# The Cerebellum

> The cerebellum learns to predict, and it learns from error; damage leaves movement intact but clumsy, on the same side as the lesion. In the classical model, parallel fibres describe the moment, the climbing fibre signals errors, and Purkinje cells act as a timed brake on deep nuclei that fire steadily. You follow the sign from a weakened synapse to a well-timed rise in output, test the Marr-Albus-Ito model against experiments that support and limit it, model adaptation to prisms, design an experiment on what drives that learning, and predict the signs and side of a cerebellar lesion.

Course: Neuroscience (Motor systems). Level: university elective. About 33 minutes.
Lesson page (link this one when citing): https://stemhop.com/lessons/cerebellum/

Builds on: Motor Units, Reflexes and Locomotion · Integration, Inhibition and Neuromodulation

## What you'll learn

- Trace the cerebellar circuit by role: mossy fibres to granule cells to parallel fibres to Purkinje cells as context, one climbing fibre from the inferior olive per adult Purkinje cell as the teaching signal in the classical model, and inhibitory Purkinje output onto deep or vestibular nuclei that also receive excitatory collaterals.
- Read a Purkinje-cell recording, telling simple spikes from complex spikes, and explain the Marr-Albus-Ito proposal, following the sign from a depressed parallel-fibre synapse to a well-timed rise in nuclear output, and the experiments that support and limit it.
- Model error-driven adaptation in people (prism throwing, visuomotor rotation, split-belt walking) and distinguish reactive from predictive adjustment.
- Predict the signs of a cerebellar lesion and their side from the region damaged and the pathway crossings, and discuss impaired predictive control as one account of incoordination.

**Damage the cerebellum and a person is not paralysed. They move, and they move badly.**

Watch someone with cerebellar damage reach for a cup. The reach overshoots, and the hand shakes as it nears the cup. Walking widens and wanders. Yet the cerebellum holds most of the brain’s neurons, about four in five in one careful count. It does not usually start movements on its own. It makes them accurate, and it learns to keep them accurate when the body or the world changes.

## Moving badly, not weakly: three puzzles

The patient at the cup raises three puzzles.

1. **Clumsy, not weak.** Why is the movement badly aimed and badly timed rather than feeble, and why does it go most wrong as the hand nears the target?
2. **Relearning.** Prism goggles shift what you see to one side. Your first throws miss, later throws hit, and when the goggles come off you miss the other way. How does the brain relearn, and why does the correction outlast the goggles?
3. **The same side.** The cerebellum’s output crosses to the other side of the brain. Why, then, does damage to the left cerebellum disturb the left arm?

One rule answers the first two puzzles, and the anatomy answers the third. The rule needs two terms. A **prediction**, here, is a learned estimate of what a command will do, used before the senses report the result. An **error** is the mismatch between the predicted or intended result and the observed one. In one line: **the cerebellum learns to predict, and it learns from error.** The next section builds the circuit that does it.

## Three roles, one circuit

Picture the circuit as a cartoon with three kinds of line. Many thin lines run across the top: these carry the **context**. One line climbs up from below: it carries the **error** (the teaching signal, in the classical model). Underneath sits an output cell that fires steadily, with a brake pressing on it: that is the **output**, and the brake is how the cortex shapes it. Every part named below fills one of these three roles.

The cerebellum sits behind the brainstem, under the back of the cerebral hemispheres. Its cortex is a thin, deeply folded sheet, and the same small circuit repeats across it. Regions differ most obviously in where the inputs come from and where the output goes, and also in cell types, firing rates and plasticity.

### Context: mossy fibres, granule cells, parallel fibres

A **nucleus**, in the brain, is a cluster of neuron cell bodies. **Mossy fibres** come from many sources. Through the pontine nuclei come signals from the cerebral cortex, including a copy of the commands it is sending; from the spinal cord and vestibular nuclei come signals about what the body is sensing. Together these are the context.

Mossy fibres excite **granule cells**, small neurons packed into the deepest layer of the cortex; most cerebellar neurons are granule cells. Each granule cell sends its axon up to the outer, molecular layer, where it splits and runs along the fold as a **parallel fibre**. Think of parallel fibres as telephone wires strung through a row of flat trees: they pass through the fan-shaped dendrites of many **Purkinje cells** in turn, and each Purkinje cell is contacted by a very large number of them.

