Map is not the territory

What it is

A map is any compressed stand-in for something larger: a metric, a model, a label, a forecast, a case narrative. Each one keeps a few properties of the thing and throws the rest away. This method makes the throwing away explicit, so you can ask whether what was left out would change the decision in front of you. It is a check on confidence rather than a ban on models, because no decision can be taken without compression.

Where it comes from

The phrase belongs to Alfred Korzybski (1879–1950), a Polish-American writer who studied how language fits fact. On 28 December 1931 he read a paper, "A Non-Aristotelian System and its Necessity for Rigour in Mathematics and Physics", before the American Mathematical Society. The occasion was the New Orleans meeting of the American Association for the Advancement of Science. He reprinted the paper as Supplement III of Science and Sanity (1933). Its argument arrives as four numbered points, of which the third is three words long: "A map is not the territory."

His illustration is plain. Take the real territory of Paris, Dresden and Warsaw, and draw a map that puts the cities in the order Dresden, Paris, Warsaw. Travelling by that map "would be misguiding, wasteful of effort", and, he adds, "in case of emergencies, it might be seriously harmful". His first point carries as much weight as his third: a map may have a structure similar or dissimilar to the structure of the territory. A good map earns its keep because the relations inside it hold outside it as well.

What it corrects

The situation is routine. A number was built for one purpose, it survived scrutiny there, and it is now being used to settle a different question. A competent person reads it, checks the arithmetic, and treats the answer as the state of the business. The dashboard shows weekly active users up 9%, so engagement improved. The failure is not laziness. It is the way a representation travels: the caveats that made the number honest stay with the analyst who built it, and the number arrives at the next meeting alone.

Ordinary care does not fix this, because care is spent on the map. Reviewers ask whether the query was right, whether the model runs, whether the slide sums. Nobody is assigned the question of what the number cannot see, and the longer the map has performed well, the less anyone thinks to ask.

How it works

1. Name the representation carrying the argument: metric, model, label, forecast or narrative.
2. State what it measures directly, and what it only stands in for.
3. List the properties it compresses away, and mark the ones that could change this decision.
4. Get one direct observation, counter-metric or boundary case from the territory itself.
5. Say how large the gap would have to be to reverse the call, then decide.

Worked example

Google Flu Trends estimated influenza in the United States from search queries. It was built by fitting the best matches among 50 million search terms to 1,152 weekly data points from the Centers for Disease Control and Prevention. The map worked, and then it drifted. It missed the non-seasonal H1N1 pandemic of 2009 completely, because much of what it had learned was the shape of winter rather than the shape of flu. In February 2013 Nature reported that the service was predicting more than double the CDC's proportion of doctor visits for influenza-like illness. Lazer, Kennedy, King and Vespignani then found that it ran high in 100 of the 108 weeks from 21 August 2011 to 1 September 2013, and overshot the 2011–2012 season by more than 50%.

The correction is the useful half of the story. The same authors showed that the CDC's own figures, two or three weeks stale, projected current flu better than the live search model: mean absolute error of 0.311 against 0.486. Combining the two beat either alone, at 0.232. The queries held a real signal, and it became trustworthy only once it was tied to a direct measurement of the thing itself.

In a Business Case Weekly case

In the Tether case, the 2018–2021 fork gives you a board seat at a moment when reserves hold receivables and other non-cash assets while holders have been told that every token is backed by a dollar. "Backed one to one" is a map of solvency. The territory is asset quality, maturity, custody, concentration and the speed at which redemptions can arrive. The method's work at that fork is to put those five properties on the page. Do that before you rank halting issuance, disclosing the gap, or continuing with attestations, and say which property would change your answer. The method does not tell you which option to pick.

In your answer

• "The figure here is X, which stands in for Y; it does not show Z."
• "The omitted property that could reverse this call is …, and it would have to be at least this large to matter."
• "Before treating this number as the business, I would check it against …, which comes from the territory rather than the model."
• "This map is accurate enough for the decision about …, and not accurate enough for the decision about …"

Common misuse

"All models are wrong" is the counterfeit. It gestures at humility, names no omission, and leaves the recommendation where it was. One question separates the real move from the imitation: after saying it, has your evidence list or your stated confidence changed? If you cannot name the missing property, the decision it would affect, and the check you would run, you have quoted Korzybski rather than used him. The opposite misuse is rejecting a map that is good enough. A model that leaves out something irrelevant here needs no repair, and saying so is also an answer.

References

• Alfred Korzybski, Science and Sanity: An Introduction to Non-Aristotelian Systems and General Semantics (1933), first edition on the Internet Archive. Supplement III is the 1931 paper, and it is short enough for one evening.
• David Lazer, Ryan Kennedy, Gary King and Alessandro Vespignani, "The Parable of Google Flu: Traps in Big Data Analysis", Science 343 (14 March 2014), 1203–1205, open copy in Harvard's DASH repository. Three pages, and the source of every figure above.
• Institute of General Semantics, "Alfred Korzybski", a biography and publication history that dates the December 1931 paper and its place in the book.