Fractalforums
Fractal Art => Fractal Image Gallery => Image Threads => Topic started by: marcm200 on September 05, 2019, 12:38:36 PM

I tried to extend the Figueiredo trustworthy Julia set algorithm (forum link (https://fractalforums.org/fractalmathematicsandnewtheories/28/juliasetstrueshapeandescapetime/2725/msg16490#new)) to the 3rd dimension using tricomplex numbers.
So basically now a pixel on the screen represents a tiny cube in space and its color is indicative of the fate of all the tricomplex numbers it encompasses. Here, colors are different from 2D, escaping being black (transparent), definitely bounded being yellow and unknown being white.
Here's the first result using Eric Baird's formula from http://www.bugman123.com/Hypercomplex/index.html#JuliaQuaternion (http://www.bugman123.com/Hypercomplex/index.html#JuliaQuaternion) and the 3d basilica version:
{x,y,z}_{new} := {x²y²z², 2xy, 2xz} + {1,0,0}
The first image below shows a small series of increasing refinements, upper row only escaping is transparent, lower row escaping and unknown are set to transparency, so one gets a view onto the bounded interior.
The 2nd image is the interior at the current computational limit at level 10 and a pointsampled version (different software, hence different coloring and size).
Limitations:
 I did not find a formula to calculate a valid escape radius for tricomplex numbers, so I just took the 2cube. Everything outside is considered definite exterior.
 Images are trustworthy in the sense that a pixel's color represents the fate of all underlying tricomplex numbers under the current escape radius. However I cannot mathematically guarantee that there are no further parts belonging to the object being outide the 2cube.
 The current implementation is only a proofofprinciple and very unoptimized, so refinement level 10 is the current limit.
 I kept dimming according to the observer's view angle as the colors of the three types of pixels are vastly different, so one does not get a false impression but retains 3d "experience".

That's an interesting exploration. IMHO in 3D even more so than in 2D.

Thanks!
Interesting, for me the 2D world is more appealing than 3D, mainly because I was disappointed when exploring the 2D Lyapunov space. I only found a couple of interesting shapes, nothing compared to the vast number of cool 2d images. Or maybe I was just not patient enough or looking in the wrong places. And partly because for me as a 3D observer looking at a 3D object I can only ever see the surface unless someone invents the 4D camera. So I guess if the surface is not very interesting  and that seemed to be the overall case in the realms I searched in  the result will be disappointing  like the tiny ant living in the complex plane looking at the Mset boundary only seeing points closer and farther away, I imagine that wouldn't be half as interesting as from our 3d view.
I am mainly interested here in seeing whether I can partition the interior of such a 3D object into disjoint subsets so when starting from a point in a set the entire orbit stays in that set.
In the 2D case then I get the basins of attraction (if there are any that is), but since here in the tricomplex case iteration functions are freely chosen for any coordinate, I have no clue as to whether that leads to attracting periodic orbits (or maybe any iteration function at all does have those?).
Maybe I get something that looks like shattered glass put together. That would be an interesting 3D shape!

The first (probably) disconnected interior I've found, again Eric Baird's formula, C=(0.4,0,0.8 ). It's interesting to see how asymmetrical the gray area is (left image). A pointsampled version shows the interior to be disconnected (or only connected with very thin filaments I missed by pointsampling).
I did a parameter walk for the formulas by Baird, Makin and Bristor (website from first post). So far I could find roughly 4 different shapes with interior: the basilica, asteroidic rocks, this disconnected one and some concave rocks.
The escape radius of R=3 has been found by using a stepapproach. I started with a very big R=2^{30} (hoping the complete objects lies in the inside), tile the world cube into 2^{30} smaller cubes and compute the bounding cubes for those tiny ones that built the sides of the world. If for a side all bounding cubes are at least 1% farther away (min distance of a corner to the origin) than the starting cube (max distance of a corner to the origin), the side is omitted. I do this for all 6 sides until no shrinking occurs anymore. Then I use that smaller R to start the process again with now a finer cube width (still 2^{30} of them though).
This led to a final R of 3. The idea being as long as the cubes keep increasing their distance after every iteration (starting from way outside going inside) I figured it'd be safe to assume one is already in the escaping region.
Although no proof this R is correct, it has at least a connection to escape time and is more justified than pure guessing I did earlier.

