Why the search for particle dark matter is really hopeless, part 2

Why the search for particle dark matter is really hopeless, part 2

Last time, I discussed the requirement for dark matter to naturally explain the observed MONDian phenomenology. That is a high standard that no dark matter model currently meets. But it is at least conceivable, and the widespread attitude in the community is that whatever needs to happen will eventually emerge from the effects of feedback in ever-higher resolution simulations. Here I attempt to explain why that is an unobtainable chimera. There will be math.

We’ve been here before. Newton proposed his universal law of gravity over three centuries ago to explain the orbits of the moon and planets. For our purposes here, it is useful to write what Newton taught us in terms of the acceleration the gravity of a mass M causes on a test mass at distance r:

gN=GMr2g_N = \frac{GM}{r^2}.

This is for a point mass; applying this to an extended mass distribution like a galaxy requires solving the Poisson equation numerically. Such details are not important here; all that really matters is that Newton gave us an equation that relates the kinematic acceleration to the mass distribution. We can observe both of these quantities.

An analogous situation occurs in galaxies. To paraphrase Newton&, everything happens as if MOND is the effective force law. There is a one-to-one relation between what we see in the distribution of stars and gas and what we get for the acceleration. This is such a direct connection that one can write down a simple equation for it:

a=ν(gN/a0)gNa = \nu(g_N/a_0)\, g_N.

Here, gN has exactly the same meaning as above. It is the acceleration that Newton predicts for the observed mass distribution. MOND amplifies this by the function ν\nu to become the acceleration a. This is the line in the plot below:

The radial acceleration relation over many orders of magnitude, from the lowest accelerations observed to the solar system. Newtonian dynamics occur at high accelerations where a = gN; MOND happens at low accelerations where a = (a0 gN)1/2. This behavior is uniquely predicted by MOND. Adapted from Lelli et al. (2024).

The function ν\nu has the properties that ν→1\nu \rightarrow 1 at high accelerations a≫a0a \gg a_0 and ν→a0/gN\nu \rightarrow \sqrt{a_0 / g_N} at low accelerations a≪a0a \ll a_0 where Milgrom’s constant is observed to have the value a0=1.2×10−10ms−2a_0 = 1.2 \times 10^{-10}\;\mathrm{m}\,\mathrm{s}^{-2}. Possible interpolation functions* are discussed by Milgrom & Sanders, but the details are unimportant here. What is important is that the acceleration depends only on the observed distribution of mass as encoded in the Newtonian acceleration gN. The excess acceleration that we normally attribute to dark matter is determined by the normal matter that we see. This is generally true for galaxies spanning properties over many orders of magnitude in luminosity, surface brightness, gas fraction, etc.

This behavior is in the data. We’ve bent over backwards to demonstrate that the radial acceleration relation is present in the data independent of any considerations specific to MOND. It is not my fault that the data look exactly like what MOND predicts. Dark matter does not naturally do that. If we entertain unnatural models, what does it need to do?

The requirement for a dark matter halo reduces to a simple equation. The observed acceleration is attributed to a dark matter halo gDM in addition to the observed baryonic material gN. Consequently, the ordinate of the RAR is a=gDM+gNa = g_{DM}+g_N. We know empirically that a=ν(gN/a0)gNa = \nu(g_N/a_0)\, g_N, so we can solve for what the dark matter is doing:

gDM=[ν(gN/a0)−1]gNg_{DM} = [\nu(g_N/a_0)-1]\, g_N.

That’s it^. The acceleration# due to dark matter can be written entirely in terms of that due to the observed baryons. Light tells mass where to go, with spooky precision. The baryonic tail wags the dark matter dog.

Here’s an example, using the same$ galaxies from before:

The rotation curves of NGC 6946 (left) and NGC 1560 (right) along with the baryonic rotation curve VN=gN rV_N = \sqrt{g_N\ r} (blue), the corresponding MOND rotation curve VM=ν gNrV_M = \sqrt{\nu\ g_N\, r} (green), and the rotation curve of the dark matter VDM=(ν−1)gNrV_{DM} = \sqrt{(\nu-1)\, g_N\, r} (red) following from the math above.

The chief lesson of this is that we don’t need dark matter; we just need to know about the RAR [ν(gN/a0)\nu(g_N/a_0)] and measure the baryon distribution that informs gNg_N. If we insist on calling the rest dark matter, then if follows from the difference via gDM=[ν(gN/a0)−1]gNg_{DM} = [\nu(g_N/a_0)-1]\, g_N. The baryon distribution determines the dark matter distribution!. Note that the the reverse is not true: the dark matter distribution appears only as a quantity derived from the baryon distribution.

There is nothing in conventional dark matter theory that explains this. Indeed, an important aspect of viable dark matter candidates is that they not interact directly with baryons, lest they be readily detected in the laboratory. And yet we have here a simple equation that tells us the two are intimately connected. So, if you want to claim that a dark matter model explains the observed phenomenology, you need to be able to derive this equation from it. This has not been done, and there is every reason to think it cannot be done.

Note that cannot be done is explicitly physics reasoning, following the reductionist principle: if you can write down a simple equation to explain an observed phenomenon, then that’s what you should do. The task then is to build a theory that yields that equation, not to ignore Occam’s razor by building Rube Goldberg simulations with dark matter and AGN feedback and star formation feedback and other ugly details we’d prefer not to discuss to sorta kinda but not really mimic what the simple equation already told you. In short,

The only way to reconcile dark matter with observations is by very precisely mimicking the predictions of MOND.

Again, this is explicitly physics reasoning. I realized this a long time ago – in the previous century – but have been too polite to say it so explicitly. I’m not the first to notice; Bob Sanders made the same point that gas-rich galaxies amount to a practical falsification of dark matter. Gas rich galaxies, like NGC 1560 above, are particularly important because their stellar mass is sub-dominant, so unimportant to the MOND fit. The one free parameter, the mass-to-light ratio, is effectively irrelevant. The predicted rotation curve follows directly from the observed gas distribution. We can’t do that with dark matter%.

Returning to Newton’s as if, we can make an analogy to the solar system. We know the solar system looks like an inverse square force law, but what if it is really an inverse cube law? The sun could have a dark matter halo that makes the force law look like inverse square because there is dark matter arranged just so as to always produce that result. And not just the sun – each and every planet would have to have its own dark matter subhalo so that its moons would also appear to obey as if the inverse square law.

Absurd? Yes, very! And yet that is exactly what we have to believe to sustain the dark matter interpretation of galaxy rotation curves. This must hold for the particular mass distribution of each and every individual galaxy. Dark matter somehow always manages to be arranged just so as to mimic MOND.

So what kinds of dark matter does this line of reasoning exclude? Basically all of them:

Fanciful illustration of the landscape of dark matter candidates by Tim Tait. I’ve crossed out all non-viable possibilities, which is all those illustrated here.

The vast majority of ideas currently being contemplated simply come up with a new particle that could provide the extra mass needed by cosmology. That might suffice for cosmology, but just adding mass does not explain the strong coupling with the observed distribution of baryons in galaxies. A viable dark matter candidate needs to do both, but the vast majority of them have been constructed in ignorance (or denial) of the need to explain the MONDian phenomenology.

It might be possible to conceive of a type of dark matter that satisfies both cosmological and galaxy dynamical requirements. Examples of such possibilities include superfluid dark matter and bipolar dark matter. These hybrid solutions are well outside the main stream of thought on dark matter candidates at present, but are more viable than any of the more commonly discussed dark sector candidates.

Of course, Milgrom already built a theory that gives us the right equation, so that seems like the obvious avenue to pursue. The remaining challenge is to build a deeper theory that encompasses both MOND and General Relativity. Searching for new particles invented in ignorance (or denial) of the observed phenomenology will inevitably prove fruitless.


&Newton preceded his Universal Law of Gravity with as if as a caution: he didn’t claim to know why it worked, and it seemed crazy at the time that one could have action at a distance, but everything happened as if the force of gravity was proportional to the product of the masses and the inverse square of the distance between them. The same holds today: we may not know why MOND works, but everything in galaxies happens as if MOND is the effective force law.

*One can write MOND as a=ν(gN/a0)gNa = \nu(g_N/a_0)\, g_N or as μ(a/a0)a=gN\mu(a/a_0)a = g_N. This choice is important to developing deeper theories. For our purposes here, it is a distinction without a difference.

^In McGaugh (2004), I generalized this to include a parameter Q that enabled a dark matter model to have a different stellar mass-to-light ratio from that preferred by MOND. That freedom has practically disappeared with the progress we’ve made with near-infrared observations and stellar population modeling, so I don’t bother to write the slightly more complicated math here – see the paper if you’re interested. It makes no difference to the argument here beyond adding the appearance of a tiny bit of flexibility where there isn’t really any.

#The circular velocity of a test particle is related to the centripetal acceleration by a=V2ra = \frac{V^2}{r}.

$One can always quibble with any one galaxy for this reason or that, but we’ve shown this works in well over a hundred galaxies, so these examples suffice to illustrate the general rule whatever the quibble.

!Sadly, the formula for gDMg_{DM} does not translate into a simple function for the dark matter density ρDM\rho_{DM} because it depends on the details of the baryon distribution, including their bumps and wiggles (as reflected in the red lines for the dark matter), which is unique to each galaxy.

%At present, nearly all conventional models relate the dark matter halo of a galaxy to its stellar mass but ignore the gas.

Why the search for particle dark matter is hopeless, part 1

Why the search for particle dark matter is hopeless, part 1

Previously, I made a list of five criteria for a convincing dark matter detection. I discussed the first four, which I hope are obvious to everyone who works in the field, but it seemed worthwhile to write them out explicitly given the hint of a detection by LZ. I left the fifth criterion – that whatever the dark matter is it should naturally explain the observed MONDian phenomenology – to be discussed separately, so that’s what we’ll do here.

Explaining the observed MONDian phenomenology is a high bar. That it does not follow naturally from dark matter should be obvious. It used to be, but apparently it no longer is, so let’s think it through. Indeed, it’s going to take two posts to get through.

First, some definitions:

Cold Dark Matter (CDM): the non-relativistic, non-baryonic dark matter that is standard to cosmology. Usually presumed to be a WIMP, it could in principle be any mass that is slow-moving and does not participate in big bang nucleosynthesis. I will nevertheless assume, as the vast majority of scientists do, that it is some new particle, if not necessarily a WIMP.

Generic Dark Matter (DM): could be any unseen mass, but here I will use it to mean non-WIMP particle DM candidates that don’t necessarily do what pure CDM predicts. For example, warm (WDM) or self-interacting dark matter (SIDM) might produce cored rather than cuspy dark matter halos. So by generic DM I mean any new particle that doesn’t interact with baryons except through gravity. This is pretty much everything that is commonly discussed in particle physics, with only a few obscure exceptions. Cold dark matter is of course a subset of generic DM, but it makes specific predictions, like cuspy halos, while generic DM is just there adding mass.

Modified Newtonian Dynamics (MOND): the hypothesis that the effective force law at low accelerations a≪a0=1.2×10−10ms−2a \ll a_0 = 1.2 \times 10^{-10}\,\mathrm{m}\,\mathrm{s}^{-2} becomes a→a0gNa \rightarrow \sqrt{a_0\,g_N}, where gNg_N is the acceleration predicted by Newton (gN=GM/r2g_N = GM/r^2 for a point mass M; more generally one must solve the Poison equation for the observed mass distribution). The Newtonian and MOND regimes are connected by an interpolation function ν\nu so that a=ν(gN/a0)gNa = \nu(g_N/a_0)\, g_N. The interpolation function is not fundamental and only matters around the transition a≈a0a \approx a_0.

Data wanting explanation

OK, so let’s see how these compare with galaxy data. I’m only going to talk about galaxy data because I’ve talked about everything else in other posts, and to offer a reflection of those who think CMB data are the only form of evidence that matters, to the exclusion of any other contrary evidence. If we adopt the same attitude, then galaxy data already suffice to falsify DM. Note my use of the generic DM there, not just CDM.

So what do we need to understand about galaxies? Their many regularities, as summarized in this table along with whether each of CDM, generic DM, and MOND provide a satisfactorily explanation:

Observation[L]CDMGeneric DMMOND
Flat rotation curvesNoSureBy construction
Inner rotation curve shapes maximal disks & diversityNoMaybe?Yes
Stellar Mass-to-Light RatiosNowhateverYes
Baryonic Tully-Fisher RelationNoNoYes
Radial Acceleration RelationNoNoYes
Central Surface Density RelationNo NoYes
Renzo’s Rule (Sancisi’s Law)NoMaybe? but no, not reallyYes
Acceleration scaleLOL no, CDM is scale freeNoYes

Satisfactory is doing a lot of work in evaluating whether each case provides a satisfactory explanation. It is easy to find claims in the literature to explain some if not all of these items, but these claims are generally not satisfactory. From the perspective of the current sociology of the field, they don’t need to be: as long as someone asserts that [insert observation] works in [C]DM, the community hears what it wants to hear, mentally checks the box, and files it under “not a problem” even when it remains very much a problem.

Rotation curve shapes: flat far out; baryon-like near-in

I’ve written about all of these things at one point or another here, but for starters, let’s consider flat rotation curves. Flat rotation curves were, of course, a primary reason why we invoked dark matter in the first place. Dark matter can certainly explain flat rotation curves at the level of “anything in excess of the baryonic rotation curve is attributable to dark matter.” This is exactly what generic DM was initially invoked to do. Part and parcel of this early construction was that dark matter was not needed at small radii. For the bright galaxies studied in the ’70s and early ’80s, the stars explained the inner parts of rotation curves quite well, a condition dubbed “maximum disk.” The inner regions were star-dominated, and dark matter only became relevant far out. That’s why I’ve given generic DM a “sure” grade: that’s what it was invoked to do.

The rotation curve (data points) of the high surface brightness spiral galaxy NGC 7331. The blue line is the Newtonian prediction for the observed stars and gas. This disk is maximal, meeting the rotation curve data at small radii for a stellar population mass-to-light ratio of 0.42M⊙/L⊙0.42\;M_{\odot}/L_{\odot} at [3.6] microns. (Nominally we expect a typical value of 0.5 in this band, so maximum disk here is already a bit on the low side in terms of stellar populations: there isn’t much leeway to decrease this further as would be necessary to accommodate a cuspy dark matter halo.) Despite the good match at small radii, the need for something extra, be it dark matter or MOND, becomes increasingly clear further out.

It is a low bar to say “any dynamical mass that’s not luminous is dark.” This, by itself, is not a satisfactory explanation: it is just an inference that when what we see falls short, then DM makes up the rest. To become a satisfactory explanation, one would need a theory that predicts this as an outcome. [L]CDM is better than generic DM in that it does make predictions.

CDM predicts NFW halos, at least when left to itself in dark matter-only simulations. These do not lead to flat rotation curves. NFW halos have rotation curves that rise steeply, turnover, then fall gradually. They are not flat over an extended region as routinely observed in real galaxies. In order to obtain the observed flatness requires a combination of baryons plus dark matter. That then requires a good theory for how the luminous galaxy forms, which is a lot more complicated than just an NFW halo. There are now lots of simulations that do this, but they generally don’t agree with each other, so we can’t claim to have a clear prediction from LCDM.

Despite their differences, detailed simulations of galaxy formation do share some common failure modes. Simulated galaxies tend to be too dark matter dominated so that (i) bright, high surface brightness objects often have submaximal disks, unlike real galaxies like NGC 7331 above or our own Milky Way. The shapes of simulated rotation curves tend to (ii) rise too fast relative to the data (those pesky cusps are hard to get rid of) and (iii) the outer rotation curves tend to fall too much, typically having a gradual decline rather than an extended flat portion. I say “tend to” and “typically” because there are a lot of simulations that share these failings, but the details differ from simulation to simulation.

