The noisy channel coding theorem in fisheries management

The channel Shannon actually meant

In 1948, Claude Shannon was working at Bell Labs on a problem telephone engineers had been fighting for decades: how much information can a wire carry before noise makes the message unrecoverable. The received wisdom was that reliability traded off smoothly against speed — send faster, expect more errors, no way round it. Shannon proved otherwise. Every noisy channel has a capacity, a hard number in bits per use. Below it, error can be pushed towards zero by clever coding. Above it, no coding scheme, however clever, saves you. The proof was constructive in spirit: it showed reliability comes from redundancy spread across many uses of the channel, or from a return path that tells the sender what actually got through, or both. Strip out both and error simply accumulates, unchecked and unmeasured.

It took engineering decades to catch up with the theorem. Reed–Solomon codes arrived in 1960. Gallager's low-density parity-check codes, published in 1963, were rediscovered and made practical only in the 1990s, alongside Berrou's turbo codes. All of them were attempts to approach a boundary Shannon had located from an armchair fifteen years earlier. The lesson that survived the engineering is more general than telephone wires: reliable reception of a message requires either repetition across independent channel uses, or feedback, or the theorem does not merely make life harder — it makes reliability impossible.

Fisheries management runs on exactly this kind of channel, and has run on it for a lot longer than anyone doing the managing tends to notice.

The stock as sender, the assessment as receiver

A fish stock does not report its own biomass. What arrives at a fisheries scientist's desk is a set of noisy proxies: catch reports from vessels with every incentive to under-declare, survey trawls that sample a few hundred stations out of an ocean basin, temperature anomaly readings that shift recruitment patterns for reasons not yet in the model, and quota filings that lag the water they describe by months. Each of these is a channel use. The stock is the sender. The stock assessment is the receiver's decoded estimate. Between them sits noise that is structured, correlated, and frequently self-inflicted — misreporting inherited from port to port, survey gear that behaves differently in warming water than the calibration assumed, effort data that gets revised after the quota it informed has already been spent.

The characteristic failure of the field follows directly from Shannon's structure, though nobody in a fisheries ministry states it that way. A quota is set on a stock assessment that is two seasons out of date. The assessment was the best available decoding of the channel at the time it was produced. But the channel kept transmitting after the decoding stopped. Recruitment shifted, a marine heatwave altered distribution, effort concentrated somewhere the model didn't expect — and the quota, fixed on last year's decode, is now reliability engineering that has silently crossed above capacity. Nobody notices until the catch data for the following season comes in showing the stock somewhere the assessment said it wouldn't be.

Three ways to run the channel

The lineage from Large Language Model to Large World Model to Large Universe Model is, on the intake axis, a lineage of how the loop between sender and receiver is closed, or left open. Fisheries management has lived through cruder versions of all three, and still runs mixtures of them today.

An assessment built once from an archived dataset — last decade's survey series, a fixed set of historical catch records, no live connection back to the water — behaves like a Large Language Model: a single block, encoded once, decoded once, with no acknowledgement path. Whatever error sat in the archive at the moment of encoding is frozen into every quota derived from it. This is not a hypothetical. Stock assessments have, more than once, been built on catch data that later turned out to be systematically misreported by an entire fleet for years, and the misreporting was baked into recruitment estimates long before anyone queried the source again.

A research survey cruise is closer to a Large World Model. For the weeks the vessel is on station, the loop is genuinely closed: the survey team can retrawl a station that looks anomalous, cross-check acoustic backscatter against net catch, adjust gear if calibration drifts. The return path exists and it earns its keep — a bad haul gets flagged and repeated rather than logged as fact. But the loop closes when the ship goes home. The stock keeps transmitting into a channel nobody is listening to for the eleven months until the next cruise, and the quota set in the interim is, again, decoding a channel that has already moved on.

A management system built around continuously ingested streams — live catch reporting, satellite-derived sea surface temperature updated daily, acoustic tags on tracked individuals, vessel monitoring system pings, port-level filings cross-checked against each other in near real time, each data point carrying a record of which vessel, which sensor, which survey produced it — is the fisheries analogue of a Large Universe Model. The loop never closes. A disagreement between this month's catch report and the temperature anomaly record can be traced to its source and reweighted rather than averaged away. Revision is continuous rather than seasonal.

The objection from Shannon's own theorem

The obvious counter comes straight from the source. Shannon also proved that for a discrete memoryless channel with known statistics, feedback does not increase capacity. A long enough block code, sent once, achieves the same asymptotic reliability as any feedback scheme. If that holds, the entire argument for closing the loop is an engineering convenience dressed up as an information-theoretic necessity — and a sufficiently exhaustive stock assessment, covering enough seasons and enough independent surveys in one large block, should do the job just as well as continuous monitoring.

The no-feedback-gain result is exactly right for a channel whose noise statistics are fixed and known in advance. Fish stocks are neither.

That is the answer, and it needs unpacking rather than asserting. Ocean noise is not memoryless. A marine heatwave correlates errors across an entire season of survey hauls simultaneously, because the fish have moved together and the gear misreads the whole cohort the same way. Reporting bias in a fleet is not independent across vessels; it is shared, inherited from the same market pressure, so ten thousand catch logs agreeing with each other is not ten thousand independent channel uses — it is one biased channel use copied ten thousand times. And even where Shannon's no-gain result would apply, it addresses capacity, not delay. A quota decision needed for the coming season cannot wait for an asymptotically long block of data to accumulate; Schalkwijk–Kailath's result on feedback channels shows that feedback drives error down doubly exponentially in the number of channel uses actually available, which matters enormously when the number available is small and the deadline is fixed by a fishing season, not by an appeal to the limit.

The objection from redundancy already present

A second objection notes, correctly, that fisheries data already has redundancy built in. The same recruitment signal shows up in trawl surveys, acoustic surveys, larval counts, and landings data, produced by independent agencies with no reason to collude. Isn't that Shannon's redundancy mechanism, already operating, without any need to keep the loop open?

It is redundancy, and it does real work — cross-validating survey methods against each other has caught real errors. But the failure mode is correlated noise masquerading as independent confirmation. Two surveys run by different agencies can still share the same gear calibration assumption, the same outdated distribution model, the same seasonal blind spot. Averaging four correlated estimates does not approach the truth; it converges, confidently, on the shared mistake. Detecting that requires going back to the channel and asking a new question of it — checking the gear against a known standard again, re-running a survey off-season, querying whether the larval count method still holds under warmer water. That act is feedback. Redundancy without it is precision without accuracy, and a fishery managed on it can be very sure of a quota that is very wrong.

The intake axis, in fisheries management as everywhere else, ends where the return path stops closing and starts staying open. Beyond that there is no further structural move to make — only better sensors, faster reporting, and more trustworthy provenance feeding the loop that is already, in principle, complete.