Refresher of LMN Resampling

1. LArTPC samplings

Detector T (ns) F (MHz) ADC (bits)
past (PD-1,MB) 500 2.000000 12
DUNE FD      
VD TDE 500 2.000000 14
VD BDE 512 1.953125 14
HD 512 1.953125 14

2. Proposal

Inside the offline WCT Noise Filter we will upsample the waveforms: \[T=\qty{512}{ns} \to T=\qty{500}{ns}\]

Comments:

  • Upsampling is information-lossless from a sampling-theory view.
    • The particular technique must take care to minimize time-domain artifacts.
  • Can be performed cost-free inside existing WCT NF frequency domain and with a “trick”.
  • Apply far-upstream to provide \qty{500}{ns} to (almost) all offline software.
    • Allows field-response and simulation to be at \qty{500}{ns}.

3. Conventional resampling \(N \to N'\)

“rational interpolation” = “integral upsample” + “integral downsample”.

  • integral upsample \(N \to N^{iu} = LN\)
  • integral downsample \(N \to N^{id} = N/M\)
  • rational interpolation \(N \to N^{ri} = RN,\ R\equiv \frac{L}{M}\)

\vfill

\small Comments

  • \scriptsize Integral downsample allows for non integer \(N/M\) by selecting every \(M^{th}\) time sample and ignoring the fact that the last selection represents fewer discarded samples than prior selections.
  • Depending on original signal bandwidth, downsampling may require a low-pass filter to avoid aliasing.

4. The LMN trick

Choose \(N\) to minimize \(L\) and \(M\) while keeping \(\frac{LN}{M} \in \mathbb{N}\).

We want to shift the Nyquist frequency by:

\[ \Delta F= F_{nyquist} - F'_{nyquist} = \frac{T-T'}{2TT'}\]

And we want that shift to be covered by an integral number \(n\) of bins of size \(\delta F = \frac{1}{NT}\).

So, we want \(n\cdot\delta F = \Delta F\) which is solved by:

\[ N = \frac{2nT'}{T-T'} = \frac{n}{3}\cdot 250\]

With the last term having plugged in DUNE numbers.

Thus for \(N\) any multiple of 250, we can assure the downsampling will be exactly integral.

5. LMN procedure

  1. Calculate a desired \(N\) and \(n\).
  2. Truncate or pad original signal to \(N\) while mitigating artifacts.
  3. Apply \(DFT\).
  4. Insert \(2n\ (R>1)\) or remove \(2n\ (R<1)\) frequency samples.
  5. Apply \(DFT^{-1}\).
  6. Fix normalization.

6. Resampling artifacts

  • Any interpolation produces artifacts in time or frequency or both.
  • Fourier-based interpolation does not distort frequency domain spectrum (other than the intentional spectral truncation/extension) but can distort time domain waveform.
    • Violating periodicity assumptions can lead to artifacts near endpoints.
    • Mitigate by padding via simple linear/cubic time-domain extension of waveform?

7. Links

Author: Brett Viren

Created: 2024-05-17 Fri 17:05

Validate