Keep every explanation and worked example on the page. Nothing moves or plays by itself, and the larger instruments load only when you ask.
Brownian motion · Laboratory preview
From wanderingto a measurable spread.
How far from its starting point might a suspended particle be? Change the time, viscosity or radius, then ask about a whole interval, not just a single position.
An executable model
The spreading laboratory
Choose an interval and ask how likely a tracer is to finish inside it. Changing a field only edits a request. The plot and table change together after you apply valid settings.
The curve shows four RMS distances on either side of zero. The interval probability uses the full unbounded model, including any part outside this picture.Read the curve as numbers, at 17 points
The accepted density at the plotted points, from the same calculation
Accepted settings 293.15 K · 1 mPa·s · radius 0.5 μm · elapsed 1 s Constants: modern SI (2019). Viscosity is a declared input, not inferred from temperature. D is calculated from these physical inputs.
One accepted calculation, in explicit units
Diffusion coefficient
0.42944 μm²/s
RMS displacement
0.92676 μm
Mean square displacement
0.85888 μm²
Selected closed interval
[−1, 1] μm
Probability inside that interval
71.943%
Density is measured per unit length; probability is a dimensionless area. A taller curve is not a larger total probability.
A tiny sphere in water wanders at random, so after a given time it could be anywhere nearby, most likely close to where it began. The spread follows a bell curve, grows as the square root of the time, and is smaller for bigger spheres and thicker liquids.
Section 4 solves the diffusion equation for particles that all start at one point: the fraction found between x and x + dx after a time t is the Gaussian law of errors, exp(−x2/4Dt) dx/√(4πDt), and the root-mean-square displacement along one axis is λx = √(2Dt). Section 3 had found D for spheres of radius P in a liquid of viscosity k, D = (RT/N)/(6πkP), so Section 5 joins the two: λx = √t √((RT/N)/(3πkP)). The instrument evaluates this with today's constants for a sphere of radius 0.5 μm in water of viscosity 1.0 mPa·s at 293.15 K: D = 0.429 μm2/s and, after 1 s, λx = 0.927 μm, with a 71.9 percent chance of lying within 1 μm of the start. Measured in a plane or in space, the distance from the start has root-mean-square values √(4Dt) = 1.31 μm and √(6Dt) = 1.61 μm. At t = 0 the distribution is a single point, which the instrument reports as that limit rather than as a curve. An optional grid solves the same equation step by step in a box with walls, and refuses a time step too long for the scheme to stay stable.
First the diffusion coefficient. Section 3 balances the osmotic push of the suspended spheres against the drag of the liquid, which Stokes gave as 6πηa times the speed for a sphere of radius a in a liquid of viscosity η. In today's symbols the balance gives D = kBT/(6πηa), with kB Boltzmann's constant and T the temperature. Here kBT = 1.381 × 10−23 × 293.15 = 4.047 × 10−21 J, and the drag factor is 6πηa = 6 × 3.1416 × 0.001 × 0.5 × 10−6 = 9.425 × 10−9 kg/s. Their ratio is D = 4.047 × 10−21/(9.425 × 10−9) = 4.294 × 10−13 m2/s, or 0.4294 μm2/s. Now the spread. Particles that all start at x = 0 are found, after a time t, spread as a bell curve with mean square 2Dt: 2 × 0.4294 × 1 = 0.859 μm2 after one second, so the root-mean-square displacement is √0.859 = 0.927 μm. The chance of lying within 1 μm of the start is the area under the bell curve between −1 μm and +1 μm. One root-mean-square distance either side holds 68.3 percent; 1 μm is 1/0.927 = 1.08 of them, and the area out to 1.08 is erf(1.08/√2) = 0.719, or 71.9 percent. Double the time and the mean square doubles, so the spread grows by √2, to 1.31 μm after 2 s; after 60 s it is √60 = 7.75 times as far, 7.18 μm. Make the sphere four times larger and D falls to a quarter, so the spread halves. In a plane, add the two independent axes: the mean square distance is 2Dt + 2Dt = 4Dt = 1.72 μm2, a root mean square of 1.31 μm, while the most likely distance is √(2Dt) = 0.927 μm and the mean distance √(πDt) = 1.16 μm. In space the mean square is 6Dt, a root mean square of 1.61 μm, and the mean distance is √(16Dt/π) = 1.48 μm. Einstein's own example takes N = 6 × 1023, water at 17 °C with a viscosity of 1.35 × 10−2 in his units (1.35 mPa·s), and spheres 0.001 mm across: λx = 8 × 10−5 cm, or 0.8 μm in a second, and about 6 μm in a minute.
