Telecom timing: masks, sources and presets
This page lists every limit the telecom-timing scenario kind checks, where each one
comes from, and what was left out. The code is src/telecom_timing.rs. The reference
scenario is scenarios/telecom-prtc-holdover-24h.toml, and the ingestion example is
scenarios/telecom-tie-ingest.toml.
Abbreviations. ITU-T is the International Telecommunication Union Telecommunication Standardization Sector. GNSS is Global Navigation Satellite System. TE is time error, the difference between a clock's time and the reference time. max|TE| is the largest absolute time error. MTIE is maximum time interval error: the largest peak-to-peak time error inside any observation window of length τ (tau). TDEV is time deviation, a root-mean-square measure of time wander at averaging time τ. cTE is constant time error (the average offset). dTE is dynamic time error (the part that moves). TE_L and dTE_L are the time error after a first-order low-pass filter with a 0.1 hertz (Hz) bandwidth; dTE_H is the part above that filter. TE_HO is time error in holdover. PRTC is primary reference time clock. ePRTC is enhanced PRTC. T-BC is telecom boundary clock and T-TSC is telecom time slave clock. PTP is the Precision Time Protocol. PPS is pulse per second. OCXO is oven-controlled crystal oscillator. CSAC is chip-scale atomic clock. CSV is comma-separated values. ADEV is Allan deviation (σ_y), the standard measure of frequency stability. ns is nanoseconds and µs is microseconds. ppb is parts per billion (a fractional frequency of 1e-9), and MHz is megahertz. Amd. is an amendment to a Recommendation, and Y.1367 and the like are the same Recommendation's number in the ITU-T Y series.
What this is. The estimators (MTIE and TDEV) are checked against the third-party
allantools package and are VALIDATED for that. The mask tables are transcriptions of
the Recommendations below. A synthetic holdover is a model built from public datasheet
figures. The kind is MODELLED overall. It is not a conformance test, and a PASS here is
not a certificate.
Sources read#
All ITU-T texts were downloaded from itu.int on 2026-09-25. The newest edition that could be read in full was used.
| Recommendation | Edition read | Status on itu.int |
|---|---|---|
| ITU-T G.8272/Y.1367, Timing characteristics of primary reference time clocks | 07/2025 | in force |
| ITU-T G.8272.1/Y.1367.1, Timing characteristics of enhanced primary reference time clocks | 2024 Amd. 1 (07/2025), a complete-text publication | in force |
| ITU-T G.8273.2/Y.1368.2, Timing characteristics of telecom boundary clocks and telecom time synchronous clocks for use with full timing support from the network | 2023 Amd. 2 (11/2025), complete text | in force |
| ITU-T G.8271.1/Y.1366.1, Network limits for time synchronization in packet networks with full timing support from the network | 2022 Amd. 3 (05/2025), complete text | in force |
| ITU-T G.8271/Y.1366, Time and phase synchronization aspects of telecommunication networks | 03/2020 and its Amd. 1 (08/2024) | in force |
Two newer amendments exist but could not be read: G.8272 (2025) Amd. 1 (08/2026) and G.8272.1 (2024) Amd. 2 (08/2026). Both are prepublished and restricted to ITU members (Telecommunication Information Exchange Service, TIES, accounts). If they change any figure below, this page and the code do not yet reflect it.
How the numbers were checked. Text extraction from these Portable Document Format (PDF) files drops the Greek letter τ and the micro sign µ, so "0.275 × 10⁻³τ" comes out as "0.275 × 10⁻³". Every row of G.8272 Tables 1 to 4, G.8272.1 Tables 1 to 5, G.8271 Table 1 and G.8271.1 Table 7-1 was therefore read again from the rendered page image. G.8273.2 Tables 7-1 to 7-5 and G.8271.1 Tables 7-2 and 7-3 were read from extracted text in which τ survived. Where a table implies continuity at a breakpoint, a unit test checks it.
Limits implemented#
τ is in seconds. An interval written "0.1 < τ ≤ 273" is open at 0.1 s and closed at 273 s, exactly as in the table, and the code keeps each end the same way.
