Hopf Bifurcation and Birth of Limit Cycles

Below about forty-seven, the water flowing past a cylinder slides by in two steady, mirror-symmetric recirculation eddies — a fixed point of the Navier–Stokes dynamics, boring and eternal. Nudge the Reynolds number past $Re_c \approx 47$ and something switches on: the wake begins to shed vortices left-right-left-right at a definite frequency, the von Kármán street that makes flags flutter and power lines gallop and Tacoma Narrows enter every engineering syllabus. Nobody drives that rhythm. The steady state itself has died, and a self-sustained oscillation has been born in its place. The Hopf bifurcation is the generic mechanism behind every such onset. The bifurcations catalogued in the sibling articles — saddle-node, transcritical, pitchfork — all move fixed points around: a single real eigenvalue crosses zero and equilibria appear, collide, or trade stability. The Hopf bifurcation is the one fundamentally different move available to a flow: a complex-conjugate eigenvalue pair $\alpha \pm i\omega$ marches across the imaginary axis with $\omega \neq 0$ , and the destabilized spiral does not fly apart — it wraps itself around a newborn limit cycle. It is the standard doorway from equilibrium to rhythm, and the first step on many roads to chaos. But the linearization only announces the crossing; it cannot say what is born. The same eigenvalue data can produce a soft, reversible onset — a stable cycle growing from zero amplitude like $\sqrt{\mu-\mu_c}$ — or a hard, hysteretic jump to a distant attractor that will not release the system until the parameter is dragged far back down. That distinction, supercritical versus subcritical, is decided by a single cubic coefficient the eigenvalues know nothing about, and for a wing developing flutter it is the difference between a design note and a crash report.

Key takeaways

A Hopf bifurcation is the birth (or death) of a small-amplitude limit cycle as a fixed point loses stability through a complex-conjugate pair of eigenvalues crossing the imaginary axis.
The defining conditions at $\mu_c$ : eigenvalues $\lambda_{\pm}(\mu) = \alpha(\mu) \pm i\omega(\mu)$ with $\alpha(\mu_c) = 0$ , $\omega(\mu_c) \neq 0$ , and transversality $\alpha'(\mu_c) \neq 0$ — a genuine crossing at nonzero frequency and nonzero speed.
Near the crossing the dynamics collapse onto a 2D center manifold where the flow is the normal form $\dot r = \alpha r + a r^3$ , $\dot\theta = \omega + b r^2$ — a pitchfork in the amplitude, dressed with rotation.
Supercritical ( $a < 0$ ): a stable cycle of radius $r_* = \sqrt{-\alpha/a} \propto \sqrt{\mu - \mu_c}$ switches on continuously — soft, reversible onset at frequency $\omega(\mu_c)$ .
Subcritical ( $a > 0$ ): an unstable cycle shrinks onto the fixed point and destroys it — hard onset, finite jump, bistability, and hysteresis bracketed by a fold of cycles.
The sign of $a$ — the first Lyapunov coefficient $\ell_1$ — is invisible to the linearization; computing it is the price of knowing whether the onset is gentle or catastrophic.
Key idea: a Hopf bifurcation is detected purely from the linearization — a complex eigenvalue pair crosses the imaginary axis transversally — but its character (soft stable cycle vs hard hysteretic jump) is a nonlinear question answered only by the cubic normal-form coefficient.

The core insight

A Hopf bifurcation is detected purely from the linearization — a complex eigenvalue pair crosses $i\mathbb{R}$ transversally — but its character (soft stable cycle vs hard hysteretic jump) is a nonlinear question answered only by the cubic normal-form coefficient.

Where this lives today

The super/subcritical distinction is now a load-bearing wall of engineering certification. Aeroelastic flutter — the Hopf bifurcation of a wing coupled to airflow — is certified by demonstrating margin between the flight envelope and the flutter boundary, and modern transonic analysis distinguishes benign supercritical limit-cycle oscillation from the subcritical case whose hysteresis makes it lethal: a gust can kick the wing across the unstable-cycle fence inside the bistable window, at flight speeds below the linear flutter boundary. The same normal-form calculus runs through power-grid engineering, where inter-area oscillations emerge by Hopf bifurcation as loading increases and the first Lyapunov coefficient decides whether the grid rings gently or jumps; through neuroscience, where the transition to epileptic seizure is modeled as a subcritical Hopf with its telltale hysteresis and where critical slowing down ( $\tau = 1/2\alpha$ diverging at onset) is mined from EEG data as an early-warning signal; and through MEMS design, where clock resonators are deliberately parked on stable limit cycles born supercritically so that amplitude is set by the device, not by noise. The theory has also kept its role as the first chapter of the chaos story. In the Ruelle–Takens picture and its experimental confirmations, a steady flow undergoes a Hopf bifurcation to a periodic state, then a secondary Hopf to a two-frequency torus, and then — rather than Landau's endless ladder — breaks into chaos. And once the Hopf-born cycle exists, it becomes the protagonist of the next article: a limit cycle in a driven or three-dimensional system can lose stability by period doubling, and that cascade has a universal ending. The Hopf bifurcation opens the door; the period-doubling cascade walks through it.

