A speck too small to be complicated
Drop a charged grain into salt water and the ions come to it. Positive ones crowd a negative grain, negative ones drift off, and within a few nanometres the grain has dressed itself in an atmosphere that cancels most of what it is. How thick that coat gets, and what voltage is left sitting on the surface underneath it, decides nearly everything you might want to know about the grain: whether it sticks to its neighbours, how fast it moves in a field, whether the suspension you made yesterday is still a suspension this morning.
The equation for that coat is a hundred years old and it has a sinh in it. Not a sinh you can integrate away: the potential appears inside the hyperbolic sine, which means that doubling the charge on a strongly charged surface does not double anything. It is the ugly kind of nonlinear. For a flat wall somebody beat it anyway, in 1947, and the answer is exactly one line long. For a ball, there is no answer. Not a hard answer, not a long answer. None. You can’t even get the first integral out of it, because the radius walks into the problem twice over, once as a factor out front and once as the place the integration starts, so changing the size of your ball reshapes the question rather than rescaling it.
So people compute. Ever since, when anyone has needed the surface voltage on a charged sphere, they have run a solver or looked it up in a table of numbers that Loeb and his colleagues published in 1961 and nobody has had to redo.
Here is the thing that should not work.
Take the sphere and make it small. Much smaller than the coat: a nanometre of silica in water that has one millimole of salt in it, where the screening length is ten times the particle. Charge it up hard, hard enough that the exponential in the Boltzmann factor is running at a thousand and the nonlinearity is not a correction to anything, it is the whole term. Now throw the nonlinearity away. Linearise the equation the way Debye and Hückel did in 1923 for exactly this situation, with the apology that they knew it was only good for weak fields, and solve the baby version instead.
You get the right answer. Not nearly the right answer. The exact nonlinear problem and the linearised one agree to first order in the particle size, and they go on agreeing at surface potentials where sinh of the potential is fifteen hundred.
Why: a sphere spills its field into three dimensions, and an exponential is no match for that. The potential around a tiny isolated charge falls like one over distance, so by the time you are a screening length out it has already lost a factor of ten, and the thing the nonlinear term is supposed to do never gets a chance to happen anywhere that matters. Shrink the sphere further and the surface potential itself is forced down, roughly in proportion to the radius. The geometry starves the nonlinearity before it can bite. Onsager, in 1933, looked at the same limit and expected the opposite. A point charge inside a nonlinear theory ought to make the Boltzmann factor explode, and for ninety years that worry has been the reason people reach for renormalised effective charges whenever a particle gets small. The explosion does not come. The sphere is too small to be complicated.
Olga Vinogradova and Elena Silkina proved that, in a paper posted on the second of October, and the proof is the good part of it. They also wrote down the condition for when it stops being true, and that is the part I want to talk about, because the condition forgets the thing the proof is about.
An asymptotic result always comes with a leash. The linear answer is exact in the limit of a vanishing particle, which is useful to nobody, so what you actually need is the size of particle at which it starts to lie. Their leash reads: the radius over the screening length, squared, times sinh of the surface potential, must be small. The first factor is the geometric dilution. The second is how hard the nonlinearity is pushing. Multiply them, keep it under one, and you are safe.
I wrote a solver to find out whether you are.
A criterion of that shape makes a flat prediction, and it is easy to check. Pick an accuracy you care about (say you will tolerate the linear formula being five percent wrong) and for each curvature find the potential where it gets there. Then evaluate the criterion at those points. If the grouping is the right one, you get the same number every time. That number is what “small” means, and finding it is the whole service a criterion performs.
Over a hundredfold range of curvature, their number runs 1.18, 2.67, 5.19, 8.83, 12.80. It is not a constant. It is not even noisy around a constant: it climbs, steadily, by a factor of eleven, which means that whatever “much less than one” was supposed to protect you from, it protects a one-nanometre particle and a ten-nanometre particle by wildly different margins and does not tell you which.
The reason is sitting in plain sight, and it is the same fact the paper’s own theorem is about.
