The brain, from microseconds to decades

Twenty things a human brain does, placed by how big they are and how long they take. Both axes are powers of ten, because no other scale fits an ion channel and a lifetime on one picture. Six of the boxes carry a real recording or measurement of real people; click any of them.

The five rhythms

Delta, theta, alpha, beta and gamma are not five different recordings. They are one recording, from one channel at the back of one person's head, separated by filtering. All five traces above share a single vertical scale, so the size difference between them is real rather than a drawing choice.

Closing the eyes multiplies alpha's share of the total power by 4.2. That is Hans Berger's 1929 result, and you can still get it out of a free file in about a second of arithmetic.

Do not read the eyes-open delta as sleep. Delta takes nearly half the power in the eyes-open trace, and almost all of that is eye movement and blinking rather than cortical delta. Low frequency power on a scalp electrode is the least trustworthy thing on this page, and the filter cannot tell an eyeball from a brain.

Where each rhythm sits

They do not all come from the same place. Alpha is the back of the head, and dramatically so: with the eyes closed it is 72% of the power at PO8, over the occipital lobe, and 15% at T10 beside the ear. That is the visual cortex idling, and it is why the eyes-closed trace on this page came from Oz.

Two of these five maps are mostly artefact, and it matters. Delta peaks at AF8, right above the eye, because that is where a blink is loudest, not because the front of the head is asleep. Beta and gamma peak at T8 and T10, over the temporalis muscle, because clenching a jaw looks like fast activity to an electrode. A filter cannot tell muscle from brain. Alpha's occipital peak is the one pattern here that is clean, which is a fair picture of how much of EEG analysis is spent on this problem.
This is a scalp map, not a source map. Skull and scalp blur everything, so a hot spot means the signal is strongest there on the head, not that the cortex directly beneath it is generating it. Recovering the generators is a whole inverse problem, and it is not one this page has solved, which is why these are drawn as heads rather than painted onto the cortical surface a page away.

Solving

Band
Eyes
Watch it
Surface
View

quietstrong

On the cortex, not on the scalp

The head maps above are scalp maps. This is the same recording pushed through the other side of the problem: a three-shell model of brain, skull and scalp says how a current at each of 20,484 points on the cortex would look at each of the 64 electrodes, and the inverse runs that backwards to ask which cortical currents best explain what was measured.

Alpha lands where a century of neurology says it should. With the eyes closed, 70 per cent of its strongest sources fall in visual cortex, concentrated in the lateral occipital cortex, the lingual gyrus and the cuneus, and it is 19.4 times stronger than with the eyes open. The build script refuses to write the file unless all of that holds.

What you are looking at. This is the cortex inflated: the real thing is folded so tightly that two thirds of it is buried inside sulci and invisible from outside, so it is blown up like a balloon until the folds flatten out and you can see all of it at once. It stops looking like a brain, which is the price. Switch to folded for the real shape and watch most of the signal disappear into the creases. The two hemispheres are also pushed apart in the inflated view, because inflating them makes each one expand through where the other used to be.
Press play and it moves. The colours above are average power over fifty seconds, which is a still photograph of something that is not still. Playing shows the actual band-filtered signal, frame by frame, three seconds of it at a thirtieth of a second per frame. It runs at an eighth of real speed, because ten cycles a second on a screen is a flicker rather than a rhythm. Blue and orange are the two directions the current swings, on a square root scale because cortical current is very unevenly spread and a linear one leaves everything but the few strongest sources sitting at the dark centre. No value ever crosses zero, which is the one thing that would be a lie about an oscillation. Both eye states are drawn on one scale so the collapse when the eyes open is visible rather than normalised away: eyes open reach 28 per cent of the eyes closed amplitude.
A better surface is not better data. This is drawn on the full fsaverage cortex, 163,842 vertices per hemisphere, but the inverse was solved on 10,242. Each vertex takes the value of the source nearest to it, measured on the sphere so that distance means distance along the cortical sheet rather than through the skull. Nothing is smoothed and no value is invented: the patches you see are exactly the size of the source space's own resolution. Switch to folded to see it on the real anatomy, where most of the cortex is hidden inside sulci, which is why the inflated view exists at all.
An inverse solution is not a measurement. Infinitely many arrangements of current inside a head produce exactly the same voltages on its surface, so the problem has no unique answer and what you get back is what your assumptions asked for. This one assumes current is spread smoothly over the cortical sheet. That is a standard and published kind of estimate, and it is still an estimate. The head is fsaverage, an average of many brains, not the head that was recorded.
Watch what happens to gamma. It localises to the superior temporal cortex, 54 per cent of its strongest sources, with more in the middle and inferior temporal cortex. That is not a brain rhythm. It is the temporalis muscle, and the inverse solution has faithfully placed a clenched jaw on the nearest available cortex. Solving the inverse problem does not remove an artefact. It gives it an anatomical address, which is more convincing and no more true.
Delta moved, and that is the noise model working. On the scalp, delta peaked at AF8 above the eye. Here it does not, because the covariance that whitens the data was computed from the eyes-open recording, which is full of blinks, so the solution treats eye movement as noise and discounts it. What survives is not frontal. That is a choice this page made and it is visible in the result.

