A triangle can reveal an inconsistency without telling you where it began.
Three radio telescopes observe the same distant source. Each pair measures a delay. Go around the triangle, adding the delays with the right signs, and some errors vanish. If a remainder survives, it tells you something important, but probably less than you first think.
We will build the triangle from signed differences, turn the same data around and check the result, then see why a perfect closure can hide a large station error. Later examples use two triangles that share a baseline, so that a seemingly independent confirmation turns out to be partly repeated evidence.
The article is about the reasoning behind closure delays in very long baseline interferometry, or VLBI. It does not replace the delay model, source imaging or calibration work. Start with the first two chapters if the observatory language is new; jump to the independent case if you already use signed baseline delays.
Geodesy subject guide · Learning that transfers into life · How X Works hub
Choose your reading route
Route 1 · Chapters 1–3: Build a signed triangle
Give each baseline a direction
See which station terms cancel and why the time tags must match.
Route 2 · Chapters 4–5: Read what a remainder means
Locate the type of error before naming a cause
Separate baseline effects from station-like effects and avoid guessing the guilty edge.
Route 3 · Chapters 6–8: Extend to a network
Add a fourth station without inventing independent loops
Use overlapping loops without counting the same information twice.
Route 4 · Chapters 9–10: Distinguish structure from calibration
Test a source explanation against observing geometry
Use source and instrument changes to test competing explanations.
Route 5 · Chapters 11–12: Make a defensible claim
Write a closure check a colleague can replay
Plan a replayable audit and try the reasoning on a different problem.
Full chapter index · Independent worked case · Questions and practical checks
Chapter index
Build a signed triangle
Give each baseline a direction
Why station terms cancel
Align observations before adding them
Read what a remainder means
Locate the type of error before naming a cause
Some structure errors are invisible to one triangle
Extend to a network
Add a fourth station without inventing independent loops
Carry uncertainty through shared baselines
See what a repeated signature can and cannot establish
Distinguish structure from calibration
Test a source explanation against observing geometry
Use calibration evidence that closure does not contain
Make a defensible claim
Write a closure check a colleague can replay
Transfer the reasoning to a survey loop
CHAPTER 1 OF 12 · Build a signed triangle
1. Give each baseline a direction
A baseline connects two observing stations. A signed delay needs an order as well as two station names. For this article, AB means “arrival label at B minus arrival label at A.” Reverse the order and the sign changes. In practice, confirm the correlator or software convention, because another system may store the opposite sign. What matters is that the convention remains consistent around the entire path.
Use a fictional timing ledger in picoseconds: A = 100, B = 140, and C = 125. The directed delays are AB = 40, BC = −15, and CA = −25. Their sum is zero. If you store AC = +25 instead of CA, the closure is AB + BC − AC, again zero. Adding AC with a plus sign would give +50 and create a problem that does not exist in the observations.
The underlying principle is familiar from a walk. A lift takes you up four floors, then down one and a half, then down two and a half. If you finish where you started, the signed changes sum to zero. The example does not require every leg to be small or every pair to agree in sign. It requires each leg to belong to the same directed route.
A useful first audit is to take a single row of real data, write the station order by hand, and compute closure both clockwise and anticlockwise. The values should have opposite signs and the same magnitude when you use the same observations. If not, inspect the orientation handling before reaching for a theory about the sky.
CHAPTER 2 OF 12 · Build a signed triangle
2. Why station terms cancel
Write dAB = tB − tA + qAB. Here t contains quantities assignable to a station at the relevant observation, and qAB is whatever this simple model leaves on the baseline. A corresponding equation holds for BC and CA. When you add all three, each station term appears once with a plus sign and once with a minus sign. Closure is qAB + qBC + qCA.
A point-source geometric delay, evaluated consistently in an idealised common-epoch picture, also has a station-difference structure: it comes from projecting each station location in the same source direction and taking a difference. The station clocks and station-local atmospheric delay terms cancel when the same station contribution appears in both adjoining observations with opposite sign. Actual geodetic processing accounts for the chosen delay and epoch conventions rather than simply substituting these teaching expressions.
Change the ledger by adding +60 ps to station B. AB rises from 40 to 100; BC falls from −15 to −75; CA stays −25. The triangle still sums to zero. Here the closure has correctly ignored a station error of 60 ps. A clock or calibration check that needed to find that error would have to use evidence that closure removes.
This is the central limit. A clean triangle does not prove that the station clocks, atmospheric modelling, source position or baseline geometry are correct. It proves only that the particular directed sum is small at the tested precision. A method is more useful when you can name the kinds of errors it deliberately cannot see.
