions.
Action S(A, ρ). Drag the dot. ✕ = stationary point.
Microstructure (Ch. 3). The Lagrangian is L = −(q/2)φ² − (r/6)φ³ + ½φφ″. Substituting the Gaussian and integrating gives
Setting ∂S/∂A = ∂S/∂ρ = 0 gives the thesis result A = −6√6·q/(5r) ≈ −2.94 q/r and ρ² = 5/q. The exact peak is −3q/r, so the Gaussian gets the height within 2%.
Pochhammer–Chree (Ch. 2). The equation is utt − uttxx − (α₁u + α₂u² + α₃u³)xx = 0. For traveling waves it becomes (1 − a₁)u − a₂u² − a₃u³ − u″ = 0 with ai = αi/c². The Gaussian's height solves a quadratic, and its width follows in closed form. The "exact" curve here is integrated numerically outward from the peak.
Inside the resonant band, the natural objects are delocalized solitary waves: a core with an oscillating tail that never quite dies, A·e−z²/ρ² + α cos(κz). They form a continuous family. On isolated curves in parameter space, the tail amplitude α is exactly zero, and that member is the embedded soliton.
The standard Rayleigh–Ritz method can't handle a trial function that doesn't decay, because the action integral diverges. Following recent extensions of the method, the thesis takes three extra steps. It averages away the oscillating double- and triple-angle terms, in the spirit of Whitham's averaged Lagrangian. It picks the tail wavenumber κ so that the terms growing linearly in z cancel. Then it varies the action with respect to α too and imposes α = 0. For Pochhammer–Chree at a₁ = 2, this yields an analytic embedded soliton. Slide the tail amplitude to move through the family, and notice that only α = 0 is truly localized.
With α = 0 and a₁ = 2, the variational equations (2.4.18)–(2.4.20) are solved by
The width works out to ρ ≈ 2.43 for every a₂. The tail wavenumber from the integrability condition, κ = √(8a₁ − 8)/(2√2) = 1, agrees with the linear-spectrum prediction √(a₁ − 1) = 1. That is a nice consistency check. Chapter 3 does the same for the microstructure PDE with a sharper, quartic core e−z⁴/ρ⁴, because a Gaussian core gives an imaginary width there.
Hearts beat, airplane wings flutter, and predator and prey populations rise and fall, all without anything outside pushing them periodically. Mathematically, those self-sustained rhythms are limit cycles, and a common way to create one is a Hopf bifurcation. A steady state is stable until you turn a dial. Then a pair of its eigenvalues crosses the imaginary axis, the steady state goes unstable, and a small oscillation is born.
Below is the textbook picture. Each dot is a possible state of the system. Move μ past zero and watch the dots stop spiraling into the center and gather onto a ring instead.
Phase plane: 400 states flowing under the equation.
Eigenvalues of the steady state in the complex plane.
The chapters below study a harder case. In a generalized (degenerate) Hopf bifurcation, a third eigenvalue sits at exactly zero along with the imaginary pair, so all three are on the imaginary axis at once. Two things follow. The usual Hopf theorem no longer applies, so I built the post-bifurcation orbits by hand with the method of multiple scales. And the trace of the Jacobian is zero, so the system has no local volume contraction to fold trajectories back. That turns out to be the key to what happens next.
Some semiconductor lasers have two sections and produce self-sustained pulsations: they flicker at microwave frequencies with no external drive, which makes them useful as oscillators. In normalized form, the carrier densities x and y in the two sections and the photon density z obey three coupled equations. I tuned the material constants so the laser sits right at a generalized Hopf point, with eigenvalues ±0.808i and 0. Then I nudged the pump rate y₀.
Choose a nudge and watch the simulation. Just below the bifurcation the laser settles onto a stable quasiperiodic orbit, a robust and desirable oscillator mode. Nudge the other way and the carrier density runs away to infinity in finite time, which reproduces Figures 4.3–4.5.
Trajectory in (x, y, z) space, slowly rotating.
