All posts

0.02–125 Hz at 250 Hz: what the bandwidth figure actually means

The passband, the Nyquist limit, and why a noise floor measured with the inputs shorted is not the noise floor you will see on a head.

0.02–125 Hz at 250 Hz: what the bandwidth figure actually means

Two numbers on the Research Kit specification carry more weight than their line length suggests: 0.02–125 Hz and 0.16 µVrms. Both are honest numbers. Both are easy to read as something they are not. This note is about the difference, because a specification you cannot interpret is not much better than one you cannot check.

125 Hz is the edge of what 250 Hz can represent

A 250 Hz sampler takes a sample every 4 ms. The Nyquist–Shannon theorem puts the highest frequency it can represent at half the sample rate — 125 Hz. Above that, a real signal does not politely disappear. It folds back into the band as an alias, where no later processing can tell it apart from a genuine low-frequency component.

So "250 Hz sampling, 125 Hz bandwidth" does not mean the device measures up to 125 Hz. It means 125 Hz is the boundary past which everything must already have been removed — by an analogue anti-alias filter, before the first sample is taken. That filter is not a brick wall. It has a corner, a slope, and a stopband, and the useful questions are:

  • where is the −3 dB point — the frequency at which the response is down by 3 dB?
  • how fast does it fall after that, and how much is left at Nyquist?
  • what happens between the passband edge and the first frequency that can alias?

If the −3 dB point sat exactly at 125 Hz, the response would still be only 3 dB down at the point where aliasing begins. A production design puts the corner lower and spends the difference on attenuation. We quote 0.02–125 Hz as the passband. The −3 dB point of the production filter is a separate number, and it belongs on the specification sheet next to the passband, not in a footnote. It has not been measured, so we do not quote one here — and we are not going to estimate it and let an estimate turn into a claim.

The low corner deserves the same treatment. 0.02 Hz is a very slow high-pass: a first-order corner there implies a time constant of about eight seconds, so a recording takes a while to settle after the electrodes go on. That is not a defect — it is what it costs to keep slow potentials — but it is a thing to plan for, not a line to skim.

0.16 µVrms is an amplifier number

Input-referred noise figures in this class are bench measurements. The inputs are shorted, the gain and filter are fixed, and the RMS is taken over a defined window. That is the right way to compare front ends: it isolates the electronics from the electrode, the subject and the room.

It is not the noise you will see on a head. Once electrodes are attached, the source impedance and the electrode–skin interface add noise of their own, and movement, sweat and mains pickup add more. A shorted-input figure is a best case, and every manufacturer who quotes one knows it.

The comparison that gets made most often is with TI's ADS1299, a common EEG front end specified at 1 µVpp in a 70 Hz bandwidth, input-referred, also with the inputs shorted. Put the two numbers side by side without conversion and 0.16 µVrms looks about six times better. It is not, for two reasons that are easy to correct for:

  • Peak-to-peak is not RMS. For Gaussian noise the ratio is roughly 6:1, so 1 µVpp is about

0.17 µVrms.

  • Bandwidth matters. Noise adds in quadrature, so a wider band contains more of it. Normalised

to noise density, 1 µVpp over 70 Hz is about 20 nV/√Hz; 0.16 µVrms over 125 Hz is about 14 nV/√Hz.

Same order of magnitude — which is what you would expect from two competent front ends, and a more useful statement than a factor that exists only because the units differ. Neither number tells you what a recording will look like. For that you need the electrode impedance, the subject and the protocol.

What a complete sheet would carry

The fix for both numbers is the same: state the conditions, and date the measurement.

  • The −3 dB point, once a production sweep exists. We are not quoting one today.
  • Input-referred noise with the inputs shorted — the amplifier baseline — with gain, filter and RMS window stated.
  • Input-referred noise with a defined source impedance, so the figure can be compared with a real electrode instead of a bench.
  • The hardware revision and the date the measurement was taken.

Until those are on the sheet, treat 0.02–125 Hz as the top of the passband and 0.16 µVrms as the best case, because that is what they are.

How to read the sheet

  • A passband is not a brick wall. Ask for the −3 dB point and the attenuation at Nyquist.
  • 250 Hz is enough for delta through gamma up to roughly 100 Hz. It is not the instrument for high-frequency oscillations above that, and we would rather say so than let a marketing table imply otherwise.
  • A shorted-input noise floor is the floor of the electronics, not of a recording. Compare it only with another shorted-input figure, in the same units, over the same bandwidth.
  • If a vendor quotes a noise figure without units or bandwidth, the number is not yet a

specification.

The full specification is on the product page. If you are comparing the kit with a particular amplifier for a study, send us the study and we will tell you honestly where the instrument fits — including when it does not.

Further reading

More posts See the specifications