CHAPTER 3 OF 12 · Build a signed triangle
3. Align observations before adding them
The cancellation proof used one B value in two places. If AB uses the station state at one instant and BC uses it at another, B does not automatically cancel. Real VLBI observations have specific time-tag conventions and can require common-epoch treatment; the International VLBI Service and published closure studies show why the timestamps of the three observables matter.
In a teaching example, B has +60 ps of extra delay when AB is measured but +67 ps when BC is measured. A and C stay as before. Then AB is 100, BC is −82 and CA is −25. The sum is −7 ps, exactly the mismatch between the two appearances of B. It would be wrong to call this a baseline hardware effect without first examining the observation timing.
Time is not the only alignment question. Were the baselines measured for the same source, on comparable spectral definitions and using compatible processing? A group delay estimated over one set of usable frequencies is not guaranteed to respond to a complicated source in the same way as a delay estimated over a different set. Label the scan and band or channel masks. A processing change that removes channels can affect what a comparison means.
If a station is missing from part of a scan, do not turn a non-simultaneous triangle into a simultaneous one by deleting the timestamps in a report. Either use a justified model that puts the measurements on the same reference, with an uncertainty for that step, or state that the triangle cannot be formed cleanly. A missing check is a better result than an invented one.
CHAPTER 4 OF 12 · Read what a remainder means
4. Locate the type of error before naming a cause
Suppose the observed delays are AB = 40, BC = −15 and CA = −13 ps. Closure is +12 ps. One explanation is that CA carries a +12 ps baseline effect. A second is that AB and BC each carry +6 ps. A third is that the data use inconsistent reference times. The triangle narrows the possibilities, but one equation does not determine three separate edge errors or validate the assumptions behind the equation.
A useful distinction is between station-like errors and nonclosing errors. An error is station-like for a particular observation set if its effect on every directed baseline can be expressed as the difference of some station values. Its physical origin does not have to be a clock. In contrast, the part that cannot be represented that way can be exposed by loops. Those are mathematical descriptions of patterns, not automatic explanations of the hardware.
For example, a cable offset in a station system can be station-like in one configuration. A processing mistake that changes the contribution of that station only on a particular baseline is not guaranteed to be station-like. Also, the algebra does not require the two classes to be disjoint physically. One incident may create a station-like component and a nonclosing component, and closure sees only the second.
Ask what diagnostic would distinguish competing stories. Try an independent calibration record for a suspected station term, a re-correlation with the same spectral mask for a processing mismatch, or another source and geometry for a structure hypothesis. Recording the next test is more informative than assigning the entire closure to the edge with the largest ordinary residual.
CHAPTER 5 OF 12 · Read what a remainder means
5. Some structure errors are invisible to one triangle
A radio source with several emitting regions does not always look the same on every baseline. The projected separation and orientation of the stations changes how the signal contributions combine. Group delay is related to how the interferometric phase varies across frequency under the processing convention. Structure can therefore introduce a delay that is not simply a difference of two station numbers.
Not all structure produces a large closure in every triangle. If two long baselines sample very similar parts of the source response and are traversed in opposite directions, their structure contributions can substantially cancel. As a toy example, let the structure terms be +18, −16 and +1 ps around a triangle. The result is only +3 ps, even though two terms are larger than 15 ps.
A pure shift of a point source illustrates another limit. With consistent geometry it changes each baseline delay according to a projection. The change can be represented by station differences and cancels in the triangle. Closure alone is therefore not an absolute source-position measurement. To constrain an astrometric direction, you need data and modelling that retain the position-sensitive component.
To test the lesson, imagine you see +3 ps for one source in one triangle. Before describing that source as point-like, look for a triangle with different projected baseline directions and lengths. If a second triangle with comparable noise shows +19 ps, the first triangle was a weak view of the effect, not a contradiction. If every tested triangle is small, report the coverage you actually had.
CHAPTER 6 OF 12 · Extend to a network
6. Add a fourth station without inventing independent loops
With four stations A, B, C and D, all six pairwise baselines allow several triangles. Store baselines in alphabetical directions. Set CABC = AB + BC − AC; CABD = AB + BD − AD; CACD = AC + CD − AD; and CBCD = BC + CD − BD. Substituting the definitions gives CABC − CABD + CACD − CBCD = 0. The four triangles cannot be four independent checks of the same simultaneous six observations.
