When the molecule does not stop

A defect's symmetry makes it rarer

A random square net with a fraction p of its sites low holds each of the thirty-five six-site shapes as a maximal cluster at a·p⁶(1−p)ᵗ per site, with a the shape's number of orientations and t its boundary. A simulated net of four million sites agrees at every filling tried. No shape ever exceeds two clusters in ten thousand sites, so at every filling a typical cluster's nearest copy lies beyond every coupling yet computed. Abundance and reach are nearly unrelated: one of the three commonest shapes binds deepest and reaches least, and the shape that reaches furthest, the compact block, loses its place to a three-by-three tree because the block has two orientations and the tree four.

Worth reading first: Two different defects rarely resonate · The run was one of thirty-five.

A square net with some of its sites lowered in energy binds electrons on clusters of those sites, each cluster a level in the gap below the host band. The run was one of thirty-five gave each of the thirty-five shapes six adjacent low sites can take its levels below the host band and the rate at which its coupling to a copy of itself decays, and two different defects rarely resonate found that clusters of different shapes keep their states to themselves unless their levels happen to agree to a part in a thousand. So a net’s six-site defects are thirty-five separate populations, each coupling only to its own kind.

Which of those populations matters depends on two things, and only one has been measured. The reach of a shape — how far its level’s coupling extends — is in hand. Its abundance is not. A shape that reaches far but occurs once in a million sites may couple less, in a real net, than a shape that reaches a little less and occurs a hundred times as often, because what decides a coupling is the distance to the nearest copy, and that distance is set by how many copies there are. The chain’s own runs showed the same trade in one dimension: runs of four, commoner and shallower, reached a band at a size twenty-six decades smaller than runs of six.

The abundance turns out to have a closed form, to be almost unrelated to reach, and to be governed by something no picture of a cluster’s level shows: how many ways the shape can sit in the net.

A shape’s count is its boundary and its orientations

Take a net in which each site is low independently with probability p. A particular set of six adjacent sites is a cluster of exactly six — a maximal cluster, not part of a larger one — when all six are low and every site touching them from outside is not. Those two sets of sites are disjoint and every site is independent, so the probability is a product. Counting every placement of a shape, its expected number per site is

ns(p)=as p6 (1−p)ts,n_s(p) = a_s\,p^{6}\,(1-p)^{t_s},

where asa_s is the number of distinct orientations the shape has in the net and tst_s is its site perimeter, the number of distinct sites adjacent to it from outside. Both are fixed by the shape’s geometry and unchanged by any of the net’s eight symmetries.

The orientations are eight divided by the number of the net’s symmetries that leave the shape unchanged. A shape with no symmetry has eight; one with a mirror line or a half-turn has four; the straight run and the two-by-three block, each with two mirror lines, have two. Summed over the thirty-five shapes the orientations come to 216, which is the number of fixed hexominoes. The perimeters run from nine to fourteen. Only one shape has nine, a square of four with a site added on two adjacent sides; only the straight run has fourteen, since a line of six is all boundary.

The formula has one consequence that needs no numbers. Every shape’s abundance rises from nothing as p6p^6, peaks, and falls as the boundary fills, and the peak sits where the derivative of p6(1−p)tp^6(1-p)^t vanishes, at

p∗=66+t.p^{\ast} = \frac{6}{6+t}.

Every six-site shape is commonest at the same kind of filling, 6/(6+t), and none ever exceeds two in ten thousand sites. The expected number per site of maximal clusters of each of the thirty-five free hexominoes, a·p⁶(1−p)ᵗ, against the fraction p of low sites, on a logarithmic scale. run: 2 orientations, perimeter 14, peak 9.9×10⁻⁶ at p = 0.300; block: 2 orientations, perimeter 10, peak 5.1×10⁻⁵ at p = 0.375; shape 33: 4 orientations, perimeter 10, peak 1.0×10⁻⁴ at p = 0.375; shape 17: 8 orientations, perimeter 10, peak 2.0×10⁻⁴ at p = 0.375. The largest any shape reaches is 2.0×10⁻⁴.
Fig. 1 The expected number per site of maximal clusters of each of the thirty-five shapes against the fraction of sites low, on a logarithmic scale, with four named shapes drawn and each one’s peak at 6/(6+t)6/(6+t) marked.

The orientations multiply the curve and do not move it. Shapes with a perimeter of ten — held by eleven shapes and 54 of the 216 orientations — all peak at three-eighths of the sites low. The run, with the longest boundary, peaks earliest, at 0.300; the one shape with a perimeter of nine peaks latest, at 0.400.

