The far triples need a diffuse ghost
Worth reading first: The uniform chain was the worst one to start from · No one distance serves three chains.
A counterpoise correction for a cluster can be assembled from pieces: each fragment’s correction from every other fragment’s functions alone, then the three-body increments that two sets of ghost functions add together. The practical question has been which triples to keep, and a distance cutoff found that the uniform light chain wanted triples reaching four separations while chains with heavy ends wanted them to stop at two. The census of every composition of light and heavy fragments then turned that around. On 127 of 136 chains of eight, keeping only the consecutive triples came within a point of the best cutoff, and the uniform light chain, where the question had started, was the single worst arrangement there was.
The explanation was a table. A triple’s increment is a calculation with only its three members present, so it depends on their kinds and offsets and nothing else, and the table said that far increments — the ones a wide cutoff is needed to catch — exist only among light fragments, because a heavy fragment’s functions are too tight to lend anything three or four separations away. That essay left a direct test: a third kind of fragment, more diffuse than the light one, whose functions reach further. If the table is right, its far increments should be larger, and the worst composition should move.
It moves, and the census was right to be suspicious of its own good news: with three kinds of fragment the consecutive rule fails far more often. But the composition that fails worst is not the one the reasoning pointed at, and the table says why.
A diffuse fragment, and a census that needs few solves
The diffuse fragment is a one-electron centre with nuclear charge 0.6, its Gaussian exponents optimised for that charge exactly as the light fragment’s were for one and the heavy fragment’s for two. With three Gaussians its exponents are 0.054, 0.245 and 1.62 against the light fragment’s 0.151, 0.681 and 4.50 — every function about 2.8 times broader in its exponent, which is an orbital that reaches about two-thirds again as far. Every composition of five, six and seven fragments from three kinds, a chain and its mirror image once, is 135, 378 and 1,134 chains.
Solving each chain fragment by fragment, as the two-kind census did, would have taken hours. It is not needed. Every pair correction depends only on its two members and their offset, and every triple increment on its three members and two offsets, so both are read from tables built once per basis size and separation; only each fragment’s full correction, with every other fragment’s ghost present, is a solve per chain. As a check, the two-kind census rebuilt from tables matches the chain-by-chain solves to a part in a billion at five and six fragments.
The rule fails for two chains in five
The measure is the one used throughout: the share of three decades of the accuracy line over which some basis size has an assembly error below the line and a correction above it, and the loss is how much less of it the consecutive triples cover than the chain’s best cutoff. On chains of seven, 670 of the 1,134 are within a point of their best with the consecutive triples, 59 per cent, and 430 lose more than two points, 38 per cent. With two kinds of fragment the same figures were 92 and 8 per cent. Pairs alone beat the consecutive triples on 33 chains. The mean covered share across all of them is 42.1 per cent with pairs, 85.8 with the consecutive triples and 90.7 with each chain’s best cutoff, so the consecutive triples are still far better than pairs on average; it is the margin to the best that has opened.
The uniform light chain, which was the worst of all two-kind chains, ranks seventh at five fragments, tenth at six and twenty-second at seven. Its loss has not changed — 27.4, 31.5 and 33.1 points, identical to the two-kind census — but twenty-one chains of seven lose more, the worst 57.7 points.
Where the consecutive triples are not the best, the best is usually close. Of the 1,134 chains of seven, 627 are best at span two, 382 at span three, 88 at four and 37 at five or six; of the 430 that lose more than two points, three-quarters are best at span three. So the failure is not that the three-body expansion stops converging when diffuse fragments are present. It is that one more separation’s worth of triples is needed far more often than with two kinds, and that the composition, rather than the chain’s length or the fragments’ count, decides when.
The trend with length runs the wrong way too. With two kinds of fragment the share within a point rose from 85 per cent at five fragments to 92 at six and seven; with three it falls, 70, 65 and 59. Longer chains have more room for a diffuse fragment to sit three or four places from a light one, which, as the table below shows, is exactly the arrangement that needs the far triples.
