I am looking for an algorithm to distribute a bunch of points (could be anywhere from a few hundred to millions) within a sphere. In this case the sphere is centered at (0,0,0).
For random points a simple method is
repeat
x:=random*diameter-radius;
y:=random*diameter-radius;
z:=random*diameter-radius;
until ((x*x+y*y+z*z)<(radius*radius));
But I want to get the points evenly spaced within the sphere and without bunching at the poles.
Any good tricks/algorithms/formulas/code snippet to accomplish this?
You could do something like this:
Put the center of your sphere at a random position within an infinite volume of evenly-spaced points, like a tetrahedral or cubic lattice.
Enumerate points in order of increasing distance from the center until you have the right number.
Rescale the selected points around the center so that the distance to the furthest point is equal to the desired radius.
If you need evenly spaced points - just place them in grid nodes.
Sphere with radius R has volume
V=4/3*Pi*R^3
so for placing N points every cell of cubic grid (perhaps you might want to use hexagonal close packing) should have volume
v=4/3*Pi*R^3/N
and edge length
l = R * (4*Pi/(3*N))^1/3
Then generate points in coordinates (a*l, b*l, c*l) where a,b,c are integers limited by -R..+R (with appropriate sum of squares).
Proposed approach is quite rough estimation and perhaps some points from N needed ones might run outside of the sphere. In this case one have to diminish cell size or use more exact value - it might be calculated using 3D analog of Gauss circle formula ()
Thanks for the explanation by joriki. The above formula has been explained in proof of the formula mentioned. But the formula has some limitations. The cube length can be 2 (-1 to 1 in all directions). From the above concept, mentioned by joriki or proof of the formula mentioned we can generalize for the cube of any length (i.e 2a (-a to a in all directions)). Here 2a is the side length of the cube.
$a^6-(a^2-x^2 )(a^2-y^2)(a^2-z^2)=x^2 a^2 (a^2-\frac{z^2}{2}-\frac{y^2}{2}+\frac{y^2 z^2}{3a^2})+y^2 a^2 (a^2-\frac{z^2}{2}-\frac{x^2}{2}+\frac{x^2 z^2}{3a^2})+z^2 a^2 (a^2-\frac{x^2}{2}-\frac{y^2}{2}+\frac{y^2 x^2}{3a^2})$
$x'=xa \sqrt{a^2-\frac{z^2}{2}-\frac{y^2}{2}+\frac{y^2 z^2}{3a^2}}$,
$y'=ya \sqrt{a^2-\frac{z^2}{2}-\frac{x^2}{2}+\frac{x^2 z^2}{3a^2}}$,
$z'=za \sqrt{a^2-\frac{x^2}{2}-\frac{y^2}{2}+\frac{y^2 x^2}{3a^2}}$,
From the above equation, we can easily separate the (x',y',z') as explained in 1. The above equation will readily find the mapping of the generalized length of the cube to a sphere. I hope this will be helpful.
Form the above equation we can also find the evenly distribution of points within the sphere by varying the grid size of the cube. By discretizing the cube into different grids and simultaneously mapping into the sphere will give the evenly distribution of points within the sphere.
Kindly ignore some typos.
C++ code for evenly distributing points within the sphere:-
for(x=-a;x<=a;x=x+n)
for(y=-a;y<=a;y=y+n)
{
for(z=-a;z<=a;z=z+n)
{
xnew=x*a*(sqrt((a*a)-(y*y)/(2)-(z*z)/(2)+(y*y*z*z)/(3*a*a)));
ynew=y*a*(sqrt((a*a)-(z*z)/(2)-(x*x)/(2)+(x*x*z*z)/(3*a*a)));
znew=z*a*(sqrt((a*a)-(x*x)/(2)-(y*y)/(2)+(y*y*x*x)/(3*a*a)));
cout<<" "<<xnew<<" "<<ynew<<" "<<znew<<endl;
}
}
}
Note:- Here in code, n is the grid size of a cube.
