The Low-Noise Neural Amplifier

What sets the noise floor

Before any transistor, physics already imposes a floor. The real part of the electrode impedance is a resistance, and any resistance at temperature T generates Johnson–Nyquist thermal noise. Its mean-square voltage grows with resistance, temperature, and bandwidth.

\overline{v_n^2} = 4\,k_B\,T\,R\,\Delta f \quad\Longrightarrow\quad v_{n,\mathrm{rms}} = \sqrt{4\,k_B\,T\,R\,\Delta f}

Thermal noise of a resistance R over bandwidth Δf. At body temperature, a 100 kΩ electrode resistance across a 10 kHz spike band gives roughly 4 μV RMS — already comparable to a small spike.

Any resistor generates random voltage noise just from heat jostling its electrons — more with higher temperature, resistance, or bandwidth. This thermal floor can't be amplified away, so it sets the ultimate limit on how small a signal you can see.

v_{n,\mathrm{rms}}: The RMS (typical size) of the thermal noise voltage.
k_B T: Boltzmann's constant times temperature — the thermal energy scale.
R: The resistance producing the noise (e.g. the electrode).
\Delta f: The bandwidth over which you collect the noise.

A 100 kΩ electrode over a 10 kHz spike band at body temperature gives roughly 4 µV RMS — as big as a small spike.

This is why electrode area and coating matter to the circuit designer: a smaller or bare metal tip has higher impedance and therefore a higher thermal floor. Volume I's electrode chemistry (PEDOT, iridium oxide) is, from the amplifier's point of view, a fight to lower R so the physics floor drops. The amplifier can never beat this floor; it can only avoid adding much on top.

Input-referred noise: the amplifier's own contribution

The amplifier adds its own noise, conventionally described as an equivalent noise source at the input so it can be compared directly with the microvolt signal. It is specified as a spectral density S_v(f) in \mathrm{nV}/\sqrt{\mathrm{Hz}}; the total noise you actually see is that density integrated over the passband.

v_{ni,\mathrm{rms}} = \sqrt{\int_{f_L}^{f_H} S_v(f)\,df}

Input-referred noise integrated over the band of interest. For a white density it collapses to (density) × √bandwidth; good neural amps reach a few μV RMS over the spike band.

Add up the amplifier's own noise across the frequency band you care about, then refer it back to the input so it can be compared head-to-head with the neural signal. For flat (white) noise this collapses to just density times the square root of bandwidth.

v_{ni,\mathrm{rms}}: Input-referred RMS noise — the amp's noise as seen at its input.
S_v(f): The noise power spectral density at each frequency f.
f_L,\ f_H: The low and high edges of the band you integrate over.
\int \cdot\, df: Summing the noise power across the whole band.

For a white density this becomes density times \sqrt{\text{bandwidth}}; good neural amps reach a few µV RMS over the spike band.

The design target is simple to state and hard to hit: keep v_{ni,\mathrm{rms}} at or below the electrode's thermal floor, so the amplifier is not the bottleneck. Widen the band and noise rises as \sqrt{\Delta f}; this is exactly why the spike band is deliberately limited to a few kHz rather than left wide open.

1/f noise, DC offsets, and the chopping trick

Two problems live at low frequency. First, transistors add flicker (1/f) noise whose density rises toward DC, right where the LFP lives. Second, the electrode–electrolyte interface has a large, drifting DC half-cell potential — tens to hundreds of millivolts — that would slam a microvolt amplifier into its rails. Both must be handled before any gain is applied.

The DC offset is removed by AC coupling or a DC servo: a high-pass corner in the sub-Hz to few-Hz range blocks the offset while passing the LFP. The 1/f noise is defeated by chopper stabilization: the signal is modulated up to a chopping frequency above the 1/f corner, amplified there where the amplifier is quiet, then demodulated back down — so the signal skips over the noisy low-frequency region entirely.

The noise efficiency factor: resolution costs power

How do you compare two amplifiers that reach different noise levels at different power? The noise efficiency factor (NEF) normalizes noise against the power spent to achieve it, referenced to an ideal single bipolar transistor (NEF = 1).

\mathrm{NEF} = v_{ni,\mathrm{rms}}\sqrt{\dfrac{2\,I_{\mathrm{tot}}}{\pi\,U_T\,4k_BT\,\mathrm{BW}}}

Steyaert–Sansen NEF: input-referred noise weighted by total supply current I_tot; U_T = k_B T / q is the thermal voltage. NEF = 1 is the physical limit of one BJT; real neural amps land around 2–4.

A fairness score for 'how little noise did you get for the power you burned?' It weighs input-referred noise against the total current the amp draws. A NEF of 1 is the theoretical best of a single transistor; real neural amps land around 2 to 4.

\mathrm{NEF}: Noise efficiency factor — lower means better resolution per unit power.
v_{ni,\mathrm{rms}}: The input-referred noise you are trying to keep small.
I_{\mathrm{tot}}: Total supply current the amplifier consumes.
U_T = k_B T / q: The thermal voltage — a fixed physical constant at a given temperature.

A lower NEF means you bought better resolution for the same power; the hard floor of one BJT is 1.

The lesson is unavoidable: for a fixed noise target, halving the noise voltage costs roughly four times the current, because noise power scales inversely with current. Since a front-end replicates this amplifier per channel, the NEF is what turns into the array's total analog power budget. When you read that a probe records N channels at some noise level, the NEF is silently multiplying by N behind the scenes.

Rejecting the common-mode world: CMRR

The body is an antenna. Mains hum, muscle activity, and stimulation all appear on every electrode roughly equally — as a common-mode signal. A differential amplifier is meant to subtract that away and keep only the difference between electrode and reference. How well it does so is the common-mode rejection ratio (CMRR), the ratio of differential gain to common-mode gain.

\mathrm{CMRR}_{\mathrm{dB}} = 20\log_{10}\!\left(\dfrac{A_d}{A_{cm}}\right)

CMRR in decibels. Neural front-ends target well above 60–80 dB; below that, line noise and stimulation leak straight into the microvolt signal.

This measures how well the amplifier ignores whatever is common to both its inputs — like mains hum — while amplifying only the tiny difference you actually want. More decibels means better rejection; below 60 to 80 dB, line noise and stimulation leak straight into your microvolt signal.

\mathrm{CMRR}_{\mathrm{dB}}: Common-mode rejection ratio, expressed in decibels.
A_d: Differential gain — how much the wanted difference is amplified.
A_{cm}: Common-mode gain — how much the shared interference leaks through.

80 dB means the amplifier suppresses common interference by a factor of 10000 relative to the signal.

CMRR is limited by mismatch: two nominally identical input transistors are never perfectly identical, and electrode impedances differ between the signal and reference sites, converting common-mode voltage into a differential error. This is why real systems care so much about balanced impedances and, in bidirectional devices, about blanking during stimulation — topics that recur when we discuss closed-loop and stimulation hardware.