the second hundred · metaphor 180
How do you come to know something too vast to ever map — a city, a field of work, a person, your own mind — when surveying the whole of it is out of the question? Not by charting it from above. By living in it, and letting where you linger be governed by what matters there.
You will never hold the whole of a great city in your head at once. What you have instead is where you have been — and if you moved honestly, spending your evenings where the life was and passing quickly through the dead blocks, then the shape of your memory comes to match the shape of the place. Not because you measured it. Because you dwelt in it in proportion to what it held.
This is a strange and powerful trick. You never see the terrain from above; you only ever stand in one spot and ask, of the next spot over, is it richer here or there? And yet, wandered long enough by a simple rule, the tally of where you spent your time becomes a faithful portrait of a country you could never survey. Mathematics found this rule and named it — Markov chain Monte Carlo — and it is how we now know spaces far too large to map.
The one rule
The walker knows almost nothing. It cannot see the terrain, cannot rank the hills, cannot even tell you where the peaks are. At each moment it stands on one spot, glances at a random spot nearby, and asks a single local question: is it higher there, or here? If the proposed spot is higher, it steps across. If it is lower, it still steps across sometimes — with a chance equal to how much lower it is. That is the whole of it. This is the Metropolis rule.
That small willingness to go downhill is everything. A walker that only ever climbs would march to the nearest peak and stop, learning one hill and calling it the world. The one that sometimes descends keeps roaming — pooling on the heights, thinning across the valleys, but never trapped. Do this long enough and a law takes hold: the fraction of your life spent in any region comes to equal that region's share of the whole terrain. Time spent becomes height. You never surveyed the country, yet your wandering drew it.
The acceptance rate — how often the proposed step is taken — is the tell of a well-tuned walk. Propose tiny steps and you accept almost all of them but crawl, decades to cross a single hill. Propose enormous leaps and nearly every one lands in some barren place and is refused, so you sit still. Between the two extremes is a stride that moves briskly and is accepted often enough — and it, uniquely, learns the whole country in a human span of time.
What to try
Everything on the panel is measured from the actual walk — the bars are literal visit counts, the acceptance rate is accepted-over-proposed, the portrait error is the honest distance between where the walker has been and the terrain it should recover. Nothing is scripted. Drag the stride down to timid and watch the walker shuffle in place: acceptance near 100%, yet after thousands of steps it has learned only the hill it started on — the far hill still bare. That is the danger of small, comfortable moves.
Now drag the stride to reckless. The walker keeps proposing wild leaps into empty terrain and refusing almost all of them — acceptance collapses, and the portrait barely changes because the walker barely moves. The tuned preset finds the stride between: brisk crossings, healthy acceptance, and — given a minute — a gold portrait that settles cleanly under the green. Press Forget & restart and notice the portrait is worst right at the start. The early wandering is a lie you have to outlast. Then look at the two hills' bars: the taller hill really does collect proportionally more of the walker's time.
The mapping
You cannot survey a city, a discipline, a marriage, or your own inner life from above; there is no vantage that holds the whole. So you do the only thing a body in a world can do — you stand somewhere, sense what is near, and move. The rule that governs your moving is the rule that decides what you will come to know. Move toward richness and linger there; pass lightly through the thin places; but keep moving. The portrait that forms in you is not a map you drew. It is a residue of where you dwelt.
This reframes what it means to understand a large thing. Understanding is not coverage — you will never cover it. Understanding is proportion: having spent yourself across a space in something like the ratio of what it holds, so that the weight of your attention matches the weight of the terrain. And it warns you where you will go wrong. A timid life, all small safe steps, knows its one hill in exquisite detail and mistakes it for the world. A reckless one, all grand leaps at nothing, never lands anywhere long enough to learn it. Knowing, like the walk, has a stride.
Read as life lessons
You cannot visit all of a vast thing. But you can spend yourself across it in the right ratio — more time where there is more — and end with a true sense of the whole from a walk that touched a vanishing fraction of it.
A walker who only ever climbs gets stranded on the first peak. The willingness to step downhill — to leave a good-enough place — is what keeps you from mistaking a local comfort for the entire country.
The portrait is worst at the start; the early wandering reflects only where you happened to begin. Knowledge has a warm-up you must discard — the first impressions are the least trustworthy ones.
In the wild
The space of possible explanations for data is unimaginably vast. MCMC walks it, dwelling on the explanations the evidence favors — and the tally becomes the posterior no one could compute directly.
Metropolis and Ulam invented it in 1953 to sample the states of a fluid — configurations of atoms too numerous to enumerate, learned instead by a random walk that lingers in likely arrangements.
The tree of how species descend has astronomically many shapes. Walkers rove the space of trees, dwelling on those the DNA supports, to draw ancestries no exhaustive search could reach.
The mapping, exactly
| Mathematics | Life |
|---|---|
| the target distribution | A country too large to survey — the true shape of a city, a field, a person — whose whole you can never hold at once. |
| height p(x) | How much a place matters — its richness, its weight, the density of what is worth attending to there. |
| a proposed step | The next move you could make — always local, always compared only against where you now stand. |
| the Metropolis rule | Go toward the richer place; sometimes leave for a poorer one anyway. The discipline that keeps you from getting stranded. |
| time spent ∝ height | Coming to know a thing by dwelling in it in proportion to what it holds — attention distributed like the terrain. |
| burn-in | First impressions, still shaped by where you happened to begin — the part of experience you must live past before it is fair. |
| poor mixing | A life of steps too timid to cross the valley — one hill known perfectly, the rest of the country never even visited. |
The honest model
The terrain is a fixed mixture of two Gaussian hills of unequal height, p(x), drawn once and never touched by the walk. The walker runs plain Metropolis: from position x it proposes x' = x + stride·ζ with ζ a standard normal, then accepts with probability min(1, p(x')/p(x)) — note it needs only the ratio of heights, never the terrain's total. Whether it moves or stays, the current position is tallied into a histogram. Nothing about the target is precomputed into that histogram; the bars are raw visit counts.
The readouts are all measured from that tally. Acceptance is accepted proposals over total. The portrait error is the total-variation distance ½·Σ|ĥᵢ − t̂ᵢ| between the normalized histogram and the normalized terrain over the bins — it starts near one and, for a well-tuned stride, decays toward the irreducible noise of a finite walk. The two hill readouts compare the walker's measured share of time on each side against the terrain's true mass there. The detailed-balance theorem guarantees the histogram's limit is exactly p; the panel simply lets you watch the guarantee arrive — and, with the wrong stride, arrive far too slowly to trust.
Where the metaphor tears
The histogram records how much time was spent where, and nothing else — not the order, not the path, not the reasons. Proportional dwelling is exactly the wrong tool for anything whose meaning is a sequence: a life, a piece of music, a friendship. Some countries are not landscapes to be weighed but journeys to be taken in order, and the tally throws the order away.
The clean theorem — visits converge to the terrain — is a promise about forever. If two rich regions are separated by a deep barren gulf, a local walker may cross it only once in millions of steps, and for any human span it will report a confident, thorough, badly wrong portrait of the one side it knows. Convergence guaranteed is not convergence witnessed. Multimodal worlds break patient wanderers.
The whole method rests on being able to weigh here against there — to feel that this spot is richer than that one. Where no such comparison is even possible, where you cannot tell whether the ground you propose is better or worse, there is no rule to follow and no portrait to form. The walk needs a terrain that pushes back; some human spaces offer no such footing.