"Fix points" and "period2 cycles"
Quick summary
Two fix points and one period2 cycle were first found in the 3D basilica by brute force and then by manual calculation.
Detailed
In a first attempt to search for fix points in 3D I used Eric Baird's formula for the Basilica {x,y,z}_{new} := {x²y²z²,2xy,2xz} + { 1,0,0 } in a bruteforce way  so no solving of nonlinear equations for now (planned for the future).
A mandatory condition for a fix point is that at any refinement level there must be a tiny cube A whose bounding cube bbx(f(A)) after one iteration intersects back with A.
So I walked the 2cube in width 2^13 * 2^13 * 2^13 pixels and checked every such tiny cube. Since there is no image construction involved, the only constraint here is time, not space.
The search returned 28 tiny cubes as possible fix points. Quite a low number compared to all 2^39 cubes present, so the approach seemed valid.
Of course, for this simple case school math and WolframAlpha sufficed to see that {0.5 + 0.5*sqrt(5), 0, 0} are the only two fix points here.
Fortunately the brute force approach returned those too (in fact 8 tiny cubes contain the actual fix points: two fix point coordinates are 0 and lie on the grid, hence two adjacent cubes share it, i.e. 2*2 and since there are two solutions for the x coordinate, that gives
a total of 8 as found, a nice afterwards confirmation of the brute force method itself). The rest of the 28 I deem will vanish when increasing refinement level.
The image below shows the basilica and the two fix points (exaggerated), each at the crossing of three perpendicular red lines (one is in the background on the right side, dimmed quite a bit). It would be interesting to see at what refinement level those actually are identified as interior.
While fix points worked quite well I tried the same approach for period2 cycles. Again with 2^39 cubes in refinement 13, 632 possible cubes involved in such a cycle were found.
Manual browsing through the cube coordinates revealed the two known fix points and two further points which appear to be around the origin and {1,0,0}. And indeed, the sequence {0,0,0} > {1,0,0} is a period 2 cycle as seen by manual calculation.
Quite interesting is, that 532 of those possible period2 cubes cluster around the fixpoint at 0.50.5*sqrt(5), 100 cluster around the origin, 8 lie around the other fix point and only 4 around 1,0,0.
Is this a hint towards superattracting (fix point at 0.61...), attracting (origin, but not the 2nd cycle point) and repelling (other fix point)? Or just a result describing the algorithmic (and not fractal) behaviour that would change at other refinements?
Since I have no idea whether something like a multiplier exists for the 3D case here to characterize cycles, I think I will just partition the interior according to orbit reachability and see what comes out.
EDIT: Correction: It must state 520 instead of 532 cubes clustering around the fix point.

Since I have no idea whether something like a multiplier exists for the 3D case here to characterize cycles, I think I will just partition the interior according to orbit reachability and see what comes out.
You can use the Jacobian derivative matrix. Its determinant will have absolute value less than one when attracting I think. Eigenvalues may also be interesting.

You can use the Jacobian derivative matrix. Its determinant will have absolute value less than one when attracting I think. Eigenvalues may also be interesting.
The product of eigenvalues should be the same value as the determinant. It is possible that (the absolute magnitude of) the determinant is less than one, but one or more eigenvalues are greater than one. That means the space around that point is overall attracted, but there may be narrow "jets" of points being repelled still.

I took your suggestions and let WA calculate the Jacobian's determinant and the eigenvalues at the 4 interesting points:
period2cycle: determinants (eigenvalues) 0 (0) and 8 (2,2,2). If I were to multiply those as in the 2D case, that would result in 0 and indicate a superattracting cycle. Hence one could expect one basin of attraction there.
the 2 fixpoints: Both are repelling: 36 (3.23,3.23,3.23) and 2.23 (1.23,1.23,1.23). Interestingly the 520 clustered tiny cubes from my last post are not around the origin (determinant 0, superattracting?) but around the fix point with Jacobian of 2.23.
It's almost like science: There is a theory (Jacobian, fix points, periods in the 3D case), a prediction (one basin for the period2 cycle)  now comes the experiment (I will need some days to finish the coding).

One partition was the correct prediction (top part of image below).
Since the result  coloring a yellow structure another color  is not very impressive, I used the cubic 2D formula (x+i*y)^3+1.25*(x+i*y)+0.025i from the 2nd Figueiredo paper with a constant z coordinate of 0 as a positive control to ensure the coloring method works correctly (lower part).
Since in the 2D case a quadratic polynomial can at most have 1 attracting basin and assuming the same for the 3D case, I will now focus my search on some higher powers, maybe trying to construct such a multiple basin 3D set by using a 2D formula with a "suitable" (not sure, how at the moment) z coordinate transformation as a start.

