This title is an example of what has come to be called Betteridge’s law. This is a relatively recent name for an old phenomenon: if a title is posed as a question, the answer is **no**. This is especially true in science, whether the authors are conscious of it or not.

Pengfei Li completed his Ph.D. recently, fitting all manner of dark matter halos as well as the radial acceleration relation (RAR) to galaxies in the SPARC database. For the RAR, he found that galaxy data were consistent with a single, universal acceleration scale, g_{+}. There is of course scatter in the data, but this appears to us to be consistent with what we expect from variation in the mass-to-light ratios of stars and the various uncertainties in the data.

This conclusion has been controversial despite being painfully obvious. I have my own law for data interpretation in astronomy:

Obvious results provoke opposition. The more obvious the result, the stronger the opposition.

S. McGaugh, 1997

The constancy of the acceleration scale is such a case. Where we do not believe we can distinguish between galaxies, others think they can – using our own data! Here it is worth contemplating what all is involved in building a database like SPARC – we were the ones who did the work, after all. In the case of the photometry, we observed the galaxies, we reduced the data, we cleaned the images of foreground contaminants (stars), we fit isophotes, we built mass models – that’s a very short version of what we did in order to be able to estimate the acceleration predicted by Newtonian gravity for the observed distribution of stars. That’s one axis of the RAR. The other is the observed acceleration, which comes from rotation curves, which require even more work. I will spare you the work flow; we did some galaxies ourselves, and took others from the literature in full appreciation of what we could and could not believe — which we have a deep appreciation for because we do the same kind of work ourselves. In contrast, the people claiming to find the opposite of what we find obtained the data by downloading it from our website. The only thing they do is the very last step in the analysis, making fits with Bayesian statistics the same as we do, but in manifest ignorance of the process by which the data came to be. This leads to an underappreciation of the uncertainty in the uncertainties.

This is another rule of thumb in science: *outside groups are unlikely to discover important things that were overlooked by the group that did the original work*. An example from about seven years ago was the putative 126 GeV line in Fermi satellite data. This was thought by some at the time to be evidence for dark matter annihilating into gamma rays with energy corresponding to the rest mass of the dark matter particles and their anti-particles. This would be a remarkable, Nobel-winning discovery, if true. Strange then that the claim was not made by the Fermi team themselves. Did outsiders beat them to the punch with their own data? It can happen: sometimes large collaborations can be slow to move on important results, wanting to vet everything carefully or warring internally over its meaning while outside investigators move more swiftly. But it can also be that the vetting shows that the exciting result is not credible.

I recall the 126 GeV line being a big deal. There was an entire session devoted to it at a conference I was scheduled to attend. Our time is valuable: I can’t go to every interesting conference, and don’t want to spend time on conferences that aren’t interesting. I was skeptical, simply because of the rule of thumb. I wrote the organizers, and asked if they really thought that this would still be a thing by the time the conference happened in few months’ time. Some of them certainly thought so, so it went ahead. As it happened, it wasn’t. Not a single speaker who was scheduled to talk about the 126 GeV line actually did so. In a few short months, if had gone from an exciting result sure to win a Nobel prize to nada.

This happens all the time. Science isn’t as simple as a dry table of numbers and error bars. This is especially true in astronomy, where we are observing objects in the sky. It is never possible to do an ideal experiment in which one controls for all possible systematics: the universe is not a closed box in which we can control the conditions. Heck, we don’t even know what all the unknowns are. It is a big friggin’ universe.

The practical consequence of this is that the uncertainty in any astronomical measurement is almost always larger than its formal error bar. There are effects we can quantify and include appropriately in the error assessment. There are things we can not. We know they’re there, but that doesn’t mean we can put a meaningful number on them.

Indeed, the sociology of this has evolved over the course of my career. Back in the day, everybody understood these things, and took the stated errors with a grain of salt. If it was important to estimate the systematic uncertainty, it was common to estimate a wide band, in effect saying “I’m pretty sure it is in this range.” Nowadays, it has become common to split out terms for random and systematic error. This is helpful to the non-specialist, but it can also be misleading because, so stated, the confidence interval on the systematic looks like a 1 sigma error even though it is not likely to have a Gaussian distribution. Being 3 sigma off of the central value might be a lot more likely than this implies — or a lot less.