Here γ = 53, Γ = 143, x₀ = 0.98730, τb = −0.757502, and the bifurcation is at y₀ ≈ 0.0141110, where b₁ = b₃ = 0 in the characteristic polynomial λ³ + b₁λ² + b₂λ + b₃. Each run starts 10⁻⁶ away from the fixed point (0.98148, 0.018519, 0.0059283) and is integrated with RK4 (Δt = 0.01) for up to 10,000 time units. With a second parameter set (4.5.2), the thesis also finds bounded chaos. Its power spectrum is broad, and its cluster fractal dimension is D ∈ (1.27, 1.32).
Real populations react to the past: animals take time to mature, and food takes time to regrow. Chapter 5 models that with a distributed delay. The predator responds to a weighted memory Q of past prey numbers, where Q̇ = a(N − Q) and 1/a is how long the memory lasts. The prey birth rate F(N) = 1 − δN and predator death rate G(P) = γ(1 + κP²) are general functions. I studied the system at a generalized Hopf point in the memory parameter a.
The conclusion surprised me. Because the trace is zero at a generalized Hopf point, the system can't dissipate. The stretching-and-folding that keeps chaos bounded in ordinary Hopf scenarios is missing, so trajectories are far more likely to escape to infinity. Try it: a little on one side gives a stable equilibrium, just across gives oscillations, and a bit farther lets the predators explode (Figures 5.2–5.4).
Prey N(t) and remembered prey Q(t), deviation from equilibrium.
Trajectory in (N, P, Q) space.
ε̃ = 427.25, α = −0.237274, β ≈ −4096, δ = 29, γ = −17, κ = 2. The equilibrium is (N₀, P₀, Q₀) ≈ (0.0041792, 0.058907, 0.0041792), and the generalized Hopf point is at a₀ ≈ 0.114764, with eigenvalues ±0.2464i and 0. Physically realistic (small, positive) parameters never satisfy b₁ = b₃ = 0, which is itself one of the chapter's findings.
Back in 1952, Alan Turing showed that diffusion can make a uniform chemical soup grow spatial patterns. Chapter 6 looks for patterns that move: traveling waves N(x − vt), P(x − vt) in a predator–prey reaction–diffusion system with general nonlinearities. Substituting a traveling wave turns the PDE into a four-dimensional ODE in the moving coordinate ζ = x − vt. Its oscillations are spatial wavetrains.
That 4-D system can undergo a double Hopf bifurcation, where two pairs of eigenvalues hit the imaginary axis together, here at frequencies ω₁ ≈ 4.05 and ω₂ ≈ 2.03. Using multiple scales, I derived the normal form that governs the two amplitudes. It predicts a set of bifurcation lines L1–L5 in the plane of two control parameters, μ₁ (a shift in D₂) and μ₂ (a shift in β). Each region has its own behavior: a uniform state, periodic wavetrains, quasiperiodic wavetrains on a 2-torus, or escape.
Click anywhere in the parameter plane or pick one of the thesis's sample points. The full four-dimensional system then runs live, and you'll see the resulting population wave sweep across the landscape.
Parameter plane (μ₁, μ₂) with bifurcation lines from the normal form. Click to simulate.
The base point is β = 20.8, α = 2.02454, d = 19.8, θ = ρ = −20.5328, v = 1, c = 13.8, z = 2.8, k₀ = −9.2, D₁ = 1, D₂ = −1. Here the four eigenvalues are ±4.054i and ±2.032i. The deviations enter as D₂ = −1 + μ₁ and β = 20.8 + μ₂. The amplitude equations are
The coefficients are α₁₁ = 0.66786, α₁₂ = 0.08732, α₂₁ = −0.16786, α₂₂ = −0.08732, a₂₀ = −27.525, a₀₂ = 317.33, b₂₀ = 55.050, and b₀₂ = −158.66. Their fixed points and stability give the lines drawn in the plane. Systems A and B, the two simpler nonlinearities, provably have no double Hopf points.
The full text is open access through UCF's STARS repository.
With thanks to Dr. Roy S. Choudhury, and to Dr. Schober and Dr. Rollins for their generous time. The interactive figures on this page were rebuilt in 2026 from the equations and parameter values in the thesis.