And none of them peaks high. The largest value any shape’s curve reaches is 2.0 × 10⁻⁴ clusters per site, for the commonest shapes at their own peak. All thirty-five together reach 3.5 × 10⁻³ per site at a filling of 0.358. A cluster of exactly six is a rare object at every filling, because six low sites must be fenced off by between nine and fourteen high ones, and the fence costs as much as the cluster.

Four million sites count them where the formula does

A formula with a product of independent probabilities in it is easy to write down wrongly, and the commonest error is the one the next paragraph tests. So the abundances are counted as well. A square net of 2,000 by 2,000 sites is generated from a seeded random sequence at each of three fillings, its low sites are joined into clusters by union–find, every cluster of exactly six is reduced to its canonical shape, and the shapes are tallied. A cluster touching the net’s edge is discarded, since part of its boundary lies outside the net, which is the decision the plane’s run census made for the same reason.

A simulated net of four million sites counts each shape where the closed form puts it, at every filling. Each shape's count of maximal six-site clusters in one seeded 2000 × 2000 net, against a·p⁶(1−p)ᵗ times the net's interior sites, at three fillings, on logarithmic axes. p = 0.2: 4620 clusters against 4675 expected, χ² 25.7 on 34 degrees of freedom; p = 0.35: 14254 clusters against 14054 expected, χ² 40.6 on 34 degrees of freedom; p = 0.5: 7204 clusters against 7256 expected, χ² 51.7 on 34 degrees of freedom. The same counts against the formula with the bond perimeter in place of the site perimeter give χ² 1757, 24225, 41634.
Fig. 2 Each shape’s count of maximal six-site clusters in a seeded net of four million sites, against the closed form, at three fillings. The dashed line is equality.

At p = 0.2 the net holds 4,620 six-site clusters against 4,675 expected; at 0.35, 14,254 against 14,054; at 0.5, 7,204 against 7,256. Shape by shape, the squared deviations summed over the thirty-five come to 25.7, 40.6 and 51.7, against 34 degrees of freedom, and no shape at any filling lies more than 2.74 standard deviations from its expected count. Nothing is fitted. The two routes share nothing but the definition of a cluster.

The error worth refusing is the perimeter. A cluster’s boundary can be counted as sites or as edges — the edges from a cluster site to a non-cluster site — and the two differ wherever one outside site touches two cluster sites, which happens at every inside corner. The edge count is what a careless reading of “every neighbour must be high” gives, and it charges an inside-corner site twice. Put into the same formula, it fails the same simulated counts by squared deviations of 1,757, 24,225 and 41,634. The agreement above is therefore a measurement that could have failed, and the site perimeter is the quantity it selects.

Abundance and reach are nearly unrelated

With the counts in hand, each population’s size can be set beside the reach measured for it. The reach here is the decay length of a shape’s coupling to a copy of itself: the slope of the logarithm of the splitting between ten and twenty-eight sites, taken along the row and along the column and averaged geometrically.

How common a shape is and how far its level reaches are nearly unrelated: the commonest reaches least, and the farthest-reaching is held back by its symmetry. Each of the thirty-five shapes at filling p = 0.35: its abundance per site against the decay length of its coupling to a copy of itself (geometric mean of row and column, from splittings at ten to twenty-eight sites). Rank correlation -0.17. shape 17: decay length 0.96, 2.0×10⁻⁴ per site; shape 33: decay length 3.96, 9.9×10⁻⁵ per site; block: decay length 4.18, 4.9×10⁻⁵ per site; run: decay length 1.85, 8.8×10⁻⁶ per site.
Fig. 3 Each shape’s abundance at p = 0.35, against the decay length of its coupling to a copy of itself. The commonest shape reaches least; the two that reach furthest are the block and shape 33.

The two quantities are close to independent, with a rank correlation of −0.17 across the thirty-five. That is the expected answer once the formula is read. Abundance depends on the orientations and the boundary and on nothing else — not on the cluster’s interior bonds, not on where its level sits. Reach depends on the depth of the shallowest level, which the census of shapes found sets the decay to within about a quarter, and depth depends on how the six sites are joined. The two are different questions about the same six sites.

The extremes are where the picture is sharp, and they run the wrong way for a band. Three shapes tie as the commonest at every filling, because they share the largest orientation count, eight, with the commonest small perimeter, ten: 6.7 per cent of all six-site clusters each at half filling. One of the three is shape 17, three by three with a square of four inside it. Its shallowest level is the deepest of the thirty-five, 1.10 below the band, and its coupling decays in 0.96 sites, the shortest reach in the census. The other two reach 1.6 sites, below the median. The net produces most often a family of defects that couple little, and the least-coupling of all is among them. The run, the shape every earlier calculation of the plane’s band was built on, is near the other end of the abundance scale: two orientations and the longest boundary make it 0.25 per cent of six-site clusters at p = 0.35 and 0.10 per cent at half filling.

