.@ Tony Finch – blog


There is a new sf series on Fox in the US called "Terra Nova". The premise is that the protagonists have travelled back 85 million years in time from 2149 to escape environmental collapse. Back then in the time of the dinosaurs the length of the day was one or two percent shorter than it is today. This led Daniel Tobias to ask the leapsecs mailing list how time should be handled by the colonists. A Mars colony would also have to answer a similar question, because they would have to cope with a day that is a couple of percent longer.

The problem is that our base unit of time does not conveniently divide the length of the day. This is inevitable however we choose to fix our units, because the length of the day is not constant. Travellers to other times or planets have to deal with an exaggerated version of the problem, but it's also true on present-day Earth. The length of the second is based on the length of the day over 100 years ago, and it is now off by about 10-8. As well as tidal slowing, there are also periodic and random variations in the length of day, which also happen to be on the order of 10-8.

There are about five ways of dealing with this problem which I'll divide into two and a half categories..

Digital schemes

When we're dividing up longer time periods into days we use calendars. We can also view schemes for dividing days into seconds as simple calendars.

Observational calendars

The simplest calendars are based on observation of astronomical phenomena. For instance in the Islamic calendar, when the new moon occurs the day counter is reset and the month counter is incremented.

Our current system of leap seconds in UTC is essentially an observational calendar, where what is being observed is the difference between mean solar time (aka UT1) and atomic time. Leap seconds are inserted to keep this difference less than 0.9 seconds.

A time or space colony could use a similar system, though their day is unlikely to be close to a round number of seconds in length, so they will probably need to switch back and forth frequently between short and long days. The disadvantage of observational calendars is that they are not predictable, so it is not possible to schedule events in the future with any precision.

Arithmetical calendars

If you have enough astronomical sophistication to measure periods of rotation and revolution accurately, you can set up a calendar with fixed arithmetic rules. For example, the Julian and Gregorian calendars.

There has been a lengthy discussion over the last ten years on the possible discontinuation of leap seconds. This would turn UTC into a very simple arithmetic calendar.

Arithmetic calendars tend to drift out of sync, either because of errors in their initial setup, or because the relevant periods are no longer what they were. Reforming a calendar to fix it is incredibly painful - think of the 350 year transition period required by the Gregorian reform.

There is a work-around available when you are dealing with a mis-match between the nominal length of day and the actual length of day, which is not available for normal calendars. Provided the difference is small enough, less than about 10-5, you can accommodate mismatches by adjusting timezone boundaries. This is easy to cope with if your timezone system is already handling random political fluctuations, and will probably happen without the need for any central co-ordination.

The difficulty with this scheme is that time of day is not a good approximation of planetary angle of rotation relative to the sun, so astronomical and navigational systems will need a source of UTC-UT1 data (aka DUT1).

Fractional schemes

In Kim Stanley Robinson's Mars trilogy the colonists stop the clocks between 24:00 and 00:00 to allow for the extra forty minutes in a Martian day beyond 24 hours. (I presume they don't actually stop the clocks inside their support systems since they still need useful telemetry logging during this "timeslip", amongst other things.) There is a similar arrangement in David Weber's Honor Harrington books.

You could perhaps allow for a partial second at the end of each day, to make the nominal length of day exactly match the actual length of day. The disadvantage is this would cause an awkward glitch in time and frequency reference broadcasts.

ETA: The fractional second idea is another trick that time-of-day calendrical systems can use which more common calendars can't, because the second is an artificial unit of time not a measurement of the position of a celestial body.

Analogue schemes

Instead of having a variable number of fixed-length seconds in a day, we can have a fixed number of variable-length seconds in a day. There are a couple of ways of doing this.

Rubber seconds

In the 1960s, time and frequency reference broadcasts were matched to the length of day using a combination of frequency adjustments and occasional jumps of 0.1 or 0.2 seconds. This allowed them to track UT2 more closely than modern UTC tracks UT1.

However it had the disadvantage of requiring difficult adjustments to the broadcast equipment make the frequency changes, and it made it more difficult for users of the broadcasts to obtain a precise reference frequency. So it was abandoned in favour of the simpler UTC scheme.

A modern variant of this is smoothed leap seconds or leap smear, where rubber seconds are used temporarily to avoid glitches caused by leap seconds.

Two timescales

All the above schemes start with a timescale based on seconds, and try to accommodate the variable length of days within this timescale (except for rubber seconds which are the other way round). Instead of trying to reconcile the irreconcilable, we could instead work with two separate timescales.

For precise time and frequency applications, establish an atomic timescale that is as stable as possible. It might be sensible to use a different base unit, say the Planck time, or rename the atomic second to say the essen, in order to avoid confusion with subdivisions of the day. Instants in this timescale should be labelled as a count of seconds since an epoch, not in YYYY-MM-DD HH:MM:SS form to avoid confusion with time of day.

For civil time of day, use UT1. All calculations involving civil time should be done inside this system, ignoring its relationship to atomic time. This timescale is not suitable for high precision applications, since there's an inherent instability of about 10-8. It retains all the properties of pre-atomic time: fixed number of seconds per day and synchronization with Earth rotation.

Time and frequency broadcasts should be based on the atomic timescale. In order to obtain civil time, these broadcasts should include the atomic time when the current civil day started, and its length in atomic seconds (and possibly also the current rate of change of the length of day).

In terms of the POSIX clock_gettime() interface, the atomic timescale roughly corresponds to CLOCK_MONOTONIC, and civil time corresponds to CLOCK_REALTIME i.e. time_t seconds since the epoch.

Back in the real world

I think the two timescales arrangement is the best way to model what is actually going on. However it is difficult to see how it could be deployed. It requires us to separate out the high-precision time and civil time functions of a lot of critical systems, such as the MSF / DCF77 / WWV / etc. time signals, GPS and other navigation systems, and NTP.

We should find out within the next year what will happen to leap seconds. I think time geeks underestimate the problems they cause by breaking deeply embedded cultural assumptions about time. The notation we use for time of day is many thousands of years old, and changing it is possibly a reform of Gregorian scope. So I tend to be in favour of abolishing them even though it'll cost a lot for astronomers to fix. Cheaper than fixing POSIX time, though.

(Previously: iCalendar is wrong.)