Graded Bounds on the Giant-Dust Tail Beyond Instrument Cutoffs

Climate models reproduce total dust optical depth reasonably well, yet they miss most of the large particles. Adebiyi et al. report that models "miss more than three-quarters of particles larger than 5 µm", even though coarse and super-coarse dust make up about 85% of atmospheric dust mass and about half of dust optical depth (A review of coarse mineral dust in the Earth system). The observations that should correct this are thin at exactly the sizes that matter. Sun photometers and satellites are tuned to particles below a few micrometres. Aircraft inlets lose large particles. Only a few campaigns with wing-mounted probes have seen particles up to 100 µm, some of them 3,500 km from Africa. The quantity that matters for climate is therefore an extrapolation. We need to know how much mass, extinction, and thermal absorption sits in the part of the size distribution beyond what an instrument can see, and it has to be estimated from a handful of direct measurements.

We propose to answer that question with a stack of nested intervals, one per level of assumption, instead of a single number. Each level states how strongly we trust a given assumption about the tail's shape: "the fitted coarse mode continues smoothly," "the tail density is convex," "the tail only decreases." For each level we compute the tightest worst-case interval on the hidden quantity. Direct giant-particle measurements then revise how much each assumption deserves to be trusted in each transport regime. The output takes the form "under what we normally believe, the missing longwave absorption lies in [a, b]; only if we also allow exceptional tail shapes can it reach c." These statements can be checked against held-out campaign data. The machinery for grading assumptions comes from ordinal ranking functions, a qualitative uncertainty calculus in which integer "surprise" ranks replace probabilities. The machinery for the intervals comes from nonparametric worst-case tail bounds.

Core Claim

Size distributions are one-dimensional in log-diameter, and the climate-relevant quantities are linear in the distribution. This makes the beyond-cutoff tail an unusually clean target for shape-constrained worst-case bounding. Lam and Mottet showed that assuming only a convex tail density lets you bound rare-event probabilities from three quantities measured in the data-rich body: the mass at the threshold, the density there, and its slope. They reported near-100% coverage on synthetic data where Generalized Pareto fits failed, and a worst-case estimate of 1.99 against a parametric estimate of 0.23 on Danish fire losses (Tail Analysis Without Parametric Models: A Worst-Case Perspective). We extend this method from interval probabilities to general linear functionals with physical kernels. We then embed it in a hierarchy of assumption sets whose credibility is graded by Spohn-style ranks (A general non-probabilistic theory of inductive reasoning).

The claim is empirical and falsifiable. After revision on training campaigns, the lowest-rank interval should reach nominal coverage on held-out long-range-transport campaigns. It should also be substantially narrower than the single most conservative bound. Log-normal continuation, which is effectively what modal climate models assume (Toward a minimal representation of aerosols in climate models: description and evaluation in the Community Atmosphere Model CAM5), and parametric tail fits should both under-cover.

Research Program

Let $f(D)$ be the volume size distribution per unit $\ln D$, observed up to a cutoff diameter $D_c$. Let $K_g(D)$ be a kernel for the target quantity $g$: volume for mass, the extinction cross-section at a visible wavelength, or the absorption cross-section at a thermal-infrared wavelength, computed from a fixed refractive index and particle shape. For assumption level $k$, the upper bound is

$$U_k(g) = \max_{f \in \mathcal{A}_k} \int_{D_c}^{\infty} K_g(D)\, f(D)\, d\ln D,$$

and $L_k(g)$ is the corresponding minimum. Here $\mathcal{A}_k$ is the set of tails that match the observed body at $D_c$ and satisfy the level-$k$ shape assumption.

The levels are nested, so $\mathcal{A}_0 \subset \mathcal{A}_1 \subset \mathcal{A}_2 \subset \mathcal{A}_3$:

  • $\mathcal{A}_0$ is a log-normal continuation of the fitted coarse mode, with uncertainty in the fitted parameters.
  • $\mathcal{A}_1$ is Lam–Mottet convexity, matching mass, height, and slope at $D_c$.
  • $\mathcal{A}_2$ is a monotone-decreasing tail with total volume capped by an independent mass constraint, such as filter or deposition-trap mass where a campaign reports it.
  • $\mathcal{A}_3$ imposes only the mass cap.

After discretizing $\ln D$, each program is a linear program, because convexity and monotonicity are linear constraints on the discretized density.

The ranks enter as follows. Each candidate tail shape $w$ receives a disbelief rank $\kappa_r(w)$ that depends on the transport regime $r$ (near-source, mid-Atlantic, long-range Caribbean). The initial rank is the smallest $k$ with $w \in \mathcal{A}_k$. A direct giant-particle observation $O$ is handled by ranking-function conditioning, $\kappa_r(w \mid O) = \kappa_r(w) - \kappa_r(O)$ for worlds consistent with $O$. The effect is concrete. If wing-probe data in the long-range regime lie outside every log-normal-consistent tail, then convex shapes that explain the data drop to rank 0 for that regime, and the log-normal worlds rise. The reported product is the family of graded intervals $[L_k, U_k]$ computed over worlds with revised rank at most $k$.

We borrow two methods from Schwind et al. The first is the idea of treating rank revision as an online learner. The second is their order-invariance property: their thresholded ranking learners are commutative in the training examples (Iterated Belief Change as Learning). Their method works on binary Hamming cubes. We replace that discrete setting with nested convex sets over a one-dimensional density, which addresses the continuous-space limitation they note. Friedman and Halpern's result that default-style inference ("normally the tail is convex") is well behaved in qualitative plausibility structures (Plausibility measures and default reasoning) justifies reading rank-0 intervals as default conclusions. Our finite, physically constrained hypothesis space is not "rich" in their sense, so we treat completeness as open and do not claim it.

