Orbitals

No one distance serves three chains

Keeping only the consecutive triples of a counterpoise assembly failed the uniform chain, and the practical alternative is a distance: keep every triple whose members lie within some span of each other. On eight fragments the uniform chain needs a span of four before triples beat pairs, and at five it beats keeping every triple, with seventy fewer calculations. The heavy-outside chain is best at two and is below pairs again by four. From six fragments on no single span beats pairs for all three arrangements, and the pair of shapes the cutoff was meant to separate enter together.

Worth reading first: Length did not rescue the consecutive triples · A repair that costs more than the whole.

A counterpoise correction for a cluster can be assembled from pieces: each fragment’s correction from every other fragment’s functions one at a time, and then the three-body increments, which add what two sets of ghost functions do together that neither does alone. Adding every triple repaired a pairwise assembly on four fragments and cost more than the calculation it stood in for. Keeping only the consecutive triples, the ones whose number grows with a chain’s length rather than as its cube, helped two of three arrangements and hurt the third. Carried to eight fragments, the failure turned out to be general: on the uniform chain the consecutive assembly settles near 48 per cent of the accuracy line against 68 for pairs alone, because the consecutive triples overshoot and the triples with a gap take it back.

That essay closed by naming the truncation a real calculation would actually use. A consecutive rule is topological; the practical one is geometric. Keep every triple whose members are all within some distance of each other, sweep the distance, and see whether there is one at which the uniform chain’s assembly passes pairs — and whether it is the same for all three arrangements.

There is one for the uniform chain, and it is not the same. No single distance serves all three chains, and the reason the lead offered for hoping one would turns out to be false on a line.

A span is a distance on a line

The chains are fragments at equal spacing on a line, so the distance between a triple’s two furthest members is a whole number of separations: its span. Keeping the triples within a distance is keeping the triples of span at most DD. At D=2D = 2 that is exactly the consecutive triples, and at D=N−1D = N - 1 it is every triple, so the sweep runs between the two assemblies already measured and reproduces both of them, cell for cell, at its ends.

The lead’s hope was specific. The shapes found to carry opposite signs on the chain of six were seen from an end fragment: ghosts one and three separations away, carrying +20.8 per cent of the three-body increment, and ghosts two and three away, carrying −11.0. A distance rule, the lead suggested, would keep the first before the second. It cannot. Both triples run from the end fragment to the fragment three away, so both have a span of three, and both have their further ghost at three. Any cutoff on distance keeps them together or drops them together. A rule that separated them would have to count the middle member’s position, and that is the topological kind of rule the distance was meant to replace.

So the question becomes simply what each span adds.

The uniform chain needs four

On 8 fragments no one span serves all three chains: the uniform one needs four, the heavy-outside one is best at two. The share of three decades of accuracy line covered by pairs plus every triple whose three fragments span at most D separations, from D = 2 (the consecutive triples) to D = 7 (every triple), on chains of 8. uniform: 2 48.1%, 3 58.4%, 4 79.9%, 5 84.1%, 6 81.6%, 7 81.3%, against pairs alone 67.9%; heavy inside: 2 77.5%, 3 77.5%, 4 77.5%, 5 77.5%, 6 77.5%, 7 77.5%, against pairs alone 68.4%; heavy outside: 2 87.9%, 3 77.9%, 4 70.1%, 5 65.0%, 6 62.3%, 7 62.3%, against pairs alone 70.8%.
Fig. 1 On chains of eight fragments, the share of three decades of accuracy line covered by pairs plus every triple up to each span, for the three arrangements, with pairs alone dashed.

On the uniform chain of eight, the covered share runs 48.1, 58.4, 79.9, 84.1, 81.6 and 81.3 per cent as the span grows from two to seven, against 67.9 for pairs alone. Span three is not enough. Span four is, by twelve points, and span five is the best of all, nearly three points above every triple together. On the heavy-inside chain nothing past span two matters, and every span covers 77.5. On the heavy-outside chain the order is reversed: 87.9 at span two, then 77.9, 70.1, 65.0, 62.3 and 62.3. Every widening of the cutoff on that chain makes it worse, and at span four it is already below pairs alone.

These are the same three behaviours the consecutive and complete assemblies showed, now laid out as curves. What the curves add is that the three do not cross at any useful place. A cutoff of four suits the uniform chain and has already cost the heavy-outside chain eighteen points; a cutoff of two suits the heavy-outside chain and costs the uniform chain twenty.

