The answer to this question depends heavily on the method that you're using to compare the two groups, as well as on the covariance structure of the data and also on whether you are conducting your comparisons, bone by bone, or constructing a composite of several bones by combining the data for the individual bones. Methods differ in their sensitivity to the ratio between N, the number of observations and p, the number of variables. As one answer said, if you are using a method that is sensitive to the number of variables relative to N, it is common to reduce the dimensionality of the data (usually, by using principal component scores rather than the shape data) when doing the analysis. It is also common to check how well the discriminant function performs by doing a resampling procedure (like a jackknife procedure), leaving out some fraction of the specimens, rerunning the analysis, and determining how many individuals are correctly classified by the discriminant function). Other methods are less sensitive than that one is. So if you are asking whether, and by how much, the means differ, you would instead do a Procrustes Anova and calculate the distance between the mean shapes. You would not measure the coefficients of variation if you are analzing shape because that variation does not depend on the mean (the variance scales with the mean for measures of size, but not for shape). There has been a lot of discussion about the potential problems posed by having "too many" variables for the number of observations in geometric morphometric analyses but an important issue to note when you read these studies is their assumptions about the covariance structure of the data. If the variables vary independently (as is almost never the case), the results from some methods (Canonical Variates Analysis/Discriminant Function Analysis, Principal Components Analysis and Beween-group PCA) depend heavily on that N/p ratio but less so when they covary. This has been discussed in the morphometrics group Morphmet so you might look there to see both some discussion of that topic as well as links to some recent papers on it.

In my own paleontological work I compare like with like. If it were me, I'd use as many bones as they all have in common. Stated another way, if all males and females are represented by the same two (or three or whatever) metacarpals, I'd focus on them. I also have to be mindful of sample size, although paleontological studies often deal with small samples.

Regarding sample sizes a common strategy is perform a PCA and use the PC scores to do subsequent analyses for instance to do a discriminant function analysis between both sexes. Really there are not rules to perform statistical analyses in geometric morphometrics using 3D data. If well is true that ideally the samples must be three times the number of landmarks in some cases the use of dimension reductions techniques such as PCA, PCoA, ICA, etc., can help specially when it is not possible obtain additional samples or even in the investigation of fossils whose sample size is usually very low.

Maybe consider sample structure and calculate coefficient of variation between at least 2 bones at https://www.easycalculation.com/statistics/coefficient-of-variance.php.

Maybe compare coefficients of variation (CV) e.g. compare sexes at https://www.medcalc.org/calc/comparison_of_coefficientsofvariation.php

Article Sexually dimorphic digits

The answer to this question depends heavily on the method that you're using to compare the two groups, as well as on the covariance structure of the data and also on whether you are conducting your comparisons, bone by bone, or constructing a composite of several bones by combining the data for the individual bones. Methods differ in their sensitivity to the ratio between N, the number of observations and p, the number of variables. As one answer said, if you are using a method that is sensitive to the number of variables relative to N, it is common to reduce the dimensionality of the data (usually, by using principal component scores rather than the shape data) when doing the analysis. It is also common to check how well the discriminant function performs by doing a resampling procedure (like a jackknife procedure), leaving out some fraction of the specimens, rerunning the analysis, and determining how many individuals are correctly classified by the discriminant function). Other methods are less sensitive than that one is. So if you are asking whether, and by how much, the means differ, you would instead do a Procrustes Anova and calculate the distance between the mean shapes. You would not measure the coefficients of variation if you are analzing shape because that variation does not depend on the mean (the variance scales with the mean for measures of size, but not for shape). There has been a lot of discussion about the potential problems posed by having "too many" variables for the number of observations in geometric morphometric analyses but an important issue to note when you read these studies is their assumptions about the covariance structure of the data. If the variables vary independently (as is almost never the case), the results from some methods (Canonical Variates Analysis/Discriminant Function Analysis, Principal Components Analysis and Beween-group PCA) depend heavily on that N/p ratio but less so when they covary. This has been discussed in the morphometrics group Morphmet so you might look there to see both some discussion of that topic as well as links to some recent papers on it.

its depend on your number of landmarks. the sample seizes should be more than double number of landmarks.for instance; if your landmarkrs are 10, your sample sizes should be 20 or more.

As Miriam expertly answered, the question depends heavily on the method that you are using to compare groups, the covariance structure and the choice of traits. I would avoid hard and fast rules. I discuss some of these issues in a (more or less) recent article that I'm leaving here in case it might interest you : Article How many more? Sample size determination in studies of morph...