Pulling the GPS Antenna on Purpose: Measuring GPSDO Holdover Drift and What Eight Hours of Coast Does to a 70 MHz Carrier

By | Thursday, October 15, 2026
0 Comment

Pulling the GPS antenna is not a fault. It is a controlled experiment, and like any experiment it is only worth running if you decide in advance what you are measuring, what you are measuring it against, and what statistic you will report. The 70 MHz carrier in your modem or upconverter is not the thing under test. It is a derived view of the 10 MHz reference, and if you report its wander as if it were an independent property you have confused the measurement with the reference.

This article is about how to run a deliberate holdover test on a GPS-disciplined oscillator, how to read the result honestly, and how to translate reference phase error into carrier phase error without overclaiming. The numbers that matter are fractional frequency, phase error in nanoseconds, and averaging time in seconds. The carrier frequency is a multiplier, not a new oscillator.

What a holdover test can and cannot tell you

A GPSDO combines a stable local oscillator with a GPS receiver optimized for timing. The receiver provides long-term accuracy; the local oscillator provides short-term stability. During normal operation the disciplining loop steers the oscillator toward GPS time. When you disconnect the antenna, the loop opens and the oscillator coasts on its own. The question is how far it walks before you reconnect.

What the test can tell you: the short-term stability of the local oscillator, the size of the initial frequency offset at the moment of disconnect, and whether the disciplining loop had settled to a stable steering value before you pulled the connector. What it cannot tell you: anything about aging over weeks or months, anything about long-term drift beyond the holdover interval, and anything about the GPS receiver’s performance while it is disconnected. Eight hours is a plausible site power event. It is not a characterization of aging.

The most useful output of a deliberate antenna-pull test is not a single stability number. It is a documented procedure and a recorded curve that can be repeated after any change to the site, the antenna, or the oscillator. The value is in the repeatability, not the first result.

Define the measurand before you disconnect

Two quantities are easy to conflate. Phase error is the time offset between your reference output and an ideal reference, measured in nanoseconds or picoseconds. Fractional frequency offset is the rate of change of that phase error, dimensionless, often expressed as parts in 109 or 1012. A holdover test produces both, and they answer different questions.

If you want to know how much the reference wanders randomly over eight hours, you want a stability statistic. If you want to know how far it has walked by hour eight, you want the phase record itself, or the frequency record with the linear trend identified and stated. Reporting a single Allan deviation number without saying whether the linear drift was removed is not auditable.

The 70 MHz carrier is downstream of the reference. If the reference has a fractional frequency error of y, the carrier has the same fractional frequency error, and the phase error in radians at the carrier is 2π times the carrier frequency times the accumulated time error. A 10 ns time error at the reference is a 10 ns time error at 70 MHz. The carrier does not amplify the time error; it only makes it easier to see if you are measuring phase at the carrier rather than at the reference. That is a measurement convenience, not a separate phenomenon.

Pick the right statistic

Allan deviation is the standard time-domain statistic for oscillator stability. NIST defines it as a non-classical statistic used to estimate stability, and notes that it is the square root of the Allan variance. The reason it exists is that the ordinary standard deviation assumes stationary, white data. A drifting oscillator is not stationary. Feed a day of frequency data from a drifting clock into an ordinary standard deviation and the answer depends on how long you measured; longer records give larger numbers, forever. Allan deviation uses the difference between adjacent frequency averages, which cancels constant drift and yields a stable, meaningful number even for non-stationary sources.

On an Allan deviation plot, stability improves as the averaging period gets longer, up to a point. NIST describes that point as the noise floor, where more averaging no longer improves the result and the remaining noise consists of nonstationary processes such as aging or random walk. The curve has a characteristic bathtub shape on log-log axes: it falls as averaging suppresses random noise, flattens at a floor, then rises again as drift and aging take over. The bottom of the curve marks the device’s best stability and the optimal averaging time for your measurement.

If the oscillator has strong linear frequency drift, the Hadamard deviation is the more honest statistic. It uses second differences instead of first, making it insensitive to linear frequency drift, and it is the right tool for rubidium and cesium standards that age. For an OCXO over eight hours, the drift term may be small enough that Allan deviation with the linear trend removed and stated is sufficient. For a rubidium standard, Hadamard is the safer choice. The point is not which statistic is superior in the abstract; it is that the statistic must match the question and the oscillator.

One more property matters for holdover. Allan deviation can be used as a predictor of time drift because, by definition, it is a measure of the expected deviation of time drift. If the measured Allan deviation at a one-day averaging time is 1 × 10−13, the prediction, within one standard deviation, is that you will be within about 9 ns tomorrow. That is the number that connects the stability plot to the phase error you will see on the carrier.

Instrumentation reality check

A holdover test without a stated reference is not a measurement. The result is only meaningful relative to something more stable than the device under test over the holdover interval. If the reference is comparable or worse, the test measures the reference as much as the device under test.