I am currently trying to write a shader in unity that draws a triangular pattern around countries in a risk-styled game if both countries are not owned by the same player (visual aid to see your borders).
Right now, I'm having an issue with making the shader set the countries properly.
It always sets country 0 to the left, and country 1 to the right - country 0 and 1 are set programically.
The line, a border, can be between 0 and 359 degrees.
How I find the countries 0 and 1 is I draw 3 points to the left and right of the midpoint of the line, one .01f, one .1f and one 1f away from the midpoints in each direction, then spin them around the midpoint to the appropriate location.
After that I do an even-odd check to see if the points are inside or outside of each country, and compare the weight results (closest gets 3 points, mid gets 2, furthest gets 1, just in case someone builds a really screwed up country that flanks the other country).
In my test map, a close to equally sliced octagon, the borders showed up correctly (after I reversed the positions of country 0 and 1 in the event the angle was over 90 and less then or equal 180). Worked without a flaw, but in other maps it doesn't work very well.
Everything but the country allocation works well, so I'm curious if anyone knows of a better way to figure out which point is to the left or a spun line, or a better conceptual way to handle this.
That above is basically when I'm doing, red being left right being blue, then I'm just checking 3 different spots then weighing in the lefts and rights found with even/odding it into the appropriate countries (one at +/- .01, the other at +/- .1 and the third 1, in case of even/odd rounding issues with closeness).
I then flip them if I find that country A is to the right, as it is on the left according to the angles I had draw. (my shader renders left first and right second, hence why I do this).
which way is left/right on a line?
From last edit is this not your case. Why not use dot product?
So if the line goes in -x direction the result is negative and if in the +x direction then the result is positive. if the result is zero that means the line goes up or down only or it is juts a point. If you need specific direction instead of left/right then use appropriate a vector instead of x axis.
dot(a,b)=a.x*b.x+a.y*b.y in 2D
dot(a,b)=a.x*b.x+a.y*b.y+a.z*b.z in 3D
Image is relevant for cases where a vector is in unit size in that case the result of dot is perpendicular projection of b into a just like on image
on which side is some point?
I think this is what you need.
As you can see if you handle line (P0,P1) and point P you want to classify as triangle then its polygon winding determines also the side of the line. So for implicit axis directions:
CW(clockwise) polygon winding means right side of the line
CCW(counter-clockwise) polygon winding means left side of the line
How to get winding? ... simply compute normal vector and get its Z coordinate. Its polarity (sign) determines winding (CW/CCW or the other way around depends on the coordinate system).
normal vector is computed as cross product of the two vertices of triangle (P1-P0)x(P-P1)
No need to compute other axises just the z so:
normal.z = ((P1.x-P0.x)*(P.y-P1.y)) - ((P1.y-P0.y)*(P.x-P1.x))
Now just do if (normal.z<0) ... else ... it should never be zero unless you call it for point on the line or the line is a point ... look here at similar question: Determine rotation direction /toward/ variable point on a circle
I'm writing a 2D game, in which I would like to have crate-like objects. These objects would move around, like real crates do. I have a hypothetical idea of how I would like to achieve that:
Basically I'd store the boxes' corners' coordinates with their force and velocity unit vectors, and in every update I'd basically do the following steps:
1. Apply the forces(gravity, from collisions, etc..) accordingly.
2. Modify velocity vector based on the force.
3. Move every corner of the box, like so:
4. I repeat nr 3. for every corner, so I get the real movement of the cube.
My questions are: Is this approach heading in the right direction? Is this theory even correct? If not, what would be the correct way to move a box around based on vectors in a 2D environment?
Just to clarify: I'm only dragging corner "A" in the picture, but I want to repeat the dragging for every other corner, with their own vectors. By "dragging" I mean the algorithm I just stated.