I guess I should give an example. Here is a nice one from the SIMBA simulations:

Figure 2 from Glowacki et al. (2020): Example galaxies from the SIMBA snapshots and their corresponding rotation curves. Galaxies have been selected to closely match galaxies in the THINGS survey, in stellar mass and SFR. Left side: the moment-0 maps for the H I , H2, and stellar distributions for the galaxy. Three viewing angles are offered, along the x, y, and z-axis of the snapshot box. The color bars are in units of solar masses M⊙. Right side: rotation curves, generated from the mass distribution of all particles making up the galaxy (giving the solid black curve), and alternatively from the velocity of the gas particles (dot–dashed blue curve).

This is good: it shows us the distribution of the component baryons by phase (atomic gas, molecular gas, and stars), and the total rotation curves*. I think these are good simulation results in that they provide a reasonable depiction of what we should expect in LCDM. They do not look like real galaxies.

Apparently they look close enough to pass casual inspection. They are qualitatively correct in that a lot of rotation curves for bright galaxies go up steeply, peak, then fall some to the flat portion (NGC 7331 is a typical example). However, the rotation curves of these simulated objects illustrate two of the three typical problems noted above: the disks are too submaximal (compare the green dashed lines to the blue line for NGC 7331) and their rotation curves fall too far before flattening out. Here the simulated rotation curves don’t really flatten out at all, a problem I’ve seen in other simulations, not just SIMBA. I find it remarkable that half a century after flat rotation curves helped found the dark matter paradigm, typical simulations in the standard model of cosmology routinely fail to satisfactorily reproduce this essential observational fact.

The third problem is that simulated rotation curves tend to rise too rapidly, which leads to the cusp-core problem. This problem is particularly pronounced in dwarf galaxies, but such low mass galaxies are not well resolved by most simulations. This issue is addressed by specialized, high resolution simulations like those described by Ruan et al. (2025), which I previously discussed in the context of Tully-Fisher. They do a good job of explaining the small scatter in the Tully-Fisher relation if you don’t look too closely into how that is accomplished. They also show example rotation curves, which exemplify the issue with the inner shape:

The rotation curve of simulated dwarf galaxy-like object r614 (red line) from the top panel of Fig. 2 of Ruan et al. (2025) compared to a real galaxy of similar mass, the low surface brightness galaxy F583-1 (data points).

The suite of simulations from which this example is drawn is one of the better ones in terms of making realistic-looking galaxy-like objects. They are descended from the family of simulations that first claimed to convert cusps into cores. That hasn’t worked out here: the rotation curve of simulated dwarf r614 looks exactly like a cuspy NFW halo. This is what LCDM generically predicts, but real galaxies don’t do that. [Same as it ever was – see Fig. 14 of McGaugh & de Blok (1998).]

Note that I am not saying the simulations are wrong. I think they often do a reasonable job of showing what we should expect for LCDM. What I am saying is that LCDM simulations don’t look like reality. If we believe the simulations get the predictions of LCDM right, then LCDM is falsified. If we don’t want to grant credence to that line of thought, then we either tweak our simulations in an effort to iterate towards something that looks less unrealistic, or simply ignore the uncomfortable aspects of the data. Those two approaches pretty much encompass the past thirty years of the field of galaxy formation.

Regardless of any notional improvements to future simulations, LCDM does not generically get flat rotation curves right – hence the “no” in the table above. I give generic DM a “sure” in this category because the bar is set lower. LCDM at least makes predictions, albeit filtered through the effects of baryonic feedback, which is treated differently by different simulations. There is no prediction for generic DM, so sure, whatever. To illustrate how low this bar is, consider the difference between fitting the data and predicting it. If presented with rotation curve data, we can fit it with any DM halo model you can imagine. In that process, we will reproduce the observed flatness by construction. That’s what a fit does. However, predicting the rotation curve is a much harder proposition. There is no reason to expect baryonic disk+DM halo to result in a flat rotation curve. Indeed, there is only a small sector of the available parameter space where that is the result; it is quite easy to build a plausible-seeming model that has either a rising or a declining rotation curve. I was surprised in my earliest modeling attempts to find just how unlikely the result was to be a flat rotation curve. Lots of subsequent simulations corroborate that without confronting the issue.

Note that MOND, like dark matter, was invented to explain flat rotation curves. That rotation curves are flat happens by construction; there is no evidentiary value in checking this box. Positive evidence does follow from more extended rotation curves than were known at the time of its invention. But aside from that relatively recent development, there’s not much to brag about in the first row of the table for any of the ideas listed there.

The different hypotheses do become distinguishable in the subsequent rows. MOND made specific, a priori predictions for the other entries in the table, and these were all realized by subsequent observations. This behavior, which I refer to collectively as the MOND phenomenology, was not anticipated by dark matter, and is not readily explained by it. Moreover, predictions matter. Even if we do come up with a DM explanation for each of the things MOND predicted a priori, why should we privilege the idea that failed to make the same predictions? This is the opposite of how science is supposed to work.

Rotation curve shapes

We’ve touched on the concept of maximum disk in bright spirals where the inner shape of the rotation curve is stipulated by the observed distribution of stars. This is not expected in LCDM, where the effects of cuspy halos and adiabatic compression combine to preclude disks from becoming maximal – hence the submaximal disks in SIMBA and other simulations. However, feedback can also play a role, so it is not entirely clear what the end results might be for high surface brightness galaxies. (The typical assertion from theory is that high mass galaxies have deep enough potential wells that the impact of feedback is minimal, but that leads to submaximal disks.)

Matters become simpler in low surface brightness galaxies, which are observed to be have large mass discrepancies. This dark matter domination extends to small radii; stellar mass has little influence on the dynamics. Consequently, one expects rotation curves to simply trace that of the dark matter halo. At a given mass, different DM-dominated galaxies reside in pretty much the same DM halo, so should have pretty much the same rotation curve. They do not.

Though long known observationally, this came as a surprise to the simulation community, who coined the term “diversity” to refer to it. They failed to notice that the diversity of rotation curve shapes correlates with the diversity of surface brightnesses: high surface brightness galaxies have rapidly rising rotation curves, low surface brightness galaxies have gradually rising rotation curves. The relation is so strong, it can be seen by eye:

Rotation curves color coded by galaxy surface brightness. There is an almost perfect rainbow variation that is plain to see: low surface brightness galaxies (blue tones) have gradually rising rotation curves; high surface brightness galaxies (red tones) have rapidly rising rotation curves – sometimes so fast that the flat velocity is already attained within 1 kpc. (Adapted from McGaugh 2020.)

The systematic variation of rotation curve shape with surface brightness is the signature of MOND. Lower surface brightness means lower stellar surface density means lower acceleration means a larger mass discrepancy. Low surface brightness galaxies are inferred to be dark matter dominated as a consequence. Though initially surprising from the conventional perspective, this is exactly what MOND predicted a priori. Milgrom discussed the expected variation in rotation curve shapes in his original series of papers. That’s what is observed above. I have not seen the same effect emerge from simulations, nor do I expect to.

Stellar population mass-to-light ratios

The mass of the stars matters to the dynamics. We observe starlight, but need to know the stellar mass for the physics. The connection between the two is encapsulated by the mass-to-light ratio of the stellar population.

Conventionally, constructing a model or fitting a rotation curve requires a minimum of three physical parameters: one for the stellar mass with its corresponding mass-to-light ratio, and two for the dark matter halo (typically a mass and length scale). That is more parameters than are required to describe the data, so there is considerable degeneracy between them. It’s like fitting a small number of points with a high order polynomial: a fit is not only guaranteed, but entire families of equally good fits are possible. Consequently, the precise value of the fitted stellar mass-to-light ratio does not provide a test; it can be whatever it needs to be and an appropriate halo fit can be obtained.

To fit a rotation curve in MOND, one has only knowledge of the RAR and a single physical# parameter, the stellar mass-to-light ratio. That provides a test. If there is nothing to MOND beyond being a weird fitting formula, then the mass-to-light ratio is a free-floating parameter that will take whatever value is necessary to provide a fit. If instead it corresponds to physical reality, then the fitted stellar mass-to-light ratios should conform to astrophysical expectations for stellar populations.

The stellar mass-to-light ratios from MOND fits (points) in the blue (left panel) and near-infrared (right panel) pass-bands plotted against B-V color (bluer colors to the left, redder to the right). From the perspective of stellar populations, one expects more scatter and a steeper color dependence in the blue band, as observed. The lines are stellar population models from Bell et al. (2003). These are completely independent, and have not been fit to these data in any way. (Adapted from Famaey & McGaugh 2012).

We know a lot about stars and their evolution. It is possible to build extremely sophisticated stellar population models. These of course have their own uncertainties, but they are good enough to provide a clear expectation for what the stellar mass-to-light ratio should be. As the figure above illustrates, MOND fits return the correct amplitude for M*/L and its variation with color. The expected amount of bandpass-dependent scatter is also apparent: more in the blue, less in the near-IR. One could not hope for better astrophysical agreement.

Tully-Fisher Relation

MOND correctly anticipated a number of striking features about the Tully-Fisher relation. People seem to obsess about the slope; MOND predicts Vf∝Mb1/4V_f \propto M_b^{1/4} while the dark matter halo version of this is V200∝M2001/3V_{200} \propto M_{200}^{1/3}. We observe the former. The latter isn’t a prediction for what we can see, so modeling is required. There exist many models, so there is no clear consensus on what LCDM predicts. It certainly didn’t predict Vf∝Mb1/4V_f \propto M_b^{1/4}, which I guess is why people obsess over this aspect of the observation and complain as the slope has converged to that value with increasing precision as the data have improved.

LCDM models that adhere to abundance matching naturally predict curvature in Tully-Fisher. This is a generic expectation that necessarily arises when mapping the curving shape of the galaxy luminosity function to the power law of the DM halo mass function. A representative example is provided by the model of Sales et al. (2017) (red line in plot). This model does OK for a range in mass, but departs from the data at both high and low mass ends.

The Baryonic Tully-Fisher relation for spiral galaxies (blue), gas rich dwarf irregulars (green), and Local Group galaxies (yellow). The straight black line is MOND. The curving red line is the model of Sales et al. (2017). There are many variations on the LCDM prediction while that of MOND works better and has remained unchanged since its introduction in 1983.

There are other aspects of Tully-Fisher that are problematic for the dark matter picture. More concerning to me than the slope is the small scatter. If one goes through the exercise of tabulating the scatter budget, one finds that it should be large. There are many sources of potential variance, including that from halos themselves. At a given mass, halos should come in a range of concentrations; that by itself exceeds the observed scatter budget, which is entirely explained by the expected amount of scatter in the mass-to-light ratios of stellar populations.

Another prediction of MOND that sits uncomfortably with dark matter is the absence of second parameter effects. Newton says V2=GM/RV^2 = GM/R; both mass and radius matter to the velocity. Yet radius is not observed to be relevant, only the mass, as predicted by MOND: Vf4=a0GMV_f^4 = a_0GM. Explaining this with dark matter requires suppressing the dependence on radius, which in turn requires an unnatural fine-tuning between the luminous and dark matter components. It was this that made me start to doubt dark matter before I was cognizant of MOND. When I learned that MOND – and only MOND – predicted this absence of second parameter effects is when I started to take it seriously.

Radial Acceleration Relation

The naive expectation for rotation curves in LCDM follows from placing an exponential disk in an NFW halo. This is the second thing I tried, before the turn of the century, after first trying then-standard pseudo-isothermal halos. This doesn’t work to reproduce the Radial Acceleration Relation (RAR) (or the Mass Discrepancy-Acceleration Relation as I called it back then).

My inability to explain the RAR wasn’t for lack of trying. If it were natural, it would follow naturally, and I would have said so in the ’90s. It did not because the RAR is not natural in terms of dark matter, only in terms of MOND. This used to be obvious to everyone. For example, at a 2006 conference in New Mexico, I said let’s ignore MOND and carefully explained how the data led to what we now call the radial acceleration relation without reference to MOND. As soon as I finished, Anatoly Klypin SHOUTED: “WE DON’T HAVE TO EXPLAIN MOND!”

But we do have to explain the data, which was the whole point of my talk and the paper it was based on (McGaugh 2004). The RAR is in the data, whether you know about MOND or not. Plausible dark matter models could do any number of things that are not that. It’s not my fault that instead the data look like MOND. So, to explain the data, you do have to explain why dark matter looks like MOND (whether you realize it or not), which, as Anatoly correctly grasped, makes no sense whatsoever.

Can a dark matter model be contrived to look like MOND? That was the charge to the galaxy formation community: I wrote McGaugh (2004) to explicitly lay out what we had to do to explain the data conventionally. Nobody took up the challenge. Most of the citations to that paper are MOND-related, it did not draw much response from the intended audience (unless it was a decrease in invitations to speak at conferences like the one in New Mexico). It was old news to me – I had spent several years expanding and vetting the data from the 1998 edition – but I guess it was still too soon to register with most scientists.

The biggest uncertainty in 2004 was the mass-to-light ratios of the stellar populations. I tried a number of prescriptions for these. The RAR was apparent in each; only the best-fit acceleration scale and the amount of scatter varied. So I left that as an “out” – one didn’t have to adopt the same M*/L that MOND fits found (though you need to be close; see above) and wrote the mathematical requirements for the DM halo in all generality. Crickets.

We spent a decade improving the situation with M*/L by acquiring near-IR data. When Federico Lelli first used these to construct the modern RAR, the scatter caused by uncertainty in the previous optical M*/L collapsed to practically nil. The RAR emerged as a tight empirical relation, a result so remarkable that it is in effect a Law of Nature.

This time (2016) the galaxy formation community took notice and fell all over itself to draw the wrong conclusion. Instead of evincing angry denial like Klypin, numerous papers suddenly asserted that the RAR happens “naturally” in LCDM models. If it were so natural, why had the community denied it so hard for so long? This assertion of naturalness is the greatest abuse of scientific language in modern times: there is nothing remotely natural about it; we had already demonstrated this. Those lessons had already been forgotten (if ever learned), but in the intervening time, people had been busily tweaking their models to make them more like data, so if you didn’t know any of the relevant history, or how badly previous models had failed, the assertion that it is all natural might seem reasonable. I guess. It really is not.

In order to build a galaxy formation model, one has to construct a model for both the visible galaxy and its parent DM halo. One of the huge unknowns in this game is how to assign the galaxy we see to a dark matter halo we do not. The standard approach in the 1990s was to assume a constant relation between the mass of the observed galaxy and that of its host dark matter halo. Mo et al. (1998) did this (“the masses of disks are a few percent of those of their haloes”); I made the same assumption in my modeling efforts. It gradually became clear that this could not work, and subsequently abundance matching became the norm for making this assignment. This collapses the dynamic range in halo mass relative to that in galaxy stellar mass, which makes models look like galaxies all live in nearly the same DM halo, which goes in the right direction.

One claim that the RAR is natural in LCDM comes from Navarro et al. (2017). They add exponential disks to NFW halos, duplicating the exercise I tried twenty years earlier. It appears to work better now because they adopt abundance matching to assign DM halos to galaxies rather than the original assumption of a constant ratio of baryons to dark matter that we all made in the hopes of avoiding this becoming a rolling free parameter. The reduced dynamic range in halo mass that follows from abundance matching has the benefit of collapsing models to approximate the RAR, at least over a finite dynamic range. This is illustrated by this plot I made four years ago but I don’t think I’ve published previously:

The radial acceleration relation (top panels) traced by kinematics (red & blue points) and gravitational lensing (orange, cyan, & open points) for early (triangles) and late type (circles) galaxies. Also shown are LCDM models (left) and MOND (right) and their residuals (bottom panels). See text for details.