"Two partitions in true 3D"
The first tricomplex iteration formula with 2 basins of attraction/partitions:
{x,y,z}_{new} := {1.25*x+x^{3}3xy²z² , 1.25*y+3*x²*yy^{3} , z^{3}y^{3}x^{3} } + { 1065353, 0, 2097152 } * 2^{25}
During the construction process I learnt that the iteration formula should resemble a complex multiplication. If the formula is arbitrarily chosen, all I get, are some strange shapes, usually with spikes in all directions that do not give rise to interior points at the currently computable refinements.
The formula here was constructed using the cubic complex polynomial z^3+A*z + c from the 2nd Figueiredo paper (known to have two basins of attraction for A=1.25 and c=0.025i) and adding/varying a zcontaining term at some positions to get a true 3D shape and not a stacked version as before.
Very interesting are the smaller spots at the end of the arms having a different color. Maybe there is a series of ever smaller getting objects of alternating color?
Preanalysis for periodic points resulted in 3 possible fix points and 9 period2 points. So the set was deemed valid and computed in various refinement levels to see a partition coloring occuring.
The image shows the two basins at level 11. Axis range is 4..+4 in all directions. The escape radius R=4 was found by the step approach I posted some time ago (start R=2^{30}).

A walk through the function family (based on complex cubic polynomial multiplication):
{x,y,z}_{new} := { A*x+x^{3}3xy²B*z², A*y+3*x²*yy^{3}, z^{3}y^{3}x^{3} } + { Cx, Cy, Cz }
I introduced the factor B as kind of a dampening number (I read here in the forum about relaxed Newton), so limiting the influence of the 3rd dimension while still having a true 3D object and not just a stacked version of a 2D slice.
First image A=2, B=0.75, C={1.2,0,0} resembles a bit a skeleton of a snail  if moluscae had bones. Unfortunately no interior cells at level 11. The pointsampled version (upper left) breaks into pieces, but I like the snail better.
The second image A=1.5, B=0, C=0 is actually only the 2nd image overall with two basins of attraction/partitions.
Looking at this and the first 2partition image A=1.25, B=1, C={ 0.03175,0,0.0625 }, it seems that a rotational symmetry is mandatory (?) for more than one partition. Maybe there is a way to test that symmetry beforehand, but I currently do not know how since the symmetry pertains to the fate of the iterated point and not the point iterate itself.
But maybe symmetry of the iteration function itself is verifiable using some elaborate system of equations. Tthe only idea I got so far is: Take a rotation axis, project any point to it (normal vector), rotate it and look if the destination point is rotated back to the original one, but that's hard enough (I've never done much in 3D graphics), but currently I'm just doing a bruteforce parameter walk on my second computer.

Walking through the family TRICZ4B (I have to assign names now for my software and my memory, this one is based on complex power 4 exponentiation, adding the damping factor for the xcoordinate and the standard sum at the z):
{x,y,z]_{new} := { A*x+x^{4}6*x²*y²B*z² , A*y+4*x^{3}*y4*y^{3}*x , z^{3}y^{3}x^{3} } + { Cx, Cy, Cz }
The bruteforce approach found quite a number of parameters with multiple basins.
The example below has three partitions  it looks as if there's liquid falling down a water fall. The bottom row depicts the "evolution" of the object at increasing refinement levels (upscaled to level 10 screen width, so some look boxy), showing how the colored components grow together, new small colored regions appear  just like the 2D blobs of black that are frequently detected when interior cells first emerge.
Some question I cannot answer so far:
Is 3 the maximum number of basins that can be found for a degree 4 polynomial in a tricomplex number setting?.
Do the individual components finally touch at one point and are separated by exterior elsewhere as in the 2D Julia sets for geometric different components of the interior?
Or do they behave more like basins of attraction for Newton's method, where the channels touch each other more in a line manner (with some nonroot converging regions in between maybe)?
It looks like they touch (as the pointsampled version suggests, upper right) in an area, but since by design of the Figueiredo algorithm for the 2D case there are always gray cells between distinct interior components, so one can never know. I assume this is also true for the 3D case here.
Interestingly, from the pointsampled version I wouldn't have guessed it to have different basins at all  I was looking for rotationally symmetrical ones like I posted before.

