Different time stamps using datestr and datetime

Hi,

I have used accumarray to minutely median average a vector, s_temp using a datenum vector time_temp:

[Yr,Mt,Dy,Hr,Mn,~] = datevec(time_temp(1));
t_zero = datenum(Yr,Mt,Dy,Hr,Mn,0);
[Yr,Mt,Dy,Hr,Mn, ~] = datevec(time_temp(end));
t_end = datenum(Yr,Mt,Dy,Hr,Mn+1,0);
tbin = (t_zero:1/24/60:t_end).';
[~, idx] = histc(time_temp, tbin);
sz = [max(idx) 1];
s_median = accumarray(idx, s_temp, sz, @median, NaN);

Then I have tested the timetable function to do the same (out of curiosity):

ts = timetable(datetime(time_temp,'ConvertFrom','datenum'),s_temp);
TS = retime(ts,'minutely',@median); 

The estimated median values unfortunately disagree occassionally, and I have realized that the problem is related to the datetime function. It seems like it rounds the datenum value to the seventh decimal place. See the example below:

RunDate=7.372784756944444e+05;
datestr(RunDate)
datetime(RunDate,'ConvertFrom','datenum')

Is there a way to come around this problem or anybody else who has encountered it?

Grateful for any help!

Answer 1

[Walter Roberson]

The best numeric approximation for that minute is the next representable number after RunDate, roughly 1e-10 larger. You would not be able to see the difference in low bit with any of the default numeric formats, and would need to use num2strexact from the file exchange for Windows, or use fprintf with a larger number of decimal places for Mac or Linux. In short, your literal constant for RunDate would need one more digit to force the bottom bit to be correct when converted to binary.

RunDate is

737278.475694444379769265651702880859375

The best datenum for the minute is

737278.4756944444961845874786376953125

[Walter Roberson]

You could consider using

dateshift(t, 'minute', 'nearest')

[Ylva Ericson]

Thanks again Walter, it works very well with dateshift:

ts = timetable(dateshift(datetime(time_temp,'ConvertFrom','datenum'),'start','second','nearest'),s_temp);
TS = retime(ts,'minutely',@median);

This also made it clear that the accumarray code had some flaws too. However, if the time_temp and tbin are recalculated and rounded to seconds (the original data are in whole seconds..) before binning with histc the results will agree.

Best,

Ylva