What does a phase-comparison setup actually resolve? A documented measurement of a GPSDO using an SDR-based phase comparator against a better GPSDO showed an RMS phase difference of 14 ns with occasional differences up to 80 ns, described as in line with the usual performance of a non-timing-grade GNSS receiver’s timepulse. In that same measurement, for averaging times greater than 100 s the GPS disciplination kicked in and the Allan deviation decreased by an order of magnitude per decade, while between 1 s and 100 s the Allan deviation was around 1 × 10−10, dominated by the TCXO’s stability. That is a useful reference point for what a modest GPSDO looks like when you measure it properly. It also shows where the measurement system’s own noise floor sits: if your phase comparator has 14 ns RMS noise, you cannot resolve a 1 ns holdover drift without averaging, and you need to say so.

The reference should be at least an order of magnitude more stable than the device under test over the holdover interval. If you are testing an OCXO-based GPSDO over eight hours, a rubidium standard or a cesium standard is a defensible reference. If you are testing a rubidium GPSDO over eight hours, you need something better than another rubidium. If you do not have a better reference, you can still run the test, but you must report the reference’s own stability alongside the result and state that the measurement is reference-limited.

Reading the curve

The Allan deviation plot for a GPSDO in holdover has three regions. At short averaging times, the curve falls as white frequency noise averages out. The slope tells you the noise type: a slope of τ−1/2 is white frequency noise, thermal-noise limited, and averaging helps. A flat section is flicker frequency noise, the intrinsic floor, and averaging beyond it buys you nothing. An upturn at τ+1/2 signals random-walk FM, usually driven by temperature and environment. A steep τ+1 ramp is deterministic aging, which should be removed before analysis.

For an eight-hour holdover, the tau of interest runs from a few seconds to about 28,800 seconds. If the oscillator is ovenized and the ambient is stable, the curve may still be falling or flat at the end of eight hours. If the oscillator is not ovenized, or the oven is marginal, you may see the upturn within the holdover window. That upturn is the signal that temperature is now the dominant term, and no amount of post-processing will remove it.

The disciplining loop changes the picture. During normal operation, the GPS receiver corrects the oscillator’s long-term drift. When you disconnect the antenna, that correction stops. The Allan deviation at long tau will rise because the disciplining is gone. In the documented SDR measurement, the disciplination kicked in for tau greater than 100 s and the Allan deviation decreased by an order of magnitude per decade. In holdover, that decrease is absent. The curve you measure during holdover is the free-running oscillator, not the disciplined system.

What NIST’s 60-day GPSDO characterization shows

NIST characterized four GPSDOs simultaneously using the same antenna over two 60-day measurement intervals, measuring both the 1 pps timing output and the 10 MHz frequency output. The devices included two with OCXO time bases and two with rubidium time bases. The results are worth reading before you trust a marketing holdover number.

NIST states plainly that GPSDO performance differs by a significant amount from device to device, that some models are not suited for all applications, and that manufacturer specifications might provide only a rough indication of actual performance. In the test with known coordinates, none of the devices appeared to lose lock during the 60-day test and nearly all readings were within ±50 ns of UTC(NIST). That is the disciplined performance, not holdover.

The self-survey results are more directly relevant to holdover testing. In the second 60-day interval, each GPSDO performed its own antenna survey. Altitude errors from self-surveyed coordinates correlate directly to time errors at a ratio that can approach 3 ns per meter. One device’s 23.2 m altitude error resulted in a mean time offset of −57.62 ns with respect to UTC(NIST), while the other devices’ mean time offsets were much smaller: −10.33 ns, −8.84 ns, and 1.02 ns. The lesson for holdover testing is direct: if the receiver has not completed a good position survey before you disconnect, you are measuring the survey error, not the oscillator. A holdover test run before a good survey is not a holdover test.

NIST also notes that the estimated uncertainty of the initial delay calibration can overlap the uncertainty introduced by the error in the antenna coordinates. That is a reminder that the cable delay, the splitter delay, and the survey altitude all contribute to the time offset you measure. Record them.

A defensible holdover test procedure

The following procedure is a recommendation, not a standard. It is written for a teleport or hub where an eight-hour holdover is a plausible power event and where the reference distribution feeds modems, upconverters, and test equipment.

Before the disconnect:

  • Confirm the receiver has completed and stored a position survey. Record the surveyed coordinates and the altitude. If the survey is recent, note the date. If the receiver is in position-hold mode, confirm the hold coordinates match the survey.
  • Record the reference used for the measurement and its own stability specification or measured curve over the holdover interval. If the reference is not at least an order of magnitude better than the device under test, state that the measurement is reference-limited.
  • Record the cable delay from the antenna to the receiver, including any splitter or distribution amplifier. NIST’s test used an 18.29 m coaxial cable with a measured delay of 73 ns and a splitter with a group delay of 3 to 4 ns, for a combined estimate of 81 ns. Your numbers will differ; the point is to have them.
  • Record the ambient temperature at the oscillator if it is not in a controlled environment. If the oscillator is ovenized, note the oven status if the equipment reports it.
  • Log the phase difference between the device under test and the reference at a rate fast enough to resolve the shortest tau of interest. For an eight-hour holdover with tau down to 1 s, a 1 Hz logging rate is adequate. For tau down to 0.1 s, log at 10 Hz.