People have become more careful in making error estimates, which ironically has made matters worse. People seem to think that they can actually believe the error bars. Sometimes you can, but sometimes not. Many people don’t know how much salt to take it with, or realize that they should take it with a grain of salt at all. Worse, more and more folks come over from particle physics where extraordinary accuracy is the norm. They are completely unprepared to cope with astronomical data, or even fully process that the error bars may not be what they think they are. There is no appreciation for the uncertainties in the uncertainties, which is absolutely fundamental in astrophysics.

Consequently, one gets overly credulous analyses. In the case of the RAR, a number of papers have claimed that the acceleration scale isn’t constant. Not even remotely! Why do they make this claim?

Below is a histogram of raw acceleration scales from SPARC galaxies. In effect, they are claiming that they can tell the difference between galaxies in the tail on one side of the histogram from those on the opposite side. We don’t think we can, which is the more conservative claim. The width of the histogram is just the scatter that one expects from astronomical data, so the data are consistent with zero intrinsic scatter. That’s not to say that’s necessarily what Nature is doing: we can never measure zero scatter, so it is always conceivable that there is some intrinsic variation in the characteristic acceleration scale. All we can say is that if is there, it is so small that we cannot yet resolve it.

Posed as a histogram like this, it is easy to see that there is a characteristic value – the peak – with some scatter around it. The entire issue it whether that scatter is due to real variation from galaxy to galaxy, or if it is just noise. One way to check this is to make quality cuts: in the plot above, the gray-striped histogram plots every available galaxy. The solid blue one makes some mild quality cuts, like knowing the distance to better than 20%. That matters, because the acceleration scale is a quantity that depends on distance – a notoriously difficult quantity to measure accurately in astronomy. When this quality cut is imposed, the width of the histogram shrinks. The better data make a tighter histogram – just as one would expect if the scatter is due to noise. If instead the scatter is a real, physical effect, it should, if anything, be *more* pronounced in the better data.

This should not be difficult to understand. And yet – other representations of the data give a different impression, like this one:

This figure tells a very different story. The characteristic acceleration does not just scatter around a universal value. There is a clear correlation from one end of the plot to the other. Indeed, it is a perfectly smooth transition, because “Galaxy” is the number of each galaxy ordered by the value of its acceleration, from lowest to highest. The axes are not independent, they represent identically the same quantity. It is a plot of x against x. If properly projected it into a histogram, it would look like the one above.

This is a terrible way to plot data. It makes it look like there is a correlation where there is none. Setting this aside, there is a potential issue with the most discrepant galaxies – those at either extreme. There are more points that are roughly 3 sigma from a constant value than there should be for a sample this size. If this is the right assessment of the uncertainty, then there is indeed some variation from galaxy to galaxy. Not much, but the galaxies at the left hand side of the plot are different from those on the right hand side.

But can we believe the formal uncertainties that inform this error analysis? If you’ve read this far, you will anticipate that the answer to this question obeys Betteridge’s law. No.

One of the reasons we can’t just assign confidence intervals and believe them like a common physicist is that there are other factors in the analysis – nuisance parameters in Bayesian verbiage – with which the acceleration scale covaries. That’s a fancy way of saying that if we turn one knob, it affects another. We assign priors to the nuisance parameters (e.g., the distance to each galaxy and its inclination) based on independent measurements. But there is still some room to slop around. The question is really what to believe at the end of the analysis. We don’t think we can distinguish the acceleration scale from one galaxy to another, but this other analysis says we should. So which is it?

It is easy at this point to devolve into accusations of picking priors to obtain a preconceived result. I don’t think anyone is doing that. But how to show it?

Pengfei had the brilliant idea to perform the same analysis as Marra et al., but allowing Newton’s constant to vary. This is Big G, a universal constant that’s been known to be a constant of nature for centuries. It surely does not vary. However, G appears in our equations, so we can test for variation therein. Pengfei did this, following the same procedure as Mara et al., and finds the same kind of graph – now for G instead of g_{+}.