The data here are kinematic data from our own work plus weak gravitational lensing from Brouwer et al. (2021). Both have been improved since I made this, in particular with the reanalysis of the lensing data by Mistele et al. (2024). Rather than remake the plot, it is informative of how the data improve over time. The lensing data are more reliable where the target galaxies are most likely to be isolated (filled points); less reliable data beyond this region are shown as open points.

Two sets of LCDM models are shown in the left panels. Both have four lines that correspond to the four mass bins used in the weak lensing analysis. The pink lines are the simplest possible: exponential disks in NFW halos. This is identical to the model of Navarro et al. (2017) except that I’ve used Kratsov’s abundance matching relation because this works better for this purpose. The green lines from Li et al. (2022) are similar, but additionally include the effects of adiabatic compression. I omit more complicated models from simulations because the plot becomes incomprehensibly busy. They don’t improve matters.

Over the range 10−11≲gobs≲5×10−9ms−210^{-11} \lesssim \mathrm{g}_{\mathrm{obs}} \lesssim 5 \times 10^{-9}\;\mathrm{m}\,\mathrm{s}^{-2}, the LCDM models appear to match the observed RAR. This much supports the claim that LCDM explains the RAR. Is it satisfactory? Is it sufficient?

If one zooms in, as seen in the residuals in the lower panel, the lines for the individual models segregate. This is not obvious in the RAR itself, where the logarithmic scale necessary to display the large dynamic range of the data compresses and conceals such unpleasant details. There is no such mass segregation in the data: all galaxies fall on the same RAR, within the uncertainties.

Another detail that becomes apparent in the residual diagram (which is mysteriously absent in all the papers I’ve seen claiming this works in LCDM) is that the models do not have the right shape. From high to low accelerations, the pink lines dip below the data, then above, then below again. At low accelerations, they dive very much below: the models cease to even come close to working at gobs≲10−11ms−2\mathrm{g}_{\mathrm{obs}} \lesssim 10^{-11}\;\mathrm{m}\,\mathrm{s}^{-2}. Their utility is limited to a narrow dynamic range: they are neither satisfactory nor sufficient.

The models considered here do not probe the high acceleration end of the RAR. This is an artifact of the simplicity of the exponential disk plus NFW halo model. Such models never reach high accelerations because those are the province of bulge components. Without those, the models end in little hooks: they are turning away from the RAR at the highest acceleration plotted. There’s no reason not to plot to higher acceleration, but it would make the models look worse – the hooks would grow in the wrong direction. I’ve made the conscious choice – as I’ve noted others have – not to show the ugly parts of this model. The difference is that I don’t pretend it works by virtue of not looking where it doesn’t work.

The models from Li et al. (2022) that include adiabatic compression are represented by the green lines. These do better than the simple exponential disk plus NFW halo models insofar as the are less bendy. But they still segregate by mass, and dive to lower accelerations even faster.

I’ve tried various simulations in this plot but I have yet to find one that does well. The turn down at low accelerations occurs at large radii where feedback effects are usually said to be minimal, so that is not a likely solution. Indeed, the problem is generic. Low accelerations probe the outer regions of CDM halos where the robust prediction from simulations is a density profile that scales as ρCDM∝r−3\rho_{\mathrm{CDM}} \propto r^{-3}, but the data scale as ρDM∝r−2\rho_{\mathrm{DM}} \propto r^{-2} as required to produce a flat rotation curve. Consequently, the mismatch between CDM and the RAR at low acceleration is serious.

In contrast, MOND does fine. It explains the data at all observed acceleration scales, not just a narrow window. The agreement now is better than is depicted with these older data, but this is already good. The only objection one might raise is that the lowest accelerations obtained from gravitational lensing for early type galaxies are offset from the prediction of MOND. (The open triangles sit above the zero residual line in the lower right panel.) This is largely rectified in subsequent analysis, as an experienced astronomer might expect, but this didn’t prevent people from leaping at this mild discrepancy as evidence against MOND. It would be, if it were true. It would also be bad news for LCDM, as these same data are even more discrepant from the models at left. This is a good example of the bias that infects the field: any discrepancy from MOND is seen as a win for dark matter – even when dark matter suffers the same problem (or worse).

Central Surface Density Relation

In MOND, the observed mass determines the kinematics through the modified Poisson equation. Just as lower accelerations mean larger mass discrepancies, so too do lower surface densities mean lower accelerations. Consequently, one expects a relation between the characteristic surface density of a galaxy and the amplitude of the mass discrepancy it evinces. In ignorance of that, Erwin de Blok and I discovered that low surface brightness galaxies were dark matter dominated. It wasn’t just “gee, there’s a lot of dark matter here;” the amplitude of the discrepancy correlates with central surface brightness.

This can be quantified by the central surface density relation:

Fig. 2 from Lelli et al (2016). The central dynamical mass surface density as a function of the central stellar surface density (left) and stellar mass (right). Points are color coded by morphological type. The dashed line shows the 1:1 relation expected in the absence of dark matter; a fit to the data (dotted line) is consistent with the prediction of MOND.

This relation involves the central regions of galaxies. High surface brightness galaxies are far enough in the Newtonian regime (GΣ⋆≳a0G \Sigma_{\star} \gtrsim a_0) that they follow the one-to-one line: what you see is what you get. Lower surface brightness galaxies diverge from the one-to-one line, having systematically lower accelerations (GΣ⋆<a0G \Sigma_{\star} < a_0) and larger mass discrepancies (i.e., they appear to need more dark matter) as surface brightness decreases. There is no reason for this behavior in dark matter, for which the central surface density of stars is not a parameter of obvious relevance.

The central surface density relation is both natural and expected in MOND. I am not aware of a dark matter model that addresses this relation. The correlation with mass (right panel above) is often discussed, and is considered to be deeply meaningful by some. However, the correlation in the right panel occurs entirely because stellar mass correlates with surface brightness, albeit imperfectly, so the right panel is a blurry version of the left panel. The surface density of stars is the important parameter here, not their total mass.

Renzo’s rule

There are other ways to see the connection between kinematics and the distribution of observed baryons. Indeed, a connection was recognized empirically by one of the great 21 cm observers, Renzo Sancisi. To quote from his abstract from IAU 220:

In the inner parts of spiral galaxies, of high or low surface brightness, there is a close correlation between rotation curve shape and light distribution. For any feature in the luminosity profile there is a corresponding feature in the rotation curve and vice versa. This implies that the gravitational potential is strongly correlated with the distribution of luminosity: either the luminous mass dominates or there is a close coupling between luminous and dark matter.

Sancisi (2003)

The shape of the rotation curve correlates with the shape predicted by Newton for the baryons:

Two galaxies exemplifying Renzo’s rule. Images of both galaxies are shown together with their rotation curves (open points), Newtonian rotation curves for the observed stars and gas (gray lines), and the corresponding MOND rotation curves (blue lines). In the high surface brightness spiral NGC 6946, the compact central bulge component predicts a sharp inner bump in the rotation curve that declines before it rises again. In the gas-rich dwarf NGC 1560, there is a broad feature in the gas that is reflected in the rotation curve. (Adapted from Fig. 21 of Famaey & McGaugh 2012; images added.)

This phenomenology is common enough for Prof. Sancisi to notice through the experience of looking at lots of data. Indeed, he said it to me in person in 1995 when we had a heated debate about the interpretation of the Tully-Fisher relation. The correspondence he noted implies that baryons are important to the mass budget, in both high and low surface brightness galaxies. In contrast, the absence of surface brightness-dependent residuals in the Tully-Fisher relation implied that LSB galaxies were certainly sub-maximal, and HSB galaxies as well if one desires to avoid fine-tuning.

Our discussion was heated in the best way. We weren’t angry, just engaged. We understood both things could not be true, but apparently they were. I redoubled my efforts to look for residuals from Tully-Fisher, ultimately finding none after many years of checking. Renzo held off on publishing his statement (for eight years!) while he looked at more LSB galaxies. Both results were sustained at the level expected for astronomical data despite the contradiction they encapsulate.

The correspondence of features (sometimes referred to as “bumps and wiggles”) in galaxy luminosity profiles and rotation curves makes sense in HSB galaxies if they are maximal in the sense discussed above. If instead dark matter dominates, there is no reason for the rotation curve to reflect the features in the baryons. Indeed, it cannot: the observed bumps and wiggles are typically features like spiral arms that require disk self-gravity to support them. The dynamics of quasi-spherical dark matter halos simply can’t support features like that – the phase space of particle orbits is too different. Even if they could, which they can’t, there would be no reason for the correspondence: the DM halo might have its own features, which if its mass dominates, could be anywhere, not where we see features in the stars.

This issue becomes more pronounced in LSB galaxies. Since the mass discrepancy is large at all radii in LSB galaxies, the mass of the DM halo always dominates. There should never be a correspondence of features, and yet there are.

The only theory that satisfactorily explains both the correspondence of features in Renzo’s rule and the absence of surface brightness-dependent residuals from Tully-Fisher is MOND.

The acceleration scale

Threading throughout the data is the acceleration scale a0=1.2×10−10ms−2a_0 = 1.2 \times 10^{-10}\;\mathrm{m}\,\mathrm{s}^{-2} suggested by Milgrom. This shows up in the data in many ways. Three of these are summarized in this plot from the review by Lelli (2022):

Three indicators of the acceleration scale in galaxy data: the BTFR (left), the central sruface density relation (center), and the RAR (right). All three indicate the same acceleration scale within the uncertainties, and are well described by MOND (red lines). Adapted from Fig. 5 of (Lelli 2022).

The BTFR shows that all galaxies share the acceleration scale aBTFR=Vf4/(GM)a_{\mathrm{BTFR}} = {V_f^4}/{(GM)}. The acceleration scale aRARa_{\mathrm{RAR}} is where the RAR bends away from Newtonian behavior. Similarly, the surface density of the bend in the central surface density relation, ΣCSDR\Sigma_{\mathrm{CSDR}}, marks the acceleration aCSDR=2πGΣCSDRa_{\mathrm{CSDR}} = 2 \pi G \Sigma_{\mathrm{CSDR}}. These relations need not exist at all in a universe dominated by dark matter. Even if they do for some reason, there is no reason to expect they all share the same acceleration scale. Yet within the uncertainties, they do: aBTFR=aRAR=aCSDR=a0a_{\mathrm{BTFR}} = a_{\mathrm{RAR}} = a_{\mathrm{CSDR}} = a_0.

Cold dark matter is scale free, so an acceleration scale is not natural to it. That doesn’t prevent people from asserting it is. Kaplinghat & Turner (2002) do just that. The absurdity$ of this claim was immediately pointed out by Milgrom (2002), but this has not precluded either Kaplinghat or Turner from continuing to assert that they’ve explained that which they have not remotely explained. It’s more like a political talking point than science.

Exceptions

The table above lists some of the rules that the vast majority of galaxies obey. I emphasize vast majority because relations like Tully-Fisher would be unrecognizable if there were lots of exceptions. It wouldn’t be a relation at all. We need to understand the rules the majority obey – which we do not in the context of dark matter – before obsessing about potential exceptions. To do the latter is a classic case of missing the forest for the trees.

There are always exceptions in astronomy. These come in several flavors: (i) goofs, (ii) apparent exceptions that actually prove the rule, and (iii) genuine exceptions that falsify a theory (anomalies). The first category is extraordinarily common in astronomy, the second can take a long time to recognize, and the third are rare if they exist at all. New examples in each category are always being discovered, so it is a Sisyphean task to check every case.

The most common flavor of exception in astronomy is the straight-up goof (i). These are so common, the experienced astronomer is rightly skeptical of a story that lacks exceptions. In the case of MOND, there have been many galaxies that appear to be problematic but later turn out not to be. They’re just goofs. There have been so many examples of this, I have lost count. Over a hundred, easily.

The ultra-faint dwarf galaxy Boötes III provides a recent example of an ordinary goof. Ultrafaint dwarfs often have velocity dispersions well in excess of that predicted by MOND, but it is challenging to measure dispersions as low as predicted. In this case, Ting et al. (2026) “measure a velocity dispersion of σv = 1.69 +1.03/−0.85 km s−1, about six times smaller than the previously reported 10.7±3.5 km s−1.” The larger number has been known for a while and is strongly inconsistent with MOND; the new measurement is pretty much bang on what MOND predicts. Consequently, one can see how it is easy to leap to the wrong conclusion, especially with confirmation bias standing behind you ready to push as you look out over the cliff.

Apparent exceptions (ii) are outliers that appear to violate a rule, and do so with high significance. In these cases, it is the rule that is not always universal. Relations like the BTFR and RAR are predicted by MOND and hold for the vast majority of galaxies. There are, however, some genuine outliers.

One can imagine building a DM model that reproduces these things. (I’m not even requiring it to be satisfactory for the moment.) Then, when one encounters an object that falls off these relations, you might conclude MOND is wrong and the DM model is better because DM tolerates outliers better (often by virtue of having no good reason for the relation in the first place). Some objects happen not to follow the statistical happenstance of the DM model, while you can’t turn off the force-law in MOND, so the rule must always apply. Right?

The problem with this reasoning is that MOND does predict exceptions. Objects for which the external field is important should deviate from the main relation that holds for isolated galaxies. I have seen many cases where what looks like a failure of MOND is actually a corroboration once you consider the external field. Such cases are a problem for our hypothetical DM model: once we construct something up that explains a relation like the BTFR, outliers can happen, but should not depend on the acceleration of from an external field. The external acceleration from some external source simply has no role to play in dark matter models.

Finally, there are genuine anomalies (iii) that would indeed falsify MOND. Most examples that I have seen that seemed to be in this category have later turned out to be in one of the first two categories: simple goofs or predicted exceptions. Some cases are still under litigation, and new examples will be discovered in the future. Do I know that all these will all work out? No, of course not. Have I seen this movie enough times to have a well-founded empirical expectation that they probably will? Yes, I have.

Dark Matter as extra, invisible mass is practically falsified

Look at the table above again. Time after time after time, new observations have been made that were not predicted by dark matter but were by MOND. Dark matter simply does not succeed in predicting galaxy dynamics the way MOND does. That’s very telling.

Many scientists choose to ignore MOND’s predictive success, or have sought to downplay its importance. When they bother to do the latter, they usually restrict themselves to a single item, often in apparent ignorance of the rest of the table. For example, a common talking point is “MOND only fits rotation curves!” Never mind that “only” here is a falsehood. The “mere” ability to fit rotation curves is already a highly non-trivial result. Each and every galaxy has its own unique mass distribution and poses a separate test of MOND, just as every planet and every comet and every asteroid test Newton’s law of gravity in the solar system. We’ve made hundreds of these tests: the effective force law in galaxies is MOND.

More generally, it isn’t just that the relations listed in the table above were correctly predicted by MOND and not by dark matter. It is also that the post hoc dark matter explanations are not satisfactory. Sometimes they are mutually contradictory.

The only way to save dark matter is to contrive for it to look like MOND. It is hard to think of a more damning statement about an otherwise nebulous hypothesis. Dark matter could be anything, do anything, and yet the one thing it has to do is mimic the very precise and comparatively inflexible predictions of its chief rival.

The needle in the dark matter haystack. Galaxies might lie anywhere in this voluminous, multiparameter space, but in practice they inevitably seem to reside in the negligibly small part of the volume that “looks like” MOND.