The physical picture

Take any system that spirals into rest — a damped pendulum, an RLC circuit, a swaying bridge deck — and imagine a knob that slowly turns the net damping down through zero. While damping is positive, every disturbance winds inward and dies: the fixed point is a stable focus, eigenvalues $\alpha \pm i\omega$ with $\alpha < 0$ . At zero net damping the spiral neither grows nor shrinks. Push past zero and the spiral unwinds outward — but outward to where? Exponential growth is a statement about the linearization, and the linearization is only the truth near the origin. A few spiral turns out, the nonlinear terms wake up. If they oppose the growth, the trajectory settles onto a closed loop of definite amplitude: a limit cycle has been born, and the system has crossed a supercritical Hopf bifurcation. Oscillation has switched on the way dusk switches on — continuously.

Figure

Two Hopf bifurcations — the supercritical square-root branch and the subcritical fold with its bistable window

The diagram above is the entire taxonomy. On the left, the supercritical case: below $\mu_c$ a stable focus, above it an unstable focus wearing a stable cycle whose radius grows as $\sqrt{\mu-\mu_c}$ — infinite slope at onset, then leveling off. That square-root law is the fingerprint of a soft Hopf onset; vortex-shedding amplitude behind a cylinder obeys it so cleanly that hydrodynamicists use it to measure $Re_c$ by extrapolation. On the right, the subcritical case: the cubic term now feeds the instability instead of taming it, so the branch of cycles bends backwards over the still-stable focus. Those backward cycles are unstable — they form the basin boundary guarding the fixed point, a fence that tightens as $\mu \to \mu_c^-$ and vanishes exactly at the crossing. Past $\mu_c$ there is nothing local to catch the trajectory, and the system leaps to whatever remote attractor exists — here, the large-amplitude branch born at a fold.

Figure

The birth of the cycle in the phase plane — destabilized focus wrapped by the newborn attracting loop

In the phase plane the supercritical story looks like the pair above: a focus that dies and delegates its stability to the loop around it. Watching only the fixed point, you would call this catastrophic destabilization; watching the whole plane, almost nothing happened — trajectories that used to end at a point now end on a small circle traversed at frequency $\omega(\mu_c)$ . That gentleness is the physical meaning of "supercritical": the new state is infinitesimally close to the old one at onset. The deepest way to see the mechanism is to notice that the amplitude has its own one-dimensional life. In the rotating frame of the oscillation, the radius obeys $\dot r = \alpha r + a r^3$ — exactly the pitchfork normal form of the sibling article Saddle-Node, Transcritical, and Pitchfork Bifurcations, restricted to $r \geq 0$ . A Hopf bifurcation is a pitchfork of the oscillation amplitude, dressed with steady rotation $\dot\theta = \omega$ . Everything you know about pitchforks transfers: the square-root branch, the sub/supercritical dichotomy from the cubic sign, the hysteresis when the subcritical branch folds back. What is new is only that the bifurcating "state" rotates — so the pitchfork's new equilibria are not equilibria at all, but a periodic orbit.

The eigenvalue crossing — what the linearization sees

Fix a planar (or center-manifold-reduced) system $\dot{\mathbf{x}} = \mathbf{f}(\mathbf{x}, \mu)$ with a fixed point that we hold at the origin as $\mu$ varies. Its Jacobian eigenvalues near the bifurcation form a complex-conjugate pair, and everything the linear theory can say is packed into how that pair moves:

\lambda_{\pm}(\mu) = \alpha(\mu) \pm i\,\omega(\mu) \qquad (1)

where: $\mu$ — the control parameter (Reynolds number, drive gain, wind speed); $\alpha(\mu)$ — the growth rate, negative for a stable focus; $\omega(\mu)$ — the rotation rate of the local spiral. For the trace-determinant classification behind these words, the sibling article Classification of Fixed Points is the reference; here the pair sits just off the real axis, in spiral territory. Three conditions make the crossing a genuine Hopf bifurcation:

\alpha(\mu_c) = 0, \qquad \omega(\mu_c) \neq 0, \qquad \left.\frac{\mathrm{d}\alpha}{\mathrm{d}\mu}\right|_{\mu_c} \neq 0 \qquad (2)

where: the first condition is the crossing itself; the second demands that the pair crosses away from the real axis — at zero frequency the pair would collide at the origin of the $\lambda$ -plane and the bifurcation would be the very different Bogdanov–Takens double-zero; the third (transversality) requires the crossing to happen at nonzero speed in $\mu$ , so the eigenvalues genuinely pass through the axis rather than kissing it and retreating. Under these conditions the Hopf theorem guarantees a one-parameter family of periodic orbits branching from $(\mathbf{0}, \mu_c)$ , with amplitude $\propto \sqrt{|\mu - \mu_c|}$ and period $\to 2\pi/\omega(\mu_c)$ . Note what is not in the list: nothing about nonlinear terms. Equations (1)–(2) certify that a cycle is born; they are structurally incapable of saying on which side of $\mu_c$ it lives or whether it attracts. That information enters through the normal form.