Keeping each corner's coordinate and speed makes no sense as you would be storing lots of redundant information. Boxes are rigid objects, which means that there are constraints that must be satisfied at any time instant, namely the distance between any two given corners is fixed. This also translates to a constraint that links the velocities of all four corners and so they are not independent values. With rigid bodies the movement of any point is the sum of two independent movements - the linear movement of the centre of mass (CM) and the rotation around a fixed axis - often, but not always, chosen to be the one that goes through the CM. Hence you only need to store the position and the velocity of the crate's CM (which coincides with the geometric centre of the crate) as well as the angle of rotation and the rate of rotation around the CM.
As to the motion, the gravity field is a constant vector field and hence cannot induce rotation in symmetric objects like those rectangular crates. Instead it only produces accelerated vertical motion of the CM. This is also what happens due to all external forces - one has to take their vector sum and apply it to the CM. Only external forces whose direction does not go through the CM give torque and so cause rotation. Such forces are any external pushes/pulls or reaction forces that arise when crates collide with each other or hit the ground / a wall. Computing torque due to external forces is easy but computing reaction forces could be quite involving process because of the constrained dynamics that has to be employed. Once the torque has been computed, one has to divide it by the moment of inertia of the create in order to get the angular acceleration. Often it is more convenient to use another axis and not the one that goes through the CM - Steiner's theorem can be employed in this case in order to compute the moment of inertia around that other axis.
To summarise:
all forces, acting on the create, are first added together (as vectors) and the resultant force (divided by the mass of the create) determines the linear acceleration of the CM;
the torque of all forces is computed and then used to determine the angular acceleration around a given axis.
See here for some sample problems of rigid body motion and how the physics is actually worked out.
Given your algorithm, if by "velocity vector" you actually mean "the velocity of CM", then 1 would be correct - all corners move in the same direction (the linear motion of the CM). But 2 would not be always correct - the proper angle of rotation would depend on the time the torque was applied (e.g. the simulation timestep), and one has to take into account that the lever arm length changes in between as the crate rotates.
I'm trying to teach myself some machine learning, and have been using the MNIST database (http://yann.lecun.com/exdb/mnist/) do so. The author of that site wrote a paper in '98 on all different kinds of handwriting recognition techniques, available at http://yann.lecun.com/exdb/publis/pdf/lecun-98.pdf.
The 10th method mentioned is a "Tangent Distance Classifier". The idea being that if you place each image in a (NxM)-dimensional vector space, you can compute the distance between two images as the distance between the hyperplanes formed by each where the hyperplane is given by taking the point, and rotating the image, rescaling the image, translating the image, etc.
I can't figure out enough to fill in the missing details. I understand that most of these are indeed linear operators, so how does one use that fact to then create the hyperplane? And once we have a hyperplane, how do we take its distance with other hyperplanes?
I will give you some hints. You need some background knowledge in image processing. Please refer to 2,3 for details.
2 is a c implementation of tangent distance
3 is a paper that describes tangent distance in more details
Image Convolution
According to 3, the first step you need to do is to smooth the picture. Below we show the result of 3 different smooth operations (check section 4 of 3) (The left column shows the result images, the right column shows the original images and the convolution operators). This step is to map the discrete vector to continuous one so that it is differentiable. The author suggests to use a Gaussian function. If you need more background about image convolution, here is an example.
After this step is done, you have calculated the horizontal and vertical shift:
Calculating Scaling Tangent
Here I show you one of the tangent calculations implemented in 2 - the scaling tangent. From 3, we know the transformation is as below:
/* scaling */
for(k=0;k<height;k++)
for(j=0;j<width;j++) {
currentTangent[ind] = ((j+offsetW)*x1[ind] + (k+offsetH)*x2[ind])*factor;
ind++;
}
In the beginning of td.c in 2's implementation, we know the below definition:
factorW=((double)width*0.5);
offsetW=0.5-factorW;
factorW=1.0/factorW;
factorH=((double)height*0.5);
offsetH=0.5-factorH;
factorH=1.0/factorH;
factor=(factorH<factorW)?factorH:factorW; //min
The author is using images with size 16x16. So we know
factor=factorW=factorH=1/8,
and
offsetH=offsetW = 0.5-8 = -7.5
Also note we already computed
x1[ind] = ,
x2[ind] =
So that, we plug in those constants:
currentTangent[ind] = ((j-7.5)*x1[ind] + (k-7.5)*x2[ind])/8
= x1 * (j-7.5)/8 + x2 * (k-7.5)/8.