ITU-T G.8272 (07/2025): PRTC-A and PRTC-B, locked mode#
| Quantity | PRTC-A | PRTC-B | Source |
|---|---|---|---|
| max|TE| | 100 ns | 40 ns | clause 6.1 |
MTIE (clause 6.2):
| Mask | Limit | Interval | Source |
|---|---|---|---|
| PRTC-A | 0.275 × 10⁻³ τ + 0.025 µs | 0.1 < τ ≤ 273 | Table 1 |
| PRTC-A | 0.10 µs | τ > 273 | Table 1 |
| PRTC-B | 0.275 × 10⁻³ τ + 0.025 µs | 0.1 < τ ≤ 54.5 | Table 2 |
| PRTC-B | 0.04 µs | τ > 54.5 | Table 2 |
TDEV (clause 6.2):
| Mask | Limit | Interval | Source |
|---|---|---|---|
| PRTC-A | 3 ns | 0.1 < τ ≤ 100 | Table 3 |
| PRTC-A | 0.03 τ ns | 100 < τ ≤ 1 000 | Table 3 |
| PRTC-A | 30 ns | 1 000 < τ < 10 000 | Table 3 |
| PRTC-B | 1 ns | 0.1 < τ ≤ 100 | Table 4 |
| PRTC-B | 0.01 τ ns | 100 < τ ≤ 500 | Table 4 |
| PRTC-B | 5 ns | 500 < τ < 100 000 | Table 4 |
Clause 6.2 also states that the minimum measurement period for TDEV is twelve times the integration period (T = 12τ). The kind applies that rule to every TDEV it reports, so TDEV is only given for τ up to one twelfth of the record. G.8272 states no holdover requirement for a PRTC.
ITU-T G.8272.1 (2024) Amd. 1 (07/2025): ePRTC#
Locked mode:
| Quantity | Limit | Interval | Source |
|---|---|---|---|
| max|TE| | 30 ns | — | clause 6.1 |
| MTIE | 4 ns | 0.1 < τ ≤ 1 | Table 1 |
| MTIE | 0.11114 τ + 3.89 ns | 1 < τ ≤ 100 | Table 1 |
| MTIE | 0.0375 × 10⁻³ τ + 15 ns | 100 < τ ≤ 400 000 | Table 1 |
| MTIE | 30 ns | τ > 400 000 | Table 1 |
| TDEV | 1 ns | 0.1 < τ ≤ 30 000 | Table 2 |
| TDEV | 3.33333 × 10⁻⁵ τ ns | 30 000 < τ ≤ 300 000 | Table 2 |
| TDEV | 10 ns | 300 000 < τ < 1 000 000 | Table 2 |
ePRTC-A holdover, time error (clause 8.2.1, Table 3). L is the time the ePRTC was locked before the loss, in days, and t is the time since the start of holdover, in seconds. The limit rises linearly from 30 ns to 100 ns over the holdover period H.
| Locked duration L | Holdover period H | Limit on |Δx(t)| |
|---|---|---|
| L < 6 days | 0 < t ≤ 70 000 s (0.81 days) | 30 + 1.000 × 10⁻³ t ns |
| 6 days ≤ L ≤ 40 days | 0 < t ≤ L · 86 400 s | 30 + 70 t / (L · 86 400) ns |
| L > 40 days | 0 < t ≤ 3 456 000 s (40 days) | 30 + 2.025463 × 10⁻⁵ t ns |
ePRTC-A holdover, wander (clause 8.2.2):
| Quantity | Limit | Interval | Source |
|---|---|---|---|
| MTIE | 4 ns | 0.1 < τ ≤ 1 | Table 4 |
| MTIE | 0.11114 τ + 3.89 ns | 1 < τ ≤ 100 | Table 4 |
| MTIE | 0.0375 × 10⁻³ τ + 15 ns | 100 < τ ≤ 10 000 | Table 4 |
| TDEV | 1 ns | 0.1 < τ ≤ 10 000 | Table 5 |
Clause 8.3.1 sets the default maximum holdover time error, max|TE_HO|, at 100 ns. The kind uses it as a default time-to-exceed budget.
ITU-T G.8273.2 (2023) Amd. 2 (11/2025): T-BC and T-TSC#
| Quantity | Class A | Class B | Class C | Class D | Source |
|---|---|---|---|---|---|
| max|TE| (unfiltered) | 100 ns | 70 ns | 30 ns | for further study | Table 7-1 |
| max|TE_L| (0.1 Hz low-pass) | — | — | — | 5 ns | Table 7-2 |
| cTE range | ±50 ns | ±20 ns | ±10 ns | for further study | Table 7-3 |
| dTE_L MTIE, constant temperature | 40 ns | 40 ns | 10 ns | for further study | Table 7-4, m ≤ τ ≤ 1 000 |
| dTE_L TDEV, constant temperature | 4 ns | 4 ns | 2 ns | for further study | Table 7-5 |
In Table 7-5, classes A and B are written m < τ ≤ 1 000 and class C is written m ≤ τ ≤ 1 000. m is 1/16 s for a 16 packet-per-second PTP stream and 1 s for a 1 PPS output. The kind treats the input as a 1 PPS time interval error, so m = 1 s. Table 7-3 Note 1 says cTE is estimated by averaging the time error over 1 000 s. The kind reports the worst absolute mean of consecutive 1 000 s blocks.