### The teaching signal: one climbing fibre

The second input is the **climbing fibre**. Climbing fibres come from the inferior olive in the medulla, and they cross the midline, so each olive supplies the opposite half of the cerebellum. Each adult Purkinje cell receives one climbing fibre, which winds up its dendrite and makes many contacts along it. Young Purkinje cells start with several and keep one, as [synapse elimination](https://stemhop.com/lessons/synapse-elimination/) explains; how the layers form in the first place belongs to [reelin and reeler](https://stemhop.com/lessons/reelin-and-reeler/).

### The output: a brake on nuclei that are already running

**Purkinje cells are the only output cells of the cerebellar cortex, and they release GABA.** They inhibit neurons of the deep cerebellar nuclei. Humans have four pairs, usually named as three: fastigial, interposed (the globose and emboliform nuclei together) and dentate. In the vestibulocerebellum some Purkinje axons skip the deep nuclei and inhibit the vestibular nuclei of the brainstem directly.

If the cortex only inhibits, where does any output come from? From the nuclei themselves. Many nuclear neurons fire steadily, driven by their own pacemaking and by excitatory collaterals from mossy and climbing fibres on their way in; Purkinje inhibition slows or pauses that firing.

**So the nuclei carry the output, and the cortex sculpts it by timed inhibition.** A Purkinje cell is a foot on the brake of a car whose engine is already running. Ease off the brake, and the car speeds up.

*Figure (AI-generated illustration).* Parallel fibres from granule cells bring context to Purkinje cells, one climbing fibre brings what the classical model treats as the error signal, and the Purkinje cell shapes the output by inhibiting a deep nucleus neuron that already fires steadily. Image: a cut through cerebellar cortex: mossy fibres excite granule cells whose parallel fibres pass through the fan-shaped dendrites of Purkinje cells, a single climbing fibre winds up one Purkinje dendrite, and Purkinje axons inhibit a steadily firing deep nucleus neuron.

Local interneurons add more inhibition: Golgi cells inhibit granule cells, and basket and stellate cells inhibit Purkinje cells. In mouse slices, Purkinje cells also inhibit a subset of interneurons in the inner molecular layer (16% of 371 sampled cells; Halverson and colleagues, 2022), so the cortex is not a strictly one-way chain.

> **The one idea to carry through this lesson.** The cerebellum learns to predict, and it learns from error. Parallel fibres tell each Purkinje cell what is happening, including a copy of the command being sent. In the classical model, a well-timed climbing-fibre signal that the result came out wrong can weaken the parallel-fibre synapses that were active at the right moment (the window depends on the region and on the conditions). Purkinje cells inhibit the output nuclei. With other inputs unchanged, weaker drive to a Purkinje cell means more nuclear output at that point the next time the same context comes round. Parallel-fibre activity without a paired climbing-fibre signal can strengthen the same synapses, so a correction can run either way.

This is the classical **Marr-Albus-Ito** account of the cerebellar cortex. Beyond the synapse experiments, its best behavioural evidence comes from raising the gain of an eye reflex (explained below), with supporting recordings during adaptation to a changed load. Prism adaptation needs the olivocerebellar system, but the synapse that changes there has not been identified.

## Reading the two signals

An electrode placed beside a Purkinje cell picks up two waveforms.

- **Simple spikes** are ordinary action potentials, fired continuously at tens of spikes per second, with a rate set by the cell’s intrinsic excitability and by its excitatory and inhibitory inputs; parallel-fibre input changes that rate.
- **Complex spikes** are large spikes followed by smaller wavelets, one for each climbing-fibre event, at around one per second. After each complex spike the simple spikes pause.

Read a made-up recording. In 2 s you count 130 simple spikes and 2 complex spikes, each followed by a 25 ms gap.

**Predict.** Work out the two rates first: one signal comes out far faster than the other. Which carries most of the cell’s moment-by-moment output? Which looks better suited to marking the moments when something went wrong?