Since j(also k) is an integer between 0 and 15 inclusive (the width and the height of the image are 16 pixels), (j-7.5)/8 is just a fraction number between -0.9375 to 0.9375.
So I guess (j+offsetW)*factor is the displacement for each pixel, which is proportional to the horizontal distance from the pixel to the center of the image. Similarly you know the vertical displacement (k+offsetH)*factor.
Calculating Rotation Tangent
Rotation tangent is defined as below in 3:
/* rotation */
for(k=0;k<height;k++)
for(j=0;j<width;j++) {
currentTangent[ind] = ((k+offsetH)*x1[ind] - (j+offsetW)*x2[ind])*factor;
ind++;
}
Using the conclusion from previous, we know (k+offsetH)*factor corresponds to y. Similarly - (j+offsetW)*factor corresponds to -x. So you know that is exactly the formula used in 3.
You can find all other tangents described in 3 implemented at 2. I like the below image from 3, which clearly shows the displacements effect of different transformation tangents.
Calculating the tangent distance between images
Just follow the implementation in tangentDistance function:
// determine the tangents of the first image
calculateTangents(imageOne, tangents, numTangents, height, width, choice, background);
// find the orthonormal tangent subspace
numTangentsRemaining = normalizeTangents(tangents, numTangents, height, width);
// determine the distance to the closest point in the subspace
dist=calculateDistance(imageOne, imageTwo, (const double **) tangents, numTangentsRemaining, height, width);
I think the above should be enough to get you started and if anything is missing, please read 3 carefully and see corresponding implementations in 2. Good luck!
I have an interesting problem here I've been trying to solve for the last little while:
I have 3 circles on a 2D xy plane, each with the same known radius. I know the coordinates of each of the three centers (they are arbitrary and can be anywhere).
What is the largest triangle that can be drawn such that each vertex of the triangle sits on a separate circle, what are the coordinates of those verticies?
I've been looking at this problem for hours and asked a bunch of people but so far only one person has been able to suggest a plausible solution (though I have no way of proving it).
The solution that we have come up with involves first creating a triangle about the three circle centers. Next we look at each circle individually and calculate the equation of a line that passes through the circle's center and is perpendicular to the opposite edge. We then calculate two intersection points of the circle. This is then done for the next two circles with a result of 6 points. We iterate over the 8 possible 3 point triangles that these 6 points create (the restriction is that each point of the big triangle must be on a separate circle) and find the maximum size.
The results look reasonable (at least when drawn out on paper) and it passes the special case of when the centers of the circles all fall on a straight line (gives a known largest triangle). Unfortunate i have no way of proving this is correct or not.
I'm wondering if anyone has encountered a problem similar to this and if so, how did you solve it?
Note: I understand that this is mostly a math question and not programming, however it is going to be implemented in code and it must be optimized to run very fast and efficient. In fact, I already have the above solution in code and tested to be working, if you would like to take a look, please let me know, i chose not to post it because its all in vector form and pretty much impossible to figure out exactly what is going on (because it's been condensed to be more efficient).
Lastly, yes this is for school work, though it is NOT a homework question/assignment/project. It's part of my graduate thesis (abet a very very small part, but still technically is part of it).
Thanks for your help.
Edit: Heres a new algorithm that i came up with a little while ago.
Starting at a circle's centre, draw a line to the other two centres. Calculate the line that bisects the angle created and calculate the intersections between the circle and the line that passes through the centre of your circle. You will get 2 results. Repeat this for the other two circles to get a total of 6 points. Iterate over these 6 points and get 8 possible solutions. Find the maximum of the 8 solutions.
This algorithm will deal with the collinear case if you draw your lines in one "direction" about the three points.