A very interesting TRICZ4B Julia set (A=0.25, B=0.75, c={ 1.2,0,0 } ).
The 3d version (left) does not show interior cells at level 11, it looks like a ship's bull's eye.
The pointsampled version (middle, side view at bottom) of the nonescaping points suggested that the object is planar (or very thin, like the one with TylerSmith's formula forum link (https://fractalforums.org/fractalmathematicsandnewtheories/28/brandnewmethodfortrue3dfractals/3118/msg17143#msg17143)). So I figuered there will never be interior 3D cells.
So I was thinking  maybe the coordinates of that 3D object could be transformed in a way that it lies completely in a coordinate plane  and then I could use the 2D TSA algorithm.
Checking the nonescaping points revealed that y was always zero. So I plugged that into the TRICZ4B formula, computed the bounding box and used it in the 2D TSA version (right). Since this was the first formula system that was not derived from a complex exponentiation but rather independent x,y coordinate transformations I checked levels 10 (already black) to 16 whether black and white do not touch directly  which they fortunately did not.
I was very lucky that it was so simple  y being constant 0. I would have had no idea how to transform an arbitrarily oriented plane where the object resided so that one coordinate would finally be constant zero (or detect such a plane in the first place).
It would be interesting to see the parameterspace object consisting only of seed values where the Julia set is planar.

"Maybe a filter?"
To get color into 3D objects I often use trafassel's direct RGB summation method. When I looked at some pointsampled images, I encountered a very remarkable resemblance with one of the twobasin pictures I posted before.
First image below:
Left side, trustworthily computed interior, colored by basin of attraction.
Right side: Pointsampled version, colored by RGB summation (see technical details below).
The versions look almost identical (some colored spots in the origin's region for the pointsampled object). So the summation in that pictuzre mimicks almost perfectly the shapes of the two distinct partitions of the TSA image.
Now I was wondering: Could this be used as a filter?
I started a new walk through a function family (degree 5, ongoing). Here is the first example (2nd image below) where I expect partitions (differently RGB colored parts separated by exterior from one another). Unfortunately at level 11 I could not yet detect interior cells.
But I like the shape of the pointsampled version, at the lower part it looks like an opened pea pod.
Is this resemblance just a coincidence in one (a couple?) of parameter settings? Or can one expect more  maybe it's even a subset of all partionable 3D Julia sets that share common orbit characteristics?
Technical details for my implementation of direct RGB summation:
 For every pixel I start with an RGB sum {0,0,0}. In each iteration red sum_{i}=red sum_{i1} + x_{i}/{x_{i},y_{i},z_{i}}]², green/blue accordingly with y and z)
 After max it, an interior point's sums are shifted into the range 0..255 by adding or substracting 255 repeatedly.
 The final values were used directly as an RGB value.

One of the most interesting shapes I have found so far  here using David Makin's tricomplex formula.
Very twisty, looks almost like a band used by professional rhythmic gymnasts, but unfortunately pointsampling shows an empty (Cantor dust?) set  beginning to disintegrate after only 9 (nine!) iterations (lower right, notyetescaped points).

The first 3d Julia set showing multiple basins that has been found in a fully automated manner by a random walk through a degree 4 function family (TRIC4CPLXMULT), that is based on a standard complex exponentation (x+y*i)^4 (details below).
{x,y,z}_{new} := { e*x^4+f*x^2*y^2+g*y^4 + h*x+j*z^4 , l*x^3*y+m*x*y^3 + k*y , n*z^3+o*y^3+p*x^3 } + { Cx, Cy, Cz }
Julia sets were automatically computed and judged by the direct RGB sum color filter I wrote about in one of my previous posts. In case there were at least two vastly different colors (judged by R, G, B independently being above 100, so 8 combinations in total), the set was actually saved.
The image below is: e=2.3 f=4.5 g=0.8 h=2.1 j=1.1 k=1 l=0.2 m=0.7 n=0.2 o=0 p=0, C={ 0.6,0,0.6 }, at level 10 where interior cells emerged, computed in the 2cube.
Technical note:
 Parameters ep take real values and are mostly the coefficients of the complex exponentiation, randomly disturbed (+ <= 5 in 0.1 increments).
 A linear xoldterm was introduced into the new x part (italic, variable h  preliminary experiments showed that this was a good way to add structure) and a dependence on zold (variable j) so every axis influences every other.
 Similarly I introduced a yoldterm (variable k) into the new ycoordinate.
 As for the zterm I use most of the time a sum of degree1 or below powers of the old coordinates.
 The { Cx, Cy, Cz }parameter was split into (4x4x4) values from 1.2 to +1.2 in each direction.
 I did not check whether the whole set is contained within the 2cube, however interior cells if found are mathematically guaranteed correct no matter the escape radius. If a cell does not have a path to outside the 2cube, the cell surely has now path outside a larger cube, so it will always be determined as interior if using the correct (larger) escape radius.
 White and gray cells however do not need to be accurate in this case as a cell might be prematurely judged as white in too small an escape radius. For those I need a valid escape radius  e.g. using the jump method I posted some time ago.