> **Check your prediction.** > **rate = count / time**
>
> Simple spikes: 130 / 2 s = 65 Hz. Complex spikes: 2 / 2 s = 1 Hz. Time paused: 2 × 25 ms = 50 ms, which is 50 / 2000 = 2.5% of the record.
>
> The simple spikes carry most of the cell’s ongoing output. Complex spikes are far rarer, and in the classical model their main job is to teach. Their low average rate does not by itself rule out other roles: some recordings find complex spikes that carry predictive information during movement. Each complex spike marks when its climbing fibre fired, and it can also send a short burst of spikes down the axon.

## Following the sign: how an error becomes a correction

**Film transcript: Following the Sign**

### Film transcript: Following the Sign

How does one small change at a synapse become a correction that arrives on time? David Marr and James Albus proposed that the climbing fibre is a teacher. But which way would it change a synapse? Marr guessed stronger. Albus guessed weaker. In 1982, Masao Ito and Masanobu Kano tested it in rabbits. Compare three conditions. Parallel fibres alone. No lasting depression. Climbing fibre alone. No lasting depression either. The two together. Afterwards the parallel-fibre response is smaller, and it stays smaller for at least an hour. Under these conditions, only pairing weakened the synapse. This is cerebellar long-term depression, the sign Albus had chosen.

The input to a brake gets weaker. Does the nucleus fire more, or less? More, with other inputs unchanged. The foot presses the brake less, so the nucleus, already running, speeds up. Two negatives make a positive. But which parallel fibres are weakened? It depends on the region. In the flocculus, it is those active about a hundred and twenty milliseconds before the climbing fibre. That matters, because a visual error reaches the flocculus roughly a tenth of a second after the mistake. So the fibres active just before carry the context in which the mistake was made. Next time that context arrives, the learned change can act before the late visual error does. The correction need not wait for feedback.

What does learning look like in a real cell? Monkeys returned a handle to a set position, and then the load changed. Climbing-fibre events show as big complex spikes; ordinary output spikes are simple spikes. In some recorded cells, complex spikes grew more frequent as the monkeys adapted. Simple-spike responses fell, and stayed lower. That pattern fits the rule, though it does not measure the synapse directly. But depression alone can only release the brake. What if the output needs to fall instead?

One way: strengthen the same synapses, so the brake presses harder. Parallel-fibre activity without a paired climbing fibre can do that. Other circuit changes can also lower output. In this one simplified channel, weakening alone can only raise the output, with other inputs fixed. Strengthening can reverse that change. The rule has limits. The teacher works in some tasks and not others. Mice without detectable depression still learned the tested tasks. Climbing fibres also signal reward. The takeaway: in this model, a well-timed error signal can weaken the eligible synapses. With other inputs fixed, less brake means more output, ready before the late feedback next time.

From the film

- In rabbit experiments, only parallel fibres paired with the climbing fibre produced a lasting depression of the synapse, which stayed smaller for at least an hour.
- With other inputs unchanged, a weaker input to the Purkinje brake means more output from the nucleus.
- In mouse slices of the flocculus, depression is strongest for fibres active about 120 ms before the climbing fibre, so a learned change can act before late visual feedback arrives.
- The rule has limits: mice with no LTD in standard slice tests still learned the tested tasks, and climbing fibres also signal reward.

Learning needs one more piece: a rule for the synapse where context and error meet. David Marr (1969) and James Albus (1971), working independently, proposed that the climbing fibre is a teacher: when it fires, the parallel-fibre synapses active with it on the same Purkinje cell should change. Marr guessed that they strengthen. Albus guessed that they weaken.

### The test

In 1982 Masao Ito and Masanobu Kano stimulated parallel fibres in the rabbit cerebellum and measured the response they evoked. Their result comes down to three conditions.

- **Parallel fibres alone.** No lasting depression.
- **Climbing fibre alone.** No lasting depression either.
- **The two together.** Afterwards the parallel-fibre response is smaller, and it stays smaller for at least an hour.

Under these test conditions, only pairing produced a long-lasting depression of parallel-fibre to Purkinje transmission. This is cerebellar **long-term depression (LTD)**, the sign Albus had chosen. Its calcium and protein kinase C machinery belongs to [calcium signaling](https://stemhop.com/lessons/calcium-signaling/) and [PLC, IP3, DAG and PKC](https://stemhop.com/lessons/plc-ip3-dag-pkc/); LTD as a general form of plasticity is in [synaptic plasticity](https://stemhop.com/lessons/synaptic-plasticity/).

