Local Maxima

An atlas of historical co-presence, and the arithmetic of who wins

Abstract

Across recorded history, certain places and periods held an unusual concentration of people who turned out to matter. The question of which one held the greatest concentration sounds empirical and is not, because "concentration" is underdetermined until you say whether you mean number, rate, proximity, interaction, breadth, or consequence. This playground holds sixty-six such clusters, from Uruk around 3200 BCE to two provisional present-day entries, and makes the definition of density an adjustable parameter rather than a hidden one. Three of its six evaluative axes can be recomputed from quantities that are not judgements, which lets the model check itself: proximity against the footprint in kilometres, per-capita intensity against the resident population, and interaction against the count of documented pairwise relationships. Mean disagreement with the independently made judgement is 11 per cent over fifteen calibration cases, with two failures retained because they are informative. The central result is negative and worth stating plainly: no cluster wins under every reasonable weighting, and the ones that come closest do so by being uniformly strong rather than by being extreme on any axis.

---

1. The question and why it resists an answer

Popular history is fond of a certain image: Einstein, Gödel, von Neumann, and Oppenheimer in the same small New Jersey town; Leonardo and Michelangelo taking rival commissions in the same Florentine year; a house on the Leutragasse in Jena where a movement was invented over five days in November. The image is compelling because it converts an abstract fact, that these people were contemporaries, into a concrete one: they could have walked to each other.

The trouble begins when you try to rank the images. Counting famous names rewards London in 1862 and Paris in 1911 and punishes Jena, which had about four and a half thousand inhabitants. Adjusting for population inverts that completely. Requiring documented interaction demotes Vienna in 1913, which is famous precisely for a coincidence rather than a conversation. Requiring disciplinary breadth demotes Bell Labs and Los Alamos, which are famous for the opposite. Requiring a small physical radius demotes Silicon Valley, which was a twenty-kilometre corridor.

Each of those is a defensible reading of "density". None of them dominates the others. The honest response is not to pick one and present the result as a finding, but to expose the choice and show what turns on it.

2. What is in the dataset

Each cluster carries:

The caveat field is not decoration. Almost every celebrated cluster circulates in an inflated version: Fichte had already left Jena, Hegel had not yet arrived, and Goethe lived in Weimar; Aristotle was born fifteen years after Socrates died; Archimedes was based in Syracuse and not Alexandria. Recording the exclusion beside the roster is what separates an atlas from a poster.

3. The model

3.1 The composite index

Six axes, weighted and renormalised, discounted for evidence:

L(c) = ( sum_k  w_k * s_k(c) ) * ( 1 - pi * (5 - sigma(c)) / 4 )

with   sum_k w_k = 1

where sigma(c) is the evidence rating and pi the evidence penalty. At the top of the evidence axis the discount is exactly one: a cluster is never penalised for being well documented, only for being poorly documented. The result stays on the same 1 to 5 scale as its inputs, which keeps it interpretable.

3.2 Proximity, from the footprint

The mechanism that makes physical closeness matter is not distance but unplanned repeated contact. Allen's work on research laboratories found the probability of two people communicating regularly falling off sharply with separation, effectively vanishing beyond a few tens of metres inside a single building. Scaled up to a town, the same shape suggests a plateau followed by decay rather than a simple inverse law: inside a walkable core, you meet people whether or not you meant to; outside it, meeting requires an intention, and intention is a filter.

P(f) = 5 * exp( -max(0, f - 1.5 km) / 14 km )

The plateau is set at 1.5 km, about twenty minutes on foot. The decay length of 14 km is anchored on the two extremes of the corpus and on nothing else. Across all sixty-six clusters this predicts the independently judged proximity axis to a mean error of 4 per cent, with a single conspicuous failure: the Palo Alto to Menlo Park corridor, where the model reads a 20 km footprint as near-isolation while the judgement allows for the fact that the corridor was traversed constantly by car.

3.3 Per-capita intensity, from the population

lambda(c) = roster size / population * 100,000

C(c)      = 2.502 + 1.086 * log10( lambda(c) )

The rate spans four orders of magnitude, from about 200 catalogued people per hundred thousand in Jena to about 0.17 in New York in 1965, so the map is logarithmic. Its two constants are fitted on Jena and London and on no other case, which makes every other cluster an out-of-sample prediction. Mean error against the judged axis is 18 per cent over the sixty-four clusters that carry a population figure.