Not the diffuse chain: the mixture
The worst chain of seven has two diffuse fragments at one end, three at the other and two light ones between. Its pairs alone cover 5.8 per cent of the line; its consecutive triples 28.3; and its triples up to span three 86.0, the best it can do, before wider triples take some back. A span of three is worth fifty-eight points to this chain, where for the light chain a span of four is worth thirty-three.
The uniform diffuse chain is not like that at all. It loses 18.5 points and ranks 106th: its consecutive triples cover 40.7 per cent, span three is worse, 24.7, and only every triple together reaches 59.3. It is a poor chain for any truncation, but not the one where the consecutive rule is furthest from the best. The uniform heavy chain, as before, has nothing beyond span two and loses nothing.
The table, with a third kind
The reason is in who is corrected and who is lending, the distinction the assembly that counted one share twice first found mattered when a heavy and a light centre shared a correction unequally. With both ghosts on one side, three and four separations away, the median increment is of the accuracy line for three light fragments, for three diffuse ones — larger, as expected — and for a light fragment with two diffuse ghosts, five times the all-light value and more than three times the all-diffuse one. At one and four separations the light fragment with diffuse ghosts is fifty times the all-light value. But a diffuse fragment corrected by light ghosts has a median far increment of , smaller than the all-light one.
That has a reading in what a ghost lends. A counterpoise correction measures how much a fragment’s energy improves when another fragment’s functions are made available to it. A diffuse fragment’s own basis already reaches far, so ghost functions from its neighbours add little that it does not have; a light fragment’s basis is tighter, and diffuse ghosts at three and four separations still overlap it and still add something. The far increments the consecutive rule misses are therefore largest when a compact fragment borrows from diffuse ones, and the chains that fail worst are the ones that put compact fragments among diffuse ones at the right distances.
Counted on the chain itself
The table describes one triple at a time. The worst chain can be asked the same question directly: of all its triple increments that reach three separations or more, summed in magnitude over every basis size and separation, how much belongs to each kind of fragment being corrected?
On the worst chain of seven, the two light fragments carry 76 per cent of that far-triple magnitude and the five diffuse fragments 24 per cent. On a chain with two diffuse fragments among five light ones, the light fragments carry 90 per cent. The totals say the same thing from the other side: the worst chain’s far triples add up to 24.2 accuracy lines over the grid, the all-diffuse chain’s to 14.8 and the all-light chain’s to 7.4. The diffuse chain has twice the light chain’s far triples; the mixture has more than three times, and most of them are corrections to its two light fragments.
That is also why the worst chains have their diffuse fragments in blocks at the ends. Each light fragment in the middle of the worst chain has diffuse fragments on both sides, two on one and three or four on the other, so both far shapes — one ghost on each side, and both on one side — come with diffuse functions in them. A light fragment at an end would see diffuse ghosts on one side only. The arrangement that maximises what light fragments borrow from diffuse ones at a distance is light in the middle and diffuse round it, and that is the chain at the top of the list.
How the failures are spread
Grouped by the number of diffuse fragments, the chains of seven make the same point. With none or one, eight per cent lose more than two points, as in the two-kind census. With two, 34 per cent; with three or four, 61 and 62 per cent; with five, 42; with six, one chain in eight; and the single chain with seven is the uniform diffuse one. The failure peaks where the chain is about half diffuse, which is where light and diffuse fragments sit together most often. Among the 72 chains made only of light and diffuse fragments, 49 lose more than two points.
The worst chains at each length share a shape: diffuse fragments in a block, light or heavy fragments next to it, and very often a second diffuse block at the other end. Every one of the twenty worst chains of seven has a diffuse fragment, and eighteen also have a light one; the other two put heavy fragments next to diffuse ones, which has a smaller version of the same effect. Their best span is three in most cases, so the cure is usually cheap — one separation beyond the consecutive triples — but nothing in the composition rule of thumb the two-kind census suggested would have prescribed it.
How the census was solved
Each fragment’s counterpoise correction from any set of ghosts is the one-dimensional calculation of every essay in this sequence: one electron per fragment in an optimised set of one to six s Gaussians, the fragment’s own nucleus present and every ghost’s charge zero, at separations of 1.6, 2, 2.5 and 3 bohr. Nuclear charges are one for light, two for heavy and 0.6 for diffuse fragments, and each kind’s exponents are optimised for its own charge. For each basis size and separation a table holds every fragment’s energy alone, every pair correction by kinds and offset, and every triple increment by kinds and offsets; a chain’s pair sum and its triple sums by span are added up from the table, and its full corrections are solved directly. The covered share and the loss are read as in the two-kind census.
