track: sets of levels, mean of numeric variable and distribution of outcomes @param row

translate into notation of other encoding @param standardEncodings (older encoding that we are adapted from, without any vector->vector translations)

essentially the same loop structure as VariableEncodings.formatSoln. So double calculating here confirms the results are coming out correctly. @param

essentially the same loop structure as VariableEncodings.formatSoln. So double calculating here confirms the results are coming out correctly. @param

@param base how many indices to skip (use to get per-outcome data, so base will be a multiple of encoding dimension) @param x vector soln @return map

@param base how many indices to skip (use to get per-outcome data, so base will be a multiple of encoding dimension) @param x vector soln. @return eff

can use- but approximation adds noise in gradient and such @param r

run complete optimization @param f vector function to maximize @param x initial start @param maxRounds maximum number of rounds to try (approximate) @

Invariant: key set calls all keys ever observed (even those with net-zero count) @param key @param delta

Invariant: key set calls all keys ever observed (even those with net-zero count)

Take data from row and paste into vec @param row @param vec

a (or its negation) is symmetric positive semi-definite TODO: more sensitive relative norm conditions

@return true if want regularization (all but const terms should set this)

only write summing terms (fx,gx,hx) into Text,WritableDouble format @param v @param context @throws InterruptedException @throws IOException