What symmetry decides

The floor under ammonia's fit is a theorem

Ammonia's fitted force field misses its eight lines by 0.46 per cent, and every judgement about which fields fit equally well has treated that misfit as noise. Most of it is not. Within each symmetry species the product of the ND₃/NH₃ frequency ratios is fixed by the masses alone, so every harmonic field predicts the same two products — and the observed lines give products 1.9 and 1.3 per cent smaller. That forces 86 per cent of the misfit on any field whatever. Charge only the rest to noise and the band of fields the fit cannot tell apart narrows 2.6-fold, and with the published three degrees of freedom the bond-angle verdict that crossed nine tenths inside it stays below the line, at 0.867 to 0.899.

Worth reading first: The worst case was the loosest constant · The partner reads the other displacement.

A symmetry label’s verdict on a distortion is read off mode shapes, and mode shapes come from a force field fitted to frequencies. Walked along the direction its frequencies see least, ammonia’s five-constant field keeps every frequency to within the fit’s own half per cent while the share of a bond-angle distortion that goes into the degenerate bend runs from 0.840 to 0.923 — across the nine-tenths line at which a label settles a coordinate. Since then the verdict has been chased through Coriolis constants and through a partner field that reproduces NH₃’s bands exactly and exchanges two definitions of a distortion.

All of it rested on one statistical choice that the essay making it flagged. “As good as the best” was defined by taking the best fit’s own misfit as noise: a field is indistinguishable from the published one when its sum of squared misfits is no more than the best’s times 1+1/ν1 + 1/\nu, with ν\nu the number of distinct lines less the number of constants. That essay said the criterion charges all of the misfit to noise, when most of it is the harmonic model failing, and it guessed where the failure lived: in the umbrella, the vibration whose double well no harmonic field can represent.

The failure is not in the umbrella, and most of it is not a model failure the fields could share differently. It is forced by a theorem, the same for every field, and it can be computed exactly.

A product that no field can move

When an isotope is substituted the force field does not change and the masses do. Within one symmetry species the frequencies are the roots of a secular equation whose determinant is the product of a kinetic determinant and a force-constant determinant, so the product of the squared frequencies in that species is det⁡G⋅det⁡F\det G \cdot \det F. Take the ratio between two isotopologues and the field cancels:

∏i∈ΓωiDωiH=det⁡GΓDdet⁡GΓH.\prod_{i \in \Gamma} \frac{\omega_i^{D}}{\omega_i^{H}} = \sqrt{\frac{\det G_\Gamma^{D}}{\det G_\Gamma^{H}}} .

This is the Teller–Redlich product rule, and it holds species by species. It was the exact statement behind water’s isotope shifts, where the familiar square root of two failed for every mode. Its consequence for a fit has not been drawn. Every harmonic field for ammonia, however its constants are chosen, predicts the same product of ND₃/NH₃ ratios in the E species and the same product in A1A_1.

Every field gives the same two products of isotope ratios, and the observed lines give two smaller ones. For every member of ammonia's family of fields along the stretch–bend constant, the product of the ND₃/NH₃ frequency ratios in the E species and in A₁: 0.533762 and 0.542935 on every member. The observed lines give 0.523758 and 0.535966, short by 1.89% and 1.29% in the logarithm.
Fig. 1 For every member of ammonia’s family of fields along the stretch–bend constant, the product of the ND₃/NH₃ frequency ratios in the E species and in A1A_1, with the products of the observed lines dashed.

The walk confirms it. Along twenty-two members of the family, from a stretch–bend constant of −1.1 to +0.2 mdyn/rad, and for the E block’s partner field, the E product is 0.533762 and the A1A_1 product 0.542935, identical to seven figures, which is the precision of the frequencies themselves. The observed lines the field is fitted to give 0.523758 in E and 0.535966 in A1A_1. They are short of the theorem by 1.89 and 1.29 per cent in the logarithm.

What the theorem forces on any field

If the products cannot move, the misfit they imply cannot be removed, and its least possible size follows in closed form. In one species there are four distinct lines, two for NH₃ and two for ND₃, and the field’s products differ from the observed by a fixed factor. The cheapest way to distribute that factor, measured by the sum of squared relative misfits the fit minimises, is to spread it evenly: each ND₃ line too high and each NH₃ line too low by a quarter of the logarithmic deficit. In E that is 0.47 per cent a line, and since the fit counts each degenerate pair twice, it weighs double; in A1A_1, 0.32 per cent.