ITU-T G.8271.1 (2022) Amd. 3 (05/2025): network limits at reference point C#
All three are deployment case 1 and apply after a first-order 0.1 Hz low-pass filter.
| Clause | max|TE_L| | MTIE limit | Interval | Source |
|---|---|---|---|---|
| 7.3.1 (class 4 applications) | 1 100 ns | 100 + 75 τ ns | 1.3 < τ ≤ 2.4 | Table 7-1 |
| 277 + 1.1 τ ns | 2.4 < τ ≤ 275 | Table 7-1 | ||
| 580 ns | 275 < τ ≤ 10 000 | Table 7-1 | ||
| 7.3.2 (enhanced network) | 600 ns | 37.39 + 9.7 τ ns | 1.3 < τ ≤ 2.4 | Table 7-2 |
| 55.27 + 2.25 τ ns | 2.4 < τ ≤ 20.2 | Table 7-2 | ||
| 90.22 + 0.52 τ ns | 20.2 < τ ≤ 211.11 | Table 7-2 | ||
| 200 ns | 211.11 < τ ≤ 10 000 | Table 7-2 | ||
| 7.3.3 (PRTC in the access network) | 100 ns | 0.0475 τ + 25 ns | 1 < τ < 400 | Table 7-3 |
| 44 ns | 400 < τ ≤ 10 000 | Table 7-3 |
Table 7-3 leaves τ = 400 s itself without a limit (one row ends "< 400", the next starts "400 <"). The kind reports no limit there rather than choosing one. Reference point A (clause 7.1) takes its limits from G.8272, so the PRTC masks above cover it.
Budgets (time to exceed)#
| Budget | Value | Source |
|---|---|---|
| ePRTC default max|TE_HO| | 100 ns | G.8272.1 (2024) Amd. 1 (07/2025) clause 8.3.1 |
| Network holdover allocation, short GNSS interruption | 400 ns | G.8271.1 (2022) Amd. 3 (05/2025) Table V.1, failure scenario (b), "re-arrangements and holdover in the network". An example allocation, not a requirement. |
| Network limit at reference point C | 1 100 ns | G.8271.1 clause 7.3.1. This is a filtered quantity; as a budget it is compared unfiltered. |
| End application, accuracy class 4 | 1.5 µs | G.8271 (03/2020) Amd. 1 (08/2024) Table 1, class 4 |
The classic "1.5 µs end to end" figure. It is G.8271 Table 1's class 4 time error requirement, stated against a common reference at the end application. G.8271.1 Table V.1 shows how a network can allocate it in normal operation: 100 ns for the PRTC/telecom grandmaster, 200 ns for the dynamic time error of the chain, a constant time error for the clocks and links, 1 100 ns in total at reference point C, then 250 ns for rearrangements and short holdover in the end application and 150 ns for the end application's own noise, 1 500 ns at reference point E. The same table's failure scenario (c), "long holdover periods, e.g., 1 day", gives 3 350 ns at C and 3 500 ns at E (its Note 5 says exceeding 1 500 ns may degrade service). So neither Recommendation states a "1.5 µs over 24 hours of holdover" requirement, and the kind does not implement one.
Not confirmed or not implemented#
Nothing below is implemented as a mask. None of it was filled in by estimate.
- Marked "for further study" in the Recommendations: G.8273.2 class D max|TE|, cTE, dTE_L MTIE and TDEV, dTE_H and holdover; ePRTC-B holdover (G.8272.1 clause 8.2.1); ePRTC-A holdover MTIE and TDEV beyond 10 000 s (Table 4 Note 1, Table 5 Note 2); ePRTC phase discontinuity (clause 7); T-BC/T-TSC phase/time holdover with both the PTP and the physical-layer inputs lost (G.8273.2 clause 7.4.2.1); the TDEV network limit at reference point C (G.8271.1 clauses 7.3.1 to 7.3.3); network limits for deployment case 2 at point C.
- Transcribed but not implemented in this version: G.8273.2 Table 7-6 (MTIE with variable temperature: 40 / 40 / 10 ns up to 10 000 s), Table 7-7 (dTE_H peak to peak: 70 / 70 / 30 ns, which needs a 0.1 Hz high-pass filter), Table 7-10 (holdover MTIE with physical-layer frequency assistance), relative time error (Tables 7-8 and 7-9), and the dTE_H limits of G.8271.1 clauses 7.3.2 (50 ns) and 7.3.3 (70 ns).
- Not applied: the moving-average filter of at least 100 samples that G.8272 and G.8272.1 state for PTP interfaces. The input is treated as a 1 PPS time interval error, which those Recommendations measure without a filter.
- Unreadable: the two restricted 2026 amendments named under "Sources read".
- Editorial oddity: in G.8272.1 Amd. 1 the head of Table 3 carries an inserted "TDEV limit [ns] 1 / 0.1 < τ ≤ 10 000" row in amendment markup, above the time-error columns. It is not treated as part of the time-error envelope. The same 1 ns TDEV is Table 5 and is implemented there.
