The uniform chain was the worst one to start from
Worth reading first: No one distance serves three chains · Length did not rescue the consecutive triples.
A counterpoise correction for a cluster of fragments can be assembled from pieces rather than computed whole: each fragment’s correction from every other fragment’s functions taken one at a time, then the three-body increments, which add what two sets of ghost functions do together that neither does alone. Keeping every triple costs more than the calculation the assembly stands in for, so a real calculation keeps only some, and the question has been which. Keeping the consecutive triples, the ones whose three members span two separations, failed on a uniform chain at every length tried. A cutoff on distance found that the uniform chain wanted triples reaching four separations, the chain with heavy fragments at its ends wanted them to stop at two, and the chain with heavy fragments inside did not care.
Three arrangements, three answers. The essay that found them left the obvious continuation: a chain that is none of the three, with light and heavy fragments in some pattern, sits somewhere between the curves, and the question is whether its best cutoff can be read off its composition before the sweep, or has to be found every time by running it.
The chains are cheap enough that the question does not need a sample. Every composition of five to eight fragments, each light or heavy, with a chain and its mirror image counted once, is twenty, thirty-six, seventy-two and 136 chains, and each can be swept from the consecutive triples to every triple. The census answers a question nobody had asked, and answers the one that was asked only by halves.
Almost every chain is served by the triples that were given up on
The measure is the one the assembly essays have used since the pairwise assembly first counted one share twice: the share of three decades of the accuracy line, from a thirtieth of a kilocalorie a mole to thirty, over which some basis size has an assembly error below the line and a correction above it. For each chain, the loss is how much of that line the consecutive triples cover less than the best cutoff that chain has.
On eight fragments, 126 of the 136 chains are best at span two, and 127 are within one point of their best there. The mean covered share across all 136 is 86.0 per cent with the consecutive triples, 86.6 with each chain’s own best cutoff, 83.6 with every triple, and 66.1 with pairs alone. The same holds at every length: within a point on 17 of 20 chains at five fragments, 33 of 36 at six and 66 of 72 at seven. Across all the ways of arranging two kinds of fragment, the cheapest triple rule is close to the best one.
And the chain on which it fails worst is the uniform one, at every length. It loses 27.4 points at five fragments, 31.5 at six, 33.1 at seven and 36.0 at eight, and the next worst chain at each length is a long way behind: a light chain of five with a heavy centre in the middle at 18.0, then 9.4, 13.3 and 14.2. Pairs alone beat the consecutive triples on only three chains of the 136, and the uniform chain is one of them.
That changes how the earlier result should be read. The overshoot that sank the consecutive triples was found on the uniform chain, and the uniform chain was studied first because it is the obvious first case: every fragment the same, nothing to choose. It is also the case for which the consecutive rule is least suited of any arrangement of these fragments. The verdict that the consecutive triples are a poor truncation was a verdict on the worst composition, and a light chain with alternating heavy fragments goes from 39.4 per cent with pairs to 92.9 with the consecutive triples, the best result in the census.
Where the heavy fragment sits matters more than that it is there
The census says the uniform chain is unusual. The plainest way to see what makes it so is to break its uniformity once, by making one fragment heavy and moving that fragment along the chain.
A heavy fragment at one end makes span two the best cutoff outright: 84.6 per cent, falling to 68.5 at span three and staying near 70 after that, with no loss at all from the consecutive rule. One place in, the best cutoff is three, at 84.7, and the consecutive triples lose 5.9 points. Two places in, the best is four, at 73.2, and they lose 6.7. At the middle of the chain the best is two again, but every cutoff is poor, 66.3 at best against 84.6 with the heavy fragment at the end.
So one substitution moves the best cutoff from two to three to four and back to two as it travels four places. A rule that read the cutoff off the number of heavy fragments would give one answer for all four of these chains, and they want three different ones. Neither does the distance a light fragment is from the nearest heavy one sort them: the chain with its heavy fragment at the middle has its light fragments closer to a heavy one, on average, than the chain with it two places in, and it wants a shorter cutoff, not a longer one. The behaviour is not monotone in any reading of the composition that is simpler than the composition itself.
A triple knows only its three members
There is one thing that can be said about every chain before solving any of them, and it is exact. A triple’s increment is the corrected fragment’s correction with two ghosts present, less its corrections with each ghost alone. Each of those is a calculation with only the corrected fragment and the named ghosts in it; the rest of the chain is not there. So a triple’s increment depends on its three members’ charges and their two offsets, and on nothing else in the chain. Every chain’s three-body sums are sums of entries from one table, with eight patterns of charge and as many shapes as there are offsets.