The span does not grow with the chain

A cutoff that worked on eight fragments might simply be the right cutoff for eight, and a longer chain might want a longer one. The uniform chain can be read at every length.

On the uniform chain the span at which triples overtake pairs does not grow with the chain. The uniform chain's covered share against the largest span kept, at 5, 6, 7, 8 fragments, with each length's pairs-alone share as a short dash at the left. 5: 45.6%, 63.4%, 73.0% against pairs 60.0%; 6: 47.1%, 62.7%, 78.5%, 78.5% against pairs 63.1%; 7: 48.8%, 62.6%, 81.9%, 81.9%, 80.3% against pairs 65.8%; 8: 48.1%, 58.4%, 79.9%, 84.1%, 81.6%, 81.3% against pairs 67.9%.
Fig. 2 The uniform chain’s covered share against the largest span kept, at five, six, seven and eight fragments, with each length’s pairs-alone share marked at the left.

It does not. At five fragments span three already passes pairs, 63.4 against 60.0; at six, seven and eight it falls short, 62.7, 62.6 and 58.4 against 63.1, 65.8 and 67.9. Span four passes at every length, 73.0, 78.5, 81.9 and 79.9. The span at which triples overtake pairs is four from six fragments onwards and is not moving. That is the useful result of the sweep: on a uniform line of these fragments, the three-body correction needs triples reaching four separations, and nothing further out is worth much.

On the heavy outside chain the span at which triples overtake pairs does not grow with the chain. The heavy outside chain's covered share against the largest span kept, at 5, 6, 7, 8 fragments, with each length's pairs-alone share as a short dash at the left. 5: 77.4%, 78.3%, 70.8% against pairs 54.2%; 6: 80.9%, 78.7%, 57.0%, 56.2% against pairs 59.7%; 7: 84.4%, 79.5%, 65.0%, 59.4%, 59.1% against pairs 66.5%; 8: 87.9%, 77.9%, 70.1%, 65.0%, 62.3%, 62.3% against pairs 70.8%.
Fig. 3 The heavy-outside chain’s covered share against the largest span kept, at five to eight fragments, with pairs alone marked at the left.

The heavy-outside chain falls with every widening at every length, and the place where it drops below pairs moves the wrong way. At five fragments even every triple covers 70.8 against pairs’ 54.2. At six, seven and eight, span four is already under pairs. A cutoff chosen to suit the uniform chain therefore makes the heavy-outside chain’s assembly worse than doing nothing past pairs, and the gap does not close with length.

Where the increment lives

The curves have a simple reading once the three-body increment is split by span.

Where the three-body increment lives: almost all in span two, except when the heavy centres are outside. On chains of 8, the median share of each cell's whole three-body increment carried by triples of each span. uniform: 2 1.294, 3 0.010, 4 -0.243, 5 0.000, 6 0.003, 7 0.000; heavy inside: 2 0.993, 3 0.008, 4 0.000, 5 -0.000, 6 -0.000, 7 0.000; heavy outside: 2 0.602, 3 0.263, 4 0.113, 5 0.021, 6 0.000, 7 0.000. A share above one is an overshoot that wider triples take back.
Fig. 4 On chains of eight, the median share of each cell’s whole three-body increment carried by the triples of each span, for the three arrangements.

On the uniform chain the consecutive triples carry 129 per cent of the whole increment, by the median over the grid of basis sizes and separations. Span three carries almost nothing, one per cent. Span four carries −24 per cent, and spans five and beyond nothing. So the consecutive triples overshoot by about a quarter and the span-four triples take it back. That is why span three does not help: there is nothing in it to help with. On the heavy-inside chain span two carries 99 per cent and nothing else carries anything, which is why every cutoff gives the same answer. On the heavy-outside chain the shares are all positive and fall away, 60, 26, 11 and 2 per cent, an increment spread across the whole chain that a cutoff can only truncate.

That makes the heavy-outside chain’s behaviour look like a contradiction at first: if every span adds a positive share, keeping more of them ought to help. It does not because of the order above the triples. The complete assembly’s failure on that chain was traced to the four-body remainder, which reaches 2.70 kilocalories a mole at eight fragments. Each widening of the cutoff brings the three-body sum closer to its total, and the total is itself further from the exact correction than the consecutive part alone. A truncation that keeps less of an overshooting order is better than one that keeps more of it. The median shares are also only medians: the covered share is set by each basis size’s worst cell, and the worst cells are not the median ones.