Can a dark matter model be contrived to look like MOND? Many have claimed to do this, even asserted that it happens “naturally.” This latter assertion is the greatest abuse of scientific language in modern times: there is nothing remotely natural about it. If it were natural, it would follow naturally, and I would have found that and said so in the ’90s. A corollary is that the people who claim now to explain it naturally would not have spent the intervening decades loudly denying that it could be so before they suddenly knew it all along. Lord knows I’ve tried to build acceptable DM models, and there are always more (increasingly silly) things to try, so when I say that I do not think it can be done, at least not with conventional flavors of DM, it is a highly nontrivial statement^.

When the dark matter paradigm took off, we had lots of indications of mass discrepancies. Flat rotation curves required more mass than they eye could see; the gravitating cosmic mass density exceeded that attributable to baryons. The vast majority of dark matter candidates currently being contemplated are informed by that perspective. As such, they simply come up with a new particle that could provide the extra mass.

That approach was sufficient in the 1980s, but since then new shit has come to light.

Many items in the table above constitute new stuff that has come to light since the DM paradigm was established. It isn’t just that rotation curves are flat and therefore we need more mass. It is also that the amplitude of the flat rotation speed correlates with the baryonic mass (the BTFR). And that the kinematic acceleration correlates with that predicted by Newton for the observed distribution of baryons (the RAR). And the rest of the table. New shit has come to light! These things did not inform the development of the dark matter paradigm, and are not explained merely by invoking invisible mass. We need to reexamine the paradigm in that new light.

The BTFR and the RAR and the other observations detailed above all point to a strong connection between baryons and dark matter, completely the opposite of the original mandate. Extra, non-interacting, invisible mass might still suffice for cosmology, but explaining galaxy dynamics is more demanding. Just adding invisible mass does not explain the strong coupling with the observed distribution of ordinary baryons. A viable dark matter candidate needs to do this, but the vast majority of them have been constructed in ignorance (or denial) of the need to explain the MONDian phenomenology.

I can imagine a non-generic DM candidate might be engineered to satisfy all needs. There are candidates that attempt to do this, like dipolar dark matter or superfluid dark matter. This post is long enough, so I won’t say more about those here. What I am convinced will not work is relying on feedback to somehow change the essence of CDM to mimic MOND. That’s a pipe dream akin to profiting from stealing underwear.

? = feedback? (Underwear gnomes from South Park.)

The observed kinematics of galaxies as described above suffice by themselves to provide a practical falsification of dark matter as widely conceived – either CMD specifically or generic extra-mass DM, excepting only highly specific ideas that build-in MOND-like behavior. (I say a practical falsification because outright falsification is impossible with unseen entities.) There are other, independent reasons to doubt DM, like the apparent absence of the expected dynamical friction due to dark matter. There are of course reasons to persist with DM, but it needs to be more than some random particle that just adds mass to the universe.

It is not possible to make progress in the context of dark matter without addressing the observed phenomenology described above as a whole rather than the common, piecemeal, divide-and-deny approach. We will not consider alternative theories until we agree that dark matter, as currently conceived, is wrong. We are stuck in a rut, and have been for a long time.


*It also shows the rotation curve that would be measured from the gas, which differs from the actual rotation curve in many simulations. This seems to have led many simulators to disbelieve observed rotation curves, but this is a simulation problem not an observational problem. We know how to recognize and correct for non-circular motions in the data. These corrections are usually small and the data never look like these simulations. It’d be really obvious in the velocity fields of real galaxies were messed up in this way. (Real galaxies are often messed up, just not in the ways seen in LCDM simulations.)

#There can also be nuisance parameters like distance and inclination. These are measured independently and ideally would be held fixed, but of course there are uncertainties on these values so one can marginalize over the range of allowed possibilities. The same is true in DM halo fits.

$Milgrom’s rebuttal is quite thorough, so I don’t want to pile on. But the acceleration scale appears in diagrams like the RAR that plot an observed kinematic acceleration (gobs\mathrm{g}_{\mathrm{obs}}) against one obtained from the observed baryons (gbar\mathrm{g}_{\mathrm{bar}}). In terms of dark matter, gobs=gbar+gDM\mathrm{g}_{\mathrm{obs}} = \mathrm{g}_{\mathrm{bar}} + \mathrm{g}_{\mathrm{DM}}. K&T talk about the dark matter; but have no model for the baryons, which are essential to both axes of the RAR. They make a hand waving argument that a0a_0 corresponds to the edge of galactic disks, which it does not – that’s what low surface brightness galaxies had already taught us by then. So they don’t explain the acceleration scale at all.

^I’m sure galaxy formation theorists can be found who will sanguinely assure us that all that needs to be done can be done, it’s just a matter of tweaking this or that. So why haven’t they succeeded in thirty years of effort? So far, they’ve mostly just come up with more elaborate versions of models I already tried and rejected as obviously unworkable. Making them more elaborate doesn’t make them less wrong.

Extended Tully-Fisher relations

Extended Tully-Fisher relations

Previously I had alluded to some of the major projects I’ve been working on. One has come to fruition and can be found on the arXiv and in the Astrophysical Journal&. It has taken many years to assemble the data in this paper, during which time the models purporting to explain some of it have evolved considerably while consistently failing to address the real problems they raise. There is a lot to explore, so it will take more than one post.

Here I start with the empirical basis: the stellar mass and baryonic Tully-Fisher relations. The Tully-Fisher relation was originally discovered as a relation between luminosity and linewidth in rotationally supported galaxies – spirals and irregulars. It immediately proved useful as an extragalactic distance indicator. As such, it was instrumental in breaking the impasse in the Hubble constant* debate (back when it was 50 vs. 100, not 67 vs. 73), and it remains useful in this role.

Physically, the obvious interpretation was that luminosity is a proxy for stellar mass and linewidth*^ is a proxy for rotation speed. This is correct. Of the various rotation speeds one can define and measure, the one that works best, in terms of minimizing the scatter in the relation, is the flat rotation speed measured in the outer parts of extended rotation curves. See Stark et al. (2009) and Trachternach et al. (2009) for further examples. The scatter is basically a function of data quality.

On the mass axis, converting measured flux to luminosity to mass is a bit dicier, as we need to know the distance for the first step and the stellar mass-to-light ratio for the second. There is inevitably some intrinsic scatter in the mass-to-light ratio of a stellar population. While I don’t doubt that luminosity is a proxy for stellar mass, improving on it is hard to do: there are many instances in which simply assuming a straight mapping of light to mass can be as effective as applying fancier population models. We might^ finally be getting past that, so it is worth discussing a bit.

The procedure to convert starlight into stellar mass involves the construction of stellar population models that use the color(s) or spectral energy distribution of a galaxy to infer the types of stars that make the light. This is a long-argued subject; suffice it to say there are a number of points where it can go wrong. The most obvious is the IMF; the initial spectrum of masses with which stars are born. Most of the light we see from galaxies is produced by its higher mass stars, which are disproportionately bright (there is a steep scaling of stellar luminosity with mass). But most of the mass is locked up in low mass stars that contribute little to the total luminosity. So we are, in effect, using the light of the few to represent the mass of the many. That would go badly wrong if we don’t know the relative mix, i.e., the shape of the IMF. This has been the subject of much research, and over many decades has been narrowed down pretty well. While I hope that this is almost settled, the specter of the IMF lurks as a menace to all stellar mass determinations.

There is a lot else we need to know to build a stellar population model. This includes such essentials as the spectra of individual stars of each and every type and stellar evolution as a function of mass and composition including exotic phases like the asymptotic giant branch. There are a lot of places where this can go badly wrong, and sometimes^% does. So I wouldn’t say we know how to do this perfectly, but we have become pretty good at it.

Converting light to mass suffices to plot the stellar mass Tully-Fisher relation. That accounts for most of the baryonic mass of high mass spirals, but it ignores the mass of the interstellar gas. This can be appreciable in lower mass systems. Indeed, the standard issue dwarf galaxy in the field is more gas than stars:

Figure 1 from McGaugh et al. (2019): The gas and stellar masses of rotating galaxies. Blue points are galaxies in the SPARC database (Lelli et al. 2016b) and the gas rich galaxies discussed by McGaugh (2012). The location of the Milky Way is noted in red (McGaugh 2016): it is a typical bright spiral. Grey points are the sample of Bradford et al. (2015). The line is the line of equality where M* = Mg.

With measurements of mass and rotation speed, we can construct the Tully-Fisher relation:

Figure 4 from McGaugh et al. (2019): The stellar mass (left) and baryonic Tully-Fisher relation (right). Data from Lelli et al. (2016b) and McGaugh (2012) are shown as blue points if both axes are measured with at least 20% accuracy; less accurate data are shown in grey. The latter include cases for which the rotation curve does not extend far enough to measure Vf, in which case the last measure point is used. These cases are systematically offset to lower velocity. Inclination uncertainties and distance errors also contribute to the scatter. The better the data, the tighter the relation. The location of the Milky Way is noted in red (you are here).

The stellar mass Tully-Fisher relation is a good correlation by the standards of extragalactic astronomy. The majority of studies in the literature are restricted to massive% galaxies, mostly those with M* > 1010 M☉ where stars dominate the baryonic mass budget so the omission of gas is not obvious. As we look to lower masses, the relation bends and the scatter increases. That this happens right where gas starts to become important to the mass budget suggests that we’re missing an important component, and voila – a nice, continuous relation that is linear in log space is restored when we plot the baryonic mass Mb = M*+Mg. Indeed, the data are consistent with a simple power law

Mb=AVf4M_b = A \, V_f^4

with A = 50 M☉ km-4 s4. The intercept A has consistently been measured within 10% of this value over the past couple of decades. That this is an integer power law so that the intercept has real physical units is intriguing. That doesn’t happen in most astronomical scaling laws, which are usually more happenstance, like the mass-luminosity relation for main sequence stars.

Why limit ourselves to rotationally supported galaxies? Let’s plots every known type of gravitationally bound extragalactic object, from the smallest ultrafaint dwarfs to the largest clusters of galaxies. Note that I’ve flipped the axes to accommodate the huge dynamic range in baryonic mass, roughly twelve (12) orders of magnitude. This is like having gnats at one end of the scale and blue whales at the other. On that scale, a person is a regular galaxy like the Milky Way.

Figure 3 from McGaugh et al. (2026): Extended Tully-Fisher relations plotting the flat-equivalent circular velocity of extragalactic systems as a function of stellar mass (top panel) and baryonic mass (bottom panel). Data for rotationally supported galaxies are depicted by circles; squares represent pressure supported systems. The blue circles are galaxies with directly measured distances, Vf from rotation curves, and stellar masses from WISE photometry from Duey et al. (2026, in preparation). Green circles are gas-rich galaxies (Mg > M*; Stark et al. 2009; Trachternach et al. 2009; Bernstein-Cooper et al. 2014; McNichols et al. 2016; Iorio et al. 2017; Namumba et al. 2025; Xu et al. 2025) not already in Duey et al. (2026). Yellow points are Local Group galaxies, both spirals and dwarfs (McGaugh et al. 2021); gray squares are ultrafaint dwarfs (Lelli et al. 2017). Lensing results for early- and late-type galaxies (Mistele et al. 2024a) are shown as pink squares and magenta circles, respectively. Red squares are clusters of galaxies (Mistele et al. 2025), and purple squares are groups of galaxies (McGaugh et al. 2026). The orange line is the BTFR fit only to rotating galaxies over a more limited range (about three orders of magnitude in baryonic mass, from Mb ~ 4 x 108 to 4 x 1011 M☉) by McGaugh (2005).

One improvement from twenty years ago, aside from the greater number of objects and the increase in dynamic range, is the accuracy of the mass measurements. I tried a number of prescriptions for the stellar mass-to-light ratio in McGaugh (2005), which resulted in a range of possible slopes. Now we just use the stellar mass from precise population models (Duey et al. 2025) and recover my best estimate from back then. The room to dodge the obvious conclusion about the slope of the relation by complaining about the choice of stellar mass estimator – a popular course of action back then – is gone. Another technical issue we’ve spent a lot of effort working on is how to put all these very different systems on the same scale of Vf. I won’t elaborate on this here: if you’re interested in that level of detail, you can go read the paper and references there in. If we got this wrong, it would add to the scatter in the relation, and/or create offsets between different types of data.

Both of the extended Tully-Fisher relations, that in stellar mass (top panel) and that in baryonic mass (bottom panel, the extended BTFR) are good correlations. That in baryonic mass is clearly better in the sense that it is tighter over a larger dynamic range. From small dwarf galaxies (Mb ~ 5 x 105) to groups of galaxies (5 x 1012 M☉), the data are consistent with a single power law (Mb ~ Vf4) for all systems with remarkably little scatter. Outside this range, the data for both the lowest and the highest mass systems deviate from a straight line towards higher mass at a given flat velocity. I don’t put much credence in the smallest systems as I think there is little chance that their measured velocity dispersions are representative of their equilibrium gravitational potential. For all practical purposes, our knowledge runs out as we hit the regime of ultrafaint# dwarfs. The deviations of the most massive systems, clusters of galaxies, are more difficult to dismiss.

Restricting our attention for the moment to the range where a single power law suffices to describe the data, we note that there is not much scatter in the BTFR. Some of it is from random uncertainties; these dominate most studies and lead to a lot more scatter than seen here: these data are very good. We can account for the known observational errors and subtract off their contribution to estimate the intrinsic scatter in the relation. This is the variance of the data from a perfect line. The intrinsic scatter for the best data (the WISE-SPARC sample of Duey et al. 2026) is about 0.11 dex in mass – about what we expect$ for stellar populations. That doesn’t leave much room for other sources of scatter, so the underlying physical relation has to be very tight indeed: essentially perfect over the range 5 x 105 < Mb < 5 x 1012 M☉.

Scatter will also occur if our mass budget is incomplete. We can see this in the transition from the stars-only relation to the BTFR. There is a lot of scatter in the stellar mass Tully-Fisher relation around 107 < Mb < 109 M☉. Galaxies in this mass range are sometimes star-dominated and sometimes gas-dominated. The gas fraction is all over the place. This shows up as scatter in the stellar mass Tully-Fisher relation. That’s not real; it is a sign that we’ve missed an important mass reservoir. This is cured when we add in the gas mass, which is dominated by atomic gas (HI to spectroscopists and astronomers). That this addition removes the scatter and restores a single power law relation strongly suggests that there are no further substantial reservoirs** of baryonic material that we’re missing.

This logic applies to other systems as well. Bright spirals do not need much correction because their baryonic mass is dominated by stars. Their stellar mass Tully-Fisher relation is pretty much already their BTFR.

Perhaps this applies to clusters of galaxies as well? There was a huge correction from stars-only to stars plus gas. The gas in this case is the hot, ionized plasma of the intracluster medium (ICM) that belongs to the cluster itself and not any individual galaxy within it. That goes most of the way to close the gap between the stars-only cluster data and the extrapolation of the BTFR fit to individual galaxies, but not all the way. So perhaps we are still missing an important baryonic mass component? It happened before – we didn’t know about the ICM for decades after Zwicky first identified the missing mass problem in clusters – so perhaps there are still more baryons to discover there.

It could also be that the apparent offset occurs because we’ve failed to put clusters on the same Vf scale as galaxies. This is not easy to do, and we’ve spent a lot of time worrying about it. I don’t think this is what’s going on, though it would make my life a lot simpler if it were. Different indicators – dynamics vs. ICM hydrostatics vs. gravitational lensing – can give somewhat different answers, but not in a way that “fixes” the problem: I see no viable path in which the offset turns out to be a simple difference in the way the depth of the gravitational potential is measured. I would love to be wrong here, but I’m not dismissing the offset for clusters as I am for ultrafaint dwarfs (which don’t do lightly).

Perhaps the extrapolation of the BTFR from individual galaxies to clusters is simply not appropriate. They’re very different kinds of systems, after all. To dig into that, we need some theoretical perspective – why does the observed power law happen? Should we expect different systems to share the same BTFR?