**Predict.** After LTD, the same parallel fibres excite the Purkinje cell less. The Purkinje cell inhibits the nucleus. When that context comes round again, does the nucleus fire more or less?

> **Check your prediction.** More, with other inputs unchanged. The foot presses the brake less, so the nucleus, already firing, speeds up: two negatives make a positive.

### The timing: correcting before the error

Which parallel fibres are depressed? In mouse slices, the best interval between parallel-fibre and climbing-fibre activity for LTD differs between cerebellar regions (Suvrathan and colleagues, 2016). In the flocculus, which steers eye movements, LTD is strongest for parallel fibres active about 120 ms *before* the climbing fibre. That matters, because an error is sensed late: a visual error reaches the flocculus through climbing fibres roughly a tenth of a second after the eyes went wrong. So the fibres active shortly before the climbing fibre carry the context in which the mistake was *made*, not the one in which it was noticed.

So when the same context recurs, the learned change can act before the delayed visual error arrives. That is anticipatory control, a correction from a learned estimate rather than from waiting for feedback, and it is what “the cerebellum predicts” means in this lesson. Its exact timing depends on the downstream circuit, and whether the cerebellum stores the predicted sensory result, the corrective command or both is still debated.

### What a learning Purkinje cell looks like

Gilbert and Thach (1977) recorded Purkinje cells while monkeys returned a handle to a set position after a load displaced it. Then the load was changed, and the monkeys had to adapt to it. Picture a **raster**, a plot with one row per trial, with simple spikes as thin ticks and complex spikes as heavy marks. In some of the recorded cells:

- Before the load changed, complex spikes came at their usual low rate.
- While the monkeys adapted, complex spikes became more frequent.
- Simple-spike responses fell, and stayed lower.

That pattern fits the rule: more error signals, then a lasting drop in the cell’s response. It does not measure the synapse directly.

### A worked example: the vestibulo-ocular reflex

The **vestibulo-ocular reflex** (VOR) turns the eyes against a head turn so that the world stays still on the retina (see [hearing and balance](https://stemhop.com/lessons/hearing-and-balance/)). It is the classic test bed. Its **gain** is eye-speed amplitude divided by head-speed amplitude. For a distant target, a gain of 1, with the eyes turning as fast as the head in the opposite direction, keeps the image still; a mouse whose eyes reach 8 degrees/s for a 10 degrees/s head turn has a gain of 0.8.

The **flocculus**, part of the vestibulocerebellum, tunes this reflex through the vestibular nuclei. Put magnifying glasses on. Seen through them, the world sweeps past faster with each head turn, so with the old gain the eyes move too little and the image slides across the retina, blurring vision. That **retinal slip** is the error, and the gain must rise. In Ito’s account, for the horizontal reflex:

1. Retinal slip drives the climbing fibres to the flocculus.
2. The parallel fibres that carried the head-turn signal shortly before can be depressed.
3. During the ipsiversive phase of each head turn (rotation toward the recorded side), floccular Purkinje cells now fire less and inhibit their vestibular-nucleus targets less.
4. That phase-specific change enlarges the reflex’s compensatory drive to the eyes, and over training the gain climbs toward the new value the glasses require.

This is **gain-increase learning**.

**Predict.** Now swap in glasses that shrink the world. The eyes move too much. Which way must the gain change, and could LTD alone make that change?

> **Check your prediction.** The gain must fall. In the one-channel brake cartoon, LTD only releases the brake, so it cannot lower that channel’s output; strengthening the same synapses would press the brake harder. The real reflex is phase-specific, though: less Purkinje firing in the other phase of the head turn could in principle lower the gain, and other circuit sites could contribute. So the direction of learning alone does not tell you whether LTD or strengthening did the work; that takes experiments like those below. This is **gain-decrease learning**, a correction in the other direction.