The normal form — what the nonlinearity decides

Near the crossing, all the essential dynamics happen on a two-dimensional center manifold (the sibling article Bifurcations and the Loss of Stability walks a center-manifold reduction in full). On that plane, combine the two real coordinates into a complex one, $z = x + iy$ , so the linear part becomes a single complex multiplier. A sequence of near-identity polynomial changes of variables then strips away every nonlinear term that can be stripped, leaving the Poincaré normal form of the Hopf bifurcation:

\dot z = (\alpha + i\omega)\,z + (a + ib)\,|z|^2 z + \mathcal{O}(|z|^5) \qquad (3)

where: $z$ — the complex amplitude of the critical mode; $a$ — the real cubic coefficient whose sign decides everything qualitative; $b$ — the cubic frequency shift, which makes the oscillation period depend (weakly) on amplitude. The quadratic terms are gone not by assumption but by theorem — Derivation A shows why $|z|^2 z$ is the unique cubic survivor. Writing $z = r e^{i\theta}$ splits (3) into an amplitude equation that knows nothing of the phase and a phase equation slaved to the amplitude:

\dot r = \alpha\, r + a\, r^3, \qquad \dot\theta = \omega + b\, r^2 \qquad (4)

where: $r = |z|$ — the oscillation amplitude; $\theta = \arg z$ — the phase, advancing at rate $\omega$ plus a small amplitude-dependent detuning $b r^2$ . This decoupling is the technical heart of Hopf theory: a two-dimensional bifurcation problem has collapsed to a one-dimensional flow in $r$ , and one-dimensional flows we understand completely. Setting $\dot r = 0$ at nonzero radius gives the newborn cycle:

r_* = \sqrt{-\alpha/a}, \qquad \text{real precisely when } \alpha \text{ and } a \text{ have opposite signs} \qquad (5)

where: $r_*$ — the limit-cycle radius in normal-form coordinates. For the supercritical sign choice $a < 0$ , the cycle exists on the unstable side $\alpha > 0$ and is stable; for $a > 0$ it exists on the stable side $\alpha < 0$ and is unstable. Since transversality gives $\alpha \approx \alpha'(\mu_c)(\mu - \mu_c)$ near the crossing, equation (5) is the square-root law in its usable form:

r_* \propto \sqrt{|\mu - \mu_c|} \qquad (6)

where: the proportionality constant is $\sqrt{\alpha'(\mu_c)/|a|}$ . An experimentalist plots measured oscillation amplitude squared against the control parameter and gets a straight line through $\mu_c$ — the standard way onset thresholds are extracted from vortex-shedding, flutter, and laser data. For a concrete system $\dot x = f(x,y)$ , $\dot y = g(x,y)$ brought to coordinates where the Jacobian at criticality is the pure rotation $\begin{pmatrix} 0 & -\omega \\ \omega & 0 \end{pmatrix}$ , the cubic coefficient has a closed — if inelegant — expression, the first Lyapunov coefficient:

\ell_1 = \frac{1}{16}\big(f_{xxx} + f_{xyy} + g_{xxy} + g_{yyy}\big) + \frac{1}{16\omega}\big[f_{xy}(f_{xx} + f_{yy}) - g_{xy}(g_{xx} + g_{yy}) - f_{xx}g_{xx} + f_{yy}g_{yy}\big] \qquad (7)

where: subscripts denote partial derivatives evaluated at the fixed point at $\mu_c$ ; $\ell_1 < 0$ means supercritical, $\ell_1 > 0$ subcritical. The formula looks unmotivated but is nothing more than the averaging of Derivation A carried out on a general cubic — the first line collects the terms that survive averaging directly, the second the quadratic terms that feed into cubics at second order. (Despite the name, $\ell_1$ has nothing to do with Lyapunov exponents — the FAQ untangles the collision.) The canonical example with everything explicit, used throughout this article, is

\dot x = \mu x - y - x(x^2 + y^2), \qquad \dot y = x + \mu y - y(x^2 + y^2) \qquad (8)

where: $\mu$ — the control parameter with $\mu_c = 0$ ; the system is engineered so that the polar reduction is exact, with $\alpha = \mu$ , $\omega = 1$ , $a = -1$ , $b = 0$ — a supercritical Hopf you can solve by hand (Derivation B).

Which equation answers which question

you want	use	notes
to detect the bifurcation	conditions (2) on the eigenvalues (1)	pure linear algebra at the fixed point
the newborn cycle's radius and stability	normal form (4), radius (5)	needs the cubic coefficient
soft or hard onset	sign of $\ell_1$ , eq. (7)	$\ell_1 < 0$ soft, $\ell_1 > 0$ hard
amplitude vs parameter fit	square-root law (6)	plot $r^2$ vs $\mu$ : straight line hits zero at $\mu_c$
the oscillation frequency at onset	$\omega(\mu_c)$ from (1)	the cycle inherits the spiral's rotation rate
hysteresis window of a subcritical onset	fold of cycles, Derivation C	needs the quintic term

Derivation AHopf normal form and the birth of the cycle

We are deriving the polar normal form (4) from a generic planar system whose fixed point carries eigenvalues $\alpha \pm i\omega$ with small $\alpha$ — and from it, the limit cycle (5) with its stability. The strategy is the classic three-step of local bifurcation theory: complexify so the linear part becomes one multiplier, remove every removable nonlinear term with near-identity changes of variables, and read the surviving resonant term in polar coordinates. The step to watch is the removal argument — why the quadratic terms can always be transformed away while one particular cubic cannot. That single surviving term is where all the physics of soft versus hard onset hides, so it pays to see exactly why it is untouchable.

Step 1. Write the system as linear part plus nonlinearity, $\dot{\mathbf{x}} = J\mathbf{x} + \mathbf{N}(\mathbf{x})$ , with the Jacobian brought to the rotation-plus-growth block

J = \begin{pmatrix} \alpha & -\omega \\ \omega & \alpha \end{pmatrix}

Why this block: any real $2\times 2$ matrix with complex eigenvalues $\alpha \pm i\omega$ is similar to it; the change of basis costs nothing and makes the rotation explicit.