Theory is something I’ve studiously avoided in this post: the possibility that there are baryons that remain to be discovered in clusters can be inferred empirically. All the other data line up, so why not clusters? But unless and until these hypothetical additional baryons are discovered, that’s just one possibility. How likely this possibility seems to be diverges rapidly once we overlay a theoretical preference, which I will leave to future posts. (I did warn it would take more than one.)


&This paper appears in ApJ volume 1001. The literature has grown quite a bit since I started contributing to it in volume 342. The Astrophysical Journal was founded in 1895. So I’ve been contributing to it for a little over a quarter of its temporal existence, but nearly twice the number of volumes have been published in that shorter time. It’s no wonder none of us can keep up.

*Indeed, Tully & Fisher’s “preliminary estimate of the Hubble constant is H0 = 80 km/s/Mpc” remains correct to this day, within the uncertainties (hard to estimate at the time, but roughly ±10 km/s/Mpc).

*^There appears to be an irreducible intrinsic scatter in the linewidth: it is not a perfect proxy for rotation speed. Linewidths are observationally easier to obtain than resolved, extended rotation curves, so the numbers of galaxies in samples using linewidths can be very large without ever approaching the quality provided by resolved interferometric observations. Bigger samples are not necessarily better.

^I emphasize might here because the community seems to have moved towards reporting stellar masses as if we observe these rather than the luminosities and colors/SEDs that the mass estimates are based upon. The latter are data – observed quantities – while stellar masses are a derived quantity that is inevitably model dependent. This doesn’t stop being true just because we decide to invest a lot of faith in our models.

*^The Sloan Digital Sky Survey provides stellar masses based on models that are known to be wrong in the near infrared. Since SDSS itself is entirely optical, one might not notice. If one mixes SDSS data with near-IR data, one will get the wrong answer.

%This is a classic selection effect. Brighter objects can be seen at a much greater distance than dim ones, so probe a much larger volume. Consequently, their raw numbers always dominate surveys even if their number density is low. Stars are a great example: most of the stars you can see at night are intrinsically luminous: bright stars that are rather far away. Mundane, low mass stars do not stand out even when nearby.

#This isn’t for lack of observations of ultrafaint dwarfs, it’s the underlying assumptions.

$No amount of information suffices to perfectly specify the stellar mass that produces an observed luminosity and SED (spectral energy distribution/set of colors), so one always expects at least some intrinsic scatter in the stellar mass-to-light ratio. I’ve seen estimates that range from 0.1 – 0.2 dex for near-IR colors. That’s as good as it can get as there is always some transient population (e.g., AGB stars) that produce an amount of light that depends on the star formation rate some time ago, not what we measure now. Optical colors are worse in the sense of having more intrinsic scatter, as they are more susceptible to the comings and goings of bright but short-lived stars whose numbers fluctuate with the stochastic star formation rate. Finding 0.11 dex intrinstic scatter is pretty much as good as it can get. (By dex we mean the scatter in log space.)

**We noted this effect in the original BTFR paper to argue that it was unlikely that we were missing substantial amounts of molecular gas (H2), which was a concern at the time. Flash forward, and we were right: the molecular gas mass is almost always a distant third behind stars and atomic gas in the baryonic mass budgets of individual galaxies. Nowadays, the concern is about the mass of baryons in the circumgalactic medium (CGM). That’s getting ahead of the story, which I’ll save for a future post. For now, it suffices to note that any baryonic mass in the CGM is far beyond the radius where the flat velocity is measured, so is not relevant to the sums here.

The odd primordial halo of the Milky Way

The odd primordial halo of the Milky Way

The mass distribution of dark matter halos that we infer from observations tells us where the dark matter needs to be now. This differs form the mass distribution it had to start, as it gets altered by the process of galaxy formation. It is the primordial distribution that dark matter-only simulations predict most robustly. We* reverse-engineer the collapse of the baryons that make up the visible Galaxy to infer the primordial distribution, which turns out to be… odd.

The Gaia rotation curve and the mass of the Milky Way

As we discussed a couple of years ago, Gaia DR3 data indicate a declining rotation curve for the Milky Way. This decline becomes more steep, nearly Keplerian, in the outskirts of the Milky Way (17 < R < 30 kpc). This is may or may not be consistent with data further out, which gets hard to interpret as the LMC (at 50 kpc) perturbs orbits and the observed motions may not correspond to orbits in dynamical equilibrium. So how much do the data inform us about the gravitational potential?

Milky Way rotation curve (various data) including Gaia DR3 (multiple analyses). Also shown is the RAR model (blue line) that was fit to the terminal velocities from 3 < R < 8.2 kpc (gray points) and predates other data illustrated here.

I am skeptical of the Keplerian portion of this result (as discussed at length at the time) because other galaxies don’t do that. However, I am a big fan of listening to the data, and the people actually doing the work. Taken at face value, the Gaia data show a Keplerian decline with a total mass around 2 x 1011 M☉. If correct, this falsifies MOND.

How does dark matter fare? There is an implicit assumption made by many in the community that any failing of MOND is an automatic win for dark matter. However, it has been my experience that observations that are problematic for MOND are also problematic for dark matter. So let’s check.

Short answer: this is really weird in terms of dark matter. How weird? For starters, most recent non-Gaia dynamical analyses suggest a total mass closer to 1012 M☉, a factor of five higher than the Gaia value. I’m old enough to remember when the accepted mass was 2 x 1012 M☉, an order of magnitude higher. Yet even this larger mass is smaller than suggested by abundance matching recipes, which give more like 4 x 1012 M☉. So somewhere in the range 2 – 40 x 1011 M☉.

The Milky Mass has been adjusted so often, have we finally hit it?

The guy was all over the road. I had to swerve a number of times before I hit him.

Boston Driver’s Handbook (1982 edition)&

If it sounds like we’re all over the map, that’s because we are. It is very hard to constrain the total mass of a dark matter halo. We can’t see it, nor tell where it ends. We infer, indirectly, that the edge is way out beyond the tracers we can see. Heck, even speaking of an “edge” is ill-defined. Theoretically, we expect it to taper off with the density of dark matter falling as ρ ~ r-3, so there is no definitive edge. Somewhat arbitrarily,** we adopt the radius that encloses a density 200 times the average density of the universe as the “virial” radius. This is all completely notional, and it gets worse, as the process of forming a galaxy changes the initial mass distribution. What we observe today is the changed form, not the primordial initial condition for which the notional mass is defined.

Adiabatic compression during galaxy formation

To form a visible galaxy, baryons must dissipate and sink to the center of their parent dark matter halo. This process changes the mass distribution and alters the halo from its primordial state. In effect, the gravity of the sinking baryons drags some dark matter along# with them.

The change to the dark matter halo is often called adiabatic compression. The actual process need not be adiabatic, but that’s how we approximate it. We’ve tested this approximation with detailed numerical simulations, and it works pretty well, at least if you do it right (there are boring debates about technique). What happens makes sense intuitively: the response of the primordial halo to the infall of baryons is to become more dense at the center. While this makes sense physically, it is problematic for LCDM as it takes an NFW halo that is already too dense at the center to be consistent with data and makes it more dense. This has been known forever, so opposing this is one thing feedback is invoked to do, which it may or may not do, depending on how it really works. Even if feedback can really turn a compressed cusp into a core, it is widely to expected to be important only in low mass galaxies where the gravitational potential well isn’t too deep. It isn’t supposed to be all that important in galaxies as massive as the Milky Way, though I’m sure that can change as needed.

There are a variety of challenges to implementing an accurate compression computation, so we usually don’t bother: the standard practice is to assume a halo model and fit it to the data. That will, at best, given a description of the current dark matter halo, not what it started as, which is our closest point of comparison with theory. To give an example of the effect, here is a Milky Way model I built a decade ago:

Figure 13 from McGaugh (2016): Milky Way rotation curve from the data of Luna et al. (2006, red points) and McClure-Griffiths & Dickey (2007, gray points) together with a bulgeless baryonic mass model (black line). The total rotation is approximately fit (blue line) with an adiabatically compressed NFW halo (solid green line) using the procedure implemented by Sellwood & McGaugh (2005). The primordial halo before compression is shown as the dashed line. The parameters of the primordial halo are a concentration c = 7 and a mass M200 = 6 x 1011 M☉. Fitting NFW to the present halo instead gives c = 14, M200 = 4 x 1011 M☉, so the difference is appreciable and depend on the quality and radial extent of the available data.

The change from the green dashed line to the solid green line is the difference compression makes. That’s what happens if a baryon distribution like that of the Milky Way settles in an NFW halo. The inferred mass M200 is lower and the concentration c higher than it originally was – and it is the original version that we should compare to the expectations of LCDM.

When I built this model, I considered several choices for the bulge/bar fraction: something reasonable, something probably too large, and something definitely too small (zero). The model above is the last case of zero bulge/bar. I show it because it is the only case for which the compression procedure worked. If there is a larger central concentration of baryons – i.e., a bulge and/or a bar – then the compression is greater. Too great, in fact: I could not obtain a fit (see also Binney & Piffl and this related discussion).

The calculation of the compression requires knowledge of the primordial halo parameters, which is what one is trying to obtain. So one has to guess an initial state, run the code, check how close it came, then iterate the initial guess. This is computationally expensive, so I was just eyeballing the fit above. Pengfei has done a lot of work to implement a method that iteratively computes the compression and rigorously fits it to data. So we decided to apply it to the newer Gaia DR3 data.

Fitting the Gaia rotation curve with adiabatically compressed halos

We need two inputs here: one, the rotation curve to fit, and two, the baryonic distribution of the Milky Way. The latter is hard to specify given our location within the Milky Way, so there are many different estimates. We tried a dozen.

Another challenge of doing this is deciding which data rotation curve data to fit. We chose to focus on the rotation curve of Jiao et al. (2023) because they made estimates of the systematic as well as random errors. The statistics of Gaia are so good it is practically impossible to fit any equilibrium model to them. There are aspects of the data for which we have to consider non-equilibrium effects (spiral arms, the bar, “snails” from external perturbations) so the usual assumptions are at best an approximation, plus there can always be systematic errors. So the approach is to believe the data, but with the uncertainty estimate of Jiao et al. (2023) that includes systematics.

For a halo model, we started with the boilerplate LCDM NFW halo$. This doesn’t fit the data. Indeed, all attempts to fit NFW halos fail in similar ways for all of the different baryonic mass models we tried. The quasi-Keplerian part of the Gaia rotation curve simply cannot be fit: the NFW halo inevitably requires more mass further out.

Here are a few examples of the NFW fits:


Fig. A.3 from Li et al. (2025). Fits of Galactic circular velocities using the NFW model implementing adiabatic halo contraction using 3 baryonic models. [Another 9 appear in the paper.] Data points with errors are the rotation velocities from Jiao et al. (2023), while open triangles show the data from Eilers et al. (2019), which are not fitted. [The radius ranges from 5 to 30 kpc.] Blue, purple, green and black solid lines correspond to the contributions by the stellar disk, central bar, gas (and dust if any), and compressed dark matter halo, respectively. The total contributions are shown using red solid lines. Black dashed lines are the inferred primordial halos.

LCDM as represented by NFW suffers the same failure mode as seen in MOND (plot at top): both theories overshoot the Gaia rotation curve at R > 17 kpc. This is an example of how data that are problematic for MOND are also problematic for dark matter.

We do have more freedom in the case of dark matter. So we tried a different halo model, Einasto. (For this and many other halo models, see Pengfei’s epic compendium of dark matter halo fits.) Where NFW has two parameters, a concentration c and mass M200, Einasto has a third parameter that modulates the shape of the density profile%. For a very specific choice of this third parameter (α = 0.17), it looks basically the same as NFW. But if we let α be free, then we can obtain a fit. Of all the baryonic models, the RAR model+compressed Einasto fits best:


Fig. 1 from Li et al. (2025). Example of a circular velocity fit using the McGaugh19$$ model for baryonic mass distributions. The purple, blue, and green lines represent the contributions of the bar, disk, and gas components, respectively. The solid and dashed black lines show the current and primordial dark matter halos, respectively. The solid red line indicates the total velocity profile. The black points show the latest Gaia measurements (Jiao et al. 2023), and the gray upward triangles and squares show the terminal velocities from (McClure-Griffiths & Dickey 2007, 2016), and Portail et al. (2017), respectively. The data marked with open symbols were not fit because they do not consider the systematic uncertainties.

So it is possible to obtain a fit considering adiabatic compression. But at what price? The parameters of the best-fit primordial Einasto halo shown above are c = 5.1, M200 = 1.2 x 1011 M☉, and α = 2.75. That’s pretty far from the α = 0.17 expected in LCDM. The mass is lower than low. The concentration is also low. There are expectation values for all these quantities in LCDM, and all of them miss the mark.


Fig. 2 from Li et al. (2025). Halo masses and concentrations of the primordial Galactic halos derived from the Gaia circular velocity fits using 12 baryonic models. The red and blue stars with errors represent the halos with and without adiabatic contraction, respectively. The predicted halo mass-concentration relation within 1 σ from simulations (Dutton & Macciò 2014) is shown as the declining band. The vertical band shows the expected range of the MW halo mass according to the abundance-
matching relation (Moster et al. 2013). The upper and lower limits are set by the highest stellar mass model plus 1 σ and the lowest stellar mass model minus 1 σ, respectively.

The expectation for mass and concentration is shown as the bands above. If the primordial halo were anything like what it should be in LCDM, the halo parameters represented by the red stars should be where the bands intersect. They’re nowhere close. The same goes for the shape parameter. The halo should have a density profile like the blue band in the plot below; instead it is more like the red band.


Fig. 3 from Li et al. (2025). Structure of the inferred primordial and current Galactic halos, along with predictions for the cold and warm dark matter. The density profiles are scaled so that there is no need to assume or consider the masses or concentrations for these halos. The gray band indicates the range of the current halos derived from the Gaia velocity fits using the 12 baryonic models, and the red band shows their corresponding primordial halos within 1σ. The blue band presents the simulated halos with cold dark matter only (Dutton & Macciò 2014). The purple band shows the warm dark matter halos (normalized to match the primordial Galactic halo) with a core size spanning from 4.56 kpc (WDM5 in Macciò et al. 2012) to 7.0 kpc, corresponding to a particle mass of 0.05 keV and lower.

So the primordial halo of the Milky Way is pretty odd. From the perspective of LCDM, the mass is too low and the concentration is too low. The inner profile is too flat (a core rather than a cusp) and the outer profile is too steep. This outer steepness is a large part of why the mass comes out so low; there just isn’t a lot of halo out there. The characteristic density ρs is at least in the right ballpark, so aside from the inner slope, the outer slope, the mass, and the concentration, LCDM is doing great.

What if we ignore the naughty bits?

It is really hard for any halo model to fit the steep decline of the Gaia rotation curve at R > 17 kpc. Doing so is what makes the halo mass so small. I’m skeptical about this part of the data, so do things improve if we don’t sweat that part?

Ignoring the data at R > 17 kpc allows the mass to be larger, consistent with other dynamical determinations if not quite with abundance matching. However, the inner parts of the rotation curve still prefer a low density core. That is, something like the warm dark matter halo depicted as the purple band above rather than NFW with its dense central cusp. Or self-interacting dark matter. Or cold dark matter with just-so feedback. Or really anything that obfuscates the need to confront the dangerous question: why does MOND perform better?


*This post is based on the recently published paper by my former student Pengfei Li, who is now faculty at Nanjing University. They have a press release about it.

&A few months after reading this in the Boston Driver’s Handbook, this exact thing happened to me.