### The rule runs both ways

That reverse change exists. With suitable stimulation, parallel-fibre activity without a paired climbing-fibre signal produces long-term potentiation (LTP) at parallel-fibre synapses (Lev-Ram and colleagues, 2003; Coesmans and colleagues, 2004). Mice whose Purkinje cells lack the enzyme PP2B lose this LTP, and also a form of excitability plasticity, and learn poorly in the tested tasks (Schonewille and colleagues, 2010). In the flocculus, climbing fibres are tuned to the direction of image slip: slip one way raises their rate above its baseline of about one per second, and slip the other way lowers it. That two-way signal could push the same synapse either way. In a single channel with other inputs fixed, LTD can only release the brake, and LTP can reverse that change and reset the synapse; across opposing channels or phases of a movement, LTD alone could still shift behaviour either way.

### Where the rule stops

The Marr-Albus-Ito model is the standard account, but not the whole story. Three limits matter here.

- **The teacher can be gated.** In two monkeys learning to adjust the VOR, a complex spike predicted the next trial’s change in the cell’s output during gain-increase learning, but not during gain-decrease learning. In mice, driving climbing fibres with light ([optogenetics](https://stemhop.com/methods/optogenetics/)) during head rotation was enough to drive gain-increase learning, and the tested pairing did not drive gain-decrease learning (Kimpo and colleagues, 2014). The same signal taught in one direction of learning and not the other.
- **Learning without LTD.** Three mouse lines with no LTD in standard slice tests learned VOR adaptation, [eyeblink conditioning](https://stemhop.com/methods/rodent-behavioural-assays/) and a stepping task normally (Schonewille and colleagues, 2011), though modified protocols later induced LTD in slices from two of these lines (Yamaguchi and colleagues, 2016). Plasticity is thought to occur at several sites, including the deep nuclei, which may hold part of the long-term memory. Where the VOR memory is stored, in the flocculus or in the vestibular nuclei it controls, was argued for years.
- **Climbing fibres carry more than movement errors.** In mice trained on a rewarded task they also signalled reward expectation, delivery and omission (Kostadinov and colleagues, 2019).

So Marr-Albus-Ito is the classical model: well supported for some tasks, gated by the task, and not the only place the cerebellum learns.

## Adaptation in people

**Film transcript: Prisms: Learning You Can Watch**

### Film transcript: Prisms: Learning You Can Watch

Can you watch the cerebellum learn? Put on prism goggles, and throw. The goggles shift what you see to the right. So the first throws miss to the right. Each miss teaches a little. Throw by throw, the misses walk in toward the target. How much each error teaches sets how fast this curve falls. Call it k, the fraction of the error corrected per throw. Now take the goggles off. Where does the next throw land?

To the left. The learned correction is still being applied, to a world that has changed back. That opposite miss is the after-effect. It shows that the mapping between seeing and moving was recalibrated. Real after-effects are usually smaller, partly because people drop a deliberate aiming strategy once the goggles are off. In people with damage to the olivocerebellar system, adaptation at this task was impaired. In the model, a smaller k means smaller steps. The curve flattens, and after the same number of throws the after-effect is smaller too. The model does not say which part is damaged. One account: lose the cerebellum's prediction, and the arm leans on slow feedback.

Movements can overshoot and swing near the target. Now the side. A patient has a lesion in the left cerebellar hemisphere. Which arm is more likely to show tremor? The left arm. Cerebellar output crosses to the right thalamus, which relays to the right motor cortex. Then the corticospinal tract crosses back, to the left side of the spinal cord. Two crossings cancel. The takeaway: the cerebellum helps learn predictions from error. Damage can leave a limb leaning more on late feedback, one cause of clumsiness, on the same side.

From the film

- In the prism task, errors shrink throw by throw, and the opposite miss after the goggles come off shows that the mapping between seeing and moving was recalibrated.
- In people with damage to the olivocerebellar system, prism adaptation was impaired; in the model, a smaller k gives a flatter curve and a smaller after-effect.
- One account is that losing the cerebellum's prediction leaves the arm leaning on slow feedback, which can cause overshoot and swing.
- Two crossings cancel, so a left cerebellar lesion usually disturbs the left arm.

Return to the prism goggles of the second puzzle. The opposite miss once they come off, the **after-effect**, shows that the mapping between seeing and moving was recalibrated. Part of the change is sensory: it can include shifts in where you see and feel your arm, not only in the motor command. Martin and colleagues (1996) used this throwing task and found adaptation impaired in people with damage to the olivocerebellar system.