Einstein wrote k for the viscosity and P for the radius. He kept R and N apart because N was the unknown the paper hoped to measure: Section 5 ends by solving for it, N = (t/λx2)(RT/(3πkP)), and by hoping that a researcher will soon decide the question. Perrin's measurements of 1908 and 1909 confirmed the square-root law and gave N near 7 × 1023. The three-dimensional root mean square is Einstein's too: he remarks that the total displacement has root-mean-square value λx√3, which is √(6Dt). The two-dimensional radial laws and the grid are later aids, not in the paper; the grid's refusal above DΔt/Δx2 = 1/2 is a property of the explicit numerical method, not of the physics. Stokes's law for the drag on a sphere dates from 1851.
One setup, three observation times
Calculated from the same accepted diffusivity
Elapsed time
RMS displacement
1 s
0.92676 μm
10 s
2.9307 μm
60 s
7.1786 μm
The same spread, as a 2D or 3D radius
A radius is not a signed coordinate: it is never negative, and the growing circumference (2D) or surface area (3D) of positions at a given distance changes which average is largest.
Radial moments at the accepted diffusivity and elapsed time
Quantity
2D radius
3D radius
Mean radius ⟨r⟩
1.1615 μm
1.4789 μm
RMS radius √⟨r²⟩
1.3106 μm
1.6052 μm
Most likely radius
0.92676 μm
not modeled
The mean 2D radius ⟨r⟩ = √(πDt) is smaller than the RMS radius √⟨r²⟩ = √(4Dt): the extra factor of r in the 2D density weights larger radii more heavily than a signed 1D coordinate does, so the two averages of the same spread disagree.
The RMS 2D radius √⟨r²⟩ = √(4Dt) is the square root of the mean squared distance from the start, not the mean distance itself. Squaring before averaging always weights the tail more than averaging the radius directly.
The most likely 2D radius, where the density p(r, t) = (r/2Dt) exp(−r²/4Dt) peaks, is √(2Dt): the same number as the 1D RMS displacement. The growing circumference of available positions at radius r (proportional to r itself) exactly cancels the falling Gaussian density near the start, moving the peak away from the origin.
The mean 3D radius ⟨r⟩ = 4√(Dt/π) counts positions on a growing sphere (area proportional to r²), pulling the average distance from the start out further than the 2D circumference case.
The RMS 3D radius √⟨r²⟩ = √(6Dt) reflects three independent coordinate directions, each contributing its own 2Dt to the mean squared displacement (compare the 1D case's single 2Dt).
What this model leaves out
Dilute spherical tracers in a homogeneous Newtonian liquid; no drift or particle interactions.
Low Reynolds number and times long compared with momentum relaxation are assumed, not checked.
The analytic curve describes an unbounded line; the numerical grid has reflecting walls.
These are consequences of a model, not measurements or evidence that the model describes nature.
No FrankenSim artifact is used here. A verified historical transcription and the full critical edition are still in preparation.
Show the calculation owner and source identity
Density, interval probability and displacement come from src/physics/reference/diffusion/distributions.ts. The grid comes from src/physics/reference/diffusion/ftcs.ts. One worker operation assembles all views.
This is the function that produced the current snapshot.
In words
On an unbounded line the probability density is a Gaussian whose width grows as the square root of elapsed time.
