Think about two ways to turn on a light. A dimmer brightens smoothly from nothing as you turn the knob. A wall switch jumps from off straight to fully on. A neuron can do both — and which one it does is decided by a hidden tipping point in its equations called a bifurcation. Push the input a little at a time and, at one special value, the neuron's whole behaviour flips from "silent" to "firing over and over."
The knob here is the input current $I$. Below a critical value $I^{*}$ the neuron sits quietly at a stable resting voltage; above $I^{*}$ it fires repeatedly. That changeover value $I^{*}$ is the bifurcation point. For the FitzHugh–Nagumo neuron loaded in the sim, the math gives $I^{*}\approx 0.34$ — cross it and the resting state stops being stable.
To be precise: the neuron is a two-variable system in voltage $V$ and a recovery variable $w$. The resting state is stable only while the Jacobian $J$ has eigenvalues with negative real parts. A Hopf bifurcation happens when a pair of complex eigenvalues crosses into positive territory — exactly when $\operatorname{Tr}(J)=0$ while $\det(J)>0$. At that instant a small oscillation (a limit cycle) is born and the neuron starts to spike. The sliders let you feel this: $I_{\text{ext}}$ walks you across the bifurcation, while $a$, $b$, and $\tau$ reshape the nullclines and actually move $I^{*}$ itself. Some neurons are "dimmers" (Type I, via a saddle-node/SNIC, firing turns on at $0$ Hz) and some are "switches" (Type II, via Hopf, firing jumps to a nonzero minimum rate).
Try this in the sim above: (1) Slowly drag $I_{\text{ext}}$ from $-0.5$ up past about $0.34$ and watch the red marker leap off the blue resting branch onto the green limit-cycle band — that leap is the Hopf onset. (2) Open the model dropdown and choose "SNIC (Type I)," then compare the f–I curve: firing now grows smoothly from $0$ Hz like a dimmer, instead of jumping. (3) Raise $\tau$ and press Compute again — the bifurcation point shifts, proving $I^{*}$ depends on the neuron's parameters, not just on how hard you push it.
§ 03
Equation Derivation — Bifurcation Theory for Neurons
▸ Codimension-1 Bifurcations in Excitable Systems
A bifurcation is a qualitative change in dynamical behaviour as a parameter crosses a critical value. In neuroscience, four codimension-1 bifurcations govern how neurons begin and end firing.
Near a bifurcation, the full high-dimensional dynamics reduces to a low-dimensional normal form via centre manifold reduction. SN normal form: \(\dot x = \mu + x^2\). At μ < 0: two fixed points (stable + unstable). At μ = 0: they collide. At μ > 0: both disappear, trajectory escapes to limit cycle. Hopf normal form (complex): \(\dot z = (\mu + i\omega)z - (a+ib)|z|^2 z\). At μ < 0: stable spiral. At μ = 0: purely imaginary eigenvalues. At μ > 0: unstable spiral + limit cycle.
STEP 2 — Saddle-Node Bifurcation (Type I)
Two fixed points exist for I < I_SN: one stable (rest), one unstable (threshold). As I increases to I_SN, they collide and annihilate. For I > I_SN: no fixed points, trajectory circles on limit cycle. The f–I curve: period T → ∞ as I → I_SN⁺ because the trajectory slows near the "ghost" of the former saddle-node. Scaling: T ≈ π/√(I − I_SN), so f ∝ √(I − I_SN). This gives f = 0 at onset → Type I. Continuous f–I curve, high dynamic range.
STEP 3 — Supercritical Hopf (Type II)
The stable fixed point loses stability when a pair of complex conjugate eigenvalues crosses the imaginary axis. At the bifurcation, eigenvalues = ±iω₀. Supercritical: the nonlinear term a < 0, so a stable limit cycle of small amplitude √μ emerges continuously. Firing rate at onset = ω₀/(2π) > 0 → Type II excitability. The f–I curve has a discontinuous onset (jumps from 0 to f_min). Example: standard FitzHugh-Nagumo with typical parameters undergoes supercritical Hopf at I ≈ 0.34.
