Compare recording methods and their resolution tradeoffs

You now know what the signal is. A spike is one neuron's fast voltage event, the digital-ish pulse from BCI-1.2, and the slow summed currents of many neurons chattering at their synapses are the other thing worth reading, the activity from BCI-2.1. Knowing what the signal is does not tell you how to catch it. That is a separate and surprisingly brutal engineering problem, and the reason it is brutal is physics, not a lack of cleverness. Every method for recording the brain is a forced compromise, and this lesson is about deriving those compromises from a single fact: voltage from a source fades with distance, and the skull is in the way.

What is actually there to record

Start with what a neuron leaves in the space around it. When a neuron fires, ions rush across its membrane (that is the action potential discharging the membrane capacitor from S9.1), and that moving charge sets up a tiny voltage in the salty fluid outside the cell. Two facts about that voltage decide everything downstream.

First, it is small and it fades fast with distance. The extracellular signature of a single spike is tens to a couple hundred microvolts right next to the cell body, and it falls off roughly with distance, so an electrode more than about a tenth of a millimeter away can no longer tell that one neuron's spike from its neighbors'. Get close and you read one neuron cleanly. Back off and the individual voices smear together.

Second, there are two frequency bands riding on top of each other. The fast, sharp part (above a few hundred cycles per second) is the spikes of nearby cell bodies. The slow part (below that) is the local field potential, or LFP: the summed, slow, synaptic activity of a whole neighborhood of neurons, hundreds of micrometers to a couple millimeters across. Spikes tell you which individual cells fired. The LFP tells you what the local crowd is doing on average, and because synaptic currents are slower and more spread out (recall BCI-2.1), the LFP is a lower-resolution, wider-reach signal even when you record it with the very same electrode.

Hold those two facts. Every method below is really just a different answer to one question: how close can your sensor get, and what does the stuff between the sensor and the neuron do to the signal on the way?

Intracortical microelectrodes: one neuron at a time

The most direct answer is to put the sensor inside the cortex, a hair's breadth from the neurons. An intracortical microelectrode is a needle-thin conductor (often a silicon shank or a fine wire, frequently packed into an array of dozens or a hundred tips, such as the well-known Utah array) pushed a millimeter or two into the tissue. Invasive means exactly this: it penetrates the brain.

Because the tip sits within that tenth of a millimeter, it resolves single-unit spikes, the isolated firing of individual neurons, and it also reads the LFP of the immediate neighborhood. This is the highest resolution any method offers, in both senses you will meet formally below: it pinpoints where the signal comes from (single cells) and it catches when to the millisecond. If you want to decode the fine, fast intent behind a hand movement, this is the only signal that even contains that information at the neuron level.

The cost is coverage and risk. Each array samples a patch of cortex the size of a few grains of rice, so you read a hundred or two neurons out of tens of billions. And putting metal through brain tissue is the highest-risk option: surgery, infection risk, and a slower biological problem worth naming honestly. The immune system treats the electrode as a wound. Glial cells wall it off in scar tissue and nearby neurons retreat, so over months to years the signal on many tips degrades. This is why you must separate a lab demo from a shipping product. Intracortical arrays have let paralyzed people in research trials move cursors and robotic arms and spell out speech, which is genuinely astonishing, but a durable, plug-and-forget consumer implant does not exist yet, and chronic signal stability is one of the hard reasons why.

ECoG: a grid on the surface

Back the sensor off just slightly and you get electrocorticography, or ECoG. Here a flexible sheet of electrodes rests directly on the surface of the cortex, under the skull and usually under the protective membrane around the brain, but it does not penetrate the tissue. Nothing is pushed into neurons.

That small retreat changes what you can read. The electrodes now sit a millimeter or more above the cell bodies, past the distance where individual spikes are distinguishable, so ECoG almost never gives you single units. What it gives you is population activity at millimeter spatial resolution: each contact reads the summed signal of a local column of cortex, sharp in time and reasonably sharp in space, far better than anything from outside the skull. This is a real, proven human technology, not a lab curiosity, because neurosurgeons already place ECoG grids to map seizure origins in epilepsy patients before surgery. That clinical use is a temporary diagnostic placement, though, not a chronic BCI you go home with, which is again the demo-versus-product line.