The overshoot is one shape, seen from one side

The span tells where the increment lives; the shape inside a span tells why. Split by shape on the uniform chain of eight, with each shape’s median share taken over the same grid of cells on its own, the consecutive overshoot is almost entirely one arrangement: a fragment with both ghosts on the same side of it, one and two separations away. That shape carries a median of 157 per cent of the whole increment. The other consecutive shape, a fragment with one ghost on each side, carries 1.4 per cent. A fragment in the middle of a consecutive triple gets almost no three-body correction at all; a fragment at the end of one gets more than the whole.

The likely reason is in what the two ghosts do, though it is a reading of the numbers rather than a separate calculation. Two ghosts on one side both extend the fragment’s basis in the same direction, and their functions overlap each other as well as the fragment, so what they add together is not what they add apart. Two ghosts on opposite sides extend the basis in two directions that barely interact, and the pair adds very nearly the sum of its parts. The shape that takes back the overshoot at span four has the same character: both ghosts on one side, three and four away, carrying a median of −22 per cent, with the shapes that straddle the fragment again small.

So the cancellation the uniform chain needs is between one-sided pairs at different distances, and a cutoff has to reach the second of them to find it. A span cutoff gets there at four. A cutoff that could be written in terms of one-sided pairs alone would get there more cheaply, but that is again a rule about arrangement, not about distance.

A rule centred on the corrected fragment

A counterpoise correction is assembled fragment by fragment, so a second rule suggests itself: for each fragment, keep the triples whose two ghosts both lie within a radius of it. The two rules keep different shapes at the same number. A triple with a ghost one separation on each side of the corrected fragment has a span of two and a radius of one; a triple with ghosts one away on one side and two away on the other has a span of three and a radius of two.

Centred on the corrected fragment, the cutoff gets worse before it gets better. On the uniform chain of 8, the covered share with triples kept by span (all three fragments within D separations) and by radius (both ghosts within r of the corrected fragment). By span: 2 48.1%, 3 58.4%, 4 79.9%, 5 84.1%, 6 81.6%, 7 81.3%. By radius: 1 59.6%, 2 42.3%, 3 56.2%, 4 77.8%, 5 84.1%, 6 81.6%, 7 81.3%. Pairs alone 67.9%.
Fig. 5 On the uniform chain of eight, the covered share with triples kept by span and by radius, against the cutoff, with pairs alone dashed.

The radius rule is worse at every small cutoff, and it is worse in a way that is instructive. At a radius of one it keeps only triples with a ghost on each side of the corrected fragment, and covers 59.6 per cent. At a radius of two it adds every shape whose ghosts lie up to two away on either side, the end fragments’ consecutive triples among them, and covers 42.3, less than at a radius of one and far less than pairs. Only at a radius of four does it pass pairs, at 77.8. By the median share, the triples at a radius of exactly two carry 155 per cent of the increment and those at a radius of four take back a third of the whole, which is the overshoot of the span rule made larger by bringing in more of the overshooting shapes before any of the correcting ones.

So a rule that looks more natural, because it asks of each fragment only what is near it, is the poorer truncation. The span rule is better because it keeps whole compact clusters, and the compact clusters are where the cancellation happens.

What each cutoff costs

The reason to truncate at all is cost, and the count of triple calculations kept at each span is easy to state: on a chain of eight there are 168 fragment-and-ghost-pair choices, and spans two, three, four and five keep 18, 48, 84 and 120 of them.

Paid for in triples: the uniform chain's best cutoff costs half of them and beats keeping all. Covered share against the number of fragment-triple calculations kept, out of 168 on a chain of 8, for each span cutoff and each arrangement. uniform: 18 kept 48.1%, 48 kept 58.4%, 84 kept 79.9%, 120 kept 84.1%, 150 kept 81.6%, 168 kept 81.3%; heavy inside: 18 kept 77.5%, 48 kept 77.5%, 84 kept 77.5%, 120 kept 77.5%, 150 kept 77.5%, 168 kept 77.5%; heavy outside: 18 kept 87.9%, 48 kept 77.9%, 84 kept 70.1%, 120 kept 65.0%, 150 kept 62.3%, 168 kept 62.3%.
Fig. 6 Covered share against the number of triple calculations kept on chains of eight, for each span cutoff and each arrangement, starting from pairs alone at zero.

For the uniform chain, half the triples at span four buy a covered share of 79.9 per cent, within a point and a half of all of them, and 120 at span five buy 84.1, more than all 168 achieve. For the heavy-inside chain, eighteen buy everything there is. For the heavy-outside chain, eighteen buy 87.9 and every one added after that costs coverage. If the arrangement is known, each chain has a best cutoff and it is cheap. If it is not known, there is no cutoff that is safe, and the safe choice is pairs alone, which none of the three chains would choose.