Choose three triangles that all use A. For this complete four-station example, their independent directions are enough to describe the nonclosing part of the six baseline values. More generally, in a connected graph with m measured edges and n stations, the number of independent loop directions is m − n + 1. A chain has no loop. Adding one edge that closes a path makes one independent loop possible.
Suppose D is connected only through BD, perhaps because the remaining baselines failed. The ABC triangle can be formed, but no loop checks the BD edge. A table that flags “network closure checked” without showing participation could mislead a reader into thinking every station and baseline was tested. Show which edges are in loops and which are bridges.
Some networks have a useful loop that is not a triangle. If AB, BC, CD and DA are available but AC and BD are missing, a four-edge loop can still test consistency of the directed observations. The claim must match the actual graph, time and source coverage. Adding an edge helps only if it contributes a usable observation with the required alignment.
CHAPTER 7 OF 12 · Extend to a network
7. Carry uncertainty through shared baselines
For a triangle with independent baseline errors of standard deviations 3, 4 and 5 ps, the variance of the closure is 9 + 16 + 25 = 50 ps², so the standard deviation is about 7.07 ps. A +12 ps closure is about 1.7 of these standard deviations. That ratio is a description under the assumptions, not a universal flag threshold.
Now use two closures CABC and CABD sharing AB. Suppose all six underlying errors in a complete four-station example are independent with variance 16 ps². Each triangle has variance 48 ps². Both use AB with a plus sign, so their covariance is 16 ps². If both closures look large, some of that agreement may come from the same noisy measurement.
The difference CABC − CABD removes AB. Its variance is 48 + 48 − 2 × 16 = 64 ps². The sum keeps AB twice and has variance 128 ps². If you treat the triangle values as independent, you get 96 ps² for both and mishandle both conclusions. Reverse one triangle’s orientation and the covariance changes sign; the physical information does not.
The underlying observations can also be correlated by processing and shared signal contributions. Do not assume that an independent-edge model describes every experiment. If a component really is identical and station-like, closure can remove it, but an estimated covariance can still have shared baseline or spectral contributions. Retain the observation and loop matrices so a reviewer can trace those choices.
CHAPTER 8 OF 12 · Extend to a network
8. See what a repeated signature can and cannot establish
A single faulty edge affects every loop containing that edge, with signs dictated by orientation. Suppose an added +9 ps on AB is the only nonclosing disturbance in the four-station model. In the three A-centred triangles, the closure changes are +9, +9 and 0 ps. This pattern is consistent with an AB disturbance. It is also consistent with other edge patterns that differ by station-like terms.
A concrete alternative is to add a station-like +10 ps at B on top of that disturbance. AB gains another +10, while BC and BD each change by −10. The pattern of triangle changes remains +9, +9, 0. Thus even perfect measurements of every closure do not recover the full edge error vector. They recover only its effect in the loop directions.
With an explicit assumption that one edge alone is wrong, comparing signatures may suggest where to look. State the assumption and inspect alternatives, especially in a sparse network. A missing triangle can make two different edge explanations observationally indistinguishable. A candidate diagnosis can be useful for prioritising a recheck without being strong enough to publish as a cause.
A practical check is to simulate one plausible error at a time through the actual loop design, including missing baselines and the covariance you used. Identify which signatures are equal, opposite or almost indistinguishable at the noise level. Only then decide whether another scan, another station, or an entirely different calibration measure would help.
CHAPTER 9 OF 12 · Distinguish structure from calibration
9. Test a source explanation against observing geometry
An extended source explanation becomes stronger when a residual pattern follows the source and projected baseline geometry, rather than merely the calendar or the name of one telescope. Compare a source over several orientations. Then compare other sources observed with comparable equipment and processing. A source-based pattern need not have a single constant value across these views.
A fictional observing night has two sources. Source X has repeated closures around +18 ps on one set of projected baselines and near zero on a different set. Source Y is near zero in both groups. This is suggestive, but check that X was not observed only with a different channel mask or under lower signal strength. If these factors changed together, the evidence does not isolate source structure.
A cleaner test uses a source model or independent imaging to predict how phase varies on the baselines, and then processes those predictions into comparable delays or closures. Reserve some observations for checking the prediction. A model fitted to the same closure values can reproduce them without predicting another orientation. Closure-informed imaging and closure validation can be related without being the same test.
Keep limits in view. Low signal strength, spectral ambiguities, differing group-delay conventions or non-simultaneous data can affect the result. A source can also have a structure contribution that is mostly station-like or cancels for the particular geometry. A near-zero plot is not a promise that all structure effects are negligible for geodetic coordinates.