Mathematics
Implementation
exportfunctiongaussianPropagator(x:number,t:number,D:number):Evaluation{if(!Number.isFinite(x)||!validDt(D,t))returnoutside("probabilityDensity","gaussianPropagator","finite x, D >= 0, t >= 0","Use finite position and nonnegative diffusivity and time.",);if(D===0||t===0)returnpointMass("probabilityDensity","gaussianPropagator");constsigma=scale(D,t);constz=x/sigma;if(!(sigma>0)||!Number.isFinite(sigma))returnoutside("probabilityDensity","gaussianPropagator","representable spread","The spread is outside the numerical range.","numerical",);returnnumber("probabilityDensity","gaussianPropagator",Math.exp(-Math.log(sigma)-0.5*Math.log(2*Math.PI)-0.5*z*z),undefined,undefined,true,);}
Audited TypeScript reference evaluator: the owner on this device, or the host fallback for a FrankenSim capability.
This function computes the listed outputs when it runs.
In words
The probability of landing between two points is the integral of that Gaussian.
Mathematics
Implementation
/** Closed intervals include the atom at t=0, including the interval [0,0]. */exportfunctionintervalProbability(x1:number,x2:number,t:number,D:number):Evaluation{constowner="intervalProbability";if(![x1,x2].every(Number.isFinite)||x1>x2||!validDt(D,t))returnoutside("intervalProbability",owner,"x1 <= x2, D >= 0, t >= 0","Use ordered finite interval endpoints and nonnegative diffusivity and time.",);if(D===0||t===0)returnnumber("intervalProbability",owner,x1<=0&&x2>=0?1:0);if(x1===x2)returnnumber("intervalProbability",owner,0);consts=2*Math.sqrt(D)*Math.sqrt(t);if(!(s>0)||!Number.isFinite(s))returnoutside("intervalProbability",owner,"representable spread","The spread is outside the numerical range.","numerical",);constlo=x1/s;consthi=x2/s;constwidth=(x2-x1)/s;constmid=lo+width/2;letp:number;// Very narrow intervals need direct integration, even when both tails are small.if(width>0&&width*(1+Math.abs(mid))<0.01){constnodes=[0.1834346424956498,0.525532409916329,0.7966664774136267,0.9602898564975363];constweights=[0.362683783378362,0.3137066458778873,0.2223810344533745,0.1012285362903763];letsum=0;for(leti=0;i<4;i++){constoffset=(width/2)*nodes[i]!;sum+=weights[i]!*(Math.exp(-((mid-offset)**2))+Math.exp(-((mid+offset)**2)));}p=(width/(2*Math.sqrt(Math.PI)))*sum;}elsep=lo>=0?0.5*(tail(lo)-tail(hi)):hi<=0?0.5*(tail(-hi)-tail(-lo)):0.5*(central(hi)-central(lo));if(p<0||p>1)returnoutside("intervalProbability",owner,"probability range","Numerical evaluation did not produce an admissible probability.","numerical",);returnnumber("intervalProbability",owner,p,undefined,undefined,true);}
Audited TypeScript reference evaluator: the owner on this device, or the host fallback for a FrankenSim capability.
This function computes the listed outputs when it runs.
In words
Advance a one-dimensional density with an explicit three-point Laplacian. Refuse when the stability ratio exceeds one half.