Step 2. Complexify with $z = x + iy$ . The linear part collapses to $\dot z = (\alpha + i\omega) z$ , and the nonlinearity becomes a polynomial in $z$ and $\bar z$ : terms $z^2$ , $z\bar z$ , $\bar z^2$ at quadratic order, $z^3, z^2\bar z, z\bar z^2, \bar z^3$ at cubic order. Why complexify: the linear flow is now multiplication by $e^{(\alpha + i\omega)t}$ — pure growth-with-rotation — so each nonlinear monomial acquires a definite rotation rate against which resonance can be judged.

Step 3. Attempt to remove a monomial $c\, z^p \bar z^q$ by the near-identity substitution $z = w + h\, w^p \bar w^q$ . Matching orders, the removal works when $h$ can be chosen finite, and the obstruction is the denominator $\lambda - (p - q)\, i\omega$ evaluated at criticality $\lambda = i\omega$ : removal fails precisely when $(p - q) = 1$ . At quadratic order, $p - q \in \{2, 0, -2\}$ — all removable. At cubic order, $z^2\bar z = |z|^2 z$ has $p - q = 1$ — resonant, unremovable. Why the resonance condition is physical: a term $z^p\bar z^q$ rotates with phase $e^{i(p-q)\omega t}$ , while $z$ itself rotates as $e^{i\omega t}$ . Only a term rotating exactly in phase with $z$ can push the amplitude coherently, orbit after orbit; every other term oscillates against the reference rotation and averages to zero. The removal calculation is the rigorous version of that averaging.

Step 4. Keep the resonant survivor with complex coefficient $a + ib$ :

\dot z = (\alpha + i\omega)\, z + (a + ib)\,|z|^2 z + \mathcal{O}(|z|^5)

Substitute $z = r e^{i\theta}$ , so $\dot z = (\dot r + i r \dot\theta)e^{i\theta}$ and $|z|^2 z = r^3 e^{i\theta}$ ; cancel $e^{i\theta}$ and separate real and imaginary parts:

\dot r = \alpha r + a r^3, \qquad \dot\theta = \omega + b r^2

Why the split matters: the amplitude equation is autonomous — a one-dimensional flow — while the phase merely accumulates. The entire two-dimensional bifurcation is now the 1D pitchfork problem in $r$ .

Step 5. Find the nonzero fixed amplitude: $0 = \alpha r_* + a r_*^3$ gives $r_*^2 = -\alpha/a$ , which exists when $\alpha$ and $a$ have opposite signs. Its stability follows from the 1D derivative:

\frac{\mathrm{d}}{\mathrm{d}r}\big(\alpha r + a r^3\big)\Big|_{r_*} = \alpha + 3 a r_*^2 = \alpha - 3\alpha = -2\alpha

Why this settles both cases at once: supercritical means $a < 0$ , cycle exists for $\alpha > 0$ , and $-2\alpha < 0$ — attracting. Subcritical means $a > 0$ , cycle exists for $\alpha < 0$ , and $-2\alpha > 0$ — repelling. The cycle's stability is always opposite to the focus it encircles. *Verified: reproduces the standard supercritical normal form and the $-2\alpha$ Floquet rate; matches Strogatz §8.2 and Guckenheimer–Holmes §3.4.*

Go deeper · Derivation B · 2 minThe exactly solvable supercritical Hopf

Derivation BThe exactly solvable supercritical Hopf

The goal is to watch every abstract claim of Derivation A come true in a system with no hidden approximations: the canonical example (8), $\dot x = \mu x - y - x(x^2+y^2)$ , $\dot y = x + \mu y - y(x^2+y^2)$ . The strategy is to convert to polar coordinates directly — no center manifold, no averaging — because the system is engineered so the angular dynamics decouple exactly. The point to watch is the cancellation of the cross terms in $x\dot x + y\dot y$ : it is exact here, which is precisely what makes this the textbook benchmark against which numerical Hopf computations are tested.

Step 1. Compute $r\dot r = x\dot x + y\dot y$ . The rotation terms cancel in the sum ( $-xy + yx = 0$ ), leaving

r\dot r = \mu(x^2 + y^2) - (x^2 + y^2)^2 = \mu r^2 - r^4

Why the cancellation: the rotational part of the vector field is perpendicular to the radial direction, so it cannot change $r$ — only the $\mu$ -growth and the cubic saturation act radially.

Step 2. Divide by $r$ to get the exact amplitude equation, and compute the phase from $\dot\theta = (x\dot y - y\dot x)/r^2$ :

\dot r = \mu r - r^3, \qquad \dot\theta = 1

Why this is literally the normal form: comparing with (4), $\alpha = \mu$ , $a = -1$ , $\omega = 1$ , $b = 0$ , with no error term at all — the $\mathcal{O}(|z|^5)$ of the general theory is exactly zero here.

Step 3. Below threshold, $\mu < 0$ : then $\dot r = r(\mu - r^2) < 0$ for every $r > 0$ , so every trajectory spirals into the origin — a globally attracting stable focus, no cycle anywhere.

Step 4. Above threshold, $\mu > 0$ : the amplitude flow has roots $r = 0$ (now repelling, since $\dot r \approx \mu r > 0$ for small $r$ ) and $r_* = \sqrt{\mu}$ . The circle $r = \sqrt{\mu}$ , traversed at $\dot\theta = 1$ with period $2\pi$ , is a limit cycle obeying the square-root law with $\mu_c = 0$ .