During the disconnect:

  • Disconnect the antenna at a known time and record it. Note whether the receiver reports loss of lock and when.
  • Continue logging phase and, if available, the oscillator’s steering voltage or DAC value. The steering value at the moment of disconnect is the initial condition for the holdover.
  • Record ambient temperature throughout. If the temperature moves by more than a degree or two, note it; that movement will appear in the phase record.
  • Do not adjust anything. The test is a coast, not a tuning exercise.

After the reconnect:

  • Record the time of reconnect and how long the receiver took to reacquire lock and settle. The settling time is part of the operational picture.
  • Compute Allan deviation from the phase record. If the frequency record shows a clear linear trend, compute Hadamard deviation as well, or remove the linear trend from the frequency data and state that you did so.
  • Report the reference, its stability, the logging rate, the holdover duration, the ambient temperature range, and whether the survey was completed before the disconnect. A result without these is not auditable.

Interpreting the result for a 70 MHz carrier

The 70 MHz carrier’s phase error during holdover is the reference’s time error multiplied by 2π times 70 MHz, expressed in radians, or simply the same time error expressed in nanoseconds if you are measuring phase at the carrier. There is no separate carrier stability to characterize. If the reference has walked 100 ns over eight hours, the carrier has walked 100 ns. If the reference has a fractional frequency offset of 1 × 10−11, the carrier has the same fractional offset, which at 70 MHz is a frequency error of 0.7 mHz.

What matters for the link is whether that phase error is within the modem’s tracking range and whether it accumulates faster than the modem can correct. For a typical satellite modem, the carrier recovery loop can track a few hundred Hz of frequency offset and a phase ramp that stays within the loop’s pull-in range. A 0.7 mHz offset at 70 MHz is negligible for the modem. A 100 ns phase step is a small fraction of a symbol at typical symbol rates. The holdover drift of a decent OCXO over eight hours is unlikely to break the link on its own. The risk is not the oscillator; it is the combination of oscillator drift, temperature, and whatever else is happening on the reference distribution.

That is why the test is worth running. Not because the answer is likely to be alarming, but because the answer is specific to your site, your oscillator, and your reference. A confidently wrong number is worse than a long derivation. The derivation here is: measure the reference, state the reference, compute the statistic that matches the question, and report the result with its uncertainty. The carrier will follow.

FAQ

How long should I hold over? Eight hours is a defensible duration for a teleport or hub because it maps to a plausible site power event. It is not long enough to characterize aging. A single eight-hour run says something about short-term holdover; it says nothing about weekly or monthly drift. If you need to know about aging, you need a longer test or a manufacturer’s aging specification, and you should state which you are relying on.

Can I use another GPSDO as the reference? Only if it is significantly more stable than the device under test over the holdover interval. Two identical GPSDOs compared against each other will show the combined instability of both, and you cannot separate them without additional information. If you must use a comparable reference, report the reference’s stability and state that the measurement is reference-limited.

What if I do not have a rubidium or cesium reference? You can still run the test, but the result will be limited by your reference. A GPSDO that is itself disciplined by GPS can serve as a reference during holdover if its own holdover performance is known and better than the device under test. If it is not, the test measures the reference as much as the device. State the limitation.

Should I remove the linear trend before computing Allan deviation? If the frequency record shows a clear linear trend, yes, and state that you did so. If you do not remove it, the Allan deviation at long tau will be dominated by the drift term and will not represent the random instability. The Hadamard deviation is an alternative that is insensitive to linear drift by construction.

Does the 70 MHz carrier need a separate holdover test? No. The carrier is derived from the reference. If you want to verify that the distribution chain from the reference to the carrier does not add phase noise or drift, that is a different test: measure the phase difference between the reference and the carrier at the point of use, with the reference as the common source. That test characterizes the distribution, not the oscillator.

What about temperature? If the oscillator is not ovenized, or if the oven is marginal, temperature is likely to be the dominant term over eight hours. Record ambient temperature and look for correlation in the phase record. If the phase record moves with the temperature record, you have found your limiting factor, and no amount of post-processing will remove it.

How do I know if the receiver had a good survey? Check the reported coordinates against a known survey if you have one, or against a long average of position fixes. NIST’s data shows altitude errors from self-survey can reach tens of meters and translate to time errors at up to 3 ns per meter. If the survey is uncertain, the holdover test is measuring the survey, not the oscillator.

Sources

No related posts.