The measurement filter#
G.8273.2 and G.8271.1 define their filtered limits through a first-order low-pass filter with a 0.1 Hz bandwidth. The kind applies the discrete smoother y_k = y_{k−1} + α (x_k − y_{k−1}), α = 1 − exp(−2π · 0.1 · Δt), started at the first sample. It is an approximation of test equipment, not a copy of any instrument. It is applied only when the sample interval Δt is 1 s or less. A coarser record is reported NOT-EVALUATED against a filtered mask instead of being filtered badly.
Oscillator presets (MODELLED)#
Every figure is the datasheet's own, read on 2026-09-25.
| Preset | Device and document | ADEV maxima used (τ: σ_y) | Aging | Temperature |
|---|---|---|---|---|
ocxo |
Microchip (Vectron) OX-208, 10 MHz; datasheet Rev 12-1-2021 | 1 s: 5e-12; 10 s: 8e-12; 100 s: 1e-11; 1 000 s: 5e-11 | ±0.15 ppb/day after 72 h (f ≤ 10 MHz) | ±0.4 ppb, 0 °C to +70 °C, referenced to +25 °C |
rubidium |
Microchip 8040C Rubidium Frequency Standard, standard performance; DS00003047A (2/20) | 1 s: 3.0e-11; 100 s: 3.0e-12 | <5e-11/month after 30 days | temperature coefficient <3E-10, 0 °C to 50 °C |
caesium |
Microchip 5071A, high-performance tube; DS00002980D (5/25) | 1 s: 5.0e-12; 10 s: 3.5e-12; 100 s: 8.5e-13; 1 000 s: 2.7e-13; 10⁴ s: 8.5e-14; 10⁵ s: 2.7e-14; 5 days: 1.0e-14; 30 days: 1.0e-14 | no rate published; lifetime change ≤5.0e-14 | ±8.0e-14 versus environment (0 °C to 50 °C, humidity, magnetic field, shock) |
csac |
Microchip SA.45s CSAC, options 001 and 003; DS00002985D (5/23) | 1 s: 3e-10; 10 s: 1e-10; 100 s: 3e-11; 1 000 s: 1e-11 | <9e-10/month after 30 days | ±5e-10, −10 °C to 70 °C |
Document links: OX-208, 8040C, 5071A, SA.45s.
Left out on purpose: the 8040C's 10 s row, which the datasheet prints as "<1.0 x 10^11" (a misprinted exponent), and the 5071A's rows below 1 s. The OX-049 was considered for the OCXO and not used, because its aging row cannot be read unambiguously from its misaligned table.
How the figures become a model. These are this engine's choices, and each one is reported in the run output:
- Noise fit. The white, flicker and random-walk frequency-noise levels are a non-negative least-squares fit to the ADEV maxima, then scaled up until the model is at or above every point. The datasheet figures are maxima, so the model envelopes them.
- Flicker floor. The floor is never below the ADEV at the longest listed τ. A datasheet stops listing where its maker stops promising improvement, so the model does not let the clock improve beyond it. This can only make a holdover worse.
- Aging. A monthly figure is converted at 30 days per month and applied as a linear frequency drift from the moment of loss. The caesium preset has no aging.
- Temperature. The stated bound is read linearly over half the operating span (for example ±5e-10 over −10 °C to 70 °C gives 1.25e-11 per kelvin), then driven by a sinusoidal temperature profile whose amplitude and period are scenario inputs. Nothing on a datasheet says the dependence is linear; this is a worst-case reading.
- Before the loss the clock is disciplined and its time error is white phase noise with a stated 1-sigma value.
What the reference scenario shows. At the committed inputs (rubidium, ±2 K daily temperature swing, 24 h holdover after 1 h locked) the record reaches a largest absolute time error of 603.2 ns. It crosses the 100 ns ePRTC budget 8 206 s after the loss and the 400 ns allocation after 25 713 s, and stays inside 1 100 ns and 1.5 µs for the whole day. It fails the ePRTC-A holdover envelope 3 627 s after the loss, which is expected: that requirement assumes a caesium-class reference. The temperature term alone can reach 660 ns, so at these inputs it matters more than aging (72 ns at the end of the day).
Validation#
tests/telecom_timing_reference.rs checks the kind's MTIE and TDEV against
allantools 2024.06 on tests/fixtures/telecom_timing/holdover_te_series.csv, a
2 048-sample CSAC holdover series with aging, flicker and temperature. On 17 MTIE and
12 TDEV averaging factors, MTIE agrees exactly and TDEV agrees to 1.1e-15 relative. The
same curves are read back out of a full run that ingests the CSV. This validates the
estimators on that record. It does not validate the synthetic holdover, the masks or
the presets, which are MODELLED rows in docs/VERIFICATION-MATRIX.md.