Every ECE curriculum spends a semester making sure you understand one fact deeply enough that you can't unlearn it: if you sample a signal below twice its highest frequency component, the information above that threshold doesn't get lost — it gets misrepresented as something else. Aliasing isn't silence. It's a lie that looks like data.
I think about this constantly now that most of my work is data pipelines instead of circuits, because the exact same failure shows up, wearing a different name and getting caught a lot less often.
Downsample a time series before you've thought about what frequency content actually matters, and you don't get a smaller, cleaner version of the same signal — you get a different signal that happens to share some of the original's shape. Aggregate transaction data to a coarser bucket without checking what pattern lives at the finer grain, and you can manufacture a trend that doesn't exist, or erase one that does. Nyquist doesn't care whether you're sampling a sine wave or sampling user behavior. The math is the same math. The failure is the same failure: information discarded before you could find out you needed it, replaced by something that looks plausible.
What ECE actually teaches, underneath the specific theorem, is a discipline: before you throw anything away — a frequency band, a row, a column, a decimal of precision — you have to be able to state what you're assuming still holds after it's gone. Most pipelines I've seen skip that step. Resampling, deduplication, and feature selection get treated as engineering hygiene, not as claims about the world that can be wrong.
a concrete example from an actual pipeline — a specific case where a resampling or aggregation decision changed the answer, ideally with the before/after made explicit.
The pivot from ECE to data science reads, on paper, like a change of subject. It isn't. It's the same question asked about a different kind of signal: what does discarding this cost you, and how would you know if it did.