The published field misses every line in the direction the product rule forces, and by about as much. Each observed line's relative misfit under the published field (bars), beside the misfit the product rule forces when the deficit in each species is spread equally over its lines with the sign that closes it (dots): NH₃ E 1691: -0.67% against -0.47%; NH₃ E 3577: -0.29% against -0.47%; ND₃ E 1191: 0.66% against 0.47%; ND₃ E 2660: 0.27% against 0.47%; NH₃ A₁ 1022: -0.19% against -0.32%; NH₃ A₁ 3506: -0.45% against -0.32%; ND₃ A₁ 749: 0.19% against 0.32%; ND₃ A₁ 2564: 0.47% against 0.32%. The forced part is 86% of the published sum of squares.
Fig. 2 Each observed line’s relative misfit under the published field, as a bar, beside the misfit the product rule forces when each species’ deficit is spread evenly over its lines.

The published field does very nearly that. Every NH₃ line is calculated low and every ND₃ line high, in both species, which is the sign pattern the theorem dictates and no other mechanism would produce so uniformly. The degenerate bend carries the most, −0.67 per cent in NH₃ and +0.66 in ND₃; the degenerate stretch −0.29 and +0.27. In A1A_1 the umbrella is the best-fitted pair in the molecule, −0.19 and +0.19, and the symmetric stretch takes the rest, −0.45 and +0.47. Where the bars differ from the even split, the field has traded misfit between lines of one species in a way the rest of the data rewards; it cannot change the sum.

The total forced misfit is a residual of 0.429 per cent, against the published 0.463. In the sum of squares, 85.7 per cent of ammonia’s misfit belongs to the theorem, and only the remaining 14.3 per cent is something different fields could distribute differently.

That disposes of the umbrella guess directly. A harmonic model of a vibration that tunnels through a barrier ought to fit it badly, and the guess was natural. But the umbrella’s lines are fitted to a fifth of a per cent, and they are what holds the A1A_1 bending constant: fitted without them, the field leaves that constant free, the umbrella’s predicted frequency falls to almost nothing, and the remaining lines fit no better: the degenerate bend is still 0.65 per cent out in each isotopologue, in the directions the E product forces.

The band, re-read

The band of indistinguishable fields was defined by a noise estimate, and the noise estimate was the published misfit. If most of that misfit is a theorem’s, it is not noise, and every field carries it identically. The fair criterion charges only the excess above the floor to noise: a member is as good as the best when its sum of squares exceeds the floor by no more than the best’s excess over the floor times 1+1/ν1 + 1/\nu, keeping the same ν=3\nu = 3 the published band used.

Charge only the unforced misfit to noise and the band of fields narrows by a factor of 2.6. The fit's residual against the held stretch–bend constant, with the published limit 0.535% giving the band -1.000 to -0.634, and the limit 0.474% that charges only the misfit above the product rule's floor (0.429%) to noise, giving -0.899 to -0.760.
Fig. 3 The fit’s residual along the held stretch–bend constant, with the published limit that treats all of the misfit as noise, the limit that takes out the product rule’s floor first, and the floor itself.

The residual’s valley does not change; only the line across it does. With all of the misfit as noise the limit is a residual of 0.535 per cent and the band runs from −1.000 to −0.634 mdyn/rad. With the floor taken out the limit is 0.474 per cent and the band runs from −0.899 to −0.760, 2.6 times narrower, and still containing the published −0.831. Along the band the other constants move with the held one: the stretching constant, the bending constant and both couplings trade against it as the loose direction described, and all of them now trade over a range 2.6 times smaller.

The choice of ν\nu is a statistical convention and the band depends on it. What does not depend on any convention is the floor itself, and the fact that the published criterion measured the band against a noise level 86 per cent of which was not noise.

The verdict, re-read

With the forced floor taken out, the bond-angle verdict stays below nine tenths across the whole band, by a hair. The share of a least-cost bond-angle distortion that goes into ammonia's degenerate bend, against the held stretch–bend constant. Across the published band (-1.000 to -0.634) it runs 0.840 to 0.923; across the floor-corrected band (-0.899 to -0.760) 0.867 to 0.899.
Fig. 4 The share of a least-cost bond-angle distortion that goes into ammonia’s degenerate bend, along the held stretch–bend constant, with both bands shaded and the nine-tenths line dashed.

Across the published band the bond-angle verdict ran 0.840 to 0.923, and nine tenths was crossed at a stretch–bend constant near −0.76. Across the narrower band it runs 0.867 to 0.899. Along the whole walk the verdict is now below the line: the label does not settle the bond-angle coordinate under any field the corrected criterion admits. The margin is 0.0012 at the band’s edge, which is smaller than anything this sequence has quoted as decisive, and the honest reading is that the theorem moved the verdict from “on either side” to “on one side, just”.