From the few random trials i have attempted using CAD software to figure out the geometries for me, this method seems to outperform all other methods previously stated However, it has already been proven to not be an optimal solution by one of Victor's counter examples.
I'll code this up tomorrow, for some reason I've lost remote access to my university computer and most things are on it.
I've taken the liberty of submitting a second answer, because my original answer referred to an online app that people could play with to get insight. The answer here is more a geometric argument.
The following diagram illuminates, I hope, what is going on. Much of this was inspired by #Federico Ramponi's observation that the largest triangle is characterized by the tangent at each vertex being parallel to the opposite side.
(source: brainjam.ca)
The picture was produced using a trial version of the excellent desktop program Geometry Expressions. The diagram shows the three circles centered at points A,E, and C. They have equal radii, but the picture doesn't really depend on the radii being equal, so the solution generalizes to circles of different radii. The lines MN, NO, and OM are tangent to the circles, and touch the circles at the points I,H, and G respectively. The latter points form the inner triangle IHG which is the triangle whose size we want to maximize.
There is also an exterior triangle MNO which is homethetic to the interior triangle, meaning that its sides are parallel to that of IHG.
#Federico observed that IHG has maximal area because moving any of its vertices along the corresponding circle will result an a triangle that has the same base but less height, therefore less area. To put it in slightly more technical terms, if the triangle is parameterized by angles t1,t2,t3 on the three circles (as pointed out by #Charles Stewart, and as used in my steepest descent canvas app), then the gradient of the area w.r.t to (t1,t2,t3) is (0,0,0), and the area is extremal (maximal in the diagram).
So how is this diagram computed? I'll admit in advance that I don't quite have the full story, but here's a start. Given the three circles, select a point M. Draw tangents to the circles centered at E and C, and designate the tangent points as G and I. Draw a tangent OHN to the circle centered at A that is parallel to GI. These are fairly straightforward operations both algebraically and geometrically.
But we aren't finished. So far we only have the condition that OHN is parallel to GI. We have no guarantee that MGO is parallel to IH or that MIN is parallel to GH. So we have to go back and refine M. In an interactive geometry program it's no big deal to set this up and then move M until the latter parallel conditions are met (by eyeballs, anyways). Geometry Expressions created the diagram, but I used a bit of a cheat to get it to do so, because its constraint solver was apparently not powerful enough to do the job. The algebraic expressions for G, I, and H are reasonably straightforward, so it should be possible to solve for M based on the fact that MIHG is a parallelogram, either explicitly or numerically.
I should point out that in general if you follow the construction starting from M, you have two choices of tangent for each circle, and therefore eight possible solutions. As in the other attempted answers to the question, unless you have a good heuristic to help you choose in advance which of the tangents to compute, you should probably compute all eight possible triangles and find the one with maximum area. The other seven will be extremal in the sense of being minimal area or saddle points.
That's it. This answer is not quite complete in that it leaves the final computation of M somewhat open ended. But it's reduced to either a 2D search space or the solution of an ornery but not humongous equation.
Finally, I have to disagree with #Federico's conclusion that this confirms that the solution proposed by the OP is optimal. It's true that if you draw perpendiculars from the circle centers to the opposite edge of the inner triangle, those perpendiculars intersect the circle to give you the triangle vertex. E.g. H lies on the line through A perpendicular to GI), but this is not the same as in the original proposed solution (which was to take the line through A and perpendicular to EC - in general EC is not parallel to GI).
I've created an HTML5 canvas app that may be useful for people to play with. It's pretty basic (and the code is not beautiful), but it lets you move three circles of equal radius, and then calculates a maximal triangle using gradient/steepest descent. You can also save bitmaps of the diagram. The diagram also shows the triangle whose vertices are the circle centers, and one of the altitudes. Edit1: the "altitude" is really just a line segment through one of the circle centers and perpendicular to the opposite edge of the triangle joining the centers. It's there because some of the suggested constructions use it. Edit2: the steepest descent method sometimes gets stuck in a local maximum. You can get out of that maximum by moving a circle until the black triangle flips and then bringing the circle back to its original position. Working on how to find the global maximum.