Step 5. Check its stability by the 1D derivative: $\frac{\mathrm{d}}{\mathrm{d}r}(\mu r - r^3)\big|_{\sqrt\mu} = \mu - 3\mu = -2\mu < 0$ — attracting, at exactly the $-2\alpha$ rate predicted in general. Why one can even write the transient in closed form: $\dot r = \mu r - r^3$ is a Bernoulli equation; substituting $u = r^{-2}$ makes it linear, giving $r^2(t) = \mu r_0^2 e^{2\mu t} / \big(\mu + r_0^2(e^{2\mu t} - 1)\big)$ — every trajectory relaxes onto $r_* = \sqrt{\mu}$ on the timescale $1/2\mu$ , which diverges at onset. Critical slowing down is visible in the exact solution. *Verified: direct substitution reproduces the polar pair; the Bernoulli solution satisfies $r(0) = r_0$ and $r \to \sqrt{\mu}$ ; matches Strogatz §8.2's benchmark example.*

Go deeper · Derivation C · 2 minThe subcritical case — fold of cycles, bistability, and hysteresis

Derivation CThe subcritical case — fold of cycles, bistability, and hysteresis

Now the dangerous sign. We take the amplitude equation with a destabilizing cubic, $\dot r = \mu r + r^3 - r^5$ , and derive the full bifurcation set: where cycles exist, where they collide, and why the observable consequence is a hysteresis loop with two different switching thresholds. The quintic term is not decoration — with $a = +1 > 0$ the cubic feeds the growth, and something at higher order must eventually saturate it or the model predicts escape to infinity. The step to watch is the fold: two cycles are born together out of nothing at $\mu_{sn} = -1/4$ , a saddle-node of periodic orbits, before the fixed point itself has any idea a bifurcation is coming.

Step 1. Find the nonzero steady amplitudes: $0 = \mu r + r^3 - r^5$ with $r \neq 0$ gives the quadratic in $r^2$

r^4 - r^2 - \mu = 0 \quad \Rightarrow \quad r_*^2 = \tfrac{1}{2}\big(1 \pm \sqrt{1 + 4\mu}\big)

Why a quadratic: amplitude equations in $r$ odd to quintic order always reduce to quadratics in $r^2$ — the fold structure below is generic, not special to this model.

Step 2. Read off existence. Real roots require $\mu \geq -1/4$ . For $-1/4 < \mu < 0$ both roots are positive: two coexisting cycles, the larger (upper sign) stable and the smaller (lower sign) unstable — check by the sign of $\mathrm{d}\dot r/\mathrm{d}r$ , which alternates along the sequence of roots $0, r_-, r_+$ . For $\mu > 0$ only the large root survives.

Step 3. Identify the fold. At $\mu_{sn} = -1/4$ the two cycles merge at $r^2 = 1/2$ and annihilate: a saddle-node (fold) of limit cycles. For $\mu < -1/4$ the only attractor is the origin. Why the fold matters physically: sweeping $\mu$ upward from below, a large-amplitude oscillation comes into existence at $\mu = -1/4$ — but the system, sitting happily at the stable origin, does not know it. The oscillation is waiting, invisible, behind the unstable cycle.

Step 4. Cross the Hopf point. In the window $-1/4 < \mu < 0$ the origin is still linearly stable ( $\dot r \approx \mu r$ with $\mu < 0$ ), but its basin is only the disk $r < r_-$ bounded by the unstable cycle — and $r_- \to 0$ as $\mu \to 0^-$ . At $\mu = 0$ the unstable cycle has shrunk onto the origin and hands it its instability. For $\mu > 0$ any disturbance, however small, grows and the trajectory runs outward until the quintic catches it at the large cycle: the amplitude jumps discontinuously from $0$ to $r_+ = \sqrt{(1 + \sqrt{1+4\mu})/2} \approx 1$ at onset.

Step 5. Sweep back down. On the large cycle, decreasing $\mu$ through $0$ changes nothing — the large branch exists and is stable down to $\mu = -1/4$ . Only at the fold does the oscillation die, collapsing to the origin from finite amplitude $r = \sqrt{1/2} \approx 0.71$ . The switch-on threshold ( $\mu = 0$ ) and the switch-off threshold ( $\mu = -1/4$ ) differ: hysteresis, with a bistable window in between where the observed state depends on history. Verified: the roots and fold location follow from the quadratic in $r^2$ ; the stability alternation matches the 1D flow criterion; reproduces the standard subcritical Hopf model of Strogatz §8.2.

Figure

The hysteresis loop of the subcritical Hopf — different switch-on and switch-off thresholds bracket the bistable window

Limiting cases

The normal form $\dot z = (\alpha + i\omega) z + (a + ib)|z|^2 z$ organizes every regime:

Case	Result	Meaning
supercritical, $a < 0$	stable cycle $r_* = \sqrt{-\alpha/a}$ for $\alpha > 0$	soft onset; amplitude $\propto \sqrt{\mu-\mu_c}$ , reversible sweep
subcritical, $a > 0$	unstable cycle for $\alpha < 0$ ; jump at $\mu_c$	hard onset; hysteresis window down to the fold of cycles
degenerate, $a = 0$	Bautin (generalized Hopf) point	quintic decides; the boundary where super turns sub in a 2-parameter family
$\omega(\mu_c) \to 0$	not Hopf — Bogdanov–Takens double zero	the pair collides on the real axis; a different, richer bifurcation
$\alpha'(\mu_c) = 0$	tangential touch, no crossing	eigenvalues graze the axis; cycles need not branch
far past onset, $r$ large	normal form invalid	truncation at cubic/quintic order holds only near $(\mathbf{0}, \mu_c)$
frequency at onset	$\omega_{\text{cycle}} \to \omega(\mu_c) \neq 0$	the newborn cycle oscillates at the spiral's frequency, not slowly