"Periodicity"
I implemented the periodicity procedure from the 2D polynomial TSA (method described here: forum link (https://fractalforums.org/fractalmathematicsandnewtheories/28/juliasetstrueshapeandescapetime/2725/msg17021#msg17021)) to see whether I could also find immediate basins in the tricomplex case.
Below are the two most "complicated" sets I encountered thus far: 3 fix points in the steak image, and a period 2 cycle in the basilicalike (Makin's formula) (left the full basins of attraction, right the immediate basins of the periodic points  colors are not necessarily comparable between left and right image).
Interesting is the similarity between the 3D basilica and the 2D polynomial form z²1 as they both show a period 2 cycle reaching to the negative axis.
I wonder whether there are longer cycles than period 2 (does anyone know parameters for such a case?)

"Chlorophyllic DNA"
If DNA were built from chlorophyll I think it would look like this (left). Quite interesting from a topological view with all those holes.
Doug Bristor's formula shows a period3cycle (upper right). The interior still needs a long way to look like the pointsampled version (lower right, direct RGB sum colored).
The parameter was found by another bruteforce walk, this time I computed small pointsampled 3D sliced Msets for a given formula and selected values that were almost on the boundary, but just inside (one layer of alsoinside voxels around the chosen one, then at least one escaping).
I hope to get longer cycles by computing larger Msets and then trying to find seed values from deep within bulblike structures (3Dcenters of hyperbolic components?).

"A period4 cycle"
Moving the seed out along the xaxis while keeping y,z=0,0 I hoped to hit smaller and smaller copies of the main Mset (if they exist, that is, which I don't know) thereby increasing the period of the original basilica (period2). Since the 1,0,0version is a very fast computable set with early emerging interior cells, I hoped that this might stay valid further out.
That led to this nice attracting period4 cycle Julia set using David Makin's formula at c={ 42614129 , 0 , 0 ] * 2^25 ~ {1.27,0,0}.
Cycle being attracting
This was deduced by algorithmic properties of the detection method without the use of the Jacobian.
Every geometrically isolated interior object (a "blob") was tested where its members point to in one iteration. It turned out that for a given blob all cells pointed to only one specific target blob  so no spreading out.
If I start with blob A and move to its target blob, to that one's target blob and so on and come back to A at some point, this sequence of blobs forms a cycle.
Since there are only finitely many detected blobs, starting with a yellow blob B and following its target blob and so on must eventually lead to a blob D that has already been visited before. The blob sequence between the first and the 2nd occurance of D then forms a cycle.
The sole cycle detected is the turquois period4 one, so every yellow blob finally lands in that cycle.

"Chinese tower"
Many of the interior showing sets I could compute so far look more or less like fields of asteroids of some shape floating around in space (and I can't go to higher refinement levels yet to "close the gaps"). So I was pleasantly surprised that this one has quite a different shape.
A period2 cycle at level 11 in a degree5 family (based on complex exponentiation (x+i*y)^5 with the usual additions):
{x,y,z}_{new} := { f*x+h*z^{2} + x^{5}10x^{3}*y^{2}+5x^{3}*y^{2} , e*y+5x^{4}*y10x^{2}*y^{3}+y^{5} , z^{2}y^{2}x^{2} } + { Cx, Cy, Cz }
here with e=f=h=0, C={ 0,71875,0.71875,1.71875 }.

