IT PAPER DEEP-READ · PAPER 36

A Mathematical Theory of Communication

Claude Shannon · Bell Labs · 1948

中文 →

What did this paper do?

The signal bars on your phone, sending a text, a photo getting quietly shrunk as it saves, 4K video that doesn't stutter — all of it traces back to one 1948 paper and one person: Claude Shannon. He did two epoch-making things: he gave "information" a unit of measure — the bit — and he proved two iron laws that no communication system can escape. This single paper founded the field of information theory and became the bedrock of the entire digital age.

An analogy first

Picture a number-guessing game: I've picked a number from 1 to 8, and you may only ask yes/no questions. The smart move is "Is it bigger than 4?" — whatever the answer, it cuts the possibilities in half, so three questions always nail it. Shannon said: the answer to one such yes/no question is exactly 1 bit of information. Information isn't the sentence itself — it's how many possibilities it rules out, how much uncertainty it removes. Being told something you were already sure of carries almost no information; something genuinely surprising carries a lot.

What was new: information could finally be "weighed"

Before him, the telegraph, the telephone, and radio each had their own ad-hoc tricks, but nobody could say what "information" even was, or whether you could measure it. Shannon stripped the problem clean away from those details — whether you're sending speech, text, or a picture. He didn't care what a message means; he cared only about which one it was, chosen out of all the possible messages. Once you think that way, every kind of communication becomes the same math problem, and information finally has a common scale.

How he pulled it off

The first iron law is about compression. Most of any message is predictable "filler" you could guess anyway — in English a q is almost always followed by a u, and you can drop a few letters and still read it fine. Shannon proved that once you squeeze out all that predictable redundancy, there is a definite floor to how small a source can be packed — go smaller and you start losing content. That floor is the theoretical ceiling behind today's zip, JPEG, and MP3.

The second law is even more startling — it's about fighting noise. Everyone then believed: a line has static, so to make fewer errors you must send slower, and to make almost no errors you must slow almost to a stop. Shannon proved instead that every line has a fixed "speed limit" (he called it capacity), and as long as you stay under that limit, some clever code must exist that pushes the error rate to nearly zero — even over a very noisy line. The trick is to spread the information across a long stretch and keep the "code words" for different messages far enough apart that the noise's random jostling gets averaged out and corrected. That overturned the old belief that faster inevitably means more errors.

What it gave us

The "bit" became the universal currency of information: speech, text, images, video, even DNA can be measured with the same ruler and stored and sent the same way — which is exactly the theoretical premise of "everything can be digitized." That speed limit also set engineers worldwide a clear target, and for more than half a century each generation of error-correcting codes has been racing to approach this "Shannon limit."

One honest caveat: Shannon proved that such a wonderful code must exist, but he never told you what it looks like or how to build it — humanity took nearly fifty more years to construct practical codes that hug that limit.

Remember it in one line

Shannon turned information into something you can weigh (the unit, the bit, = removing one 50/50 uncertainty), and proved two iron laws: every source has a compression floor; every line has a speed limit, and as long as you stay under it, even a noisy channel can be corrected down to nearly no errors. This paper is the cornerstone of the digital world.

Want Shannon's communication-system diagram, the formulas for entropy and capacity, and that "Shannon limit"? → switch to the Deep version