**This goes back to BBKS in 1986 when the bedrock assumption was that the universe had Ωm = 1, for which the virial radius was 188 times the critical density. 200 was close enough, and stuck, even though for LCDM the virial radius is more like an overdensity close to 100, which is even further out.

#This is one of many processes that occur in simulations, which are great for examining the statistics of simulated galaxy-like objects but completely useless for modeling individual galaxies in the real universe. There may be similar objects, but one can never say “this galaxy is represented by that simulated thing.” To model a real galaxy requires a customized approach.

$NFW halos consistently perform worse in fitting data than any other halo model, of which there are many. It has been falsified as a viable representation of reality so many times that I can’t recall them all, and yet they remain the go-to model. I think that’s partly thanks to their simplicity – it is mathematically straightforward to implement – and to the fact that is what simulations predict: LCDM halos should look like NFW. People, including scientists, often struggle to differentiate simulation from reality, so we keep flogging the dead horse.

%The density profile of the NFW halo model asymptotes to power laws at both small and large radii: ρ → r-1 as r → 0 and ρ → r-3 as r → ∞. The third parameter of Einasto allows a much wider ranges of shapes.

Einasto profiles. Einasto is observationally indistinguishable from NFW for α = 0.17, but allows many other shapes.

$$The McGaugh19 model user here is the one with a reasonable bulge/bar. This dense component can be fit in this case because we start with a halo model with a core rather than a cusp (closer to α = 1 than to the α = 0.17 of NFW/LCDM).

Progressive Approximations in Mass Modeling

Progressive Approximations in Mass Modeling

I have said I wasn’t going to attempt to teach an entire graduate course on galaxy dynamics in this forum, and I’m not. But I can give some pointers for those who want to try it for themselves. It also provides some useful context for fans of Deur’s approach.

The go-to textbook for this topic is Galactic Dynamics by Binney & Tremaine. The first edition was published in 1987, conveniently when I switched to grad school in astronomy. It was already a deep and well-developed field at that time; this is a compendium of considerable scientific knowledge.

Fun story: a colleague in a joint physics & astronomy department once complained to me that she wanted to develop a course in galaxy dynamics, which is a staple of graduate programs in astronomy & astrophysics. However, there was a certain senior colleague who objected, saying that since it was astronomy, it couldn’t possibly be a rigorous course worthy of a full semester graduate course. This is a casual bias that astronomers often encounter when talking to physicists, many of whom have attitudes about the subject that were trapped in amber sometime in the Jurassic. I suggested that she walk into his office and drop a copy of Galactic Dynamics on his desk from on high, as (1) it would make a hefty impact, and (2) no one who so much as skims this book could persist in this toxic attitude.

She later reported that she had done this, and it had worked.

Galactic Dynamics is not a starter book. It is the textbook we use when teaching the graduate course that this is not. A useful how-to guide for the specific material I’ll discuss here is provided by Federico Lelli. In brief, to model the gravitational potential of an observed distribution of matter, we can make one of the following series of approximations:

This is a slide I sometimes use to introduce mass modeling in science talks as a reminder for expert audiences.

All science is an approximation at some level. The most crude approximation we can employ here is to imagine that all of the mass resides at a central point. In this limit, the potential is simply

V2 = GM/R

where V is the orbital speed of a test particle on a circular orbit, G is Newton’s constant, M is the mass, and R is the distance from the point mass. Galaxies are not point masses, so this is a terrible approximation, as can be seen by the divergent V ~ R-1/2 behavior as R → 0 (the dotted line above).

The next bad approximation one can make is a spherical cow: assume the mass is distributed in a sphere that is projected as the image we see on the sky. This at least incorporates the fact that the mass is not all concentrated at a point, so

V2 = GM(R)/R

acknowledges that the mass M is spread out as a function of radius. This is a spherical cow. Since we cannot see dark matter, we almost always assume it to be a spherical cow.

For the luminous disk of a spiral galaxy, a common approximation is the so-called exponential disk:

Σ(R) = Σ0 e-R/Rd

where Σ0 is the central surface density of stars and Rd is the scale length of the disk – the characteristic size over which the surface brightness declines exponentially. This can be integrated by parts to obtain an expression for the enclosed mass M(R) which I leave as an exercise for the eager reader. This provides a handy analytic formula, the rotation curve of which is illustrated above by the dashed line.

Spiral galaxies are fairly thin when seen edge-on, so the spherical cow is not a great approximation. In a classic paper, Freeman (1970) solved the Poisson equation for the case of a razor-thin exponential disk, where one meets modified Bessel functions of the first and second kind (denoted “ikik” above). These must be solved numerically, but one can make a tabulation for use with any choice of disk mass and scale length. Such a thin disk is illustrated by the grey line above for a choice of stellar mass and scale length appropriate to NGC 6946.

The spiral galaxy NGC 6946, aka the fireworks galaxy.

Spiral galaxies are not razor thin of course. We only see a projected image on the sky, so for a galaxy like NGC 6946, we may have a good measurement of its azimuthally averaged light (and presumable stellar mass) distribution Σ(R) but we have no idea how thick it is. Here, we have to make an educated guess based on observations of edge-on galaxies. A ballpark average is R:z = 8:1, but some galaxies are thicker and others thinner, so this becomes an approximation with an associated uncertainty. This uncertainty cannot be unambiguously eliminated; it is one of the known unknowns that comprise the inevitable systematic errors in astronomy. Fortunately, allowing for a finite thickness only takes the harsh edge off of the thin disk case, and the assumption one chooses makes little difference to the result (compare the lines labeled thick and thin above).

The exponential disk formula Σ(R) is an azimuthal average over an image like that of NGC 6946. This approximation captures none of the spiral structure: it only tells us about the average rate at which the surface brightness falls off. It also imposes a smooth shape on that fall off that our eyes can see is not necessarily a great approximation. So the next level of approximation is to solve the Poisson equation numerically for the observed surface brightness profile, Σ(R), not just the exponential approximation thereto. This is the blue line in the bottom right graph above.

There are important differences between using the numerical solution for the observed light distribution and the exponential disk approximation. This has been known since the 1980s, but the analytic expression is so convenient that people need an occasional reminder not to trust it too much. Jerry Sellwood felt the need to provide this reminder in 1999:

Small apparent differences in the shape of the mass profile (left) correspond to pronounced differences in the rotation curve (right). I chose the example of NGC 6946 in part because the exponential approximation for it is pretty good. Nevertheless, the details matter, so the best practice is to build numerical mass models, as we did for SPARC.

Building numerical mass models is tractable for external galaxies, where we can see the entire light distribution. It is not possible for our own Milky Way, since we are located within it and cannot see it as a whole. Consequently, the vast majority of Milky Way models rely on the exponential approximation; so far as I’m aware, I’m the only one who has built a model that attempts to get beyond this.

Numerical mass models are still an approximation. We’re assuming that the gravitational potential is static and azimuthally symmetric. Taking the next step would require abandoning these assumptions to model the spiral arms. The Poisson equation can handle that, but it becomes dicey because the arms rotate with some pattern speed (generally unknown) and may grow or dissolve or reform on some unknown timescale. The potential at any given point is time variable even in equilibrium, so we need not just a numerical solution but a live numerical simulation to keep track of it. That can be done, but it has to be done on a case by case basis, and the answer will depend somewhat on additional assumptions that have to be introduced to run the simulation, like specifying a dark matter halo.

One can generalize further to consider the full 3D potential, e.g., to allow for asymmetry in the z-direction as well as in azimuth. One can further imagine non-equilibrium processes, such as an external perturbations. There is good evidence that the Milky Way suffers both of these effects, the passage of the Large Magellanic Cloud being one obvious and apparently large perturbation. So we are in the awkward position that the Gaia data now oblige us to consider the entire run of possible effects through non-equilibrium processes in a mass distribution that is not completely symmetric in any of the three spatial dimensions, but for the main mass component we are stuck with the inadequate approximation of an exponential disk.

Geometry appears to play a crucial role in the approach of Deur to the acceleration discrepancy problem. The essential claim is that the discrepancy correlates with flattening, with highly flattened systems like spirals evincing the classic discrepancy while spherical systems like E0 galaxies showing none. Big if true!

A useful plot appears on slide 44:

Some measure of the discrepancy as a function of apparent ellipticity.

This is the one example shown that goes into the plot of many determinations of the slope a on the following slide. It being the only one, it is the only thing I have to evaluate without chasing down every other case. Looking at this, I am not inclined to do so.

At first it looks persuasive: the best fit slope is clear. There is no reason why the discrepancy should depend on the projected ellipticity of a triaxial 3D blob of stars, so this must be telling us something important. I’d be on board with that if it were true, but I’ve seen too many non-correlations masquerading as correlations to believe this one. The fitted slope is strongly influenced by the one point at large ellipticity; absent that, a slope of zero works fine. Mostly what I see here is a lot of scatter, which is normal in extragalactic astronomy. Since there are only a few points at high and low ellipticity, we don’t know what would happen if we went out and got more data. But I bet that what would happen is that the high ellipticity points would wind up looking like those in the middle: a big blob of scatter, with no significant correlation.

I’d kinda like to be wrong about this one, so I won’t even get into the theory side, which I find sorta compelling but ultimately unpersuasive. Why are gravitons confined to a disk? What happens way far out? Surely the flatness of the disk at tens of kpc is not dictating the flatness at 1000 kpc.

Surely.

Tully-Fisher from gravitational lensing

Tully-Fisher from gravitational lensing

Last time, we discussed the remarkable result that gravitational lensing extends the original remarkable result of flat rotation curves much farther out, as far as the data credibly probe. This corroborates and extends the result of Brouwer et al. They did a thorough job, but one thing they did not consider was Tully-Fisher. If the circular speed inferred from gravitational lensing remains constant, does this flat velocity fall on the same Tully-Fisher relation that is seen in kinematic data?

We set out to answer this question. Along the way, we did three new things: 1. Dr. Mistele derived an improved method for doing the lensing analysis, extending the radial range over which the data were credible. 2. He explored the criteria by which galaxies were judged to be isolated, finding a morphological type dependence on how far out one had to exclude. 3. We reanalyzed the stellar masses of the KiDS sample to be consistent with those we used when analyzing the kinematic data of SPARC galaxies. The first two are connected, as how far out we can trust the data depends on how well we can define a clean sample of isolated galaxies. The third resolved an apparent offset between early type galaxies (ETGs, aka ellipticals) and late type galaxies (LTGs, aka spirals) seen by Brouwer et al. That appears to be an artifact of stellar population modeling, as I suspected when I first discussed their result. We don’t need to do any fitting of the mass-to-light ratio; the the apparent offset between types disappears when we use use the same population models for both kinematic and lensing data.

I could write a lot about each of these, but most of it is the stuff of technical details that would be dull to many people. If you’re into that sort of thing, go and read the long science paper which is where such details reside. Here I just want to describe the Tully-Fisher result. Spoiler alert: it is the same as that from kinematics.

First off, I’m talking strictly about the Baryonic Tully-Fisher relation: the scaling between baryonic mass and the flat rotation speed. To address this, we bin the lensing data by mass. The mass of each bin is well defined by the average of the many thousands of galaxies within the bin. By far the dominant uncertainty is the systematic in stellar mass caused by stellar population modeling. We went through this with a fine tooth comb, and I’m confident we have an internally self-consistent result. That doesn’t preclude it being wrong in an absolute sense – such is the nature of astronomy – but we can at least make a straight comparison between kinematic and lensing data using the same best-effort stellar mass estimates.

For the velocity, we estimate the average effective rotation curve for each mass bin from the lensing data. We also split the data into morphological types to look for differences. The statistics go down when one divvies up the data like this, so the uncertainties go up, but there are enough KiDS galaxies to define four mass bins. Here are their inferred rotation curves:

Figure 1 from Mistele et al. (2024). Circular velocities implied by weak lensing for four baryonic mass bins (most to least massive from the top row to the bottom) for the whole sample (left column), for LTGs (middle column), and for ETGs (right column). The lowest ETG mass bin is not shown because it contains too few lenses. Instead we show results for lenses with spectroscopic redshifts from GAMA, without splitting by mass or type due to the small sample size (gray and white symbols). For comparison, we also show results for KiDS without splitting by mass or type (small yellow symbols). Open symbols at small radii indicate where lenses are not yet effective point masses. Light-colored symbols (not-outlined) at large radii indicate data points that may still be reliable but where the isolation criterion is less certain. The error bars show the statistical errors. Horizontal lines and the corresponding shaded regions indicate the inferred Vflat values and uncertainties that we use for the BTFR. The extent of the horizontal lines indicates the radial range we consider when calculating Vflat.

Note that the average over all KiDS data shown in the lower right bin is the data shown in the press release image in the previous post, but the x-axis is logarithmic here. The GAMA data in that bin provide an important cross-check, as these galaxies have spectroscopic redshifts. They give the same answer as the larger KiDS sample, which relies on photometric redshifts. We need the larger sample to consider finer bins in mass, which is the rest of the plot.

Another thing to note here is that all the data in all the bins are consistent with remaining flat. There are some hints of a turn down at very large radii, particularly for LTGs in the second and third row, but these are not statistically significant, and only happen where the data start to become untrustworthy. Where exactly that happens is a judgement call.

Let’s take a closer look, with a comparison to radio data:

Figure 2 from Mistele et al. (2024). The circular velocities from weak lensing (circles) compared with those from gas kinematics (diamonds). The individual galaxies illustrated here have among the most extended 21 cm rotation curves in their mass bins; the lensing data continue to much larger radii still. The error bars show the statistical error, while the gray band indicates the systematic uncertainty in the radial accelerations. Symbol colors are as in Figure 1. Open symbols at large radii indicate where lenses are not sufficiently isolated. The solid green lines indicate the circular velocities of NFW halos and baryonic point masses appropriate for each mass bin. Green crosses indicate each NFW halo’s virial radius. The light green band adds a qualitative estimate of a two-halo term contribution to the NFW halo, which may become important at large radii in case our isolation criterion is imperfect there.

Again we see that the lensing data, averaged over many galaxies, extend much further out than the rotation curve of any one individual. The x-axis is again logarithmic, so the lensing data go way further out. They trace to 1 Mpc, which is crazy far beyond the observed ends of the most extended individual galaxies. A more conservative limit is the 300 kpc estimated by Brouwer et al. Surely we can go further than that, but how much further remains a judgement call.

What should we expect? The green lines show the rotation curve we’d expect for galaxy in an NFW halo with parameters specified by the stellar mass-halo mass relation of Kravtsov et al. (2018). Not all such relations agree well with kinematic data; this is the case that agrees most closely. We have intentionally cherry-picked the relation that makes LCDM look best. And it does look good up to a point, for example in the top two mass bins out to the virial radius of the halo (tick marks). Beyond that, not so much, and not at all for the two lower mass bins. The data extend far enough out that we should see the predicted decline. We do not.

The green line only represents the expected halo of the primary galaxy. When one gets so far out, one has to worry about all the other stuff out there. We’ve selected galaxies to be isolated, so there isn’t much that is luminous. But we can only exclude down to some sensitivity limit, there might be lots of tiny dwarf galaxies whose mass adds up and starts to affect the result. And of course there can be completely invisible dark matter. The green band attempts to account for this extra stuff in the so-called 2-halo term. This is hard to do, but we’ve made our best estimate based on the LCDM power spectrum. I’m sure the 2-halo term can be adjusted, but the shape is wrong. It will take some fine-tuning to get an effectively flat rotation curve out of the 1-halo+2 halo terms. They don’t naturally do that.