Common trapA linearly unstable focus does not mean the oscillation you observe is the newborn Hopf cycle

The eigenvalue crossing is necessary for a Hopf bifurcation but certifies almost nothing about what you will measure. Three separate traps live here. First, the crossing must have $\omega(\mu_c) \neq 0$ and $\alpha'(\mu_c) \neq 0$ — a pair that collides at the origin of the eigenvalue plane is a Bogdanov–Takens point, and a pair that touches the axis without crossing spawns no cycle. Second, the stability of the emerging cycle is invisible to the linearization: in the subcritical case the branching cycle is unstable and unobservable, and the oscillation you actually see after onset is a distant attractor with finite amplitude — quoting the $\sqrt{\mu - \mu_c}$ law for it is simply wrong. Third, even a certified supercritical Hopf gives the square-root law only locally: fit $r^2 \propto \mu - \mu_c$ near onset, not across the whole parameter range, where higher-order terms bend the branch.

Worked examples

1 · Reading off a supercritical onset (confidence builder)

A feedback-controlled beam experiment is modeled near onset by the exact system (8) with $\alpha = \mu$ , $a = -1$ , $\omega = 1$ . The gain is set to $\mu = 0.04$ . Find the oscillation amplitude, its period, the relaxation time back onto the cycle after a small kick, and the amplitude if the gain is quadrupled.

Step 1. Amplitude from equation (5): $r_* = \sqrt{-\alpha/a} = \sqrt{0.04/1} = \boxed{0.20}$ in normal-form units.

Step 2. Period: the phase advances at $\dot\theta = \omega = 1$ , so $T = 2\pi/\omega = 2\pi \approx \boxed{6.28}$ time units — set by the imaginary part of the eigenvalues, essentially unchanged from the ringing frequency of the stable side.

Step 3. Relaxation onto the cycle: Derivation A gives the amplitude decay rate $-2\alpha = -0.08$ , so the time constant is $\tau = 1/(2\mu) = 1/0.08 = \boxed{12.5}$ time units — about two full oscillations. Note $\tau \to \infty$ as $\mu \to 0^+$ : critical slowing down, the universal warning sign of a nearby bifurcation.

Step 4. Quadruple the gain, $\mu = 0.16$ : $r_* = \sqrt{0.16} = 0.40$ — amplitude only doubles. The square-root law means early amplitude growth is steep near onset and stingy afterwards; to double the amplitude you must quadruple the distance past threshold.

2 · A subcritical onset with numbers — the hysteresis loop quantified (exam level)

A flutter model reduces to the amplitude equation $\dot r = \mu r + r^3 - r^5$ . (i) Locate the fold of cycles. (ii) At $\mu = -0.16$ , find all coexisting states and classify their stability. (iii) Compute the amplitude jump when $\mu$ crosses $0$ from below, and the amplitude from which the oscillation collapses when $\mu$ is swept back down. (iv) State the hysteresis window.

Step 1. Steady amplitudes solve $r_*^2 = \tfrac12(1 \pm \sqrt{1 + 4\mu})$ (Derivation C). The fold sits where the discriminant vanishes: $\mu_{sn} = -\tfrac14 = \boxed{-0.25}$ , with $r_{sn} = \sqrt{1/2} = 0.707$ .

Step 2. At $\mu = -0.16$ : $\sqrt{1 + 4(-0.16)} = \sqrt{0.36} = 0.60$ , so $r_*^2 = \tfrac12(1 \pm 0.6)$ gives $r_+^2 = 0.80$ and $r_-^2 = 0.20$ . Three states coexist: the origin (stable, since $\mu < 0$ ), the small cycle $r_- = \sqrt{0.20} = \boxed{0.447}$ (unstable — the basin fence), and the large cycle $r_+ = \sqrt{0.80} = \boxed{0.894}$ (stable). The system is bistable; which state you find depends on history.

Step 3. Sweep up through $\mu = 0$ : the origin destabilizes and the trajectory jumps to the large branch at $r_+ = \sqrt{\tfrac12(1 + \sqrt{1})} = \sqrt{1} = \boxed{1.0}$ — a finite-amplitude leap from rest, with no warning in the small-signal linear response. Sweep back down: the oscillation persists (large branch stable for $\mu > -0.25$ ) until the fold, where it collapses from amplitude $r = 0.707$ to rest.

Step 4. Hysteresis window: $\boxed{-0.25 < \mu < 0}$ — a quarter-unit of parameter in which "is the system oscillating?" cannot be answered without knowing where it came from. For an aircraft wing this window is why flutter testing creeps up on the boundary: crossing it once may mean not coming back by easing off.

3 · The square-root law as a measuring instrument (applied)

Vortex shedding behind a cylinder switches on at $Re_c \approx 47$ via a supercritical Hopf bifurcation. If the shedding velocity amplitude is $A_1$ at $Re = 48$ , what is it at $Re = 51$ ? What does a plot of $A^2$ vs $Re$ give you?

