Thirty-eight levels pin the wrong depth
Worth reading first: The fourth constant chose a curve, not a depth · Three constants, two potentials.
A Morse curve whose range bends and stays bounded, with the stretch in bond lengths, was fitted to carbon monoxide’s first four spectroscopic constants. It matched all four with a curve that dissociates, and that curve was 5.7 per cent too deep. So were its counterparts for three hydrides, by more, and in every case the plain Morse curve, which knows only two constants, came nearer the measured depth. The essay closed on two possibilities. Either four constants are too few, being a description of a narrow skin around the minimum, so that the depth was simply not constrained. Or the family has the wrong shape where the depth is decided.
The four constants were read from carbon monoxide’s lowest four levels. But the published expansion of its vibrational levels was fitted to thirty-eight of them, up to , and the highest lies 62,252 cm⁻¹ above the lowest. Counting the zero-point level, that is 70 per cent of the way from the bottom of the well to the dissociation limit. That is no longer a skin around the minimum. A curve fitted to those levels directly has its depth under direct constraint for the first time in this line of argument, and the question is where the constraint puts it.
More levels, and the depth comes down to a floor
The fits were made on levels to , for from 5 up to 37, by least squares on the term values above . The plain Morse curve has two parameters, a depth and a range. The bounded family has those and the two shape parameters and , and it dissociates exactly when exceeds . Every fit was required to stay in that region.
The bounded family’s fits come in two stages. Up to every best fit is pressed against the edge , with near , at a depth of 12.48 eV. These eighteen levels are fitted to between 0.003 and 0.015 cm⁻¹, but the fit is not free. Asked for its best curve, the family is pushing towards the members that would turn over and never dissociate, and the constraint is what stops it. It is the same preference the linear family showed when carbon monoxide’s measured fourth constant chose its turning-over fit, and it has a reason. A curve that will eventually flatten can bend the way the lowest levels want without paying for it higher up.
From the fits leave the edge, and the depth comes down: 12.44 eV at , 12.18 at 30 and 12.01 at 37. Each is fitted to under a tenth of a wavenumber until the last, which is fitted to 0.34 cm⁻¹ rms with a worst level at 1.1. At thirty-eight levels the bounded family is still 7.0 per cent too deep, deeper than the four-constant member’s 5.7.
The plain Morse curve goes the other way. Fitted to six levels it is 1.5 per cent shallow, and as levels are added its depth climbs through the measured value near and on to 11.46 eV at 37, 2.1 per cent deep. Its residual climbs with it, to 17.5 cm⁻¹ rms over the thirty-eight. It has two parameters and cannot follow the levels, and that is why its depth moves.
The pattern each curve leaves
The residuals show what each fit is doing. The free bounded fit stays within about a wavenumber everywhere and has no pattern worth naming. The four-constant member, which was held only to the bottom, starts at a few hundredths and runs away above , to cm⁻¹ at . Its levels come too low, the signature of a well that is too wide near the top. The plain Morse curve fitted to all thirty-eight levels swings from near to near 30 and back to . That is the shape of a fit spreading its error across the range it cannot match. The plain Morse curve from the two lowest constants alone, off the scale, misses by 655 cm⁻¹.
The one curve with no pattern is the deepest of them, and nothing in its residuals warns of it.
The valley
A fitted depth on its own does not say how firmly the levels hold it, and the answer to the question left by the four-constant essay is in that firmness. So the bounded family was held at a range of depths, from 11.0 to 12.4 eV, with its range and two shape parameters fitted to all thirty-eight levels at each.
The residual is a narrow valley. At 12.0 eV it is 0.43 cm⁻¹, essentially the free fit’s. Two tenths of an electronvolt shallower it is 6.4, at 11.6 it is 12.8, and at the measured 11.22 eV it is 21.4 cm⁻¹, sixty-four times the free fit. On the deep side the valley wall is gentler, 0.67 at 12.2 and 0.95 at 12.4, where the fit is back on the edge of dissociation. Held at the measured depth, the bounded family fits the levels worse than the plain Morse curve does free.
So the four-constant essay’s first possibility is answered, and not in the way it hoped. The depth is not unconstrained. Thirty-eight levels hold it inside a valley two tenths of an electronvolt wide. The valley’s floor is simply not at the measured depth, and it is further from it than the floor four constants gave.