The table says where the far increments come from. With both ghosts on one side, three and four separations away, a triple of three light fragments has a median increment of of the accuracy line. Make the further ghost heavy and it falls sevenfold; make both ghosts heavy and it falls 250-fold; make the corrected fragment heavy too and it is , four hundred thousand times smaller. A heavy fragment’s functions are tighter, because they were optimised against a nuclear charge of two, so as a ghost it lends less — the same borrowing of another atom’s functions that makes a counterpoise correction necessary in the first place, read in reverse — to a neighbour three or four separations away, and two such ghosts lend still less together. The far triples that the uniform chain needs to cancel its overshoot exist only between light fragments.
The near triples go the other way. With ghosts one and two separations away on one side, a heavy corrected fragment with two light ghosts has a median increment of , seven times the all-light triple’s. A heavy fragment’s own description is the one that most needs borrowed functions, and two light neighbours on one side supply them in a way that does not add. That is a large consecutive increment with nothing far away to cancel it, and it is why a heavy fragment at the end of a chain suits the consecutive rule so well.
Rebuilt member by member from the table, every span’s three-body sum for a sample of chains agrees with the chain’s own to . The table is not an approximation to the census; it is the census’s three-body half, factored.
Composition does not name the exceptions
The table explains the uniform chain. It does not, on its own, say which other chains need more than the consecutive triples, and two readings of composition that look as though they should were tested.
The first is the longest run of light fragments, the obvious reading of the table: far triples live among light fragments, so a long light run ought to want a long cutoff. Its rank correlation with the loss across the 136 chains is −0.22, the wrong sign. The single-substitution chains show why: with the heavy fragment at the end the run is seven and the loss is nothing; two places in, the run is five and the loss is 6.7 points. The second reading is closer to the mechanism: the median, over the grid, of the far triples’ sum against the consecutive triples’ sum, which needs the triple solves but not the exact correction. Its rank correlation is −0.09. A chain of four light, three heavy and one light fragment has a far sum 3.07 times its consecutive sum and loses nothing; the chain with heavy pairs at both ends and a light run of four between them has a ratio of 2.83 and loses 14.2 points.
The reason neither works is visible in the shape of the measure. The loss is not set by how large the far increments are; it is set by whether they cancel something the consecutive triples overshot, at the basis size and separation where the assembly is closest to failing. That is a property of the signs of a few dozen entries at a few cells, and it does not reduce to a count.
A convergence rule is a bet on the next shell
A calculation that cannot afford the exact correction can still watch its own triples. The rule a careful practitioner would use is a convergence rule: widen the cutoff, span by span, until the next span adds less than some fraction of the triples kept so far, and stop there. It needs no reference, and it can be applied cell by cell, so that each basis size and separation chooses its own span.
On the uniform chain the rule works. It widens until the span-four shells have cancelled the overshoot and stops, at spans from two to six depending on the cell, and the covered share goes from 48.1 to 84.1 per cent, as good as the best single cutoff. It gains more than a point on six chains, the uniform one among them. On 42 of the 136 it does worse than simply keeping the consecutive triples, and on the chain with heavy fragments at both ends it does worse by 24.5 points: 63.3 per cent against 87.9. That chain’s far shells are small but never smaller than five per cent of what is kept, so the rule keeps widening towards every triple, which on that arrangement is further from the answer than the consecutive part alone. Over the census its mean is 84.0 per cent, below the consecutive triples’ 86.0. Loosening the threshold to ten or twenty per cent does not change the count much: 41 and 36 chains are made worse.
So the convergence rule is a bet that a shrinking shell means a converging sum. On the uniform chain the bet pays because the shells really do settle. On a chain whose three-body order overshoots as a whole, a shrinking shell is a sum heading steadily towards the wrong total, and the rule cannot tell the two apart without the reference it was invented to avoid.
The chains that need more
Twenty chains across the four lengths lose more than two points with the consecutive triples, eight of them at eight fragments. Their longest light runs go from two to eight, so the table’s condition for far triples — light fragments three and four separations apart — is necessary and nowhere near sufficient: at eight fragments, 97 chains with a light run of two or more lose nothing at all. In every case but one the best span is three or four, one or two separations beyond the consecutive rule; the exception is the uniform chain of eight, which wants five. And in seven of the twenty a lost point for the consecutive rule is a gain for pairs alone: the uniform chain at each of the four lengths, and three chains with heavy pairs at both ends and a light run between. On the other thirteen the consecutive triples are worth having even where they are not the best cutoff.