STEP 4 — Subcritical Hopf and Hysteresis
If the nonlinear term a > 0 in the Hopf normal form: an unstable limit cycle exists BELOW the bifurcation point. When the fixed point loses stability, the trajectory jumps to a large pre-existing stable limit cycle. This creates bistability: both stable rest AND stable firing coexist for a range of I below I_Hopf. The system shows hysteresis: I to start firing ≠ I to stop firing. The standard Hodgkin-Huxley model undergoes a subcritical Hopf bifurcation — this is why HH has Type II excitability with a slight bistability near threshold.
STEP 5 — Bifurcation at Firing Offset
Firing ends via: (1) SN of limit cycles — the stable and unstable limit cycles collide and vanish (fold of limit cycles, often subcritical); (2) Homoclinic orbit — the limit cycle grows until it touches a saddle point and becomes a homoclinic orbit. The combination of onset and offset bifurcations gives 4 fundamental neuron classes: SN/SN (Type I, no hysteresis); Hopf/Homoclinic (Type II, no hysteresis); Hopf/SN of LC (Type II, with hysteresis/bistability); SNIC/SNIC (Type I, infinite-period).
▸ Worked Example — Finding the FHN Hopf Bifurcation Point
FHN model with a=0.7, b=0.8, τ=12.5. Jacobian at fixed point (V*, w*):
🧠 ConceptualWhat is the practical difference between Type I and Type II excitability?▼
Type I neurons (SNIC/SN bifurcation) can fire at arbitrarily low frequencies near threshold — the f–I curve starts from 0 Hz and increases continuously. This makes them excellent frequency encoders: firing rate faithfully encodes stimulus intensity over a wide dynamic range. Type II neurons (Hopf bifurcation) jump from silence to firing at a minimum frequency (~30–100 Hz). They are better coincidence detectors: they preferentially fire when inputs arrive simultaneously within a narrow time window. Type II neurons also synchronise more readily in networks. Most cortical excitatory neurons are Type I; fast-spiking PV interneurons and many auditory neurons are Type II.
Key takeaway: Type I = frequency encoder (SNIC/SN, f from 0 Hz); Type II = coincidence detector (Hopf, f jumps to minimum onset frequency). The onset frequency is the diagnostic.
🔬 SimulationWhat does the bifurcation diagram show on its axes?▼
The bifurcation diagram plots long-term behaviour vs. I_ext (x-axis). For each I: (1) solid blue line = stable fixed point voltage V* (neuron resting); (2) dashed red = unstable fixed point (threshold); (3) solid green lines bounding a region = maximum and minimum of the stable limit cycle (neuron firing). The bifurcation point is where these behaviours switch. A fold (S-shaped curve) indicates bistability — the neuron can rest or fire at the same I. Move the I_ext slider to observe the real-time operating state highlighted on the diagram. Switch model type (dropdown) to compare Hopf (discontinuous onset) vs SNIC (continuous onset from 0 Hz).
Key takeaway: The bifurcation diagram is the complete "map" of neural states — silence, firing, bistability, hysteresis — all visible at a glance as a function of stimulus current.
🌍 AppliedHow does bifurcation theory explain epilepsy and deep brain stimulation?▼
Epileptic seizures correspond to network bifurcations: normal brain = stable asynchronous fixed point; seizure onset = network crosses Hopf bifurcation into synchronous limit cycle oscillation. Seizure termination = the oscillation loses stability (fold of limit cycles or SNIC). Anti-epileptic drugs work by shifting the bifurcation point: Na⁺ channel blockers (carbamazepine) raise I_bif for Type II neurons. Deep brain stimulation (DBS) for Parkinson's (130 Hz pulses) works by desynchronising the pathological beta-band (13–30 Hz) limit cycle in the subthalamic nucleus. The DBS frequency is chosen to push the STN network from a synchronous limit cycle back to an asynchronous state, not to stimulate specific neurons. Bifurcation theory predicts the optimal stimulation frequency analytically.