The checks run where the figures are drawn. The two-kind census rebuilt from the tables, at five and six fragments, gives every chain’s pairs and every span’s covered share as the chain-by-chain solves did, to . The uniform light chain keeps its published loss at every length. And the refusal tests the table’s construction: a table that filed each triple by its two distances only, ignoring which side of the corrected fragment each ghost sits on, must fail to reproduce a chain, because an end fragment and an interior one see the same distances on different sides. It fails by far more than the tolerance.
What three kinds on a line cannot say
Three kinds are three points on a continuum. A real cluster’s fragments differ continuously in how far their functions reach, and the finding that the mixture fails worst is a statement about two kinds a factor of about two apart in reach. Whether the failure grows steadily with the contrast between fragments, or peaks at some ratio, is not measured.
A line, again. On a line the distance between two fragments is their separation in the chain, and the far triples here are far in both senses. In three dimensions a diffuse fragment can be near a light one that is many fragments away in any ordering, and no single distance served three chains even on a line.
Seven fragments. The three-kind census stops at seven, where it is 1,134 chains; eight would be 3,321. The trend with length, falling from 70 to 59 per cent within a point, has three points.
The figure of merit is a worst cell. As before, the covered share is set by the worst cell at each basis size, and a chain can lose many points through one cell; the diffuse fragment’s correction is largest at the shortest separation and the largest basis, and that is plausibly where these losses live.
A table that predicted the wrong chain and the right cause
The two-kind census closed by saying that the uniform chain was the worst case because its far triples were all light, and it offered a test. The test confirmed the cause and refuted the prediction drawn from it. Far triples do grow with diffuse fragments, and the consecutive rule does fail more; but the worst chain is not the most diffuse one, because a diffuse fragment is not much helped by its neighbours’ functions. The table had all of that in it before the census was run: the prediction looked at one column and the answer was in another. A census of 1,134 chains was needed to see the failure’s size and its spread, but its location was readable from twenty-seven patterns, and that is the more useful fact for anyone choosing a cutoff for a cluster that has never been computed.
The same lesson has appeared before in the correction that turned out to be two functions and in the repair that cost more than the whole: a counterpoise assembly is not described by how large its pieces are, but by which pieces cancel which, and which fragments are lending to which. On a chain of three kinds, the triples that matter are the ones in which a compact fragment borrows from diffuse neighbours at a distance.
Still open: a contrast in reach, and the cells that decide
The obvious open question is the contrast. With a diffuse charge of 0.6 the worst chains mix light and diffuse fragments; with a charge closer to one the two kinds would merge and the census would return to the two-kind result, and with a charge much smaller the diffuse fragment’s own correction might dominate instead. A sweep of the diffuse charge, reading the worst chain’s loss and the share within a point at each, would say whether the failure depends on the contrast in a way that could be read from two fragments’ basis sets before any chain is built.
The nearer question is which cell each failure is set in. Both censuses lose their points at the basis size and separation where an assembly is closest to failing, and for a diffuse fragment that is likely the shortest separation, where its functions overlap its neighbours most. If the losses concentrate in one cell, the cutoff could be chosen per cell rather than per chain, and the composition would stop mattering 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.
Shares its objects with
Essays naming at least two of the same things, that neither author linked.
- One basis size where it is worth doing — both name basis set superposition error, counterpoise, fragment method, gaussian basis, many-body expansion
- Length did not rescue the consecutive triples — both name basis set superposition error, fragment method, gaussian basis, many-body expansion
- The overshoot was one arrangement — both name basis set superposition error, counterpoise, fragment method, many-body expansion
- The line was holding the answer up — both name basis set superposition error, counterpoise, many-body expansion
Named objects
A dashed tag is an object no other essay names yet.
Basis set superposition errorCounterpoiseFragment methodGaussian basisMany-body expansion