That leaves the second possibility, the shape. A family can fit a set of levels this well and be wrong about something the levels constrain only if the levels constrain it through the family. Here the family carries the depth from the levels it can see to the dissociation limit it cannot, and how it carries it is fixed by its own form at large separation.
What the edge was saying
The first stage deserves a second look, because it is the same behaviour the shorter readings showed and now it can be seen to go away. With eighteen levels or fewer the family’s best dissociating member sits exactly on the boundary beyond which it would stop dissociating. That is not an accident of the starting point: the fits were started from three places, one of them deep inside the dissociating region, and refitted from the answer for the next number of levels, which is off the edge, and the best of them ends there every time. The lowest levels are asking for a curve that bends harder than any dissociating member can, and they get as close as they are allowed.
It was the lowest four levels’ preference that made carbon monoxide’s measured fourth constant pick the turning-over fit, and it was a guessed cubic coefficient that first showed how much of a Morse curve’s agreement near the minimum is borrowed from terms nobody measured. What changes at twenty-five levels is that the levels themselves start to reach the region where a curve that turns over would be visibly wrong. The edge stops being where the fit wants to be, climbs away from , and the depth starts to come down. That transition is the moment the data begin to see the upper half of the well, and it happens between and , about where the levels pass half the depth.
So the two readings of the four-constant essay were not alternatives. The constants were too few, and they were too few in a specific way: the bottom of the well alone prefers a curve that never dissociates, and a dissociating family pressed against that preference gives a depth that means very little. Thirty-eight levels relieve the pressure and the depth starts to mean something, and what it means is still the family’s tail.
Where the curves part
Drawn against the separation, the three curves are indistinguishable at the scale of the figure up to the energy of , a little under eight electronvolts above the minimum. Above it they part, towards 12.01, 11.46 and 11.22 eV. The bounded family’s free fit keeps rising more steeply from about 2.2 Å outward, and the curve held at the measured depth turns towards its asymptote sooner. Everything that distinguishes them lies in the third of the well that no fitted level reaches, and the residuals say that inside it the levels have a strong opinion only through the family.
The spacings past the last level
The cleanest way to see where the excess sits is the old one, a plot of the spacing between neighbouring levels against . The area under that curve down to zero spacing is the dissociation energy measured from the lowest level. Birge and Sponer used it to estimate dissociation energies from spacings that stopped well short of the limit, by extrapolating a straight line.
Below the free fit’s spacings and the expansion’s are the same to a fraction of a wavenumber, by construction. Past it they separate. At the free fit’s spacing is 912 cm⁻¹ against the expansion’s 895, carried forward as a polynomial, and the curve held at the measured depth has 857. The free fit’s 0.785 eV of excess depth, 6,330 cm⁻¹, lies entirely in the area between curves to the right of the last fitted level, because to the left of it the curves coincide.
A Morse curve’s spacings fall on a straight line, and a Morse curve with a bounded bent range bends that line only gently, because its tail is still an exponential. A real bond’s is not. Le Roy and Bernstein showed that the levels just below a dissociation limit follow a law fixed by the inverse power with which the potential approaches its asymptote, and not the straight line an exponential tail gives. So the family carries its own law from to the limit, and that law has to enclose 6,330 cm⁻¹ more area than the real spacings do. Whether a tail of the right power would account for all of it is the question this leaves open.
Every fit, in numbers
The table carries what the two stages look like in the parameters. Through eighteen levels equals to the fourth decimal; at twenty-five it is 0.142 against 0.107; at thirty-one, 0.44; at thirty-eight, 1.31 against 0.139. So the curve’s range at large separation, times its value at the minimum, goes from zero on the edge to 0.77 at thirty-one levels and 0.89 at thirty-eight. Each set of levels added moves the far range up, and each move lowers the depth. The trend is still running at , which is the honest limit of the result: the family’s depth would probably fall further with levels above , and there are none in the expansion it was fitted to.
The surprise is in the agreement
It is natural to read a residual of a third of a wavenumber over thirty-eight levels as a potential that is right. The residuals here are as good as anything a direct potential fit usually reports, and the depth that comes with them is wrong by more than a twentieth. The reason is not hidden: the levels are all in the lower two thirds of the well, and the depth is a property of the top third. What makes it easy to miss is that the levels do constrain the depth sharply. The valley is real. A fit that reports its depth with an uncertainty from the curvature of that valley would quote 12.0 ± 0.1 eV and be confidently wrong, because the uncertainty belongs to the family, not to the molecule.