The residual structure is more interesting than the mean. The model systematically under-reads clusters whose relevant population is institutional rather than municipal. Chicago in 1942 is the extreme case: divided by three and a half million residents the Metallurgical Laboratory looks negligible, when the population that actually mattered was the few dozen people who could get into the squash court under Stagg Field. The Vienna Circle, the Warsaw school, and the Kyoto school fail the same way and for the same reason. This is a real defect with a clear repair, and the repair is not available from the data as it stands, so the defect is documented rather than papered over.

3.4 Interaction, from the documented relationships

D(c) = 2E / ( n * (n - 1) )        graph density over the roster

I(c) = 1 + 4 * sqrt( D(c) )        mapped to the 1 to 5 scale

Graph density over the roster, with a square root because the marginal information in a documented link falls as the graph fills: learning that two of four strangers corresponded changes your picture of the group far more than learning that the twentieth pair in a circle of seven also did.

This axis separates the two things the atlas most needs to keep apart. Canonical density is how many consequential people were nearby. Causal density is how much they acted on one another. Vienna in 1913 has four nodes and zero edges: Freud, Trotsky, Stalin, and Hitler in one city, with no documented relationship between any pair. Jena in November 1799 has nine nodes and twenty-six edges, most of them inside a single household. A list of names cannot distinguish those. A graph can.

The link record is entered for twelve clusters and is deliberately incomplete elsewhere; where it is entered, it is complete for small circles and incomplete for large ones. That pattern is visible in the calibration: Philadelphia in 1787 and the Fifth Solvay Conference come out exact, Oxford in 1652 within 6 per cent, Jena within 12 per cent, and Cambridge in 1932 misses by 46 per cent, because fourteen people spread over four faculties generate ninety-one possible pairs and only the famous dozen are recorded.

3.5 Spread, the complement of footprint

Footprint asks how tightly one cluster sat inside itself. Spread asks how far apart the simultaneous clusters sat from each other, as great-circle distance between every qualifying pair. The two are close to independent, and together they separate two very different pictures of a dense period.

In 1932 the four qualifying clusters span 9,405 km at the widest, Cambridge to Kyoto, with a mean pair of 4,904 km: four regions working in parallel, connected by journals rather than by streets. Restrict the window to a five-year band around 1927 at a one-kilometre cutoff and the survivors sit inside a few hundred kilometres of central Europe. Same corpus, same model, opposite conclusion about what kind of thing was happening.

Distances are computed on the sphere. The map draws them on an equirectangular projection, which stretches area away from the equator, so the picture and the number disagree by design and the number is the one to trust.

3.6 The co-presence field

The playground's name comes from its central image. Each eligible cluster is placed on a nonlinear time axis as a Gaussian bump whose height is its composite index and whose width is the screen space its period occupies. The bumps add. The resulting curve is a field over historical time whose local maxima are the moments when several clusters, not necessarily one, were active at once.

The nonlinearity of the axis is a deliberate distortion. On a linear axis from 3300 BCE to 2026, every cluster after 1500 would fall inside the rightmost eighth of the width, which would be a statement about the calendar rather than about history. Four segments give the recent centuries room proportional to their record density.

4. What the weightings do

Six presets are supplied, each of which is a different answer to "what does density mean":

The last two are the instructive ones. Both co-presence and output run into a ceiling: a cluster is in this atlas because it produced something and because it held remarkable people, so a dozen entries sit at or near 5.0 on both. Weighting an axis heavily does not help when that axis cannot tell its top candidates apart. The ordering that survives is then decided by the axes the weighting claimed not to care about, which is a general hazard of composite indices and not a quirk of this one.

5. Results

Under the balanced weighting with a 25 km cutoff and an evidence floor of 2.0, the leading clusters are Cambridge in 1932, Jena in 1799, the Weimar Bauhaus, Los Alamos, and Bell Labs, separated by less than 0.1 on a 1 to 5 scale. That separation is smaller than any plausible uncertainty in the underlying judgements, so the correct reading is that these five are indistinguishable, not that they are ordered.