This won't work in IE because it doesn't support canvas, but most other "modern" browsers should work.
I did this partially because I found some of the arguments on this page questionable, and partially because I've never programmed a steepest descent and wanted to see how that worked. Anyways, I hope this helps, and I hope to weigh in with some more comments later.
Edit: I've looked at the geometry a little more and have written up my findings in a separate answer.
Let A, B, C be the vertexes of your triangle, and suppose they are placed as in your solution.
Notice that the key property of your construction is that each of the vertexes lies on a tangent to its circle which is parallel to the opposite side of the triangle. Obviously, the circle itself lies entirely on one side of the tangent, and in the optimal solution each tangent leaves its circle on the same side as the other vertexes.
Consider AB as the "base" of the triangle, and let C float in its circle. If you move C to another position C' within the circle, you will obtain another triangle ABC' with the same base but a smaller height, hence also with a smaller area:
figure 1 http://control.ee.ethz.ch/~ramponif/stuff/circles1.png
For the same reason, you can easily see that any position of the vertexes that doesn't follow your construction cannot be optimal. Suppose, for instance, that each one of the vertexes A', B', C' does not lie on a tangent parallel to the side connecting the other two.
Then, constructing the tangent to the circle that contains (say) C', which is parallel to A'B' and leaves the circle on the same side as A'B', and moving C' to the point of tangency C, it is always possible to construct a triangle A'B'C which has the same base, but a greater height, hence also a greater area:
figure 2 http://control.ee.ethz.ch/~ramponif/stuff/circles2.png
Since any triangle that does not follow your construction cannot be optimal, I do believe that your construction is optimal. In the case when the centers of the circles are aligned I'm a bit confused, but I guess that it is possible to prove optimality along the same lines.
I believe this is a convex optimization problem (no it's not, see below), and hence can be solved efficiently using well known methods.
You essentially want to solve the problem:
maximize: area(v1,v2,v3) ~ |cross((v2-v1), (v3-v1))|
such that: v1 in C1, v2 in C2, v3 in C3 (i.e., v_i-c_i)^2 - r_i^2 <= 0)
Each of the constraints are convex, and the area function is convex as well. Now, I don't know if there is a more efficient formulation, but you can at least use an interior point method with derivatives since the derivative of the area with respect to each vertex position can be worked out analytically (I have it written down somewhere...).
Edit: grad(area(v1,v2,v3))(v_i) = rot90(vec(vj,vk)), where vec(a,b) is making a 2D vector starting at a and ending at b, and rot90 means a positive orientation rotation by 90 degrees, assuming (vi,vj,vk) was positively oriented.
Edit 2: The problem is not convex, as should be obvious considering the collinear case; two degenerate solutions is a sure sign of non-convexity. However, the configuration starting at the circle centers should be in the globally optimal local maximum.
Not optimal, works well when all three are not colinear:
I don't have a proof (and therefore don't know if it's guaranteed to be biggest). Maybe I'll work on one. But:
We have three circles with radius R with positions (from center) P0, P1, and P2. We wish to find the vertices of a triangle such that the area of the triangle is maximum, and the vertices lie on any point of the circles edges.
Find the center of all the circles and call that C. Then C = (P0 + P1 + P2) / 3. Then we find the point on each circle farthest from C.
Find vectors V0, V1, and V2, where Vi = Pi - C. Then find points Q0, Q1, and Q2, where Qi = norm(Vi) * R + Pi. Where norm indicates normalization of a vector, norm(V) = V / |V|.
Q0, Q1, and Q2 are the vertices of the triangle. I assume this is optimal because this is the farthest the vertices could be from each other. (I think.)
My first thought is that you should be able to find an analytic solution.