"Period 8"
A very interesting Julia set using David Makin's formula, not from the shape point of view (it's just another collection of asteroids), but the first with period 8 (turquois/yellow as always, only 7 immediate basins visible, see details below) and with interior cells emerging at level 14.
David Makin's formula has the interesting property, that, when used with {x,y,0} and {cx,cy,0} it reduces to the classic quadratic z²+c case.
Makin:
{x,y,z}_{new} := { x²y²z² , 2xy , 2z*(xy) } + { Cx, Cy, Cz }
when z, Cz=constant 0:
{x,y,0}_{new} := { x²y² , 2xy , 0 } + { Cx, Cy, 0 }
So I was interested to see how the 2D corresponding Julia set behaves in the TSA.
The 2nd image shows at level 14 also a period8 cycle (yellow)  and black interior emerges at level 14 here as well. So the 2D properties translate into the higher dimension world and form a 3D object. Maybe I can do that for other interesting shapes as well.
Technical details:
 I implemented a new data structure that works especially well with sparse objects: allocating memory per row only encompassing the gray region, while still preserving direct array access and keeping memory fragmentation at minimum.
 The 8th immediate basin in the 3D version (far left) is not visible in the image here which is a trustworthily 8fold downscaled version, but it was computationally determined by the TSA. Currently my cubeviewer cannot yet handle cubes of side length 16k (working on it).
 The C value was found exploiting the reduction property looking for small Makin Mset copies and their geometric centers (as described in post #18).

"Cord loop"
This filledin Julia set shows some winding ropes, a phenomenon which I do not encounter very often. It's very filigree, as compared to the more common asteroid rocks.
I'm using David Makin's formula as a base currently  changing exponents, adding some terms etc, hoping to find new shapes while retaining its fast computation property and relative abundance of attracting periodic points.
{x,y,z}_{new} := { x^{4}y^{4}B*z² , A*x+2xy , z^{4}x²*y² } + { 0.5 , 1 , 0 } (named MakinExp4).
with A=0.25, B=1
The image below is at level 13 (tw. downscaled 4fold), level 14 would need just a bit more memory than I have available currently.
The initial pointsampled small image showed a nice color separation (direct RGB sum), which corresponds to the immediate basin found. Although by far not all sets show such a separation, it is usable as a priority queue for computation order.

A degree5 formula that shows a nice evolution of interior and its fix point: New regions appearing at every step, and, as also often seen in the 2D case, formerly nonimmediate (brown/yellow) regions getting flooded over by the immediate red/cyan basin (e.g. the objects to the left and right of the front eigth symbol). It looks a bit like a lion's cage is being closed with a red lid.
{x,y,z}_{new} := { x^{5}y^{5}B*z^{5} , x^{4}A*y²*z² , 2x*(zy) } + { 0.375 , 0.96875 , 0.96875 }
with A=0 and B=0.5
Technical details:
 Levels 9 to 11: Immediate basin = turquois, attraction basin = yellow.
 At first I couldn't compute level 12 as it reached an implementation limit: more than 2^16 gemometrically separated blobs.
 Devising a new cycle detection method (storing only the current orbit and the immediate basins of all cycles), this set was now computable.
 The current method's limit is 2^16 separated Fatou components in an orbit to a cycle. I wonder when I reach that barrier.

In the 2D images the gray region is always visible and gives a good impression of how the overall shape of the Julia set looks like.
I tried the same here in the 3D case, experimenting with transparency of the gray region.
Here's a period3 cubic triplex Julia set (magenta is the immediate basin, yellow the attraction basin at that level):
{x,y,z}_{new} := { x^{3}3xy²+0.5z² , 3yx²y^{3} , 2xz2yz } + { 0.265625 , 1.265625 , 0 }
I'm trying to find a setting that  if one has e.g. a magentacolored interior region deep inside a gray region (a "rock"), that the magenta shines through to the surface. But I'm not quite there yet.
Usually my virtual screen that goes through the cube picks colors up  either a nontransparent full color RGB value (than it stops travelling for that screen pixel). Here I'm adding small gray amounts per encountered gray cell in the journey path (i.e. gray is not fully transparent) until the interior is reached and then overlay the two color values. But I either get a gray smear (added gray is too thin) or a white area (added gray is too thick).
Is there a better way of getting a "shinethrough" result?