So ECoG is the moderate option across the board: it still requires opening the skull (moderate risk, real surgery, but no cortical penetration), it covers a moderate patch, and it delivers moderate-to-good resolution. It is the sensible middle of the road, and that is exactly what makes it useful.

EEG: the whole head, seen through a wall

Now pull the sensor all the way out to the scalp and you get electroencephalography, or EEG: metal electrodes pressed against the skin of the head, fully noninvasive, no surgery at all. It is cheap, safe, portable, and it covers the whole head at once. It is also, in space, badly blurred, and it is worth deriving why rather than just accepting it.

Between a cortical neuron and a scalp electrode sit the membranes around the brain, a layer of salty fluid, the skull, and the scalp. The skull is the villain. It is a poor electrical conductor, so it attenuates the signal (the microvolts that reach the scalp are even tinier than at the cortex, which is why EEG needs heavy amplification). Worse than attenuation, the skull spatially smears the signal. Because the current has to spread sideways to get through the resistive bone, the contribution of any one patch of cortex fans out across a wide area of scalp before it arrives, a process called volume conduction. Two consequences follow directly. Each electrode picks up an overlapping mixture from a broad region, and neighboring electrodes see heavily overlapping mixtures. The net effect is that a single scalp electrode reads the blurred, summed activity of millions of neurons at once, and its spatial resolution is on the order of centimeters.

Here is the headline of the whole track, made concrete. An intracortical implant reads individual spikes. EEG blurs millions of neurons into one smeared number. That gap is not an engineering shortfall you can amplify your way out of. It is distance and the skull mixing the sources together before the signal ever reaches the sensor.

One thing EEG keeps, though, is speed. The skull blurs where the signal came from, but it does not slow it down, so EEG still tracks activity at millisecond timing. That is why EEG is spatially coarse and temporally sharp, a combination that turns out to matter a great deal, and it is why consumer EEG headbands exist and do real if limited things (detecting broad states like drowsiness or a coarse yes/no intent) while remaining hopeless at reading any single neuron.

fMRI: reading the blood, not the spike

Functional magnetic resonance imaging, or fMRI, breaks the pattern completely, and the honest framing starts by saying what it is not. It is not electrical. It never touches the neuron's voltage at all. Instead it images the BOLD signal (blood-oxygen-level-dependent), which tracks changes in blood oxygenation. The logic is a proxy chain: active neurons burn more fuel, so local blood flow surges to resupply them, which changes the local ratio of oxygenated to deoxygenated hemoglobin, which changes the magnetic signal the scanner measures.

The upside is coverage and spatial detail unavailable to any electrical method from outside. fMRI images the whole brain, including deep structures no scalp electrode can reach, at a spatial resolution of a few millimeters. The catch is baked into the proxy. Blood flow is slow. The hemodynamic response does not peak until roughly four to six seconds after the neurons actually fired, so no matter how you process it, fMRI's temporal resolution is seconds, not milliseconds. It is the mirror image of EEG: spatially decent and temporally terrible, and it measures a downstream consequence of activity rather than the activity itself. The famous fMRI "mind-reading" headlines are real decoding results, but they are done inside a room-sized magnet, offline, and second by second, which is a very different thing from a wearable that reads intent in real time.

The triangle you cannot beat

Line the four methods up and the pattern is not four random tradeoffs. It is one triangle with three corners you are always trading between: resolution, invasiveness, and coverage.

• Intracortical: highest resolution, highest invasiveness, tiniest coverage.
• ECoG: high-to-moderate resolution, moderate invasiveness (skull opened, no penetration), moderate coverage.
• EEG: coarse spatial resolution (sharp in time), zero invasiveness, whole-head coverage, cheap.
• fMRI: good spatial resolution (terrible in time), zero invasiveness, whole-brain coverage, huge and expensive.