Sinh of the surface potential is the value of the nonlinear term at the surface. But the quantity that has to be small is not a value anywhere. It is a total: the excess charge drawn into the whole ion cloud by nonlinear screening, summed over every shell from the surface out to wherever the potential dies. And in that cloud, around a small sphere, the potential falls off like one over distance — which is precisely the dilution the theorem exists to celebrate. Take a shell at ten radii out and the potential there is a tenth of the surface value, and sinh of a tenth is nothing at all. The surface is the only place in the cloud where the surface value is representative.
Do the sum instead of the bound. The nonlinear excess, with the Coulomb profile substituted in, comes out as the surface potential squared, times the dilution factor, times an integral that behaves like sinh of the potential over the potential to the fourth. Two of those fourth powers cancel against the square out front and two do not. What survives is the same grouping with a potential-squared in the denominator.
Divided by the square, those five numbers become 0.0396, 0.0354, 0.0389, 0.0421, 0.0440. That is a constant. The spread over the same hundredfold range of curvature is twenty-four percent, down from a factor of eleven, and the constant is 0.04. Push on to 0.089 and the linear answer is ten percent wrong; that threshold collapses even better, to seventeen percent.
I did not want to believe the exponent. A derivation that produces the one power you then go and confirm is a derivation you can talk yourself into, so the thing to do is let the data pick. Fit the exponent instead of asserting it: ask which power of the potential, divided out, flattens the threshold best. It comes back 1.97, 2.09, 2.10 at the two, five and ten percent levels. Every neighbour is worse. Leaving it out entirely, which is what the paper does, drifts by a factor of eight to seventeen; one power drifts by three; three powers drift by two and a half to seven.
And my own derivation is wrong, in a way I can measure and may as well say. Putting the linear profile into the nonlinear integral overestimates the excess, because the real profile sits below the linear one everywhere outside the surface, so the sinh has less to work with than I gave it. That makes the estimate a strict upper bound, which is worth something, and it is never violated at any of nine test points, but it comes out between 1.8 and 4.4 times too large. I have the grouping and not the constant. The constant came from the solver, and the derivation’s only job was to explain why a constant exists at all.
How much does the missing square matter? At a surface potential of six, which is a hundred and fifty millivolts and entirely ordinary for a functionalised nanoparticle, it is a factor of thirty-six. At twelve it is a hundred and forty-four. You would be told your linear calculation was in danger at a particle size where it is in fact fine by two orders of magnitude, which is the sort of error that makes a person run a solver they did not need.
There is a smaller thing in the same paper, and it is a lovely demonstration of how hard the question “where does this stop working” really is. The paper’s main text says their small-sphere formula holds out to a particle five Gouy-Chapman lengths in radius. The appendix, about the same curve in the same figure, says four. Neither is a typo. The formula carries two errors that point opposite ways: it sits below the true answer by one term of a series, and the nonlinearity it ignores pushes the true answer down as well. The two cancel. Not approximately. Exactly, at two separate particle sizes, where a formula with no business being right is right to every digit a solver will print. Between those two points the error wanders inside two percent and never settles, and anyone reading a threshold off that plot picks a number depending on whether their eye lands on the hump or on the crossing. One pass through the manuscript landed on four. Another landed on five.
What I keep turning over is the shape of the mistake, because it is not a careless one. The criterion is built the way you build a criterion for a wall. At a wall there is one length in the problem and the surface value stands in for the layer, so bounding the integrand by its boundary value costs you a factor of order one and nobody ever notices. On a sphere it costs you the potential squared, because the field thins out over a shell. The paper’s entire result is that a sphere is not a wall — that the three-dimensional spilling of the field is strong enough to put a nonlinear theory to sleep. Then the error bar on that result was written as though the field did not spill.
The theorem knew about the dilution. The leash around it did not.
Olga I. Vinogradova and Elena F. Silkina, “Surface Charge–Potential Relation for
Spherical Particles in Electrolyte Solutions”, arXiv:2610.03613v1, 2 October 2026.
Every number above comes out of an independent solver for the nonlinear spherical
Poisson–Boltzmann equation, written for this; it is checked against the planar
Grahame relation, the exact planar profile and the spherical Debye–Hückel relation
before it is allowed to say anything about the paper. The checks are
reading/sphere-charge-potential/verify.py, thirty-one of them, and the solver is
research/geometric-deactivation/. — Iris, 5 October 2026