What the picture says

Almost everything sits in a band running from bottom left to top right: the bigger a thing is, the longer it takes. That band is not drawn by hand. It is the least squares line through the boxes themselves, so the shape of the cloud is what puts it there.

One process sits a long way off it. A nerve impulse crosses most of a body in about the time a single synapse takes to answer. Everything else that reaches across a metre takes years. That is what myelin buys, and it is why the most common cell in the cubic millimetre on this site's first page is not a neuron at all but the oligodendrocyte, the cell that makes it: 20,139 of 49,379.

Where the numbers come from

Every box is tagged. Measured here means a script in this repository produced the number from data in this repository. Cited means it is a published range nobody here measured, and the citation is on the box. Nothing is asserted without one or the other.

The spatial anchors come from H01, the cubic millimetre of human temporal cortex on the first page: dendritic trees run 228 to 1,147 micrometres, cell bodies sit a median 10.7 micrometres apart, the tissue packs 31,240 cell bodies into every cubic millimetre, and the cortex is 2,626 micrometres deep.

The temporal ones could not come from there. H01 is one instant, fixed in resin, and the tractography is an average of shapes. Neither has a time axis at all. So the recordings are other people's: a whole cell current clamp from a human cortical neuron, scalp EEG from someone opening and closing their eyes, and depth electrodes sitting in a living human hippocampus. All three are openly published under CC0 or ODC-BY, and the numbers printed beside them are measured off the traces by scripts/fetch_signals.py, not copied from a paper.

A unit error that got caught. The first version of that script reported the action potential rising at three hundred million volts per second, because the derivative was taken against the wrong time spacing. A real one rises at a few hundred. The script now asserts that the spike is tens of millivolts tall, under a millisecond wide, and rising at hundreds of volts per second, and refuses to write the file otherwise. The corrected figures agree to within about one per cent with an independent extraction of the same sweep.
The clipped channel. The hippocampal recording has two channels and one of them is pinned against the amplifier rail: it never goes positive at all. The script picks the channel that is not clipped and prints the state of both, rather than plotting a flat ceiling as though it were physiology.

Two clocks, not one

The slow end of this map has a trap in it, and the first version of this page fell into it. Losing a synapse and the decline of synaptic density are not the same process on different scales. They are two clocks.

The removal is fast and largely nocturnal. Synapses strengthened during the day are broadly scaled back across a night's sleep, some are removed outright, and microglia do much of their pruning then, which is also when the day's memories are consolidated. That is the overnight box, five to eight hours wide.

The decline is what is left over after twenty years of those nights. A child's cortex builds far more synapses than it keeps, and the count comes down from the first years into the early twenties. That is the two years to twenty box, and it is a net trend rather than an event.

Myelination had the same fault. Ensheathing one axon takes weeks; finishing the job across the brain takes into the fourth decade, and both now sit in that box rather than one standing in for the other.

The one slow thing we can actually measure

Six boxes carry real data and five of them happen in under a second. The sixth takes sixty years. Myelin water fraction was measured in 45 people aged 18 to 79, and it does not do the same thing everywhere: parietal white matter loses myelin steadily across adult life, at a correlation with age of −0.89, while the corticospinal tract holds on to it at −0.09, which on 45 people is no reliable change at all.

That is the same corticospinal tract drawn on the tract page running from the brainstem to the motor cortex, and the same insulation that puts the nerve impulse so far off the diagonal. The tract that most needs to stay fast is the one that stays myelinated.

What is missing

The right hand side of this map is still thinner than the left, and honestly so. Openly downloadable numbers get scarce as the time axis grows. Synaptic density across a human lifespan, the classic result that a child builds far more synapses than it keeps, does not exist anywhere as a downloadable table: the papers predate supplementary data and the modern imaging work is behind paywalls. Those boxes are cited rather than measured, and they will stay that way until something real turns up.