Every chain and every span: the covered share, with the best cutoff marked. For three arrangements at five to eight fragments, the covered share of pairs alone and of pairs plus triples up to each span from two to seven separations, with the best of each row in bold.
Fig. 7 Every arrangement at five to eight fragments: the covered share of pairs alone and of pairs plus triples up to each span, the best of each row in bold.

The table carries the claim in full. At five fragments a span of three beats pairs for all three arrangements, 63.4, 75.9 and 78.3 against 60.0, 61.2 and 54.2. At six, seven and eight no span does: every span at or under three fails the uniform chain, and every span at or over four fails the heavy-outside one. The window in which one cutoff serves everything closes at six fragments, the length from which the heavy-outside chain’s complete order first falls behind its consecutive part and its overshoot can no longer be hidden by a short chain.

How each assembly 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 in an optimised set of one to six s Gaussians, 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, at separations of 1.6, 2, 2.5 and 3 bohr, each energy the lowest root of the generalised eigenvalue problem. A triple’s increment is its correction with both ghosts less the two corrections with one each, filed by its span and by its radius from the corrected fragment, so every cutoff under either rule is read from one set of solves. The covered share is the share of three decades of the accuracy line, from 10−1.510^{-1.5} to 101.510^{1.5} kilocalories a mole, over which some basis size has an assembly error below the line and a correction above it — the measure the assembly essays have used since they began.

The arithmetic is checked where the figures are drawn. At a span of two every cell’s error is the consecutive assembly’s, and at the largest span the complete assembly’s, to 10−1210^{-12}. Both rules end at the same covered share when they keep every triple. The two one-gap shapes seen from an end share a span. And the refusal is the count on the cost axis: the number of triples kept at each span is counted a second time by sliding a window of D+1D+1 fragments along the chain, and the two counts must agree, so that the cost curve is not an artefact of the rule that defines it.

A line is the easy case

On a line, distance and topology nearly coincide. A real cluster in three dimensions has triples that are compact in space and far apart along any ordering, and a distance cutoff there keeps sets no consecutive rule could name. The finding that a span of four is enough for the uniform chain is a statement about fragments in a row. What carries over is the method — split the increment by distance, read where it overshoots and where it cancels — and the warning that a single cutoff can be right for one arrangement of fragments and wrong for another of the same size.

The arrangements are caricatures. Heavy centres inside, outside or nowhere are the three extremes of one fragment basis being larger than another’s, and real clusters mix them. A cutoff chosen for a mixed cluster would be choosing between curves that pull in opposite directions, and nothing here says where between them it would land.

The covered share is one figure of merit. It is set by the worst cell at each basis size, so a cutoff that improves the typical cell and worsens one can lose. The median shares and the covered shares do not have to agree, and on the heavy-outside chain they do not.

A cutoff is a claim about the arrangement

The repair that cost more than the whole showed that adding the next order can be worth less than the calculation it replaces. This sweep adds that the right place to stop the next order is not a property of the order. A uniform chain wants its triples to reach four separations, so that the span-four shapes can cancel the consecutive overshoot; a chain with heavy ends wants them to stop at two, because the further shapes carry it towards a complete order that is itself further from the answer. The same cutoff applied to both is the right answer for at most one of them.

One correction was already found to be two functions depending on the basis, and this is the same lesson at the level of assembly: the pieces of a many-body expansion do not have sizes that can be ranked in advance. They have signs, and the signs depend on what the fragments are. A truncation chosen without computing what it drops is a claim about those signs, made without looking.

Still open: four bodies by distance, and a mixed chain

The obvious open question is the four-body remainder on the heavy-outside chain, read the way the triples have been read here. It is what makes every widening of the triple cutoff hurt, and it doubles between five fragments and eight. Four-body increments can be split by span exactly as the triples were, and the split would say whether the remainder lives in compact quadruples, which a four-body correction at small span could repair, or is spread along the whole chain, in which case the many-body expansion for this arrangement is simply not converging in the order that is being truncated.

The nearer question is a chain that is neither. A row with heavy and light centres in a pattern — alternating, or one heavy centre in the middle of a light chain — would sit between the curves above, and the span at which it is best would say whether the right cutoff can be read off a chain’s composition before computing anything, or has to be found each time by the sweep.

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.

Named objects

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

Basis set superposition errorFragment methodGaussian basisMany-body expansionModel limit