Step 1. The square-root law (6): $A \propto \sqrt{Re - Re_c}$ , so $A_1/A_2 = \sqrt{(48-47)/(51-47)} = \sqrt{1/4} = \boxed{1/2}$ — the amplitude at $Re = 51$ is twice that at $Re = 48$ .

Step 2. $A^2$ vs $Re$ is a straight line intercepting zero at $Re_c$ — which is how $Re_c \approx 47$ is measured in practice, by extrapolating the Landau plot from comfortably-above-onset data rather than hunting the infinitesimal onset itself. The linear slope is the ratio $\alpha'(Re_c)/|a|$ of the two normal-form ingredients.

Origins · 1 minEberhard Hopf and how this was discovered

Eberhard Hopf

1902–1983 · Berlin · Leipzig · MIT · Indiana

Hopf proved the theorem that turned the onset of oscillation from an observation into a certified mechanism. Trained in Berlin among the ergodic theorists, he spent 1931–1936 at Harvard and MIT — where his work on radiative transfer produced the Wiener–Hopf factorization method, a tool now everywhere from crack mechanics to finance — before taking a chair in Leipzig. There, in 1942, he published the bifurcation theorem: under the crossing conditions (2), a family of periodic orbits of amplitude $\propto \sqrt{|\mu - \mu_c|}$ branches from the equilibrium. The result had roots he acknowledged — Poincaré had the normal-form machinery, and Aleksandr Andronov's Gorky school had described the "birth of self-oscillations" for planar systems in the 1930s, which is why careful authors say Poincaré–Andronov–Hopf. Hopf's contribution was the clean general statement and its extension beyond two dimensions. His decision to remain in Germany through the war left him professionally isolated for years; the theorem itself was carried into the English-speaking mainstream only in the 1970s, when translations and the Marsden–McCracken volume made "Hopf bifurcation" the standard name — attached, with history's usual carelessness, to the one member of the trio who insisted he had merely completed it.

History

1892

— Poincaré develops normal-form reductions and Lyapunov publishes the stability memoir; the tools of the crossing are in place decades before the theorem.

1937

— Andronov and his Gorky school classify the birth of limit cycles from a focus in planar systems — the bifurcation, in dimension two, with applications to radio self-oscillators.

1942

— Hopf proves the general branching theorem in

n

dimensions in wartime Leipzig, square-root amplitude law and all.

1971

— Ruelle and Takens propose that turbulence begins with a short chain of Hopf-type bifurcations, replacing Landau's infinite cascade and putting Hopf at the door of chaos theory.

1976

— Marsden and McCracken's volume and the Landau-plot tradition in fluid experiments canonize the supercritical square-root law as laboratory practice.

Going further

References. Strogatz, Nonlinear Dynamics and Chaos, 2nd ed., §8.2 · Guckenheimer & Holmes, Nonlinear Oscillations, Dynamical Systems, and Bifurcations of Vector Fields, §3.4 · Kuznetsov, Elements of Applied Bifurcation Theory, §3.5 · Landau & Lifshitz, Mechanics (Vol 1), §§small oscillations · Marsden & McCracken, The Hopf Bifurcation and Its Applications

Provenance. Hopf conditions, normal form, and Lyapunov-coefficient formula follow Guckenheimer & Holmes §3.4 and Kuznetsov §3.5; the exactly solvable supercritical and subcritical examples and the square-root scaling law follow Strogatz §8.2. A verifier should recompute: (i) the polar reduction of the complex normal form, (ii) the exact $\dot r=\mu r-r^3$ example and cycle stability $-2\mu$ , (iii) the subcritical fold at $\mu=-1/4$ , and (iv) the Van der Pol linearization real part $\mu/2$ and the cited vortex-shedding $Re_c\approx 47$ .

Builds on: Linear Stability Analysis and the Jacobian · Classification of Fixed Points: Nodes, Saddles, Spirals, Centers · Limit Cycles and Self-Sustained Oscillations · Bifurcations and the Loss of Stability Leads to: The Van der Pol Oscillator · Poincaré–Bendixson Theorem · Period-Doubling Cascade and the Route to Chaos · Saddle-Node, Transcritical, and Pitchfork Bifurcations

Is a Hopf bifurcation just a pitchfork bifurcation in disguise?

In the amplitude variable, literally yes — and the analogy is worth exploiting. The radial normal form $\dot r = \alpha r + a r^3$ is the pitchfork normal form restricted to $r \geq 0$ , so the branch structure, the square-root law, and the sub/supercritical dichotomy all transfer wholesale from the sibling article on 1D bifurcations. What is genuinely new is the phase: the "new equilibrium" of the pitchfork is here a state of steady rotation, $\dot\theta = \omega + br^2$ , so the bifurcating object is a periodic orbit, not a fixed point. That one difference changes the phenomenology completely — the system acquires a frequency, a Floquet theory, and the possibility of later period-doubling — none of which a true pitchfork has.

Why exactly can the quadratic terms always be removed but the cubic term

|z|^2 z

cannot?

Because removal is a resonance question. A monomial $z^p\bar z^q$ evolves with phase factor $e^{i(p-q)\omega t}$ under the linear flow, and the near-identity substitution that would remove it produces a denominator proportional to $(p - q - 1)i\omega$ . Quadratic terms have $p - q \in \{2, 0, -2\}$ , never $1$ , so their denominators are nonzero and finite coordinate changes absorb them. The cubic $z^2\bar z$ has $p - q = 1$ : its denominator vanishes, the substitution blows up, and the term is structurally stuck. Physically, it is the only low-order term that rotates exactly in phase with the oscillation itself and can therefore pump or drain amplitude coherently over many periods instead of averaging away.