The ranks also yield a testable quantitative prediction. Ranks are read as orders of magnitude of an infinitesimal probability, with rank $k$ corresponding to $\epsilon^k$. If the ranks are calibrated, then the frequency with which held-out truth falls outside the rank-$k$ interval should shrink roughly geometrically in $k$. We will check this directly rather than assume it.

Finally, we propagate the graded intervals on thermal absorption and visible extinction into a simple direct-radiative-effect estimate. The aim is to ask at which rank the sign of the coarse-dust contribution becomes undetermined. This is the question that motivates the whole exercise: Adebiyi et al. argue that missing coarse dust makes models overpredict cooling (A review of coarse mineral dust in the Earth system).

Baselines

The naive reference is log-normal continuation of the observed coarse mode beyond $D_c$, the implicit behaviour of modal schemes. We will fit it to the same body data and report its point estimate with a parameter-bootstrap interval.

The standard statistical competitor is a Generalized Pareto fit to the upper body of each size distribution. The threshold will be chosen by a stability plot, not an ad hoc cutoff like the AOD > 0.7 used in Deep learning for Aerosol Forecasting.

The strongest learned competitor is Extremal Random Forests (Extremal Random Forests). This method fits conditional tail parameters with a penalized shape estimate. It is the strongest existing tool that can exploit covariates the ranking method uses only coarsely, namely transport distance, altitude, and source region. We will give it identical training campaigns, identical covariates, and a matched hyperparameter search, and apply the same leave-one-campaign-out split.

The ablation that isolates the proposal's key ingredient is a single worst-case bound with no ranking or revision. We will run it at level $\mathcal{A}_1$ (pure Lam–Mottet convexity, extended to our kernels) and at $\mathcal{A}_2$.

No reported result was found on Lacuna for any of these methods on aerosol size-tail extrapolation. The closest reported numbers are Lam and Mottet's near-100% synthetic coverage and their 1.99 versus 0.23 Danish-fire contrast. We will count the proposal as beating a baseline under two conditions. First, its rank-0 interval must meet the 90% nominal coverage target on held-out campaigns. Second, it must either be at least 30% narrower in log-width than the single-level convex bound, or reach coverage the parametric and forest methods miss.

Why not just…?

Why not just use the Lam–Mottet convex bound alone? This is the most serious threat. If a single convex bound is already narrow enough to settle the sign of the coarse contribution and covers the held-out data, the ranking layer is decoration. The single-level ablation settles it directly. The case for ranks rests on a specific expectation: near-source data will make log-normal continuation adequate, while long-range data (wing probes finding 100 µm particles 3,500 km out) will violate it. If so, one global assumption is either too loose near sources or wrong far from them. If that regime dependence does not appear, we will report the plain bound as the answer.

Why not just fit a Bayesian mixture over tail-shape classes? A Bayesian model could place prior weights on the four classes and update them. With a few dozen giant-particle profiles, however, the posterior over class weights will mostly reflect the chosen prior. The Friedman–Halpern critique (Plausibility Measures: A User's Guide) applies here: numerical weights impose precision the evidence does not support. We will still run this mixture with two reasonable priors as an extra comparison. If it matches our coverage and width and is insensitive to the prior, it is a legitimate equal alternative, and we will say so.

Why not just fit an extreme-value distribution, or a flexible heavy-tailed generator? Parametric tail fits suffer from threshold instability and coverage failure under starvation (Tail Analysis Without Parametric Models: A Worst-Case Perspective). Phase-type decoders are light-tailed at infinity (Phase-Type Variational Autoencoders for Heavy-Tailed Data). Generators such as ExGAN enforce no physical constraints (ExGAN: Adversarial Generation of Extreme Samples). Extremal Random Forests is the fair representative of this family and is included as the strongest baseline.

Why not just add size bins to the transport model? Watson-Parris et al. found that switching from modal to sectional schemes did not fix size biases, which were driven by removal processes rather than resolution (In-situ constraints on the vertical distribution of global aerosol). Tegen et al. likewise attribute undersized coarse particles to overestimated settling and scavenging (The global aerosol–climate model ECHAM6.3–HAM2.3 – Part 1: Aerosol evaluation). Our method is also complementary to model fixes. It produces an observational constraint that any revised scheme can be scored against.

Proposed Experiments

Cutoff-extrapolation benchmark. We will use size distributions from the dust campaigns synthesized by Adebiyi et al.: Fennec (near-source), SAMUM (transit), AER-D (Atlantic outflow), and SALTRACE (Caribbean arrival). Where the review indicates wing-probe coverage extends to tens of micrometres, we will artificially truncate each profile at $D_c = 10$ µm, and separately at 5 µm, to mimic photometer and inlet limits. We will then predict the beyond-cutoff volume, visible extinction, and thermal absorption, and score against the withheld measurements. Evaluation will be leave-one-campaign-out, and the decisive split trains on near-source and transit campaigns and tests on SALTRACE.

Metrics. We will report coverage at each rank, interval log-width, and interval score. We will also report the fraction of profiles for which the sign of the coarse direct radiative effect is determined at rank 0.

Calibration check. Violation rates should decrease roughly geometrically with rank, as the infinitesimal reading of ranks predicts. If they do not, the graded interpretation fails even if the intervals are useful.

Order-invariance check. We will permute the order in which campaigns are presented during revision and confirm identical final ranks.

Real-world transfer. We will apply the regime-revised ranks to AERONET inversions, whose retrieved size distributions end at 15 µm. We will compare the implied beyond-cutoff dust optical depth against DustCOMM's size-resolved loading as a coarse sanity check, not as ground truth.

Stress test. We will repeat the benchmark on the highest-loading storm profiles, selected by a threshold-stability criterion, since that is where both size-tail and event-tail scarcity bite.