Key takeaway: Epilepsy = unwanted transition to limit cycle; Parkinson's tremor = pathological synchrony. DBS treats both by controlling the network bifurcation parameters.
💡 Non-ObviousWhat is a subcritical Hopf bifurcation, and why does it create hysteresis?▼
In a supercritical Hopf, a small stable limit cycle emerges continuously as the fixed point loses stability — no jump, no hysteresis. In a subcritical Hopf, an unstable limit cycle exists below the bifurcation point I_Hopf. When the fixed point loses stability, the trajectory jumps directly to a large pre-existing stable limit cycle. This creates a bistable region below I_Hopf where both rest and firing are stable — a region where the neuron "remembers" whether it was firing. I to START firing (onset) > I to STOP firing (offset) — classic hysteresis. The standard Hodgkin-Huxley model has a subcritical Hopf. In the simulation, switch to "Subcritical Hopf" and observe the S-shaped bifurcation diagram with a fold.
Key takeaway: Subcritical Hopf = bistability + hysteresis near threshold. HH is subcritical. Supercritical Hopf = no hysteresis, smooth transition. The sign of the Lyapunov coefficient distinguishes them.
📐 ComputationalHow is a bifurcation diagram computed numerically?▼
Professional method: numerical continuation (AUTO, XPPAUT, MatCont). The algorithm: start at a known fixed point, predict the next one at parameter + δp using the tangent direction, correct via Newton's method, detect stability changes by tracking Jacobian eigenvalues, handle bifurcation points with branch switching. This finds unstable branches and limit cycles. The simulation uses a simpler "brute force" approach: for each I value, integrate the ODE for long enough and record the asymptotic extrema of V. This correctly identifies stable branches but misses unstable branches and bistable regions. For research-grade bifurcation diagrams, use XPPAUT (free, Ermentrout) or MatCont.
Key takeaway: Brute-force simulation finds stable attractors only. Numerical continuation (XPPAUT) finds unstable branches, saddle points, and bistable regions. Use both: simulation for quick insight, continuation for complete analysis.
🎓 DeepWhat is a codimension-2 bifurcation, and why does it matter for neurons?▼
A codimension-2 (codim-2) bifurcation requires two parameters to simultaneously reach a critical value — it is a bifurcation of bifurcations. The most important for neuroscience is the Bogdanov-Takens (BT) point, where a Hopf curve and a saddle-node curve meet in parameter space. Near a BT point, the neuron can exhibit Hopf oscillations, saddle-node transitions, homoclinic orbit bifurcation, and fold of limit cycles — all within a small parameter neighbourhood. Real neurons (HH, Morris-Lecar) operate near codim-2 points, which is why neuromodulators (dopamine, acetylcholine) can switch neurons between markedly different firing modes by moving the operating point near the BT point. The Codim-2 tab shows the two-parameter bifurcation diagram (fold + Hopf curves in I-b space) with the BT point marked.
Key takeaway: Neurons operating near Bogdanov-Takens codim-2 points are maximally flexible — neuromodulators can switch them between multiple firing modes with small parameter changes.
🧠 ConceptualHow does bursting arise from bifurcation theory — what is fast-slow decomposition?▼
Fast-slow decomposition (Rinzel, 1987) analyses bursting by treating the slow variable s (e.g., slow K⁺ or Ca²⁺ current) as a quasi-static parameter and plotting the bifurcation diagram of the fast subsystem as a function of s. A burst occurs as s slowly traverses a bifurcation: (1) s in quiescent zone → stable fixed point; (2) s slowly changes → fast subsystem crosses saddle-node → spiking begins; (3) spikes elevate s; (4) s crosses another bifurcation (SNIC or homoclinic) → spiking ends; (5) s recovers → next burst. The burst pattern (square-wave, parabolic, elliptic) is completely determined by which bifurcations bound the spiking region of the fast subsystem's bifurcation diagram.
Key takeaway: Bursting = slow parameter sweep through fast-subsystem bifurcation diagram. The burst type is determined by the topology of the fast-subsystem bifurcation boundaries.
