The Squares That Weren't Aimed

In the summer of 1944 the V-1 flying bomb — the doodlebug, the buzz bomb — came down on London by the thousand. It was not a precision weapon: a gyroscope held a heading and a small propeller counted off the air it had flown through, and when the count ran out the engine cut, the nose dropped, and wherever the machine happened to be became the target. People on the ground could not read it that way. They saw the map fill in — a cluster here, a street untouched there — and did what minds do with a scatter of dots: they found intent in it. Some neighbourhoods felt hunted; some felt lucky; families moved on the strength of it.

After the war an actuary named R. D. Clarke asked the question an actuary would. He took a 144-square-kilometre tract of South London, divided it into 576 equal squares of a quarter square kilometre each, and simply counted how many flying bombs had fallen in each. Then he compared the tally to the one distribution that describes pure, aimless randomness — points thrown at a plane with no memory and no target, the Poisson distribution. If the bombs were being steered, the counts should depart from it. Here is what he found. (Every number below is recomputed in your browser from Clarke's published table; nothing is asserted.)

The fit is almost uncanny. Where randomness predicts 226.7 empty squares, the war left 229; where it predicts 30.6 squares hit three times, there were 35; the worst single disagreement is barely five squares out of five hundred. Judged by the χ² criterion, 88 per cent of comparable random scatters would fit worse than the real bombs did. By every standard test the flying-bomb map is indistinguishable from a handful of rice thrown at a chessboard. Feller, reprinting Clarke's table in his textbook, drew the moral himself:

It is interesting to note that most people believed in a tendency of the points of impact to cluster. If this were true, there would be a higher frequency of areas with either many hits or no hit and a deficiency in the intermediate classes. Table 4 indicates perfect randomness and homogeneity of the area; we have here an instructive illustration of the established fact that to the untrained eye randomness appears as regularity or tendency to cluster. — William Feller, An Introduction to Probability Theory, Vol. 1, §VI.7

The clusters people saw were real as shapes and empty as evidence: scatter 537 points at random and you always get knots and voids, and the eye always reads them as design. That reading has a name — the clustering illusion — and you can manufacture it yourself.

The honest seam — random enough to hide the aim

It would be a lie to stop at "the bombs fell at random," and the venue does not tell that lie. The V-1 was not unaimed; it was badly aimed. It carried no guidance that could see London at all — two gyroscopes held pitch and yaw, a magnetic compass held the heading, and a little propeller on the nose counted off the air it had flown; when the count ran out the controls slammed and it dived. The notional aim point was Tower Bridge; early machines landed within a circle some 30 km across, and no V-1 ever hit the bridge. The campaign's impacts fell short of central London, skewed into the southern suburbs — pulled there partly by a real British deception: MI5's double agents (the Double-Cross System) over-reported the bombs falling beyond the centre, hoping the Germans would shorten the range and drop them shorter still, into the very South London boroughs Clarke later gridded. That the deception was run is established; how much it actually moved the mean impact point is debated — the German command partly cross-checked the reports against radio range-data and discounted them — so we name it, and don't overclaim it.

So what Clarke's test actually shows is subtler than "pure chance." Across a quarter-square-kilometre grid, the weapon's aiming error was so much larger than the grid that the residual scatter was, at that resolution and in that place, statistically random. The aim was real; it was simply invisible at the scale he measured — and it does not stay invisible. In 2019 two researchers (Liam and Luke Shaw, writing in Significance) rebuilt the impact map from the original London County Council bomb-damage records and re-ran the test. Inside Clarke's own southern region the Poisson fit still holds (they get p ≈ 0.18). But widen the window to the whole London County area and it shatters: p ≈ 5 × 10⁻⁶, the hits decisively denser to the south, decaying with distance from the aim point. Clarke's result was true because he had chosen a patch homogeneous enough to hide the gradient. Both statements hold at once, and holding them together is the whole point: a test that fails to reject randomness has not proven randomness — it has only failed to find structure at the scale it looked. Clarke's grid was blind to the aim; a wider grid was not. The clustering illusion is real, and so was the bombsight.

Why this is a reproduction, not a story

Clarke published a single table — how many squares received how many hits — and not the map of which square got what. So the frequency table is the primary datum, and it is the only thing under the check here; the grids you drop above are honest illustrative realisations that match that table's shape, never Clarke's own arrangement, which does not survive. From the table alone, the browser recomputes the mean (0.9323 hits per square), the full Poisson fit — reproducing William Feller's published expected counts of 226.74, 211.39, 98.54, 30.62, 7.14, 1.57 to the second decimal — the goodness-of-fit χ² ≈ 1.17, and the dispersion index ≈ 1.006. A separate engine drops 537 bombs uniformly at random 20,000 times and the distribution it grows lands on the same Poisson, with no Poisson ever assumed. The offline verifier runs all of it, plus a control proving a truly clustered field is caught, and passes 22 of 22.