The rule that falls out is that you cannot maximize all three. Push for more resolution and you pay in one of the other two: you go more invasive (get the sensor closer, past the skull, into the tissue) or you accept less coverage (a tiny sharp patch instead of a blurry whole). There is no method in the corner where resolution, coverage, and noninvasiveness are all high at once, because the same distance that buys you safe, wide coverage is the distance that blurs the signal. fMRI looks like it sits in that forbidden corner, noninvasive, whole-brain, with millimeter spatial detail, but it pays on the two axes the distance argument cannot see: its time resolution is seconds, and it reads a downstream blood-flow proxy, not the neural activity itself. Read resolution here as resolution in time and directness, not just in space.

A useful way to feel this is audio. A lavalier mic clipped to one singer's collar is the intracortical electrode: it hears that one voice in crisp detail and almost nothing else in the arena. A single microphone hung high over a packed stadium is EEG: it hears the whole crowd at once, it can tell a cheer from a boo instantly, but it can never recover one fan's words out of the roar.

Now the failure edge, because an analogy without its limit is a bug. The stadium mic mostly just gets quieter with distance, and in principle you could raise the gain or add more mics and beamform your way toward individual voices, because sound in air is a fairly recoverable linear mixture. The skull is worse than a distant mic. It does not only attenuate, it spatially low-pass filters and smears each source across many electrodes before any of them sense it, and it mixes millions of sources at once. No amount of gain and no number of extra scalp electrodes reconstructs a single neuron's spike from that, because the information was destroyed by mixing before it reached the sensor, not merely made faint. Keep the mic picture for the coverage-versus-detail intuition and drop it the moment you start imagining you could just turn EEG up until it reads single cells.

Key terms

spike (action potential)

The fast voltage pulse a single neuron fires, whose small extracellular signature fades within about a tenth of a millimeter, so only a sensor very close to the cell can resolve which individual neuron fired.

local field potential (LFP)

The slow, low-frequency part of the extracellular voltage, dominated by the summed synaptic activity of a local neighborhood of neurons rather than any single cell.

intracortical microelectrode

A needle-thin sensor pushed into the cortex, close enough to resolve single-unit spikes plus LFP. Highest resolution, tiniest coverage, highest risk.

ECoG (electrocorticography)

An electrode grid resting on the cortical surface under the skull without penetrating tissue. Millimeter-scale population signals, moderate coverage and risk, no single units.

EEG (electroencephalography)

Scalp electrodes, fully noninvasive, whose signal is smeared and attenuated by the skull so each reads the blurred sum of millions of neurons. Centimeter spatial resolution but millisecond timing.

fMRI (BOLD signal)

A noninvasive scan that images blood-oxygen changes as a slow proxy for activity. Whole-brain coverage and millimeter spatial detail, but temporal resolution of seconds because blood flow lags firing.

spatial resolution

How finely a method pinpoints where a signal came from, from single neurons (implants) to centimeters (EEG).

temporal resolution

How finely a method pinpoints when activity happened, from milliseconds (electrical methods) to seconds (fMRI's slow hemodynamic response).

Why closer wins, in one equation's worth of intuition

The whole triangle traces back to how an electric field from a source falls off with distance. Very roughly, the voltage a small current source produces drops in proportion to one over the distance to your sensor. Sit a tenth of a millimeter away and one neuron dominates what you read. Move ten times farther and its contribution is a tenth as strong, while the number of other neurons now within range has grown enormously, so any single cell is drowned out by the crowd. That is the entire reason distance costs you spatial resolution, and it is why the only way to read single units is to get physically close, which forces invasiveness. The skull then adds a second, nastier effect on top of simple falloff: because current must spread sideways through resistive bone, it acts as a spatial low-pass filter, throwing away the fine spatial pattern and leaving only the broad, slow, large-scale structure. Attenuation alone you could fight with amplification. The spatial smoothing you cannot, because filtering out the fine detail is destroying information, not just shrinking it. Distance blurs, and the skull blurs harder.