How the census was solved
Each fragment’s counterpoise correction from any set of ghosts is the one-dimensional calculation used throughout these essays: one electron per fragment, a set of one to six s Gaussians, the wrong shape for an atom but the one every real basis uses, whose exponents were optimised for the fragment’s nuclear charge, one on a light centre and two on a heavy one, the fragment’s own nucleus present and every ghost’s charge set to zero. The four separations are 1.6, 2, 2.5 and 3 bohr, and each energy is the lowest root of the generalised eigenvalue problem. Every triple’s increment is filed by its span, so each chain’s covered share at every cutoff is read from one set of solves; across the four lengths that is 264 chains and about 1.18 million solves.
The checks run wherever these figures are drawn. The three arrangements already published are rows of the census, and their covered shares at every span come out identical to the distance-cutoff sweep’s, to . The uniform chain is confirmed the worst composition at each of the four lengths. Every span’s three-body sum on four sample chains of seven fragments, at six cells each, is rebuilt from three-centre solves of the triple table and agrees to . The refusal tests that last agreement: rebuilt with one stray ghost from outside the triple left in every table solve, the sums must stop matching, and they do by far more than the tolerance. A table that matched either way would be bookkeeping rather than a statement about the chain.
What a row of two kinds of fragment cannot say
Two kinds of fragment. The census is complete for its alphabet and says nothing about a third. A fragment whose functions are more diffuse than the light one’s would make the far triples larger rather than smaller, and the uniform chain’s failure might then stop being the worst case. The table structure carries over exactly: a triple still knows only its members, and a new kind of fragment adds rows rather than changing old ones.
A line. On a line a triple’s span and its geometry are the same thing. In three dimensions a compact triple and a far one along any ordering coincide, and a cutoff on span has no single meaning.
One figure of merit. The covered share is set by the worst cell at each basis size, against an accuracy line held fixed although the whole picture is most sensitive to it, so the losses here are losses at the cells where the assembly is nearest to failing. The convergence rule’s damage is in exactly those cells; a rule judged by its typical error might look better.
Four bodies and beyond are not truncated. Every assembly here keeps all pairs and a subset of triples, and the exact correction contains every order. The heavy-ends chain’s damage under the convergence rule is a four-body story, and nothing in this census reads the four-body remainder directly.
The case that looked simplest was the hardest
There is a habit in testing a method of starting from the most symmetric case, on the grounds that if the method fails there it fails everywhere. The uniform chain was that case, and the consecutive triples failed on it, and the failure was traced to one arrangement of ghosts whose overshoot only the span-four triples cancel. Every step was correct. What was not established was that the symmetric case was representative, and the census shows it is the least representative arrangement there is: the only composition in which every far triple is an all-light triple, and therefore the one in which the far order matters most.
A correction that turned out to be two functions and a repair that cost more than the whole both had the same structure: a conclusion drawn on one configuration, found later to belong to that configuration. The difference here is that the census can be completed. There are 136 chains of eight and all of them were solved, so the statement that the consecutive triples serve nearly every arrangement is a count, not a sample.
Still open: a third kind of fragment, and the cells that decide
The obvious open question is a third kind of fragment. The table says the uniform chain is the worst case because its far triples are all light, so a fragment more diffuse than the light one would test the explanation directly: if the far increments grow, a chain of diffuse fragments should lose more than the uniform light chain does, and the worst composition should move. With three kinds the census at eight fragments is several thousand chains rather than 136, but the table is twenty-seven patterns, and the three-body half of every chain is a sum over it.
The nearer question is which cells decide. Every loss here is set at the basis size and separation where the assembly’s error is closest to the correction’s size, and the convergence rule fails in exactly those cells. Reading, for each chain, the one cell that sets its covered share would say whether the exceptions share a cell — a particular basis size at the shortest separation, perhaps — in which case the cutoff could be chosen per basis size rather than per chain, and the composition would not need to be read at all.
What links here
Computed from the collection rather than written here: the essays that point at this one.
Reads more easily once this is understood
Essays that name this one as worth reading first.
Named objects
A dashed tag is an object no other essay names yet.
Basis set superposition errorCounterpoiseFragment methodGaussian basisMany-body expansion