Is the first Lyapunov coefficient related to Lyapunov exponents?

Only through Aleksandr Lyapunov's name — the two concepts answer different questions and it is worth keeping them firmly apart. The first Lyapunov coefficient $\ell_1$ is a bifurcation-theoretic quantity: a single number computed from derivatives of the vector field at one fixed point at one parameter value, whose sign classifies the Hopf bifurcation as super- or subcritical. Lyapunov exponents are dynamical quantities: long-time averages of the exponential separation rate of nearby trajectories, defined along an orbit and used to diagnose chaos (they get their own article later in this chapter). A system can have $\ell_1 < 0$ at a Hopf point and later, elsewhere in parameter space, a positive Lyapunov exponent; neither implies the other.

The van der Pol oscillator has a Hopf bifurcation at mu = 0, but its cycle amplitude is 2, not proportional to root mu. Which statement is wrong?

Neither — van der Pol is a degenerate Hopf, and seeing why sharpens the theorem. The crossing conditions (2) hold: eigenvalues $\mu/2 \pm i\sqrt{1 - \mu^2/4}$ cross at $\mu = 0$ with nonzero speed and frequency. But the square-root law also needs the cubic coefficient to be nonzero and independent of $\mu$ at the crossing — and in van der Pol the same $\mu$ multiplies both the linear pumping and the nonlinear saturation, so the effective normal form degenerates: at $\mu = 0$ the system is conservative with a continuum of orbits, and the cycle appears at finite amplitude $2$ . Shift parameters so pumping and saturation are controlled independently and the generic $\sqrt{\mu}$ Hopf reappears. The sibling article on the van der Pol oscillator tells this story in full.

If the subcritical branching cycle is unstable and unobservable, why should anyone care about it?

Because the unstable cycle is the basin boundary — the fence between "return to rest" and "jump to large oscillation" — and its radius is the system's noise tolerance. In the bistable window a subcritical system at rest is metastable: a perturbation smaller than $r_-$ decays, one larger escapes to the far attractor. As $\mu \to \mu_c^-$ the fence radius shrinks to zero, so the tolerated kick shrinks to zero before the linear instability arrives — a wing can flutter, and a grid can lose synchrony, at parameter values where every eigenvalue still says stable. Engineering safety margins for subcritical systems are statements about the unstable cycle, which is why numerical continuation packages spend so much effort tracking orbits nobody will ever observe.

Does the Hopf theorem require exactly two dimensions?

No — it requires exactly two critical dimensions. In an $n$ -dimensional system the hypothesis is that one complex pair crosses the imaginary axis while all other eigenvalues keep strictly negative real parts. The center-manifold theorem then guarantees a two-dimensional invariant surface, tangent to the critical eigenplane, that attracts all nearby dynamics; the normal form (3) lives on that surface and everything in this article applies verbatim. This is why a Hopf bifurcation in the (infinite-dimensional) Navier–Stokes equations at $Re_c \approx 47$ behaves exactly like the planar theory: the wake dynamics near onset collapse onto two effective dimensions. If a second pair crosses nearby, the center manifold is four-dimensional and richer phenomena (torus bifurcations, mode interactions) take over.

What happens if the frequency at the crossing is zero?

Then it is not a Hopf bifurcation at all. As $\omega(\mu_c) \to 0$ the complex pair collides at the origin of the eigenvalue plane, producing a double-zero eigenvalue — the Bogdanov–Takens bifurcation, a genuinely two-parameter phenomenon. Near a BT point the phase portrait contains, in different parameter wedges, saddle-node bifurcations, a Hopf curve, and a curve of homoclinic bifurcations where the limit cycle dies by colliding with a saddle at infinite period. The practical warning: if your measured onset frequency drifts toward zero as you approach threshold, the $\sqrt{\mu - \mu_c}$ amplitude law and the finite-frequency onset of Hopf theory both fail, and the oscillation period near onset diverges logarithmically instead of staying near $2\pi/\omega$ .

Can a Hamiltonian system undergo a Hopf bifurcation?

Not the dissipative kind described here. Hamiltonian spectra are symmetric under $\lambda \mapsto -\lambda$ (and $\lambda \mapsto \bar\lambda$ ), so eigenvalues come in quadruples $\pm\alpha \pm i\omega$ or pairs on an axis — a single complex pair cannot drift across the imaginary axis leaving the rest of the spectrum stable, and the attracting limit cycle of Hopf theory cannot exist anyway because phase-space volume is conserved (no attractors, as the sibling article on conservative systems shows). What Hamiltonian systems have instead is the Hamiltonian Hopf bifurcation: two pairs, sitting on the imaginary axis, collide and split off it in a Krein collision — the mechanism behind instability onset in the Lagrange points and in symmetric tops. Same name, genuinely different animal.

Hopf Bifurcation and Birth of Limit Cycles

Where this lives today

The physical picture

The eigenvalue crossing — what the linearization sees

The normal form — what the nonlinearity decides

Which equation answers which question

Limiting cases

Worked examples

1 · Reading off a supercritical onset (confidence builder)

2 · A subcritical onset with numbers — the hysteresis loop quantified (exam level)

3 · The square-root law as a measuring instrument (applied)

Going further

Before this

Where this leads