Picture a hand on a light switch. Flick it on and off slowly and you get a steady blink — that is an ordinary neuron firing one spike at a time. Now imagine the hand instead tapping the switch in fast little flurries — rat-a-tat-tat — then resting, then flurrying again. That stop-and-go rhythm is bursting: the cell fires a tight packet of spikes, falls silent, then bursts again. Heartbeat-like pacemaker cells, sleeping thalamus, and insulin-releasing pancreas cells all talk this way.
Why the rhythm? Because two clocks run at once. A fast process makes the individual spikes (milliseconds each), while a slow process quietly keeps score in the background. Picture the slow process as a bucket that fills a little with every spike. While the bucket is low, spikes keep coming. Once it is full enough, it slams the brakes and the cell goes quiet. During the quiet, the bucket slowly drains; when it is empty enough, spiking restarts. Call the bucket level $z$. The simplest story is just: spikes raise $z$, high $z$ stops spikes, low $z$ lets them resume.
At the precise level, this is the fast–slow decomposition. Treat the slow variable $z$ as if it were a frozen knob and ask what the fast part does for each value of $z$: for some $z$ it settles to rest (silence), for others it cannot rest and keeps spiking. As the real $z$ drifts slowly up and down it carries the cell across the boundaries between those regimes — each crossing is a bifurcation, and the pair of bifurcations at burst onset and offset (fold, SNIC, homoclinic, sub-Hopf) is what gives each burster its name in the table below. The Hindmarsh–Rose equations in §03 are the tidiest three-line version of exactly this fast–fast–slow setup, with $r$ setting how slowly the bucket fills and $I$ setting overall excitability.
Try this in the sim above. Nudge Input I upward and watch how many spikes are packed into each burst change — more drive usually widens the spiking window before the slow brake catches up. Push I very high and the bursts can smear into one continuous tonic train, because the slow brake can no longer win. Then stretch the T sim window so several whole burst-and-pause cycles fit on screen, and notice that the long silent gaps are far longer than the gaps inside a burst — that two-timescale signature is bursting's fingerprint.
§ 03
Equation Derivation
▸ Bursting via Fast-Slow Decomposition
Bursting arises when a neuron has (at least) three timescales: fast (spike, ~1 ms), slow (burst envelope, ~50–200 ms), and possibly ultraslow (inter-burst, seconds). The fast-slow decomposition (Rinzel, 1987) analyses bursting by treating the slow variable s as a quasi-static parameter.
$$\boxed{\frac{dx}{dt} = y - ax^3 + bx^2 - z + I}$$
$$\boxed{\frac{dy}{dt} = c - dx^2 - y}$$
$$\boxed{\frac{dz}{dt} = r(s(x-x_R) - z)}$$
Standard parameters: a=1, b=3, c=1, d=5, I=3.25, r=0.006 (slow), s=4, x_R=−1.6. The slow variable z drives the fast (x,y) subsystem through a bifurcation cycle.
Symbol
Meaning
Timescale
Role
x
Membrane potential analogue
Fast
Spike variable
y
Fast recovery (K⁺-like)
Fast
Spike recovery
z
Slow adaptation current (Ca²⁺ or K_Ca)
Slow (r ≪ 1)
Controls burst onset/offset
r
Slow timescale (0.001–0.01)
—
r ≪ 1 ensures slow-fast separation
I
Applied current
—
Sets overall excitability
▸ Derivation Steps
STEP 1 — Why Three Timescales Produce Bursting
Bursting requires: (1) A fast spike mechanism (V and recovery variable); (2) A slow variable that accumulates during spiking (e.g., intracellular Ca²⁺ activating K_Ca channels, or a slow K⁺ channel). During a burst: the fast subsystem oscillates (spikes). The slow variable z accumulates. When z is large enough, the fast subsystem's bifurcation diagram predicts silence (z pushes the fast subsystem below its spiking threshold). During silence: z decays. When z is small enough, the fast subsystem crosses the burst onset bifurcation → next burst.
STEP 2 — Rinzel's Fast-Slow Decomposition Method
Treat z as a slowly varying parameter (freeze it). Compute the bifurcation diagram of the fast (x,y) subsystem as a function of z. Identify the spiking region (where the fast subsystem has a stable limit cycle) and the quiescent region (where it has a stable fixed point). The burst trajectory traces: (1) slow increase of z through the burst onset bifurcation → spiking begins; (2) further z increase through the burst offset bifurcation → spiking ends; (3) slow z decrease (back through the quiescent region) → next burst cycle.