Something that is easy to do is define a flat value of the rotation speed. That’s just the average over the lensing data. We exclude the points at R < 50 kpc, as the assumption of a spherical mass that we make in the lensing analysis isn’t really valid at those comparatively small scales. We tried averaging over a bunch of different ranges, all of which gave pretty much the same answer. For illustration, we show two cases: a conservative one that only uses the data at R < 300 kpc, and another that goes out to 1 Mpc. Having measured Vflat over these ranges, we can plot Tully-Fisher:

Figure 3 from Mistele et al. (2024). The baryonic Tully–Fisher relation implied by weak lensing for the entire sample (yellow symbols, left column) and for ETGs and LTGs separately (red and blue symbols, right column). The Vflat values are weighted averages of the Vc values shown in Figure 1 for 50 kpc < R < 300 kpc (first row) and 50 kpc < R < 1000 kpc (second row). Vertical error bars represent a 0.1 dex systematic uncertainty on M*/L. For comparison, we also show the best fit to the kinematic data from Lelli et al. (2019; solid gray line) and the corresponding binned kinematic data (white diamonds).

Lo and behold, we find the same Baryonic Tully-Fisher relation from lensing data as we find with kinematics. This does not surprise me, but it didn’t have to be true. It shouldn’t be true in LCDM: if we can measure out to the virial radius, we should see some indication of a decline in velocity. We have and we don’t.

We also see no statistically significant separation between ETGs and LTGs. This is important, as a theory like MOND predicts that there should be no morphology dependence: only the baryonic mass matters. Brouwer et al. did see an indication of such a split, but it was small compared to the uncertainty in stellar population models. We don’t see it when we use our own stellar mass estimates. This is particularly true in the more conservative (300 kpc) case. There is a hint of a segregation when we average out to 1000 kpc, but the statistics say this isn’t significant. Since the lowest mass bin is most affected, I suspect this is a hint that the isolation criterion is failing first for the smallest galaxies. That makes sense, as the sensitivity limit on interlopers makes the lowest mass bin most susceptible to having its signal inappropriately boosted. It also makes sense that ETGs would be affected first, as ETGs are known to be more clustered than LTGs. It is really hard to define an isolated sample of ETGs, as discussed at length by Mistele et al.

The lensing data corroborate previous kinematic work. Rotation curves are flat. The amplitude of the flat rotation speed correlates with baryonic mass as Mb ∝ Vf4. The radial acceleration relation extends to very low accelerations. These are all predictions of MOND. Moreover they are unique predictions: predictions made a priori by MOND and only by MOND. Dark matter models so far provide no satisfactory explanation*.

That hasn’t prevented people from overlooking these basic facts in order to get to the apparent if statistically meaningless difference between ETGs and LTGs. Nevermind the successes! The slight offset between ETGs and LTGs falsify MOND! Seriously: other scientists have already made this argument to me while completely eliding the successes of MOND. It’s a case of refusing to see the forest for a tree that’s a little away from the others.

I think I said something about how this would happen when I first wrote about Brouwer et al‘s lensing result. Ah yes, here it is:

MOND predicted this behavior well in advance of the observation, so one would have to bend over backwards, rub one’s belly, and simultaneously punch oneself in the face to portray this as anything short of a fantastic success of MOND.

I say that because I’m sure people will line up to punch themselves in the face in exactly this fashion.

And so it has come to pass. Sometimes human behavior is as predictable as galaxy dynamics.


*There are many claims to explain limited portions of these results, but none are satisfactory. There is no LCDM model that matches the entire dynamic range of the radial acceleration relation. See, for example, Fig. 5 of Brouwer et al. (reproduced below), which shows the MICE and BAHAMAS simulations. Neither extend into the regime that is well-constrained by kinematic data; there is no reason to think they would successfully do so and good reason to think otherwise. MICE comes nowhere close to this regime and has no baryonic physics that would allow it do even address this question. BAHAMAS comes close but appears to turn away from the kinematic data before it gets there. We’ve built our own LCDM models; they don’t work either. We can make them come close, but only over a limited dynamic range, not over the full span of the data. It isn’t good enough to only explain a limited range of the data. One has to explain the full range, and the only theory that does that so far is MOND.

Fig. 5 from Brouwer et al. showing the radial acceleration relation inferred from the MICE (red band) and BAHAMAS (orange band) simulations. Not also that in our assessment of stellar masses, the lower acceleration points translate a bit to the right.

Rotation curves: still flat after a million light-years

Rotation curves: still flat after a million light-years

That rotation curves become flat at large radii is one of the most famous results in extragalactic astronomy. This had been established by Vera Rubin and her collaborators by the late 1970s. There were a few earlier anecdotal cases to this effect, but these seemed like mild curiosities until Rubin showed that the same thing was true over and over again for a hundred spiral galaxies. Flat rotation curves took on the air of a de facto natural law and precipitated the modern dark matter paradigm.

Optical and radio data

Rotation curves shouldn’t be flat. If what we saw was what we got, the rotation curve would reach a peak within the light distribution and decline further out. Perhaps an illustration is in order:

The rotation curve (data points, left) of NGC 6946 (right). The red line shows the expected rotation curve for the detected normal matter, which includes both the stars (yellow, from 2MASS) and atomic gas (blue, from THINGS). This provides a good description of the inner rotation curve but falls short further out. The excess observed rotation leads to the need for dark matter or MOND. Also noted is the extent of the rotation curve measured optically to the effective edge of the stars (Daigle et al. 2006; Epinat et al. 2008) and that measured with radio interferometric observations of the gas (Boomsma et al. 2008).

An obvious question is how far out rotation curves remain flat. In the rotation curves traced with optical observations by Rubin et al., the discrepancy was clear but modest – typically a factor of two in mass. It was possible to imagine that the mass-to-light ratios of stars increased with radius in a systematic way, bending the red line above to match the data out to the edge of the stars. This seemed unlikely, but neither did it seem like a huge ask.

Once one gets to the edge of the stellar distribution, most of the mass has been encompassed, and the rotation curve really should start to decline. Increasing the mass-to-light ratio of the stars ceases to be an option once we run out of stars*. Fortunately, the atomic gas typically extends to larger radii, so provides a tracer further out. Albert Bosma pursued this until there were again enough examples to establish that yes, flat rotation curves were the rule. They extended much further out, well beyond where the mass of the observed stars and gas could explain the data.

How much further out? It depends on the galaxy. A convenient metric is the scale length of the disk, which is a measure of the extent of the light distribution. Some galaxies are bigger than others. The peak of the contribution of the stars to the rotation curve occurs around 2.2 scale lengths. The rotation curve of NGC 6946 extends to about 7 scale lengths, far enough to make the discrepancy clear. For a long time, the record holder was NGC 2403, with a rotation curve that remains flat for 20 scale lengths.

Twenty scale lengths is a long way out. It is observations like this that demanded dark matter halos that are much larger than the galaxies they contain. They also posed a puzzle, since we were still nowhere near finding the edge of the mass distribution. Rotation curves seemed to persist in being flat indefinitely.

Results from gravitational lensing

Weak gravitational lensing provides a statistical technique to probe the gravitational potential of galaxies. Brouwer et al. did pioneering work with data from the KiDS survey, and found that the radial acceleration relation extended to much lower accelerations than probed by the types of kinematic data discussed above. That implies that rotation curves remain flat way far out. How far?

Postdoc Tobias Mistele worked out an elegant technique to improve the analysis of lensing data. His analysis corroborates the findings of Brouwer et al. It also provides the opportunity to push further out.

Weak gravitational lensing is a subtle effect – so subtle that one must coadd thousands of galaxies to get a signal. Beyond that, the limiting effect on the result is how isolated the galaxies are. Lensing is sensitive to all mass; if you go far enough out you start to run into other galaxies whose mass contributes to the signal. So one key is to identify isolated galaxies, and restrict the sample to them. KiDS is large enough to do this. Indeed, Mistele was able to show that while neighbors+ were a definite concern for elliptical galaxies, they were much less of a problem for spirals. Consequently, we can trace the implied rotation curve way far out.

How far out? In a new paper, Mistele shows that rotation curves continue way far out. Way way way far out. I mean, damn.

The average rotation curve of isolated galaxies (blue points) inferred from KiDS gravitational lensing data. This remains flat well beyond a million light-years with no end in sight. The width of the figure is the distance between the Milky Way and Andromeda. For comparison, the rotation curve of a single galaxy, UGC 6614, is shown in red. An image of the galaxy is shown to scale centered at the origin. UGC 6614 was selected for this illustration because it has a comparable rotation speed to the KiDS average and because it is one of the largest galaxies known: the red points are already a very extended rotation curve. Image credit: Mistele, Lelli, & McGaugh 2024.

Optical rotation curves typically extend to the edge of the stellar disk. That’s about 8 kpc in the example of NGC 6946 given above. Radio observations of the atomic gas of that galaxy extend to 17 kpc. That fits within the first two tick marks on the graph with the lensing rotation curve.

UGC 6614 is a massive galaxy with a very extended low surface brightness disk. Its rotation curve is traced by radio data to over 60 kpc. It is one of the most extended individual rotation curves known. The statistical lensing data push this out by a factor of ten, and more, with no end in sight. The flat rotation curves found by Rubin and Bosma and everyone else appear to persist indefinitely.

So what does it mean? First, flat rotation curves really are a law of nature, in the same sense of Kepler’s laws of planetary motion. Galaxies don’t obey those planetary rules, they have their own set of rules. This is what nature does.

In terms of dark matter halos, the extent of isolated galaxy rotation curves is surprisingly large. Just as we come to the edge of the stellar disk, then the gas disk, we should eventually hit the edge of the dark matter halo. In principle we can imagine this to be arbitrarily large, but in practice there are other galaxies in the universe so this cannot go one forever.

In the context of LCDM, we now have a pretty good idea of how extended halos should be from abundance matching. A galaxy of the mass of UGC 6614 should live in a halo with a virial radius of about 300 kpc or less. There is some uncertainty in this, of course, but we really should have hit the edge with the lensing data. There should be some sign of it, but we see none.

One complication is the so-called 2-halo term. In addition to the primary dark matter halo that hosts a galaxy, when you get very far out, you run into other halos. Isolated galaxies are selected to avoid this to the extent possible, but eventually there will be some extra mass that causes extra lensing signal that would cause an overestimate of the rotation speed. I’ll forgo a detailed discussion of this for now (see Mistele et al. if you’re eager), but the bottom line is that it would require some unnatural fine-tuning for the 1+2 halo terms to add up to such flat rotation curves. There ought to be a perceptible feature in the transition from the primary halo to the surrounding environment. We don’t see that.

In the context of MOND, a flat rotation curve that persists indefinitely is completely natural. That’s what an isolated galaxy should do. Even in MOND there should be an environmental effect: the mass of everything else in the universe should impose an external field effect that eventually limits the extent of the rotation curve. How this transition happens depends on the density of other galaxies; by selecting isolated galaxies this effect is put off as much as possible. Hopefully it will be detected as the data improve from projects like Euclid.

The primary prediction of MOND is an indefinitely extended rotation curve; the external field effect is a subtle detail. Yet again, that is what we see: MOND gets it right without really trying, and in a way that makes little sense in terms of dark matter. Sometimes I wish MOND had never been invented so we could claim to have discovered something profoundly new, or at least discuss the empirical result without concern that the data would get confused with the theory. MOND predictions keep being corroborated, yet the community persists in ignoring its implications, even in terms of dark matter. It’s gotta be telling us something.

We have a press release about this result, so perhaps you will see it kicking around your news feed.


*We could, of course, invoke dark stars, but that’s just an invisible horse of a different color.

+There is a well known correlation between morphology and density such that elliptical galaxies tend to live in the densest environments. This means that they are more likely to have neighbors that interfere with the lensing measurement, so finding that identifying isolated ellipticals with a clean lensing signal is more challenging that finding isolated spirals comes as no surprise. Isolated ellipticals do exist so it is possible, but one has to be very restrictive with the sample.

Is the Milky Way’s rotation curve declining?

Is the Milky Way’s rotation curve declining?

Yes, some. That much is a step forward from a decade ago, when a common assumption was that the Milky Way’s rotation curve remained flat at the speed at which the sun orbited. This was a good guess based on empirical experience with other galaxies, but not all galaxies have rotation curves that are completely flat, nor can we be sure the sun is located where that is the case.

A bigger question whether the Milky Way’s rotation curve is declining in a Keplerian fashion. This would indicate that the total mass has been enclosed. That would be a remarkable result. If true, it would be the first time that the total mass of an individual galaxy has been measured. There have been claims to this effect before that have not panned out when the data have been extended to larger radii, so one might be inclined to be skeptical.

There are several claims now to see a distinctly declining rotation curve based on the third data release (DR3) from Gaia. The most recent, Jiao et al., has gained some note by virtue of putting “Keplerian decline” in the title, but very similar results have also been reported by Ou et al., Wang et al. and Sylos Labini et al. They all obtain basically the same answer using the same data, with minor differences in the error assessment and other details. There are also differences in interpretation*, which is always possible even when everyone agrees about what the data say.

Jiao et al. measure a total mass for the Milky Way of about 2 x 1011 M☉. Before looking at the data, let’s take a moment to think about that number. Most mass determinations – and there are lots, see Fig. 2 of Wang et al. – for the Milky Way have been in the neighborhood of 1012 M☉. Indeed, for most of my career, it was traditionally Known to be 2 x 1012 M☉. The new measurement is an order of magnitude smaller. That’s a lot to be off by, even in extragalactic astronomy. The difference, as we’ll see, has to do with what data we use.

The mass of stars and gas in the Milky Way is about 6 x 1010 M☉, give or take ten billion. That means that nearly a third of the total mass is normal baryonic matter that we can readily see. So the ratio of dark-to-baryonic mass is only 2.3:1, well short of the cosmic ratio of about 6:1. That’s embarrassing – especially since much of the effort in galaxy formation theory has been to explain why the baryon fraction is much less than the cosmic fraction, not much more. And here our Galaxy is an outlier, having much less dark matter for its stellar mass than everything else. It is always a bad sign when the Galaxy appears to violate the Copernican Principle.

Nonetheless, this is what we find if we look at the Gaia DR3 data. Here is a model I’ve shown before, extrapolated to larger radii with some new data added. The orange circles are the Gaia DR3 rotation curve as given by Jiao et al. For radii greater than 18 kpc, they show a clear decline consistent with a Keplerian curve for a 1.95 x 1011 M☉ point mass (dotted line), as per Fig. 9 of Jiao et al.

Milky Way model (blue line) compared with various data.

This is the first time we’ve been able to trace the rotation curve so far out with stars in the disk of the Milky Way, and the Keplerian line is a good match. If that’s all we know, then a total mass of only 2 x 1011 M☉ is a reasonable inference. That’s not all we know.

As I alluded above, a halo mass this small makes no sense in the context of cosmology. Not only is 2 x 1011 M☉ too small, the more commonly inferred dynamical mass of 1012 M☉ is also too small. According to abundance matching, which has become an important aspect of LCDM, the Milky Way should reside in a 3 or 4 x 1012 M☉ halo. So the new mass makes a factor of 2 or 3 problem into a factor a ten problem. That is too large to attribute to scatter in the stellar mass-halo mass relation. Worse, there is no evidence that the Milky Way is an outlier from scaling relations like Tully-Fisher. We can’t have it one way and not the other.

The traditional mass estimates that obtain ~1012 M☉ rely on dwarf satellite galaxies as tracers of the gravitational potential of the Milky Way. Maybe they’re not fair tracers? We have to make assumptions about their orbits to use them to infer a mass; perhaps these assumptions are wrong? It is conceivable that many of our satellites are on first infall rather than in well-established orbits. Indeed, the consensus is that our largest satellites, the Magellanic Clouds, are on first infall, and that they cause a substantial perturbation to the halo of the Milky Way. This was an absurd thought 15 years ago – the Magellanic clouds must have been here forever, and were far too small to do damage – but now this is standard lore.