A period4 attracting cycle in a rather complicated iteration formula, based on Tyler Smith's idea of two consecutive rotations. Here the first step is a power3 complex exponentation of x,y_old, followed by a classical power2 rotation with y_temporary,z_old.
The interior when emerging does not look like rocks but rather multiplearmed stars. The yellow nonimmediate basins can grow together and merge with the periodic basins, but the object looks here at level 11 already quite similar to what the pointsampled version shows.
formula CPF3:
{x,y,z}_new := { x^{3}3xy² + B*z² , 9x^{4}y²6x²y^{4} + y^{6}  z² , 6x²*yz2y^{3}z } + { Cx, Cy, Cz }

An interesting fixpoint (purple): The two "arcs" have a very thin connection, but are geometrically one object. This is actually only the 2nd example I have encountered that has very anfractuous Fatou components.
{x,y,z}_new := { x²y²z² , x^{3}  3xy² , 4yx^{3}  4xy^{3} } + 2^25 * { 20132659 , 33554432 , 6710887 }
(It's the first result of mixing different component functions between already used iteration formulae.)

A nice shape, although no interior found till level 12. It seems to have a eversmaller getting elephant trunk rolling in.
This is one of so far only a few examples where the object actually resembles a 3D twisted version of a quadratic Julia set (although parameters are not related, z²+(0.30.42i) ).
Level 13 is currently too timeexpensive.
{x,y,z}_new := { 2xy , 2xz , x²y²z² } + { 0 , 0.5 , 0 }

"Romulan warbird"
Lately I find some strange shapes, this one looks like a space ship from planet Romulus. It consists almost entirely of the large immediate basin of a fixpoint (purple), I wonder whether the small yellow notyet/nonimmediate basins will eventually all merge into it.
{x,y,z}_new := { 4xyz+2zs , 4x²y²+4xys+s²z² , x²y²2xy+z } + { 0.25 , s , 0 } with s=0.75
The x/y_new terms are based on Tyler Smith's 2step rotational formula.

This is the most completed set I've calculated thus far, until it exceeds my memory: Level 13, period2 cycle (bright green) (With my current bruteforce combinations I find a lot of fix points and period2, but so far nothing longer).
The set shows an interesting evolution. with the arcparts emerging quite early.  otherwise I might have missed them thinking "boring rocks again".
{x,y,z}_new := { x²y² , 4x²y²+4xys+s²z² , 2xzyz } + { 1 , s , 0 } with s=1

One of the most (rotationally) symmetrical sets I've found so far. I guess the symmetry stems from the y_new term, which, again, is based on Tyler Smith's 2step formula. Unfortunately it owns again just a fixpoint (red).
{x,y,z}_new := { 6x²yz2z²y² , 4x²y²+4xys+s²z² , x^{4}x²y² } + { 0.5, s ,1 } with s=0.5
It's the first set computed with a new version of juliatsa3d (to be released) which uses a low resolution reverse cell graph as in the classical 2D software. That led to quite a substantial reduction in memory transfer to the cache (and some fewer bounding box calculations). So for images not exhausting the memory completely, that works quite well as a speedup.

"Pretender"
A very nice shape of this level 13 Julia set  no interior found, and it's not clear if any will ever appear. But it looks a bit like small Enterprise star ships next to a giant space station.
On the right of the image below are three smaller greenish pointsampled pictures of the same set: nonescaping points, escape radius 100; 100 max it (upper), 125 (middle), 150 (lower).
The object "loses" interior points, but to what extent? Empty interior? Cantor dust? Some very small volumebearing miniobjects missed by pointsampling?
{x,y,z}_new := { 6x²yz2z²y² , 4x²y²+4x²ys+s²z² , x^{4}y^{4}z² } + { 0.25 , s , 0.25 } with s=1
With interior present the gray area at some point reflects roughly the shape of the underlying nonescpaing objects. So it is interesting to see how the gray disappears when presumably no interior is there.

A period1 cycle in a very strange shape. I wonder whether there is branchingout at the tips (or at the thick part) of the threads here.
I hope I can transfer in the future the periodicpoint detection method from 2D to this case to see whether here actually a fixpoint exists or if it is more like a strange attractor's set of unstable but contained orbits (by construction, every turquois voxel iterates always in the turquois connected object).
{x,y,z}_new := { x^{4}y^{4}z² , 2xzyz , y+3x²yy^{3} } + { 0.5, 0.5 , 0.25 }
The formula is quite simple and very fast to compute (double precision was still sufficient in level 13).