STEP 3 — Interspike Interval During a Burst
Within a burst, the instantaneous firing rate is not constant — it changes as z slowly varies. Near the burst onset bifurcation (SNIC or SN): the period T ≈ 1/√(z−z_onset) (long period near onset). Near the burst offset: T becomes very short (near a fold of limit cycles) or very long again (near a homoclinic orbit). This produces the characteristic parabolic shape of ISI distributions within a burst, and the specific "parabolic burster" shape where ISIs are long at burst start and end, short in the middle.
STEP 4 — Thalamic Bursting: T-Current Mechanism
Thalamic neurons burst via a completely different mechanism: low-threshold Ca²⁺ current (I_T, T-type Ca²⁺ channels). At hyperpolarised potentials (<−70 mV), T-channels de-inactivate (h_T → 1). Upon depolarisation (e.g., end of inhibitory input), the de-inactivated T-channels activate, producing a large Ca²⁺ depolarisation (low-threshold spike, LTS). The LTS triggers a burst of Na⁺ APs riding its crest. After ~100 ms, T-channels inactivate, terminating the burst. This mechanism produces the thalamic burst/tonic mode switch: at depolarised potentials (−65 mV), T-channels are inactivated → tonic spiking. At hyperpolarised potentials (−80 mV), T-channels de-inactivate → burst mode when depolarised.
STEP 5 — Interspike Interval Histogram Signatures
Bursting produces a multimodal ISI histogram: intra-burst ISIs (short, 5–30 ms) and inter-burst intervals (long, 100 ms–several seconds). A bimodal ISI histogram is the key diagnostic signature. The ratio of burst duration to inter-burst interval is the duty cycle. Parabolic bursters show an additional feature: within-burst ISI variation (shorter ISIs in burst center than at edges). The Fano factor (variance/mean of spike counts in fixed time windows) is much greater than 1 for bursting neurons, while it approaches 1 for Poisson and <1 for regular spiking.
▸ Worked Example — Hindmarsh-Rose Burst Period
HR model at standard parameters: r=0.006, s=4, x_R=−1.6, I=3.25. The slow variable z oscillates between z_min ≈ 1.5 (spiking ends, offset bifurcation) and z_max ≈ 3.5 (z begins declining, onset bifurcation).
$$\text{Burst period} \approx \frac{1}{r} \times \text{(range of z traversed)} = \frac{1}{0.006} \times \text{(time to traverse z from 1.5 to 3.5 and back)}$$
$$\tau_z = 1/r = 167 \;\text{time units}. \quad \text{At I=3.25, burst period} \approx 200\text{–}400 \;\text{time units}$$
To increase burst frequency: increase r (faster slow variable → shorter burst period) or decrease I (moves bifurcation boundaries closer). To increase spikes-per-burst: increase the range of the spiking region in z, i.e., increase I (wider spiking range → more spikes before offset bifurcation is reached).
▸ References
[Rin87]
Rinzel — A formal classification of bursting mechanisms in excitable systems, Lecture Notes in Biomathematics 71, Springer, 1987
[IZH07]
Izhikevich — Dynamical Systems in Neuroscience, MIT Press, 2007. Ch. 9 (bursting). Free: dynamicalsystems.org
[HR84]
Hindmarsh & Rose — A model of neuronal bursting using three coupled first-order differential equations, Proc. R. Soc. Lond. B 221:87–102, 1984
🔬 SimulationWhat does the Burst Diagram tab show, and how does it differ from a regular bifurcation diagram?▼
The Burst Diagram tab overlays the fast-slow decomposition analysis on the phase plane: (1) the fast subsystem's bifurcation diagram plotted as a function of z (the slow variable), with spiking and quiescent regions shaded; (2) the slow trajectory of z superimposed, showing how it traverses the bifurcation diagram. The burst onset corresponds to z crossing the SN or SNIC boundary of the fast subsystem; the offset corresponds to z crossing the homoclinic or fold-of-LC boundary. This overlay makes the burst mechanism mechanistically transparent — you see exactly which bifurcation of the fast subsystem is responsible for each burst boundary. Change r and watch the trajectory traverse the z-axis faster or slower.