There is a field this cannot reach. The E block’s partner reproduces NH₃’s bands exactly and gives a bond-angle verdict of 0.926. It is not a member of the walk, and the product rule is blind to it for the plainest reason: it has the same products as every other field. What separated it from the published field was ND₃’s E bands, which disfavour it about twofold, and its Coriolis constant, which differs by more than the band’s whole width. The floor changes nothing about that comparison, since both fields carry the same floor.

The E partner trades the two definitions of a distortion: its least-cost verdicts are the published field's pure ones. For the degenerate stretch and bend of NH₃, the share of a unit distortion that goes into the E vibration its label names, under three definitions of the distortion, for the published field and its E partner. stretch, least-cost: 0.9261 against 0.8832; stretch, pure coordinate: 0.8832 against 0.9261; bend, least-cost: 0.8832 against 0.9261; bend, pure coordinate: 0.9261 against 0.8832; stretch, Cartesian: 0.9859 against 0.5421; bend, Cartesian: 0.5164 against 0.9688.
Fig. 5 For NH₃’s degenerate stretch and bend, the share of a unit distortion that goes into the vibration its label names, under three definitions of the distortion, for the published field and for its E partner.

What the lines would need to be

The theorem also says what the data would have to look like for a harmonic field to fit them exactly. Holding NH₃’s lines, the products are restored by raising ND₃’s E lines 0.95 per cent, to 1,202 and 2,685 cm⁻¹ from 1,191 and 2,660, and its A1A_1 lines 0.65 per cent, to 754 and 2,581 from 749 and 2,564. Equivalently, NH₃’s lines could be lowered by the same factors.

The sign of the deficit says something about where it comes from. Anharmonicity lowers every observed fundamental below its harmonic frequency, and lowers a hydrogen’s further than a deuterium’s, so if all eight lines were fundamentals their ND₃/NH₃ ratios would come out larger than the harmonic rule allows, not smaller. A product that is short means the NH₃ lines sit high, or the ND₃ lines low, against what any one harmonic field can produce — which is what a set mixing harmonic frequencies for one isotopologue with observed fundamentals for the other would look like. Which of the eight numbers are which kind, the arithmetic cannot say. What the theorem settles is that they are not, all together, the harmonic frequencies of any field, and that a fit to them inherits a misfit of 0.43 per cent before any constant is chosen.

Water is the control, and it is a clean one. Its field is fitted to H₂O’s and D₂O’s three lines each with four constants, and its published residual is 0.0073 per cent, sixty times smaller than ammonia’s. Its two species give products that differ from the theorem’s by 0.036 per cent in A1A_1 and 0.003 per cent in B2B_2, fifty times less than ammonia’s E deficit, which is what eight harmonic frequencies of one molecule ought to look like. And the floor those differences force is the whole of water’s residual: the published field leaves nothing on the table that a different field could take, because the four constants have absorbed everything except what the masses fix. Water’s tiny misfit and ammonia’s large one are the same kind of thing, a theorem’s floor, and they differ fiftyfold because water’s eight numbers very nearly obey it and ammonia’s do not.

That is also why water’s band was twenty-two times narrower. Water’s noise estimate was, in effect, the product rule’s floor and nothing else, so its criterion was already measuring fields against what fields can change. Ammonia’s criterion was measuring them against a level six-sevenths of which no field could change, and the band it drew was correspondingly generous.

The product rule on ammonia, in one table. The E and A₁ products of ND₃/NH₃ frequency ratios shared by every harmonic field and given by the observed lines, the floor they force, the band of fields read both ways, the bond-angle verdict across each band, and the ND₃ lines the rule would need with NH₃'s held. E product, every field: 0.533762; E product, observed: 0.523758; A₁ product, every field: 0.542935; A₁ product, observed: 0.535966; floor residual · published: 0.429% · 0.463%; share of the sum of squares forced: 85.7%; band, all misfit as noise: −1.000 to −0.634; band, floor taken out: −0.899 to −0.760; bond-angle verdict, published band: 0.840 to 0.923; bond-angle verdict, floor taken out: 0.867 to 0.899; ND₃ E lines the rule needs: 1202, 2685 for 1191, 2660 (+0.95%); ND₃ A₁ lines the rule needs: 754, 2581 for 749, 2564 (+0.65%).
Fig. 6 The product rule on ammonia: the products shared by every field and given by the lines, the floor, both bands, the bond-angle verdict across each, and the ND₃ lines the rule would need.