Then the equations of the circles are:
(x1-h1)^2 + (y1-k1)^2 = r^2
(x2-h2)^2 + (y2-k2)^2 = r^2
(x3-h3)^2 + (y3-k3)^2 = r^2
The vertices of your triangle are (x1, y1), (x2, y2), and (x3, y3). The side lengths of your triangle are
A = sqrt((x1-x2)^2 + (y1-y2)^2)
B = sqrt((x1-x3)^2 + (y1-y3)^2)
C = sqrt((x2-x3)^2 + (y2-y3)^2)
So the area of the triangle is (using Heron's formula)
S = (A+B+C)/2
area = sqrt(S(S-A)(S-B)(S-C))
So area is a function of 6 variables.
At this point I realize this is not a fruitful line of reasoning. This is more like something I'd drop into a simulated annealing system.
So my second thought is to choose the point on circle with centre A as follows: Construct line BC joining the centres of the other two circles, then construct the line AD that is perpendicular to BC and passes through A. One vertex of the triangle is the intersection of AD and circle with centre A. Likewise for the other vertices. I can't prove this but I think it gives different results than the simple "furthest from the centre of all the circles" method, and for some reason it feels better to me. I know, not very mathematical, but then I'm a programmer.
Let's assume the center of the circles to be C0,C1 and C2; and the radius R.
Since the area of a triangle is .5*base*height, let's first find the maximum base that can be constructed with the circles.
Base = Max {(|C0-C1|+2R),(|C1-C2|+2R,(|C2-C0|+2R}
Once the base length is determined between 2 circles, then we can find the farthest perpendicular point from the base line to the third circle. (product of the their slopes is -1)
For special cases such as circles aligned in a single line, we need to perform additional checks at the time of determining the base line.
It appears that finding the largest Apollonius circle for the three circles and then inscribing an equilateral triangle in that circle would be a solution. Proof left as an exercise ;).
EDIT
This method has issues for collinear circles like other solutions here, too and doesn't work.
Some initial thoughts.
Definition Call the sought-after triangle, the maximal triangle. Note that this might not be unique: if the circles all have the same centre, then there are infinitely many maximal triangles obtained by rotation around the center, and if the centres are colinear, then there will be two maximal triangles, each a mirror image of the other.
Definition Call the triangle (possibly, degenerately, either a point or a line) whose vertices are the centres of the circles the interior triangle.
Observation The solution can be expressed as three angles, indicating where on the circumference of each circle the triangle is to be found.
Observation Given two exterior vertices, we can determine a third vertex that gives the maximal area: draw the altitude of the triangle between the two exterior vertices and the centre of the other circle. This line intersects the circumference in two places; the further away point is the maximising choice of third vertex. (Fixed incorrect algorithm, Federico's argument can be adapted to show correctness of this observation)
Consequence The problem is reduced to from a problem in three angles to one in two.
Conjecture Imagine the diagram is a pinboard, with three pins at the three centres of the circles. Imagine also a closed loop of string of length equal to the perimiter of the interior triangle, plus the radius of a circle, and we place this loop around the pins. Take an imaginary pen and imaginarily draw the looping figure where the loop is always tight. I conjecture that the points of the maximal triangle will all lie on this looping figure, and that in the case where the interior triangle is not degenerate, the vertices of the maximal triangle will be the three points where the looping figure intersects one of the circle circumferences. Many counterexamples
More to follow when I can spare time to think about it.
This is just a thought, no proof or math to go along with the construction just yet. It requires that the circle centers not be colinear if the radii are the same for each circle. This restriction can be relaxed if the radii are different.
Construction:
(1) Construct a triangle such that each side of the triangle is tangent to two circles, and therefore, each circle has a tangent point on two sides of the triangle.
(2) Draw the chord between these two tangent points on each circle
(3) Find the point on the boundary of the circle on the extended ray starting at the circle's center through the midpoint of the chord. There should be one such point on each of the three circles.
(4) Connect them three points of (3) to fom a triangle.
At that point I don't know if it's the largest such triangle, but if you're looking for something approximate, this might be it.
Later: You might be able to find an approximate answer for the degenerate case by perturbing the "middle" circle slightly in a direction perpendicular to the line connecting the three circles.