Key takeaway: The Burst Diagram superimposes the slow variable's trajectory on the fast subsystem's bifurcation diagram. Burst onset/offset = the z values where the trajectory crosses the fast subsystem's bifurcation boundaries.
🧠 ConceptualWhat is the biological function of bursting — why do neurons burst instead of firing tonically?▼
Bursting serves multiple computational functions: (1) Reliable synaptic transmission — a single spike often fails to release neurotransmitter; a burst of 3–5 spikes reliably triggers postsynaptic response. Probability of vesicle release per spike: ~0.3; for a burst of 5 spikes: 1−0.7⁵ ≈ 83% (much more reliable). (2) Signal detection — thalamic T-current bursting produces brief, intense signals that overcome background noise, acting as a 'wakeup call' for cortex. (3) Information encoding — burst duration, intra-burst firing rate, and spikes-per-burst each encode different stimulus features. (4) Oscillation generation — thalamic bursting drives cortical slow oscillations (spindles, delta). (5) Pathological states — epileptic seizures involve pathological bursting; Parkinson's tremor is driven by pathological bursting in STN.
Key takeaway: Bursting enhances synaptic reliability (batch release), serves as a cortical wake-up signal, encodes stimulus features in burst patterns, and generates network oscillations. Pathological bursting underlies epilepsy and Parkinson's tremor.
🌍 AppliedHow does the thalamic burst/tonic mode relate to sleep and anesthesia?▼
During wakefulness: thalamic neurons are tonically depolarised (~−65 mV) by neuromodulatory inputs (ACh, NA, histamine). T-channels are inactivated → tonic spiking mode → faithful relay of sensory information to cortex. During NREM sleep: neuromodulatory inputs decrease → thalamic neurons hyperpolarise (~−80 mV) → T-channels de-inactivate → burst mode → rhythmic bursting drives cortical spindles (12–15 Hz) and then delta waves (1–4 Hz). During anesthesia: drugs hyperpolarise thalamic neurons further → forced burst mode → isoelectric suppression. This mode switch is why sensory stimuli do not wake sleeping or anaesthetised subjects — the thalamus is in burst mode and no longer reliably relaying sensory information to cortex. This is also why thalamic DBS can restore consciousness after severe TBI by shifting neurons from burst back to tonic mode.
Key takeaway: Thalamic burst mode (−80 mV, T-channel de-inactivated) = sleep/unconsciousness; tonic mode (−65 mV) = wakefulness. This switch is controlled by neuromodulators and underlies the sleep-wake cycle and the mechanism of anaesthesia.
💡 Non-ObviousWhy does increasing applied current DECREASE the number of spikes per burst in some bursting models?▼
In square-wave bursters (fold/homoclinic), the burst offset occurs when the slow variable z reaches a critical level where the fast subsystem's limit cycle disappears (homoclinic bifurcation). Higher I shifts the fast subsystem's bifurcation diagram — specifically, it moves the homoclinic bifurcation boundary to a higher z value. This means z must accumulate more before the burst ends → more spikes per burst. HOWEVER, in parabolic bursters (SNIC/SNIC), higher I can move both bifurcation boundaries in the same direction, shrinking the spiking region of the fast subsystem — fewer spikes per burst. And in some models, very high I causes the neuron to fire tonically (no bursting at all) — the slow variable can no longer force the fast subsystem out of spiking. The direction of the effect depends entirely on which bifurcation type limits the burst.
Key takeaway: The effect of I on spikes-per-burst depends on the burst type (square-wave vs parabolic) and which bifurcation boundaries I moves. Increasing I can increase, decrease, or eliminate bursting depending on the model topology.
📐 ComputationalWhat is the minimum r value for valid fast-slow decomposition in Hindmarsh-Rose?▼
The fast-slow decomposition assumes τ_s ≫ τ_f — the slow variable changes negligibly during a single spike. For HR: the fast system (x,y) has period ~10 time units per spike. For z to change <10% during one spike, we need r×10 < 0.1, so r < 0.01. Standard r = 0.006 satisfies this. At r = 0.01, the decomposition is approximate but usually qualitatively correct. At r = 0.1, z changes significantly during each spike — the fast-slow analysis fails and the behavior may change qualitatively (e.g., transition from bursting to chaos). Always check: (fast period)×r ≪ 1. For the Izhikevich model, the equivalent slow variable is u with timescale 1/a — for bursting, a must be small (0.01–0.02) while the fast AP has period ~3–5 ms.