Cambridge's robustness is real but partly structural. It is a compact town during an exact breakthrough year that also happened to contain first-rank mathematics, economics, astrophysics, and philosophy, which means it is not weak anywhere. A composite index rewards exactly that, and would reward it even if a specialist would rather celebrate Göttingen for physics or Jena for intensity. Robustness under aggregation is a property of the aggregation as much as of the object.

Restricting the footprint to one kilometre removes every city in the corpus. What remains is schools, laboratories, observatories, and one six-day conference: the Bauhaus at Weimar, Bell Labs at Murray Hill, the Via Panisperna institute in Rome, the Scottish Café in Lwów, Chicago Pile-1, the Maragha observatory, and the Fifth Solvay Conference. This is the clearest single demonstration that the answer is a function of the question. At 25 km the leaderboard is a list of cities; at 1 km no city is eligible at all.

6. Limitations

The three modelled axes are checkable. The other three are not, and the composite index inherits every bias in them. Co-presence in particular measures the visibility of a cluster in the surviving record at least as much as its actual density: Uruk enters with four anonymous occupational groups because its innovators are unnamed, Timbuktu is suppressed by its archive rather than by its scholarship, and Los Alamos and Bell Labs list eight or ten people in front of thousands of technicians and human computers whose absence is not a judgement about their contribution.

Premodern population figures carry error bars that routinely exceed a factor of two. Because the per-capita map is logarithmic with slope 1.086, a factor-of-two error moves that axis by 0.33 on the 1 to 5 scale, which is enough to reorder clusters separated by less than that. Any premodern entry whose rank is not stable under that perturbation should not be treated as placed.

The two present-day entries cannot be scored on output at all, since output is a judgement about consequences that have not finished happening. Both carry evidence ratings below 2.5 and vanish from every ranking once the evidence floor passes that value, which is the correct behaviour.

Finally, the whole exercise rests on a canon. Choosing which people count as consequential is prior to every calculation here, and no amount of arithmetic downstream repairs a selection made upstream.

7. Why these clusters form at all

Reading across the corpus, the recurring mechanisms are unsurprising individually and instructive together.

Patronage concentrates instruments, translations, and salaries quickly, and dissolves just as quickly when a ruler dies: Baghdad, Córdoba, Samarkand, Herat, Prague, Florence. Small university towns work because relevant people cross paths without arranging to, at meals and in colleges and on walks, which raises the probability that an unfinished idea meets the one person able to finish it. Salons, cafés, journals, and clubs matter because they are where disciplinary status temporarily weakens; the Scottish Café, the Mbari Club, and the Viennese coffeehouse are all versions of the same institution. Laboratories convert proximity into causation by adding shared instruments, shared data, and immediate feedback, which is why eight people at Murray Hill outproduce fifty famous residents of a capital who rarely met. Exile and political rupture supply the raw material, and intellectual golden ages are frequently the receiving end of another region's catastrophe. New media lower the cost of forming a school, from manuscripts in Baghdad to printing in Wittenberg to networked computing around Palo Alto.

The last mechanism is the one the atlas can almost measure. Density is productive only where hierarchy is partially permeable: where students can contradict professors, artists can meet scientists, and foreigners can enter local networks. The interaction axis is a crude proxy for permeability, and the clusters that score highest on it are the ones where the record shows people arguing across a line that elsewhere would have held.

8. References

• Allen, T. J. (1977). Managing the Flow of Technology. MIT Press.
• Chandler, T. (1987). Four Thousand Years of Urban Growth: An Historical Census. Edwin Mellen Press.
• Friedrich Schiller University Jena. The Dawn of the Modern Age.
• Solvay Institutes. The Solvay Conferences on Physics.
• University of Cambridge, Department of Physics. Cavendish Laboratory history.
• American Institute of Physics. Heisenberg and the quantum mechanic.
• Nokia Bell Labs. Shannon and the digital age.
• City of Vienna. 1913: Hitler, Stalin, Trotsky and Freud.
• Stanford University. Mapping the Republic of Letters.