There are tracers at large radii besides dwarf satellite galaxies. The figure above shows three: globular clusters (pink triangles) and two types of stars in the halo: blue horizontal branch stars (green squares) and K giants (red squares). These are well-known parts of the Milky Way that have been with us for many billions of years, so they’ve had plenty of time to become equilibrium tracers of the gravitational potential. They clearly indicate a larger enclosed mass than predicted by the Keplerian decline traced by the Gaia rotation curve, and are consistent with traditional satellite analyses. Perhaps these data are somehow misleading, but it is hard to see how.

Gaia is great, but has its limits. It is really optimized for nearby stars (within a few kpc). Outside of that, the statistics… leave something to be desired. Is it safe to push out beyond 20 kpc? I don’t know, but I did notice this panel from Fig. 8 of Wang et al.:

Radial velocities of stars at different heights above the Galactic plane.

The radial velocity is a minor component of disk motion, where azimuthal motion dominates. However, one does need to know it to solve the Jeans equation. Having it wrong will cause a perceptible systematic error. You notice the bifurcation in the data for R > 22 kpc? That, in technical terms, is Messed Up. I don’t know what goes awry there, but I’ve done this exercise enough times for the sight of this to scare the bejeepers out of me. No way I trust any of these data at R > 22 kpc, and I hope having seen this doesn’t give me nightmares tonight.

Perhaps the uncertainty caused by this is adequately reflected in the large error bars on the orange points above. Those with R > 22 kpc are nicely Keplerian, but also consistent with a lot of things, including the blue line that successfully predicts the halo stars and globular clusters. That’s not true for the data around R = 20 kpc where the error bars are much smaller: there the discrepancy with the blue line I take seriously. But that is a much more limited affair that might indicate the presence of a ring of mass – that’s what gives the bumps and wiggles at smaller radii – and certainly isn’t enough to imply the entire mass of the Milky Way has been enclosed.

But who knows? Perhaps fifteen years hence it will be the standard lore that all galaxies reside in dark matter halos that are only twice the mass of their luminous disks. At that mass ratio, all the galactic dark matter could be baryonic. I wouldn’t bet on it, but stranger things have happened before, and will happen again.


*A difference in interpretation is largely what the debate about dark matter and MOND boils down to. There is no doubt that there are acceleration discrepancies in extragalactic objects that require something beyond what you see is what you get with normal gravity. Whether we should blame what we can’t see or the assumption of normal gravity is open to interpretation. I would hope this is obvious, but this elementary point seems to be lost on many.

A few words about the Milky Way

A few words about the Milky Way

I recently traveled to my first international meeting since the Covid pandemic began. It was good to be out in the world again. It also served as an excellent reminder of the importance of in-person interactions. On-line interactions are not an adequate substitute. I’d like to be able to recount all that I learned there, but it is too much. This post will touch on one of the much-discussed topics, our own Milky Way Galaxy.

When I put on a MOND hat, there are a few observations that puzzle me. The most persistent of these include the residual mass discrepancy in clusters, the cosmic microwave background, and the vertical motions of stars in the Milky Way disk. Though much hyped, the case for galaxies lacking dark matter does not concern me much: the examples I’ve seen so far appear to be part of the normal churn of early results that are likely to regress toward the norm as the data improve. I’ve seen this movie literally hundreds of times. I’m more interested in understanding the forest than a few outlying trees.

The Milky Way is a normal galaxy – it is part of the forest. It is easy to get lost in the leaves when one has access to data for millions going on billions of individual stars. These add up to a normal spiral galaxy, and we know a lot about external spirals that can help inform our picture of our own home.

For example, by assuming that the Milky Way falls along the radial acceleration relation defined by other spiral galaxies, I was able to build a mass model of its surface density profile. The resulting mass distribution is considerably more detailed than the usual approach of assuming a smooth exponential disk, which would be a straight line in the right-hand plot below. With the level of detail becoming available from missions like the Gaia satellite, it is necessary to move beyond such approximations.

Left: Spiral structure in the Milky Way traced by regions of gas ionized by young stars (HII regions, in red) and by the birthplaces of giant molecular clouds (GMCs, in blue). Right: the azimuthally-averaged surface density profile of stars inferred from the rotation curve of the Milky Way using the Radial Acceleration Relation. The features inferred kinematically correspond to the spiral arms known from star counts, providing a local example of Renzo’s Rule.

This model was built before Gaia data became available, and is not informed by it. Rather, I took the terminal velocities measured by McClure-Griffiths & Dickey, which provide the estimate of the Milky Way rotation curve that is most directly comparable to what we measure in external spirals, and worked out the surface density profile using the radial acceleration relation. The resulting model possesses bumps and wiggles like those we see corresponding to spiral arms in external galaxies. And indeed, it turns out that the locations of these features correspond with known spiral arms. Those are independent observations: one is from the kinematics of interstellar gas, the other from traditional star counts.

The model turns out to have a few further virtues. It matches the enclosed mass profile of the inner bulge/bar region of the Galaxy without any attempt at a specific fit. It reconciles the rotation curve measured with stars using Gaia data with that measured using gas in the interstellar medium – a subtle difference that was nevertheless highly significant. It successfully predicts that the rotation curve beyond the solar radius would not be perfectly flat, but rather decline at a specific rate – and exactly that rate was subsequently measured using Gaia. These are the sort of results that inclines one to believe that the underlying physics has to be MOND. Inferring maps of the mass distribution with this level of detail is simply not possible using a dark matter model.

The rotation curve of the Milky as observed in interstellar gas (light grey) and as fit to the radial acceleration relation (blue line). Only the region from 3 to 8 kpc has been fit; the rest follows. This matches well stellar observations from the inner, barred region of the Milky Way (dark grey squares: Portail et al. 2017) and the gradual decline of the outer rotation curve (black squares: Eilers et al. 2019) once corrected for the presence of bumps and wiggles due to spiral arms. These require taking numerical derivatives for use in the Jeans equation; the red squares show the conventional result obtained when neglecting this effect by assuming a smooth exponential surface density profile. See McGaugh (2008 [when the method was introduced and the bulge/bar model for the inner region was built], 2016 [the main fitting paper], 2018 [an update to the distance to the Galactic center], 2019 [including bumps & wiggles in the Gaia analysis]).

Great, right? It is. It also makes a further prediction: we can use the mass model to predict the vertical motions of stars perpendicular to the Milky Way’s disk.

Most of the kinetic energy of stars orbiting in the solar neighborhood is invested in circular motion: the vast majority of stars are orbiting in the same direction in the same plane at nearly the same speed. There is some scatter, of course, but radial motions due to orbital eccentricities represent a small portion of the kinetic energy budget. As stars go round and round, the also bob up and down, oscillating perpendicular to the plane of the disk. The energy invested in these vertical motions is also small, which is why the disk of the Milky Way is thin.

View of the Milky Way in the infrared provided by the COBE satellite. The dust lanes that afflict optical light are less severe at these wavelengths, revealing that the stellar disk of the Milky Way is thin but for the peanut-shaped bulge/bar at the center.

Knowing the surface density profile of the Milky Way disk, we can predict the vertical motions. In the context of dark matter, most of the restoring force that keeps stars near the central plane is provided by the stars themselves – the dark matter halo is quasi-spherical, and doesn’t contribute much to the restoring force of the disk. In MOND, the stars and gas are all there is. So the prediction is straightforward (if technically fraught) in both paradigms. Here is a comparison of both predictions with data from Bovy & Rix (2013).

The dynamical surface density implied by vertical motions (data from Bovy & Rix 2013). The dark blue line is the prediction of the model surface density described above – assuming Newtonian gravity. The light blue line is the naive prediction of MOND.

Looks great again, right? The dark blue line goes right through the data with zero fitting. The only exception is in the radial range 5.5 to 6.4 kpc, which turns out to be where the stars probing the vertical motion are maximally different from the gas informing the prediction: we’re looking at different Galactic longitudes, right where there is or is not a spiral arm, so perhaps we should get a different answer in this range. Theory gives us the right answer, no muss, no fuss.

Except, hang on – the line that fits is the Newtonian prediction. The prediction of MOND overshoots the data. It gets the shape right, but the naive MOND prediction is for more vertical motion than we see.

By the “naive” MOND prediction, I mean that we assume that MOND gives the same boost in the vertical direction as it does in the radial direction. This is the obvious first thing to try, but it is not necessarily what happens in all possible MOND theories. Indeed, there are some flavors of modified inertia in which it should not. However, one would expect some boost, and in these data there appears to be none. We get the right answer with just Newton and stars. There’s not even room for much dark matter.

I hope Gaia helps us sort this out. I worry that it will provide so much information that we risk missing the big picture for all the leaves.

This leaves us in a weird predicament. The radial force is extraordinarily well-described by MOND, which reveals details that we could never hope to access if all we know about is dark matter. But if we spot Newtonian gravity this non-Newtonian information from the radial motion, it predicts the correct vertical motion. It’s like we have MOND in one direction and Newton in another.

This makes no sense, so is one of the things that worries me most about MOND. It is not encouraging for dark matter either – we don’t get to spot ourselves MOND in the radial direction then pretend that dark matter did it. At present, it feels like we are up the proverbial creek without a paddle.

Are there credible deviations from the baryonic Tully-Fisher relation?

Are there credible deviations from the baryonic Tully-Fisher relation?

There is a rule of thumb in scientific publication that if a title is posed a question, the answer is no.

It sucks being so far ahead of the field that I get to watch people repeat the mistakes I made (or almost made) and warned against long ago. There have been persistent claims of deviations of one sort or another from the Baryonic Tully-Fisher relation (BTFR). So far, these have all been obviously wrong, for reasons we’ve discussed before. It all boils down to data quality. The credibility of data is important, especially in astronomy.

Here is a plot of the BTFR for all the data I have ready at hand, both for gas rich galaxies and the SPARC sample:

Baryonic mass (stars plus gas) as a function of the rotation speed measured at the outermost detected radius.

A relation is clear in the plot above, but it’s a mess. There’s lots of scatter, especially at low mass. There is also a systematic tendency for low mass galaxies to fall to the left of the main relation, appearing to rotate too slowly for their mass.

There is no quality control in the plot above. I have thrown all the mud at the wall. Let’s now do some quality control. The plotted quantities are the baryonic mass and the flat rotation speed. We haven’t actually measured the flat rotation speed in all these cases. For some, we’ve simply taken the last measured point. This was an issue we explicitly pointed out in Stark et al (2009):

Fig. 1 from Stark et al (2009): Examples of rotation curves (Swaters et al. 2009) that do and do not satisfy the flatness criterion. The rotation curve of UGC 4173 (top) rises continuously and does not meet the flatness criterion. UGC 5721 (center) is an ideal case with clear flattening of the rotational velocity. UGC 4499 marginally satisfies the flatness criterion.

If we include a galaxy like UGC 4173, we expect it will be offset to the low velocity side because we haven’t measured the flat rotation speed. We’ve merely taken that last point and hoped it is close enough. Sometimes it is, depending on your tolerance for systematic errors. But the plain fact is that we haven’t measured the flat rotation speed in this case. We don’t even know if it has one; it is only empirical experience with other examples that lead us to expect it to flatten if we manage to observe further out.

For our purpose here, it is as if we hadn’t measured this galaxy at all. So let’s not pretend like we have, and restrict the plot to galaxies for which the flat velocity is measured:

The same as the first plot, restricted to galaxies for which the flat rotation speed has been measured.

The scatter in the BTFR decreases dramatically when we exclude the galaxies for which we haven’t measured the relevant quantities. This is a simple matter of data quality. We’re no longer pretending to have measured a quantity that we haven’t measured.

There are still some outliers as there are still things that can go wrong. Inclinations are a challenge for some galaxies, as are distances determinations. Remember that Tully-Fisher was first employed as a distance indicator. If we look at the plot above from that perspective, the outliers have obviously been assigned the wrong distance, and we would assign a new one by putting them on the relation. That, in a nutshell, is how astronomical distance indicators work.

If we restrict the data to those with accurate measurements, we get

Same as the plot above, restricted to galaxies for which the quantities measured on both axes have been measured to an accuracy of 20% or better.

Now the outliers are gone. They were outliers because they had crappy data. This is completely unsurprising. Some astronomical data are always crappy. You plot crap against crap, you get crap. If, on the other hand, you look at the high quality data, you get a high quality correlation. Even then, you can never be sure that you’ve excluded all the crap, as there are often unknown unknowns – systematic errors you don’t know about and can’t control for.

We have done the exercise of varying the tolerance limits on data quality many times. We have shown that the scatter varies as expected with data quality. If we consider high quality data, we find a small scatter in the BTFR. If we consider low quality data, we get to plot more points, but the scatter goes up. You can see this by eye above. We can quantify this, and have. The amount of scatter varies as expected with the size of the uncertainties. Bigger errors, bigger scatter. Smaller errors, smaller scatter. This shouldn’t be hard to understand.

So why do people – many of them good scientists – keep screwing this up?

There are several answers. One is that measuring the flat rotation speed is hard. We have only done it for a couple hundred galaxies. This seems like a tiny number in the era of the Sloan Digitial Sky Survey, which enables any newbie to assemble a sample of tens of thousands of galaxies… with photometric data. It doesn’t provide any kinematic data. Measuring the stellar mass with the photometric data doesn’t do one bit of good for this problem if you don’t have the kinematic axis to plot against. Consequently, it doesn’t matter how big such a sample is.

You have zero data.

Other measurements often provide a proxy measurement that seems like it ought to be close enough to use. If not the flat rotation speed, maybe you have a line width or a maximum speed or V2.2 or the hybrid S0.5 or some other metric. That’s fine, so long as you recognize you’re plotting something different so should expect to get something different – not the BTFR. Again, we’ve shown that the flat rotation speed is the measure that minimizes the scatter; if you utilize some other measure you’re gonna get more scatter. That may be useful for some purposes, but it only tells you about what you measured. It doesn’t tell you anything about the scatter in the BTFR constructed with the flat rotation speed if you didn’t measure the flat rotation speed.

Another possibility is that there exist galaxies that fall off the BTFR that we haven’t observed yet. It is a big universe, after all. This is a known unknown unknown: we know that we don’t know if there are non-conforming galaxies. If the relation is indeed absolute, then we never can find any, but never can we know that they don’t exist, only that we haven’t yet found any credible examples.

I’ve addressed the possibility of nonconforming galaxies elsewhere, so all I’ll say here is that I have spent my entire career seeking out the extremes in galaxy properties. Many times I have specifically sought out galaxies that should deviate from the BTFR for some clear reason, only to be surprised when they fall bang on the BTFR. Over and over and over again. It makes me wonder how Vera Rubin felt when her observations kept turning up flat rotation curves. Shouldn’t happen, but it does – over and over and over again. So far, I haven’t found any credible deviations from the BTFR, nor have I seen credible cases provided by others – just repeated failures of quality control.

Finally, an underlying issue is often – not always, but often – an obsession with salvaging the dark matter paradigm. That’s hard to do if you acknowledge that the observed BTFR – its slope, normalization, lack of scale length residuals, negligible intrinsic scatter; indeed, the very quantities that define it, were anticipated and explicitly predicted by MOND and only MOND. It is easy to confirm the dark matter paradigm if you never acknowledge this to be a problem. Often, people redefine the terms of the issue in some manner that is more tractable from the perspective of dark matter. From that perspective, neither the “cold” baryonic mass nor the flat rotation speed have any special meaning, so why even consider them? That is the road to MONDness.