**Worked problem.** This thrower is made up; the task follows Martin and colleagues. Prisms make the first throw land 20 cm right of target. Assume each throw corrects a fixed fraction k = 0.30 of the remaining error and that nothing else changes.

> **error after n corrections = E0 × (1 − k)n**

Where does k come from? It is the fraction of the remaining error corrected per throw. In the simplest linear version of the guiding rule (each synaptic change proportional to the error, a fixed link from synapse to throw, nothing forgotten), the error shrinks by the same fraction every throw, and k combines the synaptic change per unit of error (the learning gain) with how strongly a change in that synapse moves the throw. It is not the size of any single synaptic change, which shrinks as the error does. A small k describes slow adaptation like that of Martin’s patients; the model cannot say which part of the pathway is damaged.

*Step 1.* With E0 = 20 cm, the error after 1 to 4 corrections is 20 × 0.7 = 14.0, then 9.8, 6.86 and 4.80 cm.

*Step 2.* To get below 2 cm, the error must fall below 2/20 = 0.1 of its start, so you need 0.7n < 0.1. Take logs: n > ln 0.1 / ln 0.7 = (−2.303) / (−0.357) = 6.46, so 7 corrections, leaving 20 × 0.77 = 1.65 cm.

*Step 3.* The learned correction is 20 − 1.65 = 18.35 cm leftward. Remove the prisms and that correction is applied to an unshifted world, so the first throw lands about 18.4 cm left. That is the after-effect in this model. In real throwers the after-effect is usually smaller than the full correction, partly because part of the correction is a deliberate aiming strategy that people drop once the prisms are off.

**Predict.** An illustrative patient learns with k = 0.05. After the same 7 corrections, will the after-effect be larger or smaller than the healthy thrower’s, and why?

> **Check your prediction.** *Step 4.* Smaller. After 7 corrections the error is 20 × 0.957 = 20 × 0.698 = 14.0 cm, and the after-effect is only 20 − 14.0 = 6.0 cm. Little was learned, so there is little left to apply once the prisms are off: after the same number of throws, slow learning leaves both a large error and a small after-effect. (With unlimited practice and no forgetting, both throwers would eventually approach full compensation.)

*Interactive on the lesson page:* Prism throwing — error, correction and after-effect

Split-belt walking separates two timescales. Morton and Bastian (2006) used a treadmill with two belts, one under each leg, and drove one leg faster than the other. People with cerebellar damage kept the quick, reactive adjustments within each step but showed disrupted predictive adaptation, the kind that leaves an after-effect when the belts are equalized. In this task, the cerebellum seemed needed for predicting more than for reacting.

### Design problem

Does cerebellum-dependent adaptation learn from a **sensory prediction error** (the mismatch between the predicted and the observed result of a command) or from having made a corrective movement? Design conditions that tell the two apart.

> **One way to do it.** Rotate the cursor that shows the hand, say by 30 degrees. In a **shooting** condition, people make fast reaches that pass through the target with no chance to correct during the movement; they see the error but make no correction. In a **pointing** condition, they may correct online. If making a correction is what teaches, shooting should adapt less. If prediction error teaches, both should adapt.
>
> Test healthy people and people with cerebellar damage. Match the size and timing of the visual error across conditions, and measure after-effects with the rotation removed and feedback withheld, so the learned change is not confused with online steering.
>
> Tseng and colleagues (2007) ran this contrast. Healthy people adapted in both conditions, and people with hereditary cerebellar ataxia adapted less than controls in both, which supports prediction error as the signal. Limits: this is behaviour in one patient group; it does not identify the synapse or show that every task uses the same signal.

## Predicting a lesion: what, which side, which region

### What goes wrong: movement without prediction

Now the first puzzle. On one account of the lesion, losing the cerebellum’s learned prediction leaves movement leaning more on sensory feedback, and visual feedback is slow, arriving a tenth of a second or more after the movement went wrong. Other pathways remain, but a system that corrects strongly using late information tends to overshoot, correct back too late again, and swing. That may help explain why the patient is clumsy rather than weak, and why the shaking grows as the hand closes on the cup, where fine corrections are needed.