You see here the same kind of trend for Newton’s constant as one sees above for the acceleration scale. The same data have been analyzed in the same way. It has also been plotted in the same way, giving the impression of a correlation where there is none. The result is also the same: if we believe the formal uncertainties, the best-fit G is different for the galaxies at the left than from those to the right.

I’m pretty sure Newton’s constant does not vary this much. I’m entirely sure that the rotation curve data we analyze are not capable of making this determination. It would be absurd to claim so. The same absurdity extends to the acceleration scale g_{+}. If we don’t believe the variation in G, there’s no reason to believe that in g_{+}.

So what is going on here? It boils down to the errors on the rotation curves not representing the uncertainty in the circular velocity as we would like for them to. There are all sorts of reasons for this, observational, physical, and systematic. I’ve written about this at great lengths elsewhere, and I haven’t the patience to do so again here. it is turgidly technical to the extent that even the pros don’t read it. It boils down to the ancient, forgotten wisdom of astronomy: you have to take the errors with a grain of salt.

Here is the cumulative distribution (CDF) of reduced chi squared for the plot above.

Two things to notice here. First, the CDF looks the same regardless of whether we let Newton’s constant vary or not, or how we assign the Bayesian priors. There’s no value added in letting it vary – just as we found for the characteristic acceleration scale in the first place. Second, the reduced chi squared is rarely close to one. It should be! As a goodness of fit measure, one claims to have a good fit when chi squared equal to one. The majority of these are not good fits! Rather than the gradual slope we see here, the CDF of chi squared should be a nearly straight vertical line. That’s nothing like what we see.

If one interprets this literally, there are many large chi squared values well in excess of unity. These are bad fits, and the model should be rejected. That’s exactly what Rodrigues et al. (2018) found, rejecting the constancy of the acceleration scale at 10 sigma. By their reasoning, we must also reject the constancy of Newton’s constant with the same high confidence. That’s just silly.

One strange thing: the people complaining that the acceleration scale is not constant are only testing that hypothesis. Their presumption is that if the data reject that, it falsifies MOND. The attitude is that this is an automatic win for dark matter. Is it? They don’t bother checking.

We do. We can do the same exercise with dark matter. We find the same result. The CDF looks the same; there are many galaxies with chi squared that is too large.

Having found the same result for dark matter halos that we found for the RAR, if we apply the same logic, then all proposed model halos are excluded. There are too many bad fits with overly large chi squared.

We have now ruled out all conceivable models. Dark matter is falsified. MOND is falsified. Nothing works. Look on these data, ye mighty, and despair.

But wait! Should we believe the error bars that lead to the end of all things? What would Betteridge say?

Here is the rotation curve of DDO 170 fit with the RAR. Look first at the left box, with the data (points) and the fit (red line). Then look at the fit parameters in the right box.

Looking at the left panel, this is a good fit. The line representing the model provides a reasonable depiction of the data.

Looking at the right panel, this is a terrible fit. The reduced chi squared is 4.9. That’s a lot larger than one! The model is rejected with high confidence.

Well, which is it? Lots of people fall into the trap of blindly trusting statistical tests like chi squared. Statistics can only help your brain. They can’t replace it. Trust your eye-brain. This is a good fit. Chi squared is overly large not because this is a bad model but because the error bars are too small. The absolute amount by which the data “miss” is just a few km/s. This is not much by the standards of galaxies, and could easily be explained by a small departure of the tracer from a purely circular orbit – a physical effect we expect at that level. Or it could simply be that the errors are underestimated. Either way, it isn’t a big deal. It would be incredibly naive to take chi squared at face value.

If you want to see a dozen plots like this for all the various models fit to each of over a hundred galaxies, see Li et al. (2020). The bottom line is always the same. The same galaxies are poorly fit by any model — dark matter or MOND. Chi squared is too big not because all conceivable models are wrong, but because the formal errors are underestimated in many cases.

This comes as no surprise to anyone with experience working with astronomical data. We can work to improve the data and the error estimation – see, for example, Sellwood et al (2021). But we can’t blindly turn the crank on some statistical black box and expect all the secrets of the universe to tumble out onto a silver platter for our delectation. There’s a little more to it than that.