That is the lesson the cubic coefficient taught at the other end of the well, where a Morse curve’s guessed cubic was right only because the curve carried every higher term with it. It is also what two coefficients that are not a potential found, and three constants with two potentials before that. What a set of measurements decides about a potential is decided through the form the potential is given.
How the fits were made
The levels are the term values implied by the published sixth-order expansion for , compiled by Huber and Herzberg from a fit to . Each trial curve is solved by a finite-difference Hamiltonian on a uniform grid from 0.6 to 3.2 bond lengths, its lowest thirty-eight eigenvalues found by Sturm bisection, at 4,000 and 8,000 points. The two are combined by Richardson’s rule, which removes the grid’s error: against the exact levels of the plain Morse curve the combination is right to 0.002 cm⁻¹ at every level, where either grid alone is off by several wavenumbers at . Fits are Levenberg–Marquardt with finite-difference derivatives, from two fixed starts and from the previous fit’s answer, then checked by refitting each from the next fit’s answer, taking the best. The measured depth is the dissociation energy from the same compilation plus the expansion’s zero-point level. The refusal is a target that is known: the thirty-eight levels of the family’s own member at the measured depth, fitted from exactly the same starts. They come back at 11.29 eV with a residual of 0.22 cm⁻¹, not at 11.22 with none, so the procedure stalls short of an exact answer by about 0.06 eV and a fifth of a wavenumber. That is its tolerance, and it is a twelfth of the 0.78 eV the measured levels put between the fit and the measurement.
Where the model stops
The expansion stands in for the measurements. The levels are those the published expansion implies, not the line positions, and an expansion fitted to is smooth where the data have scatter. A fit to the lines themselves would weight levels by their measured precision, and the high levels, measured less precisely, would count for less.
One family. Everything here is about a Morse curve with a bounded bent range. A family with a different tail, an inverse power joined to the Morse form, would carry the same levels to a different depth, and the claim is only that this family’s tail puts it at 12.0 eV.
No rotation. Only the vibrational levels at zero angular momentum are fitted. Each level’s rotational constant is an average of over that level’s wavefunction, which is not the average a harmonic field supplies, and its centrifugal distortion follows from the curvature by one relation. Both would constrain the curve’s width at each energy, which the term values do not carry, and would sharpen the test. They are also what a structure is read from, so a curve right in its levels and wrong in its rotational constants would be a bond length out of a spectrum that disagreed with itself.
The trend is still running. The depth falls with each set of levels added past . Levels above would very likely lower it further. How far is a question about data the expansion does not contain.
Who measured what
The level expansion and the dissociation energy are from Huber and Herzberg’s compilation. Morse’s curve is of 1929, Birge and Sponer’s extrapolation of 1926, and Le Roy and Bernstein’s treatment of levels near a dissociation limit of 1970. Fitting a potential directly to measured levels rather than to constants is the direct-potential-fit approach of Le Roy and co-workers, and its expanded Morse oscillator is a more flexible form of what is fitted here. What is computed here is this family’s fit by number of levels, the edge it presses against, the valley of the residual against depth, and where the excess lies.
Still open: a tail that knows its power, and the lines themselves
The obvious open question is the tail. Join the bounded family to an inverse-power tail at a separation beyond the last level’s turning point, with the power the separated atoms’ ground states require, and refit the thirty-eight levels with the joining point and coefficient free or fixed. If the valley’s floor moves to the measured depth, the 6,330 cm⁻¹ belonged to an exponential tail. If it does not, it belongs to the middle of the curve, between the last level and the tail.
The nearer question is what the lines themselves say. The expansion smooths the high levels into a polynomial, and the last few levels below are where the spacings start to bend. A fit to measured term values with their own uncertainties, or with each level’s rotational constant as a second datum, would say whether the bend the depth needs is already visible in the data and was smoothed away, or is not there to be seen.
Shares its objects with
Essays naming at least two of the same things, that neither author linked.
- A precision is a spacing — both name least-squares, model limit, underdetermination
- Adding data made it worse — both name least-squares, model limit, underdetermination
- An end effect with two signs — both name least-squares, model limit, underdetermination
- How many parameters a curve is worth — both name least-squares, model limit, underdetermination
- More coordinates than motions — both name least-squares, model limit, underdetermination
- One integer, and everything it changes — both name least-squares, model limit, underdetermination
Named objects
A dashed tag is an object no other essay names yet.
AnharmonicityDissociationLeast-squaresModel limitMorse potentialPerturbation theoryUnderdetermination