Hiya, Julie! Actually, the answer to some of your questions is in the title of the table. The Coefficient of Relationship (CoR) has, for many years, served as the basis for calculating a theoretical, expected amount of inherited DNA and DNA sharing. It remains highly useful as a baseline average. You can Google it to find numerous references, but one good one is http://www.genetic-genealogy.co.uk/Toc115570135.html.
It's a simple equation, and can not only be used to estimate DNA sharing by relationship, but also to offer a prediction of the effect of pedigree collapse and inbreeding. Another good reference for the CoR and how it's applied is from Diahan Southard: https://www.yourdnaguide.com/ydgblog/2019/7/26/pedigree-collapse-and-genetic-relationships.
In that article, she also explains the answer to another of your questions, what you referred to as column f, the expected percentage of the same DNA shared by two cousins that was inherited from the same ancestor. Scroll down to Diahan's section, "The math of pedigree collapse." In the second table there and just above it, she explains the simple CoR notion: for grandparents and 1st cousins, you get 25% of each grandparent's DNA and you have four grandparents; the numbers work out to 25% divided by 4, so 6.25%...25% of the 25%. With great-grandparents and 2nd cousins, you have 8 g-grandparents and expect to have about 12.5% of the DNA of each of them: 12.5% divided by 8 equals 1.5625%; and so on.
About centiMorgans. If I could stand on a soapbox and shout out one thing about the way we use autosomal DNA in genealogy, it would be, "The centiMorgan is only a rough estimate, folks! It's just a formulaic guesstimate!" We see far too many conversations about whether 7cM is valid while 6.5cM is not, or whether 12cM is large enough to always be bankable.
By definition, a centiMorgan is nothing but a computational, estimated prediction; we still use essentially the same equation, called the Kosambi Map Function, as developed by Damodar Kosambi 1944. It isn't just an estimate; it's a very rough estimate for two significant reasons: the first is that we're still using a 7-year-old human genome map, or assembly, for all our common genealogy reporting. And even the most current genome maps don't take into account recent discoveries regarding things like recombination hotspots in the autosomes. Our whole basis for assumptions about how frequently which areas of which chromosomes will undergo crossover may well be inaccurate. At the very least, it's dated.
Second, during meiosis and gametogenesis, males and females undergo crossover--recombination--at very different frequencies...females about 70% more often than males. And remember, we estimate a cM based on the potential of crossover at the next meiotic event...a centiMorgan isn't an estimate of what we think happened at a previous meiosis. This means that centiMorgans are calculated against either a male or female genome map. In examples I've given before, the same exact base-pair to base-pair range on the female map may be 27.6cM...on the male map only 4.2cM. Every reported cM value we see for genealogy is what's called sex-averaged. The reporting company will take that 27.6cM and 4.2cM and report the average: 11.7cM.
Further, different testing companies use different algorithms to report centiMorgans, even though they're all (at least currently) reporting against human genome map GRCh37 (we assume GRCh37.p13). AncestryDNA, for example, applies a proprietary form of computational phasing against genotype cohort modeling before they report the size of a match in cMs; many feels that, as a result, Ancestry reports smaller segments overall than other companies. MyHeritage uses what they term "stitching" to do almost the reverse: essentially using similar genotype cohort modeling but to assume that two very small segments might actually be one larger segment. And of course we can't forget that different versions of different microarray chips may be testing as few as 20% of the same SNPs. So in many instances we're using a good deal of guesswork and imputation to determine what a segment actually is even before we apply the estimated prediction of centiMorgan calculation to that segment.
Is the centiMorgan a useful tool? You bet. It's really the only way we have of estimating autosomal genetic relationship. But the cMs we see reported are far, far less precise and accurate than most believe they are. So...
"What is the source of the numbers in column d? I can see that they aren't column e divided by around 3400, but to me that's what the heading implies."
No, that column is the "Expected % Amount of Total DNA Shared Between Cousins." It's a percentage based on the CoR; it has nothing to do with centiMorgans. If you'll look again at https://isogg.org/wiki/Autosomal_DNA_statistics, you should get your answers to the remaining questions. ISOGG uses the same CoR calculations, and explains how the associated centiMorgan values are derived.
They--and most of the testing companies--start with a baseline assumption that the genome is 6800cM. For reasons described above, we know this is simply an overarching mathematical rounding: inaccurate, but useful for comparisons and it's pretty much all we have right now.
Similarly, the potential for crossover during meiosis varies wildly across the chromosomes and even portions of the same chromosome. But 1cM equates to a 1% chance that a crossover will occur between any two loci on a chromosome at the next meiosis. That works out to being very roughly about once every 1 million base pairs. One percent of 6800cM is, ta dah!, 68. Multiply the CoR estimated average percentage of sharing by 68, and you'll get the figures used by ISOGG and just about everyone else for genealogy. For example: 2nd cousins would expect to share 3.125%; 3.125x68=212.5cM.
Lastly, the final column is quite admittedly a wing-and-a-prayer SWAG. That's why it's noted as "roughly approximated" and "extrapolated." All that's going on there is use of the numbers from the Brenna Henn research paper and guessing at what might be the equivalent chance of same DNA matching by evaluating, as a proportion, both Henn's probability number of detecting any matching DNA, the expected percentage of total DNA shared, and the expected percentage of sharing the same DNA from the same ancestor. The table was done over two years ago, and admittedly I should probably have left the first two cells in that column blank. I should go back and revise that; though it will take some explanation in the footnote to explain why I did. And the Coronapocalypse still has me working 16-hour days so...maybe later.
Absolutely zero empirical studies to support the percentages in the last column. But the point--at the time--was to show that if you had only x% chance of an nth cousin sharing any detectable DNA at all with you, that we need to be cognizant that there's an even slimmer chance that any two cousins will share the same detectable DNA inherited from the same ancestor.