Re: Surprising results from tests intended to show the difference in semantics between transaction_timestamp(), statement_timestamp(), and clock_timestamp()

I'm surprised by the results that I describe below. Please help me interpret them.

Briefly, why does transaction_timestamp() report a later value than statement_timestamp() when they're both invoked in the very first statement after "start transaction". (They report identical values in an implicitly started txn.)

  transaction_timestamp() as txn_11,
  statement_timestamp()   as stm_11,
 
  t_finish       in timestamptz,
  t_start        in timestamptz,
  expected_diff  in double precision)
 
select expected_tstz_diff_error('txn_11 - stm_11 = 0',            :s1_txn_11, :s1_stm_11, :zero);

txn_11 - stm_11 = 0            : error: -0.1 ms

  err      constant double precision not null :=
             ( (extract(epoch from t_finish) - extract(epoch from t_start)) - expected_diff )*dp_1000;