The two shapes that reach furthest are the block, two by three with a shallowest level of 0.044 and a decay length of 4.18 sites, and shape 33, a three-by-three tree — a row of three with one site above its middle and one below each end — at 0.050 and 3.96 sites. They have the same perimeter of ten. They differ in symmetry. The block is unchanged by both mirror lines of the net and by a half-turn, so it has two orientations; shape 33 has one mirror line and four orientations. At every filling there are exactly twice as many of the tree as of the block, because the formula gives them the same curve times their orientation counts.

Every nearest copy is past the last computed coupling

What a population’s size buys is a shorter distance to the nearest copy. If copies were scattered independently at density n, the chance that none lies within a distance s of a given cluster would be exp⁡(−nπs2)\exp(-n\pi s^2), and half of all clusters would have a copy closer than

s1/2=ln⁡2πn.s_{1/2} = \sqrt{\frac{\ln 2}{\pi n}}.

That half-way separation is the quantity the plane’s resonance argument was built on. A net’s copies are not quite independent — a cluster’s own sites and fence exclude other clusters — so the simulated nets measure the nearest-copy distances directly as well.

At the filling where six-site clusters are commonest, every shape's typical nearest copy lies beyond the last computed coupling. For each shape at p = 0.35, the median distance between a cluster's centroid and its nearest same-shape copy in the simulated net (dots), against √(ln 2 / πn) for copies scattered independently at the shape's density (bars). Couplings were computed out to twenty-eight sites. The nearest median is 32 sites; the run's is 154. Share of clusters with a copy within twenty-eight sites: largest 40 per cent, block 8, run 0.
Fig. 4 The median distance from each cluster to its nearest copy of the same shape at p = 0.35, simulated and from independent scatter. The shaded band is where couplings have been computed.

At p = 0.35, where six-site clusters are commonest, the two agree to a few sites for every shape with enough copies to measure. The simulated medians run from 32 sites for the commonest shapes to 154 for the run, whose 42 clusters in the whole net have no copy within twenty-eight sites at all. Not one shape has its typical nearest copy inside the range where couplings were computed. At most 40 per cent of any shape’s clusters have a copy within twenty-eight sites; for the block it is 7.8 per cent.

The bound is general, not a feature of one filling. Since no shape’s abundance exceeds 2.0 × 10⁻⁴ per site, no shape’s half-way separation falls below 33 sites at any filling, and that minimum belongs to the three commonest shapes, one of which reaches least of all. Every pair the plane’s earlier essays computed at ten, twelve or sixteen sites was a close pair, drawn from the near tail of the distance distribution rather than its middle — a deliberate choice for measuring how a coupling decays, and the wrong place to read what a typical cluster sees.

Which population couples most

Reading the coupling at the half-way separation needs it at distances the calculation did not reach. The coupling at twenty-eight sites is known for every shape, and each shape’s own decay length was fitted between ten and twenty-eight; continuing the logarithm at that slope gives an estimate. It is an extrapolation, by a factor of 1.2 to 6 in distance at p = 0.35 and by far more at low fillings, and it is drawn dashed for that reason. What it can be trusted for is a ranking and an order of magnitude, since the decay lengths were found to follow the level’s depth closely and an error of a quarter in the decay length moves the answer by a quarter of its exponent.

The widest sub-band belongs to a shape nobody singled out, and it is widest near a third of the sites low. The coupling of a shape to its own copy at the median nearest-copy distance for independent scatter, as a power of ten, against the filling: computed at twenty-eight sites and continued at the shape's own decay length beyond. shape 33: -267.5, -44.7, -18.2, -10.8, -8.0, -6.7, -6.2, -6.2, -6.6, -7.6, -9.4; shape 26: -251.4, -43.1, -18.1, -11.0, -8.3, -7.2, -6.9, -7.1, -7.9, -9.5, -12.4; block: -357.9, -59.5, -24.1, -14.2, -10.4, -8.7, -8.0, -8.0, -8.6, -9.9, -12.3; shape 17: -780.6, -127.7, -50.3, -28.6, -20.2, -16.5, -15.0, -15.0, -16.3, -19.2, -24.4; run: -893.2, -163.2, -72.2, -46.7, -37.6, -35.3, -37.3, -43.7, -56.4, -80.0, -125.6 at p = 0.05, 0.1, 0.15, 0.2, 0.25, 0.3, 0.35, 0.4, 0.45, 0.5, 0.55. The widest at any filling is shape 33's, 10^-6.2 at p = 0.4.
Fig. 5 The coupling of each named shape to its own copy at the half-way separation, as a power of ten, across the filling. The strongest belongs to shape 33 from a fifth of the sites low upward, and peaks near a third.