How the floor and the bands were computed

Every field’s frequencies are the harmonic normal modes of ammonia’s valence force field — a bond stretch, an angle bend, a stretch–stretch, a bend–bend and a stretch–bend constant — at NH₃’s and ND₃’s masses, with the modes sorted into species by degeneracy rather than by position. Each member of the walk is the field refitted to all eight lines with the stretch–bend constant held, by the same simplex fit that produced the published field. The floor is the closed-form minimum of the sum of squared relative misfits subject to the two products, with each species’ deficit spread evenly over its four distinct lines and degenerate lines counted twice as the fit counts them. The narrower band’s edges are found by regula falsi on the refitted residual.

The checks run where these figures are drawn. Every member of the walk that keeps the observed assignment, twenty-two fields, and the E partner give the same two products to 10−710^{-7}. The floor is no larger than the published sum of squares, since no field can beat a theorem. The narrower band lies inside the wider one and contains the published constant. And the refusal tests that the rule is used per species: the product of every ratio together is also independent of the field, but the E product alone must differ from the A1A_1 product — it does, 0.5338 against 0.5429 — or the split into two floors would be a statement about the total mass and nothing else.

What a harmonic floor cannot say

It is a floor for harmonic fields. An anharmonic model, or a fit to harmonic frequencies derived from the observed fundamentals and their anharmonic constants, would not carry it, and the band such a fit gave might be wider or narrower than either band here. The floor is the price of fitting a harmonic model to these eight numbers, not a property of ammonia.

The noise convention remains a convention, and the count it uses is not settled. The corrected criterion still uses 1+1/ν1 + 1/\nu with ν=3\nu = 3, and the verdict’s margin of 0.0012 is smaller than the change a different count would make — and the theorem itself uses up one equation per species, which leaves ammonia one degree of freedom rather than three and puts the verdict back across the line. What is robust is the direction: removing the theorem’s share of the misfit can only narrow the band, because it lowers the floor the criterion is measured from.

The assignment is taken as given. The rule needs each line in its species, and the products here use the species the fit itself assigns: the umbrella and the symmetric stretch in A1A_1, the bend and the antisymmetric stretch in E. If one ND₃ line had been assigned to the wrong species, the products would move by far more than the deficit, and the floor would be measuring an assignment error rather than a mixture of kinds of frequency. The fit’s own agreement, every line within 0.7 per cent in the species it is given, makes that unlikely; it does not make it impossible, and the product rule is in fact the standard test of exactly that.

Ammonia’s walk is one direction. The band is along the stretch–bend constant, which was the loose direction to within a part in a thousand. Fields off that line are excluded by much larger misfits, but the partner shows that the set of exact fits is not connected to the walk, and the product rule says nothing about how many such disconnected fields there are.

A residual is a sum of a theorem and a model

The census of verdicts and every essay that has refined it read a fit’s residual as a single number: a half per cent of agreement, to be compared with the next field’s. The product rule splits it into two parts with different meanings. The part fixed by the masses is shared by every field and says only that the data are not harmonic frequencies; the part above it is what distinguishes one field from another. A band of fields is a statement about the second part, and measuring it against the sum overstates how loose a fit is by exactly the share the theorem owns. On ammonia that share is 86 per cent.

It is symmetry that makes the split exact. Without species there would be one product for the whole molecule and one floor; with them there are two products and two floors, and they can be computed independently because the secular equation factorises. The same separation that lets a label say which vibration a distortion goes into is what lets a theorem say how much of the misfit no label’s field could ever remove — degeneracy is a group theorem, and so, it turns out, is most of this fit’s error.

Still open: anharmonic targets, and other molecules’ floors

The obvious open question is the fit with the floor removed at its source. Converting NH₃’s and ND₃’s observed fundamentals to harmonic frequencies with their measured anharmonic constants would give eight numbers that obey the product rule up to the errors in those constants, and a harmonic field fitted to them would have a residual set by the model’s form alone. Its band along the stretch–bend constant, read with no theorem in the noise, would say whether the bond-angle verdict is below nine tenths by more than a hair.

The nearer question is every other fitted field. Water, boron trifluoride and methane are each fitted to two isotopologues or one, and each of their species has a product that every field shares. The floor for each is a few lines of arithmetic on numbers already in hand, and on boron trifluoride, the census’s worst case, where a whole curve of fields fits exactly, it would say whether that exactness is real or whether its single isotopologue simply gave the theorem nothing to contradict.

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.

Force constantIrreducible representationsIsotope substitutionNormal modeValence force field