I would like to generate a random smooth convex body, like a pebble or a potato (but strictly convex, that's necessary). My attempts:

ConvexHullMesh[RandomReal[1, {10000, 3}], PlotTheme -> "SmoothShading"]

enter image description here

-- clearly bad, too cube-biased;

ConvexHullMesh[ Map[# + RandomReal[.3, 3] &, Flatten[Table[{Cos[u] Cos[v], Cos[u] Sin[v], Sin[u]}, {u, -(\[Pi]/2), \[Pi]/2, \[Pi]/100}, {v, -\[Pi], \[Pi], \[Pi]/200}], 1]], PlotTheme -> "SmoothShading"]

enter image description here

-- better, but now too spherical; besides, both not sufficiently smooth.

What would be a scientific approach? What actually is a random smooth convex body?

  • 2
    $\begingroup$ Have you seen this? $\endgroup$ Dec 28, 2016 at 16:59
  • $\begingroup$ Thanks, J. M. and @user9490, probably I can use this. Still have to figure out how to make it strictly convex - once there are any concave regions , if I just make a convex hull they will produce something semiconvex, which I want to avoid. $\endgroup$ Dec 28, 2016 at 17:01
  • $\begingroup$ maybe look at this mathworld.wolfram.com/Superellipsoid.html $\endgroup$
    – george2079
    Dec 28, 2016 at 18:22
  • $\begingroup$ How fast a solution should be? Do you plan to run the pebble making function a lot of times (say, in a loop) or relatively few pebbles would be sufficient for your plans? $\endgroup$ Dec 28, 2016 at 18:43
  • $\begingroup$ @george2079 This is too regular for me :D $\endgroup$ Dec 28, 2016 at 21:17

4 Answers 4


As mentioned in the comments this answer produces results too slowly and the pebbles are not that smooth, but since I did go through with the idea (which I find interesting) I am posting the outcomes.

The idea for making a random pebble is to generate random points that would determine pebble's shape and then use a 3D quantile envelope to derive pebble's surface.

Here we generate the random points in such a way that they determine the pebble shape:

data1 =
   MultinormalDistribution[{1, 2, 
     3}, {{3, 0, 0}, {0, 1, 0}, {0, 0, 2}}], 1*10^4];

data2 =
   MultinormalDistribution[{1, 2, 2/5}, 
    0.8 {{1, 0, -1/2}, {0, 1, 0}, {-1/2, 0, 2}}], 1*10^4];

data = Join[data1, data2];

(* {20000, 3} *)

Making random variate mixtures with different distributions and parameters (means/centers, variations/correlation matrices) would bring different pebble shapes.

Some additional rotation (optional):

rmat = RotationMatrix[Pi/3., {{1, 1, 1}, {1, -1, 1}}];
data = data.rmat;

Plot the generated random points:

Block[{qs = 12}, 
  qs = Map[Quantile[#, Range[0, 1, 1/(qs - 1)]] &, Transpose[data]];
  PlotStyle -> {PointSize[0.002]}, PlotRange -> All, 
  PlotTheme -> "Detailed", 
   -> {{{0, 0, -1}, Most[qs]}, {{0, 1, 0}, qs[[{1, 3}]]}, {{-1, 0, 0}, 

enter image description here

Find the directional quantile envelope:

 qreg = QuantileEnvelopeRegion[data, 0.95, 78];

(* {4.57647, Null} *)

Discretize the obtained region in order to plot it:

 bdreg = BoundaryDiscretizeRegion[qreg];

(* {168.562, Null} *)

The command above will take less time if smaller number of directions in QuantileEnvelopeRegion are used. (The third argument.) The obtained pebble might have some very flat, angular sides.

Plot together with a sample of the points:

Block[{testData = RandomSample[data, 4000]},
    PlotStyle -> {Gray, PointSize[0.006]}], bdreg}]]

enter image description here

Just the pebble by itself:


enter image description here

Probably some further refinements or manipulations of the obtained discretized region can be made in order to derive smoother surfaces.


Another approach is to refine a coarse mesh with multiple iterations of Loop subdivision. One such implementation is here.

init = ConvexHullMesh[RandomReal[{-1, 1}, {8, 3}]];

BoundaryMeshRegion[init, MeshCellStyle -> {1 -> Black}]

Nest[LoopSubdivide, init, 6]

  • $\begingroup$ Thank you, it is definitely very efficient! Except that I don't really understand what the code is doing. For example, if it tries to approximate the existing polyhedron with a smooth shape, then it is not quite what is needed. A realistic approach would rather shrink the initial shape, similarly to the fact that, to obtain a pebble from a piece of rock, several pieces must break out. But maybe after all the final result is similar, I don't know. $\endgroup$ Jun 25, 2021 at 17:43

Although there is an accepted answer, let me describe another approach which I stumbled upon on mathoverflow, in form of the question about "derived" polyhedra there.

Start with a random collection of points in 3d; form convex hull; take barycenters of faces; iterate.

Seems like after about ten iterations a reasonably realistic pebble is obtained, even if one starts with an almost regular system of points initially.

subd[mesh_] := With[
 {vertices = MeshCoordinates[mesh], triangles = Map[First, MeshCells[mesh, 2]]}, 
 ConvexHullMesh[Map[Mean[vertices[[#]]] &, triangles]]

trim[mesh_, M_] := Nest[subd, mesh, M]

An example: start with

initmesh = ConvexHullMesh[
  {{1, 0, 0}, {0, 1, 0}, {0, 0, 1}, {-.0001, -.000002, -.000000004}, {-.000000000005, 0, -.1}}

Then trim[initmesh, 10] produces

enter image description here

I like this version since it sort of imitates the process by which pebbles are actually formed.

Another (either good or bad, depending on what you need from it) feature of this method is that it seems to be extremely sensitive to initial conditions. In my experiments, changing one of the initial coordinates by about one trillionth resulted in quite different final result.


An alternative implementation of the idea in მამუკაჯიბლაძე's answer:

abrade = Nest[ConvexHullMesh[PropertyValue[{#, 2}, MeshCellCentroid]] &, #, #2] &;


abrade[ConvexHullMesh[{{1, 0, 0}, {0, 1, 0}, {0, 0, 1}, 
   {-.0001, -.000002, -.000000004}, {-.000000000005, 0, -.1}}], 10]

enter image description here

abrade[ConvexHullMesh[RandomReal[1, {10, 3}]], 10]

enter image description here


Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge you have read our privacy policy.

Not the answer you're looking for? Browse other questions tagged or ask your own question.