The strongest typical coupling at any filling from 0.2 to 0.55 is shape 33’s. It peaks at 10⁻⁶·², at a filling of 0.35 to 0.40 where the shape’s half-way separation is 47 sites. At lower fillings, where every population is sparse and the distances are hundreds of sites, the order is set by decay length almost alone and shape 26, a tree with eight orientations, leads below a filling of 0.2.

The block, which has the longest reach and couples most strongly of all at twenty-eight sites, 1.0 × 10⁻⁴ against shape 33’s 8.4 × 10⁻⁵, is not the leader at any filling. Half as common, its half-way separation is 2\sqrt{2} times longer — 67 sites against 47 at p = 0.35 — and twenty extra sites at a decay length of about four are worth nearly two decades. The block’s two mirror lines cost it a factor of about sixty in its typical coupling. No property of the block’s level enters that factor; it is the count of ways a two-by-three rectangle can lie in a square net.

The commonest and the most studied fare worse still. Shape 17 has a typical coupling of 10⁻¹⁵ at its best filling, its short reach outweighing its numbers. The run’s best is 10⁻³⁵, at a filling of 0.30, about twenty-nine decades below shape 33’s best.

That ranking is also the order in which the populations would become bands as a sample grows. The level spacing of an N-site sample falls as 1/N, and a population’s typical pair resonates once the spacing falls below the coupling at its half-way separation, so each decade of coupling is a decade of sample size. The coupling at the half-way separation is not the population’s typical coupling in the sense of an average — an average over these distributions means nothing — but the threshold the count of resonant pairs turns on: half the population has a pair at least that strong. Shape 33’s population is the first to reach that point and the run’s is, on these numbers, never going to.

The twins rarely sit together

Two different defects rarely resonate left four groups of shapes whose shallowest levels agree to within 10⁻³ — four shapes near 0.098, two near 0.194, two near 0.394, two near 0.504 — and found that such a pair shares its state only while it sits close: shapes 6 and 30 still shared 0.47 at sixteen sites, shapes 1 and 32 had fallen to 0.08 there. Whether the twins merge their populations in a real net is a question about how often a twin sits that close, and the abundances answer it.

The accidental twins: how often each pair of near-degenerate shapes actually sits close enough to share a state. The pairs of different shapes whose shallowest levels lie within 10⁻³ of each other, at p = 0.35 in the simulated net: the four groups hold 10.9, 3.9, 7.3, 4.0 per cent of all six-site clusters. For each pair, the share of the first shape's clusters that have a cluster of the second within twelve and within sixteen sites: 8 and 11, 2.2 and 3.2 per cent; 11 and 30, 5.6 and 9.6 per cent; 6 and 11, 2.6 and 3.4 per cent; 8 and 30, 4.0 and 7.3 per cent; 6 and 8, 6.5 and 11.5 per cent; 6 and 30, 4.4 and 9.1 per cent; 1 and 32, 4.8 and 7.2 per cent; 10 and 19, 7.1 and 10.9 per cent; 4 and 16, 7.0 and 9.6 per cent.
Fig. 6 Each pair of accidental twins at p = 0.35: the share of all six-site clusters the group holds, and how often a cluster of the first shape has one of the second within twelve and sixteen sites in the simulated net.

The four groups hold between them a large share of the clusters — 10.9, 3.9, 7.3 and 4.0 per cent at p = 0.35 — but a twin partner within twelve sites is rare: from 2.2 to 7.1 per cent of a shape’s clusters, depending on the pair, and from 3.2 to 11.5 per cent within sixteen. For the pair that holds its resonance furthest, shapes 6 and 30, 4.4 per cent of shape 6’s clusters have a shape 30 within twelve sites. The accidental twins are therefore a correction of a few per cent to the populations they belong to, and that is its stated size: for most of their clusters the twin partner is as far away as any copy, and at those distances, as the half-way couplings above show, no coupling survives that could overcome a detuning of even a few millionths.

