communicoupling

Against the noise,
say it again.

Why is love said twice — "I love you. I love you." — and goodbye performed three ways at the door: the hug, the wave, the text from the train? Why does every wedding run on words the whole room already knows? Claude Shannon answered in 1948, in the paper that founded information theory, by setting meaning aside and pricing transmission itself: every channel is noisy; reliability across a noisy channel can be bought with redundancy; and each channel has a hard capacity, C = 1 − H(p), above which no code, however clever, delivers reliably.

The model is austere. A message is bits; the channel flips each bit with probability p; the sender may repeat each bit and the receiver takes the majority. Repetition drives errors down — that is the purchasable half. The ceiling is the other half: below rate C, errors can be made as rare as you please; above it, the theorem closes the door on every scheme ever built or buildable.

Below, a binary symmetric channel carrying a short message bit by bit. Turn up the noise and watch I LOVE YOU shred on the wire; buy it back with repetition; then push the rate past capacity and watch cleverness stop mattering.

noise · pchance each wire bit flips in transit
0.20
redundancy · rtimes each bit is said · majority vote decodes
message · Lwords sent · 5 bits per glyph
3 words
raw bit error
flips on the wire · measured
decoded bit error
after the majority vote · measured
rate vs capacity
1/r against the Shannon limit
The wire · one bit per send
each column is one message bit said r times · flips arrive hot · the majority vote writes the bottom rows
0 sends
this send glyphs that survived the channel
Decoded error vs redundancy · the Shannon boundary
the curve is exact binomial arithmetic at the current p · the shaded region is rate above capacity, where reliability is provably out of reach
decoded error at this p measured, this run capacity boundary · 1/r = C
Watch what happens
A noisy wire and no redundancy: one bit in five flips, so most glyphs of I LOVE YOU arrive damaged. Raise redundancy and watch the majority vote pull the message back out of the storm.
C = 1 − H(p) · H(p) = −p log₂p − (1−p) log₂(1−p) Shannon's ceiling for the binary symmetric channel — H is the uncertainty each possible flip injects. Below rate C, some code makes errors as rare as you please; above C, none exists. Repetition buys reliability at rate 1/r; the theorem says how far any purchase can go.

Meaning set aside, transmission priced

Two theorems about every noisy channel.

Shannon's first move was surgical: "the semantic aspects of communication are irrelevant to the engineering problem." What remains is a source, a channel, and a rate. The instrument's channel is the simplest one worth having — the binary symmetric channel, which delivers each transmitted bit intact with probability 1 − p and flipped with probability p, independently, indifferently. Against it stands the simplest code worth having: say every message bit r times (r odd, so votes cannot tie) and let the receiver take the majority. A single flip in a triple is outvoted; error now requires a coordinated majority of flips, which is much rarer than any one flip. At p = 0.2 a bare bit arrives wrong one time in five; a 5× majority is wrong one time in seventeen.

Reliability, then, is purchasable — and the price is rate: ninefold saying means one-ninth the distinctions per unit of talk. Shannon's deeper theorem walls the market. Define C = 1 − H(p), with H the binary entropy. Below rate C there exist codes — far cleverer than repetition — that make errors vanishingly rare while keeping the rate up. Above C there are none: no scheme, discovered or undiscovered, is reliable there. The lower plot draws the boundary live. Raise p and the shaded impossible region sweeps across the redundancy axis, swallowing code after code — and once it swallows yours, no dial on this instrument, or any instrument, buys the message back.

What to try

Three sends down a bad wire.

01

Shred, then restore

Set p = 0.20, redundancy : two glyphs in three arrive wrong and the sentence dies on the wire. Slide redundancy to , then — decoded error falls 20% → ~6% → ~2%, and I LOVE YOU climbs back out of the noise.

02

Watch the ceiling sweep

Hold redundancy at (rate 0.2) and raise p slowly. The dashed boundary in the lower plot sweeps right and, near p ≈ 0.24, swallows your code: your rate now exceeds C, and no code at that rate — yours or anyone's — can be made reliable.

03

Visit the dead channel

Push p = 0.50: decoded error pins at 50% for every r — capacity zero. Ease back to p = 0.45 and voting claws error down to ~38%: the difference between almost no channel and none at all is everything.

The channel between two people

Love is said twice because the air is noisy.

Now put people at either end. Between any two of us runs a channel, and it is never clean: the room is loud, the attention is partial, the day was long, the grief is fresh, the line is breaking up, the trust is thin. Every one of these is p. Against them we deploy, without ever having taken the course, Shannon's remedy. "I love you" gets said twice because once is one bit's width of luck. The goodbye gets said three ways — hug, wave, text — because each is a repetition down a differently-noisy line, and the beloved majority-decodes: assembles, from the versions that survived, what was meant. Important news is delivered, then written down, then confirmed the next morning. Repetition is engineering, done by heart.

Ritual is the heavy-redundancy limit. Weddings and funerals transmit into the noisiest states a person enters — weeping, terror, joy — so they use words everyone already knows, said slowly, said again, said by the whole room at once. The rate is terrible; almost nothing new arrives per minute; delivery is close to certain. That is the design brief of a vow. And capacity names the other thing everyone learns eventually: what a relationship can carry depends on its state. In fury or fresh grief, p climbs and C collapses, and some conversations exceed what the channel can bear today. The theorem leaves two honest options: slow down — spend more redundancy on fewer distinctions — or change the channel itself: take the walk, write the letter, wait. Past the ceiling, more words change nothing; only a different p does. At the far end sits the coin-flip interlocutor, p = 0.5, whose reception is independent of your transmission. Whatever you send, what arrives is already theirs. Capacity zero has one implication: through that channel, in that state, stop sending — and work on the channel.

Neighbouring instruments

What redundancy cannot buy.

Shannon's silence about meaning marks this page's border. Perfect transmission leaves understanding unpurchased: two people can exchange every bit intact and still miss each other completely. That gap is what teachback measures, by adding a return channel and checking what was actually decoded — and it is the improbability Luhmann built a theory on. And nothing here says how the bits came to mean anything in the first place; that price is set in signaling games.

The mapping

Mechanism ↔ social life.

In the modelIn the world
bitsThe distinctions you are trying to get across — that you love them, that this goodbye matters.
noise pDistraction, grief, distance, a crackling line, bad faith — whatever turns what was sent into something else.
repetition codeSaying it twice, three ways: the hug, the wave, the text from the train.
majority decodeThe listener assembling your drift from the versions that got through.
capacity CWhat this relationship, in this state, can actually carry per unit of talk.
p = 0.5The interlocutor for whom your words carry no information at all — what arrives is independent of what you sent.

Where it tears

Limits.

It prices transmission, never understanding.

Shannon's theory is meaning-blind on purpose — the blindness is what made the mathematics possible. Two people can achieve zero bit error and total mutual incomprehension: the bits arrive, the sense does not. This page's channel says nothing about that second failure, which belongs to teachback and to Luhmann.

Repetition is the crudest code in the book.

Real systems — and real languages — do far better. Hamming, turbo and LDPC codes approach capacity while keeping the rate up, and natural language carries roughly half its letters as recoverable redundancy, which is why a sentence survives missing letters. The simulation shows the phenomenon at its bluntest; the theorem's ceiling covers all codes, including every one cleverer than this.

Human noise has motives.

The binary symmetric channel flips bits independently and indifferently. Human mishearing is often neither: errors cluster on exactly the words that threaten, and a motivated listener is an adversarial channel that flips the same bit the same way every time — a failure the majority vote cannot outvote. Shannon models the storm, never the saboteur.