The short answer is âconstrain that second point to be 0 with regards to one axis, and strictly positive with regards to the otherâ.
Then the longer answer/commentary.
I know nothing about Bayesian MDS, other than I googled MDS, read the Wikipedia definition, and saw some scatter plots of points that all seem to very much want to be removed, and vaguely equidistant, from the origin (a âdonutâ distribution). So, there is a high risk of there being gaps in my understanding in my answer below. Certainly I wonder what baseline prior youâd use for these results?
The approach with fixing the four symmetries (you donât mention identity symmetry, or label switching, which is the fourth â maybe thatâs because youâre handling very few dimensions) will I suspect be guided with the intent of the model.
In the case of PCA, I would nudge towards sparsity on the resulting matrix via a single prior with high density around the origin and centered around it, thus strongly reducing the number of candidate results, and doing away in principle with rotation and translation. There will remain multiple valid, roughly as sparse and entirely asymmetrical, solutions, which can be addressed by specifying the model as a dirichlet-mixture of such independent solutions. The benefit of this approach is that it leads to solutions that are elegant to interpret as the factors are almost guaranteed to be orthogonal. Resolving the reflection symmetry is approached by fixing one point to always be strictly positive (across all candidate solutions, and all axis). However the point must be well chosen to be guaranteed to be far removed from all axis, which goes against the principle of a well-balanced orthogonal result, so there might be a need to apply this constraint on a variety of points, for different axis each. This, in addition, is also required to fix label switching (if one is bothered by it at all, as it less likely to occur within a chain). When one is looking for exploratory âfactorâ solutions to a data set, with such solutions lending themselves to critical examination, and for a measure of the relative likelihood of the candidate solutions, this approach delivers that. The identification of points to make strictly positive as to fix the identity and rotation of each factor is, however, tremendously tedious, as is the selection of the number of solutions, and dimensions, to allow.
However from my (close to in-existent) understanding of bMDS, you would appear to be seeking everything but sparsity in the result â really you would want the opposite, in what I described as a donut prior, so itâs quite likely this wouldnât work at all.
On the interventions you mention I can say this:
-
Donât nudge your constraints (any zeroes) with priors: integrate it to the model definition (theyâre âobserved constraintsâ, in a way)
-
To solve rotation, you have to constrain as many points to each axis as there are dimensions - 1 (i.e. your solution works for 2 dimensions only)
-
Centering the prior as I would do for PCA, as opposed to forcing any one point to be the origin whilst making the effective center of the prior a random variable itself (is that what youâre doing?), seem entirely equivalent approaches to me.
To summarise: your approach, with the addition of a positivity constraint, works for 2 dimensions, however Iâm unsettled by translation symmetry being a concern, as I would expect the center to have some constraints in any (if trivially) preferred solution. Perhaps only the distance between the points really matters, then I can see how fixing one point to be the origin would reduce jitter as opposed to any softer constraints imposed to all points by means of a prior.