Key takeaway: r×(fast period) ≪ 1 must hold for fast-slow decomposition validity. For HR: r < 0.01. At r = 0.1, the timescale separation is insufficient — bursting may become chaotic or the fast-slow analysis gives wrong predictions.
🎓 DeepWhat is chaos in the Hindmarsh-Rose model, and how does it arise from bursting?▼
At certain parameter values (r = 0.006, I ≈ 3.0–3.5, but also near bifurcation transitions), the Hindmarsh-Rose model produces chaotic bursting: the number of spikes per burst varies irregularly, and the inter-burst intervals are aperiodic, even though the system is deterministic. The chaos arises from the interaction between the fast and slow subsystems near a codimension-2 bifurcation: as z approaches the burst onset bifurcation, the trajectory can be near the unstable manifold of a saddle, causing sensitive dependence on initial conditions. The Lyapunov exponent becomes positive (chaos) for specific I ranges. Chaotic bursting is distinguished from Poisson-random firing by: (1) deterministic → repeatable with same initial conditions; (2) structured ISI distributions (specific patterns); (3) positive Lyapunov exponent; (4) strange attractor in 3D phase space. Chaos in neurons has been proposed as a source of variability that improves signal detection (stochastic resonance-like), but this remains controversial.
Key takeaway: HR chaos arises near codimension-2 bifurcations where fast and slow manifolds interact. Chaotic bursting is deterministic but aperiodic — distinguished from noise by positive Lyapunov exponents and a strange attractor in 3D phase space.
§ 04 Best Resources
Izhikevich — Dynamical Systems in Neuroscience, MIT Press, 2007. Free: dynamicalsystems.org
Gerstner et al. — Neuronal Dynamics, Cambridge, 2014. neuronaldynamics.epfl.ch
Strogatz — Nonlinear Dynamics and Chaos, Westview, 2015
Ermentrout & Terman — Mathematical Foundations of Neuroscience, Springer, 2010
§ 05
Misconceptions & Common Errors
Sub-block A — Conceptual Misconceptions
❌"Bursting is just fast irregular firing — the neuron fires with variable ISIs."
✅Bursting has a specific mathematical structure: bimodal ISI distribution (short intra-burst ISIs + long inter-burst intervals), structured temporal organisation, and a mechanistic origin in slow variable cycling through bifurcations. Simple irregular firing (Poisson or noisy LIF) has a unimodal exponential or gamma ISI distribution. To distinguish: plot the ISI histogram — bimodal = bursting, unimodal = irregular/regular tonic. Also plot the return map (ISI_{n+1} vs ISI_n): bursting produces a characteristic 'T-shaped' cluster (short ISIs cluster near the origin; long inter-burst ISIs form a separate cluster).
📖 Izhikevich — DSN Ch. 9; Rinzel (1987)
Sub-block B — Numerical Errors
❌Setting r too large (e.g., r=0.1) and expecting bursting — instead getting tonic firing or irregular spiking.
✅Bursting in the HR model requires r ≪ 1 for valid fast-slow separation. At r=0.1: z changes by 10% per spike — the slow variable is no longer 'slow.' The fast subsystem does not have time to complete its spiking cycle before z moves significantly — bursting gives way to aperiodic or tonic firing. Rule: r should be at least 10× smaller than the fast system's frequency (in the same time units). For HR with I=3.25, fast spikes occur every ~10 units → r should be < 0.1/10 = 0.01. Standard r=0.006 satisfies this. Always check the timescale ratio before interpreting Hindmarsh-Rose results.
🔍 Why: r too large destroys the fast-slow timescale separation — bursting disappears because z can no longer serve as a quasi-static parameter.
§ 05 References
Izhikevich — Dynamical Systems in Neuroscience, MIT Press, 2007
Gerstner et al. — Neuronal Dynamics, Cambridge, 2014. neuronaldynamics.epfl.ch
Ermentrout & Terman — Mathematical Foundations of Neuroscience, Springer, 2010
Strogatz — Nonlinear Dynamics and Chaos, Westview, 2015