§ 04 Best Resources
Izhikevich — Dynamical Systems in Neuroscience, MIT Press, 2007. Free: dynamicalsystems.org
Strogatz — Nonlinear Dynamics and Chaos, Westview, 2015. Ch. 6–8
Rinzel & Ermentrout (1989) — Analysis of neural excitability, MIT Press
Gerstner et al. — Neuronal Dynamics, Ch. 4. neuronaldynamics.epfl.ch
§ 05
Misconceptions & Common Errors
Sub-block A — Conceptual Misconceptions
❌"A bifurcation means the system becomes unstable — stability implies no bifurcation."
✅A supercritical Hopf bifurcation transitions a stable fixed point to a stable limit cycle — both before and after the bifurcation, the long-term behaviour is stable. The fixed point transiently loses stability, but the system immediately finds a new stable attractor (the limit cycle). Bifurcation = qualitative change in dynamics; stability = whether attractors exist. Supercritical transitions can be perfectly stable throughout. Instability (all eigenvalues positive, divergence to infinity) is a separate phenomenon from bifurcation.
❌"The Hopf bifurcation frequency ω₀ is fixed — the neuron fires at ω₀/(2π) Hz for all I above threshold."
✅ω₀ is the frequency at the bifurcation point only. As I increases above I_Hopf, the limit cycle frequency changes continuously — it increases with I. The f–I curve above a Hopf bifurcation is not constant. ω₀ sets the minimum onset frequency for Type II neurons. The actual firing rate at a given I must be computed by integrating the ODE numerically or via continuation — there is no closed-form expression for the frequency away from the bifurcation point (unlike the LIF model).
📖 Izhikevich — DSN Ch. 6.1
❌"Negative eigenvalues mean stable; positive mean unstable — that's all bifurcation analysis requires."
✅Eigenvalues determine local linear stability at a fixed point but don't predict the type of bifurcation or global behaviour. Hopf requires tracking whether complex eigenvalues cross the imaginary axis AND computing the Lyapunov coefficient (cubic nonlinear term) to determine sub vs. supercritical. Negative real parts confirm stability but say nothing about coexisting limit cycles (possible in subcritical region). A saddle-node requires tracking when a real eigenvalue passes through zero. Eigenvalue analysis is the starting point, not the complete answer.
❌Computing bifurcation diagrams using only one initial condition per I value — misses bistable branches entirely.
✅Use multiple initial conditions per I value: one near the resting fixed point (e.g., V = −1.2, w = −0.6 in FHN) and one in the "firing" zone (V = 1.5, w = 0.5). For each, integrate long enough (~1000 time units) and record the asymptotic max/min of V. If both initial conditions converge to different attractors at the same I, the system is bistable. Additionally, scan I in both increasing and decreasing directions — if the resulting curves differ, hysteresis is present (subcritical bifurcation). True unstable branches require numerical continuation (XPPAUT, AUTO).
🔍 Why: Single initial condition finds only one attractor. In bistable parameter regions, the result depends entirely on which basin of attraction the initial condition falls in.
❌Defining the bifurcation point as the I value where the neuron "first fires" in a noisy simulation.
✅The bifurcation point is a property of the deterministic ODE system: SN condition = ∂f/∂V = 0 simultaneously with f = 0; Hopf condition = Tr(J) = 0 with Det(J) > 0. In a noisy simulation, the neuron can fire at any I due to noise-induced crossings — the stochastic "threshold" is diffuse and I-dependent, not a sharp bifurcation point. The stochastic threshold ≠ deterministic bifurcation point. Always compute the bifurcation point analytically or via continuation on the deterministic system; report the stochastic threshold separately.
🔍 Why: In noisy simulations, the first spike occurs randomly below the deterministic bifurcation — conflating these gives incorrect bifurcation point estimates.
§ 05 References
Izhikevich — Dynamical Systems in Neuroscience, MIT Press, 2007. Free: dynamicalsystems.org
Strogatz — Nonlinear Dynamics and Chaos, Westview, 2015
Kuznetsov — Elements of Applied Bifurcation Theory (4th ed.), Springer, 2004