The same simulated nets say something about the neighbours a cluster does have. At p = 0.35, 80 per cent of six-site clusters have some other six-site cluster within twelve sites. Almost all of those neighbours are of a different shape, and different shapes, as the pair census found, detune each other by a median of 0.195 and mix by less than a hundredth. The crowding is real and nearly inert.

The run in a row was rarely a cluster of six

One earlier object deserves a second look in this light. The plane’s run census counted horizontal runs of exactly six low sites in a row, bounded by a high site at each end, at a density of x6(1−x)2x^6(1-x)^2 per site — 1/256 at half filling. It said plainly that this is a one-dimensional object put in a plane, the right one for a layered material where a defect runs along one direction, and not a two-dimensional cluster.

The closed form says how different the two objects are. A row-run is a maximal cluster only if the twelve sites above and below it are high as well, so the maximal horizontal run occurs at x6(1−x)14x^6(1-x)^{14}, and at half filling only one row-run in 4,096 is a cluster of exactly six. The rest are parts of larger clusters, whose levels are not the run’s. The earlier census was internally consistent and correctly labelled; what changes is how the result should be read for a net that is random in both directions, where the row-run’s density is not the density of any defect the net actually holds.

How the counts were made

Each of the thirty-five free hexominoes was generated by growing clusters cell by cell and reducing them under the square’s eight symmetries, as in the census of shapes. Its orientations are eight over the number of symmetries that fix it, and its site perimeter the number of distinct outside neighbours of its six cells. The closed form was checked against a scan of each curve for the position of its maximum to a part in ten thousand.

The simulated nets are 2,000 by 2,000 sites, one for each of p = 0.2, 0.35 and 0.5, each from its own seeded random sequence. Low sites were joined by union–find over nearest neighbours; clusters touching the edge were discarded; each six-site cluster was reduced to the canonical form of its cells and matched to its shape. Nearest-copy distances are between cluster centroids, computed for every cluster of every shape.

Levels and couplings are those of the shape census: host sites at energy 4 with unit hopping, cluster sites at 0, the exact Green’s function of the square net, and the splitting of the shallowest level when a second copy sits at ten, sixteen, twenty-two and twenty-eight sites between centroids along each axis. The decay lengths are straight-line fits to the logarithm of those splittings, and the half-way couplings continue the logarithm from twenty-eight sites at that slope.

What a census of isolated clusters leaves out

The populations here are clusters of exactly six. A random net at these fillings is mostly made of other things: at p = 0.35 the simulated net held 250,732 isolated low sites, 74,040 pairs and 49,721 clusters of seven or more, against 14,328 clusters of six. Every one of those binds levels of its own or none, and every one lies between the six-site clusters whose couplings are being compared. The couplings were computed for two clusters alone in a perfect host. Between two real copies, forty sites apart, there are on average hundreds of other low sites, each of which scatters the tail of the state that is supposed to carry the coupling. Whether that lengthens or shortens the reach is not known from anything here.

The half-way couplings also rest on an extrapolation, from twenty-eight sites to between 33 and several hundred. Nothing in the shape census suggests the decay changes character beyond twenty-eight — the decay lengths followed the levels’ depths across all thirty-five — but a factor of two in distance is a factor of two in the exponent’s lever, and the figures say so.

And the abundances are for site percolation with independent sites. Real alloys have short-range order: like atoms attract or repel, and the probability that a site is low depends on its neighbours. Any such correlation changes p6(1−p)tp^6(1-p)^t into something that depends on the cluster’s shape beyond its boundary count, and could reverse the block’s disadvantage or deepen it.

Still open: a cluster in a crowd, and a net with order in it

The obvious open question is the crowd. Two copies of shape 33 forty-seven sites apart in a net at p = 0.35 have, between them, a few hundred low sites in clusters of every size, and the coupling between the two copies is carried through all of them. A calculation of the coupling with the intervening sites drawn from the same random net — the two clusters fixed, the rest of the net resampled — would say whether the half-way couplings above are a fair guess or whether the crowd changes them by orders of magnitude, and it needs only the same Green’s-function method applied to a larger defect.

The nearer question is short-range order. A single parameter, the probability that a low site’s neighbour is also low beyond what independence gives, changes every shape’s abundance differently, because compact shapes have more internal neighbours than trees. A modest preference for like neighbours would favour the block and shape 17 over the trees. Whether any physically plausible amount of order hands the strongest coupling back to the block — whether its symmetry penalty can be paid by chemistry — is one exponent’s worth of algebra away.

Shares its objects with

Essays naming at least two of the same things, that neither author linked.

Named objects

A dashed tag is an object no other essay names yet.

Defect stateDimensionalityDisorderResonanceTight-binding models