Mathematics
Implementation
exportfunctionftcs1d(p:FtcsParameters,budget=REFERENCE_FTCS_BUDGET):Computation<FtcsFrames>{const{n,frames,stepsPerFrame,D,dx,dt,profile}=p;if(!integer(n,3)||!integer(frames,1)||!integer(stepsPerFrame,1))returninvalid(["n","frames","stepsPerFrame"],"Use at least three cells and positive whole-number frame and step counts.",);if(![D,dx,dt].every(Number.isFinite))return{kind:"refused",refusal:makeRefusal("nonfinite-input",{parameterIds:["D","dx","dt"]}),};if(D<0||dx<=0||dt<=0||![0,1,2].includes(profile))returninvalid(["D","dx","dt","profile"],"Use nonnegative D, positive dx and dt, and profile 0, 1, or 2.",);// This exact operation order is part of the planned upstream conformance contract.constr=D===0?0:(D*dt)/(dx*dx);if(!Number.isFinite(r)||(D>0&&r===0))returnnumerical("The stability ratio is outside binary64 range; the grid was not advanced.");if(r>0.5){constdtMax=(dx*dx)/(2*D);constdxMin=Math.sqrt(2*D*dt);constdMax=(dx*dx)/(2*dt);if(![dtMax,dxMin,dMax].every((v)=>Number.isFinite(v)&&v>0))returnnumerical("The dimensional stability repairs are outside binary64 range.");// Mathematical bounds can round to a value with r=0.5000000000000001.// Preserve dtMax in the explanation; make each offered action executable.constdtRepair=stableRepair(dtMax,false,(v)=>(D*v)/(dx*dx));constdxRepair=stableRepair(dxMin,true,(v)=>(D*dt)/(v*v));constdRepair=stableRepair(dMax,false,(v)=>(v*dt)/(dx*dx));if(dtRepair===null||dxRepair===null||dRepair===null)returnnumerical("No representable repair was found near the dimensional stability boundary.",);return{kind:"refused",refusal:makeRefusal("ftcs-unstable",{parameterIds:["dt","dx","D"]},{details:{ratio:r,limit:0.5,dtMax},rankedRepairs:[{label:"Reduce the time step to the explicit scheme's limit.",action:{parameterId:"dt",value:dtRepair},},{label:"Use a coarser spatial grid.",action:{parameterId:"dx",value:dxRepair},},{label:"Choose a smaller diffusivity; this changes the physical setup.",action:{parameterId:"D",value:dRepair},},],},),};}conststepCount=(frames-1)*stepsPerFrame;constlimited=budgetCheck(n*stepCount,8*n*(frames+2),budget);if(limited)returnlimited;constelapsedTime=stepCount*dt;if(!Number.isFinite(elapsedTime)||(stepCount>0&&elapsedTime===0))returnnumerical("Elapsed model time is outside binary64 range.");constfield=newFloat64Array(n);if(profile===0)field[Math.floor(n/2)]=1/dx;elseif(profile===1)field.fill(1,0,Math.floor(n/2));else{field[Math.floor(n/4)]=0.5/dx;field[Math.floor((3*n)/4)]=0.5/dx;}if(!field.every(Number.isFinite))returnnumerical("The initial cell density is outside binary64 range.");constderivative=newFloat64Array(n);constvalues=newFloat64Array(n*frames);values.set(field);for(letframe=1;frame<frames;frame++){if(!advanceInPlace(field,derivative,r,stepsPerFrame))returnnumerical("A cell became nonfinite or negative; no partial frames are returned.");values.set(field,frame*n);}return{kind:"accepted",data:Object.freeze({values,shape:Object.freeze([frames,n]asconst),stepCount,elapsedTime,stabilityRatio:r,boundary:"zero-flux",ownerId:"diffusion.ftcs1d",}),};}
Compare two setups side by side in your reading. Each laboratory has its own settings, worker and accepted results.
The relationship behind the instrument
D=6πηakBT,⟨x2⟩=2Dt,λx=2Dt
Temperature is T, viscosity is η, particle radius is a, and the diffusion coefficient is D. The constant kB is supplied by the explicitly named modern SI set. The RMS displacement λx is the square root of the mean squared displacement along one coordinate.
These are not a particle’s total travelled distance or instantaneous speed. For the unbounded model, the curve gives a probability density; an interval gets its probability from the area under that curve.
p(x,t)=4πDt1exp(−4Dtx2)(D>0,t>0)
At the starting time there is a point distribution, not a finite density curve. The laboratory keeps that limiting state distinct.
What the numerical comparison asks
The finite grid approximates diffusion with an explicit stepping scheme. Its diffusion number must not exceed one half. A refused timestep is a limitation of this algorithm, not a prohibition on diffusion.
r=(Δx)2DΔt≤21
The numerical box reflects probability at its walls. Once the spread reaches them, comparing it with an unbounded Gaussian is also comparing two boundary models.