### Which side: two crossings

**Ipsilateral** means on the same side; **contralateral** means on the opposite side. Cerebellar signs are typically ipsilateral, and you can work out why. Output from the left dentate and interposed nuclei leaves through the superior cerebellar peduncle and crosses in the midbrain to the right red nucleus and the right ventrolateral thalamus, which relays to right motor and premotor cortex. The corticospinal tract from the right cortex crosses back at the pyramidal decussation in the caudal medulla and descends on the left. The rubrospinal tract from the right red nucleus crosses separately, at the ventral tegmental decussation in the midbrain, and also descends on the left.

**Predict.** A patient has a lesion of the left cerebellar hemisphere. Which arm is more likely to show intention tremor?

> **Check your prediction.** The left arm. Two crossings cancel: a left cerebellar lesion disturbs the left limbs. Counting only one crossing gives the right arm, a common mistake. (The olivary climbing fibres also cross, so the right inferior olive teaches the left cerebellum.)

*Figure (AI-generated illustration).* Output from the left cerebellar nuclei crosses in the midbrain to the right thalamus and motor cortex, then crosses back at the pyramidal decussation, so a left cerebellar lesion usually disturbs the left arm. Image: a front view of the brain, cerebellum, brainstem and spinal cord with one path from the left cerebellar nuclei crossing in the midbrain to the right thalamus and right motor cortex, then crossing back at the pyramidal decussation to descend to the left arm.

### Which region: a rough map

The regions map broadly onto jobs. The zones overlap, so treat this as a rough guide.

| Region (and nuclei) | Broadly serves | Damage can produce |
| --- | --- | --- |
| Midline vermis (fastigial) | posture, the trunk and gait | an unsteady, wide-based **gait ataxia**; lesions of the caudal vermis (pyramid, uvula) can also impair gaze holding, causing **nystagmus**, rhythmic involuntary eye movements |
| Intermediate zone (interposed) | the limbs | limb signs, listed below |
| Lateral hemispheres (dentate) | output through the thalamus to motor and premotor cortex, and also to prefrontal and parietal areas | limb signs, listed below |
| Flocculonodular lobe (vestibular nuclei) | balance and eye movements | gait ataxia and nystagmus |

Damage to an intermediate or lateral hemisphere can produce **limb ataxia** (deficient coordination of movement), **dysmetria** (overshooting or undershooting a target), **dysdiadochokinesia** (difficulty with fast alternating movements), **intention tremor** (a tremor that worsens as the limb nears its endpoint), decomposition of movement, and slurred speech (dysarthria).

Intention tremor is not the rest tremor of Parkinson’s disease, which is seen when the limb is still and relaxed. Rest tremor starts with the [basal ganglia](https://stemhop.com/lessons/basal-ganglia/), though it also involves a cerebellum-thalamus-cortex loop.

Cerebellar damage is not purely motor either. In 20 patients with disease confined to the cerebellum, Schmahmann and Sherman (1998) described executive, visual-spatial and affective changes, most marked with lesions of the posterior lobe and vermis. [Eyeblink conditioning](https://stemhop.com/methods/rodent-behavioural-assays/), a cerebellum-dependent form of learning, returns in [memory systems](https://stemhop.com/lessons/memory-systems/).

## Going deeper

### Counting the neurons

In one careful count of four adult male brains, about 69 billion of the roughly 86 billion (mean 86.1, standard deviation 8.1) neurons sat in the cerebellum (Azevedo and colleagues, 2009, using the [isotropic fractionator](https://stemhop.com/methods/cell-counting-in-tissue/)).

### Does the cerebellum ever start a movement?

One exception to the usual rule: in trained mice, its output through the thalamus is needed to trigger a cued movement.

### Rates and pauses differ by stripe

**Zebrin** is a protein marker that divides the cortex into stripes. In awake resting mice, Zhou and colleagues (2014) reported pooled mean simple-spike rates of 58.3 Hz in zebrin-positive Purkinje cells and 92.9 Hz in zebrin-negative ones. In adult mice the mean pause in simple spikes after a complex spike was 20.0 ms in zebrin-negative and 34.5 ms in zebrin-positive cells (Beekhof and colleagues, 2021).

---

From STEMhop (https://stemhop.com), free lessons by Christopher Deppmann. Lesson text © the authors. Practice questions: https://stemhop.com/lessons/cerebellum/
