The theorem itself
Take a signal with no frequency component above some bound B. Nyquist's result says this signal can be reconstructed exactly from a sequence of instantaneous samples, provided those samples are taken at a rate greater than 2B. Not roughly recovered. Exactly. The theorem is a statement about information sufficiency: at or above that rate, the samples contain everything the original signal contained, and no more is needed.
Below that rate something specific and unforgiving happens. The frequency content above half the sampling rate does not simply disappear. It folds back into the range the samples can represent, and it does so indistinguishably from genuine low-frequency content. This is aliasing. A rotor spinning at 25 revolutions per second, filmed at 24 frames per second, does not vanish from the footage. It reappears as a rotor turning slowly backwards. The image is sharp. It is internally consistent. Nothing in the recorded frames indicates that the true motion has been discarded and replaced by an artefact of the sampling rate meeting the rotor's speed. A viewer who trusted the footage would report a false number with total confidence, because the footage offers no seam at which to doubt it.
This is why sampled-data engineering insists on an anti-aliasing filter: a stage placed before the sampler that removes whatever frequency content the chosen rate cannot resolve. The filter looks like a loss. It discards information. But it converts an invisible error into a visible, bounded one — you know what you cut, because you cut it on purpose, rather than finding out later that the sampler cut it for you and lied about the result. The alternative to filtering is not more information. It is the same missing information plus false confidence about its absence. And crucially, the threshold 2B is a property of the signal being observed, not a property of how carefully or eagerly the observer is watching. Wanting to catch the rotor's true speed does not raise the frame rate. Only knowing B, and sampling faster than 2B, does.
Origin
Harry Nyquist worked out the rate–bandwidth relationship at Bell Labs in 1928, addressing a concrete engineering limit: how many telegraph symbols per second a channel of a given bandwidth could carry without the symbols interfering with each other. Vladimir Kotelnikov gave a rigorous proof of the sampling theorem in 1933, and Claude Shannon supplied the canonical statement in 1948–49 as part of building information theory, drawing also on earlier interpolation work by E. T. Whittaker. The lineage is telecommunications, not physics in the abstract — the question was always "how much of this signal survives being chopped into discrete pieces and sent down a wire." Control engineering absorbed the result through digital sampled-data theory in the 1950s and 60s, at which point the anti-aliasing filter became a standard, almost reflexive design step: filter, then sample, or the sampler will hand you a plausible fiction.
The turn
Set the theorem aside from telegraphy for a moment and ask a different question: what is an intake rate, in general, for any system that claims to represent something outside itself? A Large Language Model is trained on a corpus compiled up to a cutoff and then stops. Its rate of observation on anything happening in the world after that point is zero. Zero is a rate, and Nyquist's inequality applies to it exactly as it applies to any other rate: a process changing faster than zero — which is every live process — aliases completely. The model does not report ignorance. It reports a reconstruction, fluent and structurally indistinguishable from a true account, because nothing in a purely aliased signal signals its own corruption. This is the rotor blade again, at the scale of a civilisation's news cycle instead of a helicopter.
A Large World Model improves this within a narrow band. It samples continuously and quickly while a scene is present — camera frames, lidar sweeps, tens of hertz — which is more than adequate for a grasp attempted at 30 Hz or a footstep planned half a second ahead. But the sampling stops when the episode ends, and resumes, if at all, only when a new episode begins. Between episodes, the rate is again zero. A tank corroding over eight months, a supply chain drifting over a fiscal quarter, a patient's glucose crashing at three in the morning: none of these live inside an episode's duration, so none of them are sampled at any rate greater than zero, and all of them alias into whatever the model last saw. The model is not wrong to trust its 30 Hz data. It is wrong, silently, about everything slower.
The position this points toward is one where sampling rate is not fixed by the system's construction — corpus cutoff, episode boundary — but treated as a parameter, set per stream, revised as the stream's own bandwidth becomes known, and recorded alongside whatever belief the samples produced. That record is the anti-aliasing filter's structural analogue: not a claim that everything has been resolved, but a legible account of what rate was used and therefore what frequencies were never resolvable at all. Below that arrangement, confident false reconstruction is not a risk to be managed. It is a theorem to be expected.
What this is not asking for
The obvious misreading is that Nyquist counsels sampling as fast as possible, and that the lineage argument therefore wants a system drowning in maximal throughput. Disown this directly. Oversampling an unfiltered signal still aliases everything above whatever the new, higher limit is — the problem recurs at a different scale, not less silently. And a rate chosen for its size rather than its relation to a known bandwidth produces the same false confidence, at greater cost and with more noise imported along the way. The theorem's actual content is relational: a rate is adequate only relative to a bandwidth, and that bandwidth must be established independently of the samples themselves. The demand is for a known bandwidth, a filter or its equivalent, and a record — not for volume.
Objections, taken seriously
Clever sampling defeats the naive bound. Compressed sensing reconstructs sparse signals from far fewer samples than Nyquist requires; event-triggered control samples only when an error threshold is crossed. The bound is not as hard as claimed.
True, and this genuinely narrows the claim. The 2B bound constrains rate given a signal model — bandlimited, uniformly sampled. Relax the model and the bound relaxes. But every such scheme purchases its economy with a prior: sparsity in a known basis, a Lyapunov bound, an event threshold set in advance. When the prior is wrong the failure is silent again, and arguably worse, because the reconstruction is now regularised into looking even more plausible. And event-triggered sampling still requires something watching continuously enough to detect the triggering event — the pressure has been relocated from rate to model validity, not eliminated.
Most decisions need only a coarse estimate. Sampling faster imports jitter and cost. Grid operators ran on four-second scans for decades without catastrophe.
Correct, and worth holding onto: decisions have their own bandwidth requirements, often modest, and oversampling can degrade an estimator rather than improve it. The argument is not for maximal rate everywhere. It is for choosing rate against the decision's bandwidth rather than inheriting it from how the system happened to be built. The 1996 Western Interconnection outages are the standing counter-example on the other side — a fixed, coarse SCADA scan rate near 0.2 Hz meeting inter-area oscillations at 0.2–0.8 Hz, aliased into apparent slow drift until the drift became a blackout. The lesson was not "scan faster always." It was that the rate had never been revisited against the dynamics actually present.
Sampling is always finite. "Continuous" is either rhetoric or an unreachable limit, and temporal rate is being fused with breadth of coverage to make one axis look terminal.
Conceded on both points, and the claim survives them because it was never infinitary. No real system samples in the continuous mathematical sense. The three positions differ in whether intake is closed at a cutoff, episodic within a bounded scene, or open with rate held as a live, revisable, attributed parameter. Nyquist supplies pressure toward the third only along time; the fusion with breadth and provenance is the thesis itself, not smuggled premise.
What the theorem does and does not establish
It establishes that a fixed intake rate guarantees undetectable error on every process faster than that rate, and that no downstream processing recovers what was never sampled. It does not establish that faster is always better, that continuous coverage is achievable, or that a revisable-rate architecture resolves everything it touches. It establishes a floor beneath which confidence is unearned, nothing above it about how that confidence should be used.