CHAPTER 10 OF 12 · Distinguish structure from calibration
10. Use calibration evidence that closure does not contain
Suppose telescope B had a maintenance event at midnight. All triangles still close afterward, but baseline residuals with B shift with opposite signs according to whether B is first or second. That is the signature of a possible station-like offset. A clock or cable calibration log could be informative precisely because closure was designed to discard this pattern.
Conversely, if only one triangle becomes anomalous when a spectral band is removed, review the baseline processing and effective delay definition. Reprocess a controlled subset with consistent masks and evaluate what changed. Do not force a clock explanation merely because the word “delay” appears on the report. The test should discriminate among mechanisms that could actually produce the observed sign pattern.
Compare raw delays, post-fit residuals and closure with the model choices labelled. A fit may absorb some effects into station parameters, and changes to weighting may change the shape of residual plots. A triangle formed from residuals only retains the clean cancellation interpretation when the subtracted model terms and processing conventions support it.
A useful handover makes one bounded statement at a time. For instance, “The common-epoch closure rises on triangles containing AB after the band-mask change; the sign pattern is consistent with an AB-related processing effect, but the present observations do not separate that from all source-dependent explanations.” That leaves the next investigator a clear direction rather than an unsupported verdict.
CHAPTER 11 OF 12 · Make a defensible claim
11. Write a closure check a colleague can replay
Save the source identifier, scan times, baseline orientation, reference epoch, delay convention, channel mask, uncertainties and any transformation applied before the sum. State whether the input contains observations or residuals. Keep the signed terms, not only a histogram of closure magnitudes. Without the signs a colleague cannot reconstruct a shared-edge test or see a station-dependent mismatch.
Start with one reproducible triangle. Then map its neighbours: identify which triangles share each edge, which rely on different projected geometry, and which were excluded for missing coverage. If you collapse the results into one average, preserve the per-source and per-configuration checks as well. Opposite signed effects can average to nearly zero without disappearing.
Separate the acceptance question from the observation. If your goal is to identify a processing change, the decision rule may rely on a before–after comparison with matched data. If your goal is to quantify a source contribution, you need a model comparison across geometry. The same numerical closure can be interesting for both questions but conclusive for neither.
When uncertainty is large, write what would improve it. Another triangle sharing the same noisy edge might add less than a different scan with different geometry. A source model may explain part of the signal, but a separate station log could still reveal a larger station-like change. A measured closure becomes useful when it guides the next choice.
CHAPTER 12 OF 12 · Make a defensible claim
12. Transfer the reasoning to a survey loop
Consider three level marks. The measured height differences along A to B, B to C and C to A are +1.20, +0.35 and −1.53 m. Closure is +0.02 m. The number tells you the differences do not fit a perfectly consistent set of heights. It does not identify which leg is wrong. A datum shift of +5 m at all three marks would not change any leg or the closure.
Now change the failure mode. Imagine the height assigned to B is too large by 0.04 m in both derived differences. A to B increases by 0.04 m, B to C decreases by 0.04 m and the loop sum is unchanged. The specific mechanism differs from a VLBI clock, but the algebra is the same: the quantity you wanted to test is in a direction the loop discards.
Take the reverse transfer back to the telescopes. You now see three small closures for a session and a station coordinate shift compared with an independent solution. There is no paradox. The loop test and the coordinate comparison are sensitive to different error patterns. The right next question is which additional measurements have independent sensitivity to the disputed coordinate shift.
Keep a short rule: a nonzero closure is evidence of inconsistency under the stated model; zero closure is not proof of absolute accuracy; a pattern across loops is stronger when its covariance and coverage are known. Apply that rule to new data before turning the result into a statement about a clock, a telescope, or a source.
Independent worked case and evidence checks
For one fictional common-epoch scan, store AB = +43, AC = +30, AD = +54, BC = −8, BD = +22 and CD = +20 ps. Compute CABC = 43 − 8 − 30 = +5 ps; CABD = 43 + 22 − 54 = +11 ps; CACD = 30 + 20 − 54 = −4 ps; and CBCD = −8 + 20 − 22 = −10 ps. Check that 5 − 11 − 4 − (−10) is zero.
A colleague says the +11 ps triangle proves AB is high. Test that claim by adding a +10 ps station-like offset at B. The new AB is +53; BC becomes −18; BD becomes +12. Recompute each triangle. All four closures are unchanged even though three baselines moved. The test cannot establish AB’s absolute bias.
Now pretend the same three A-centred closures have independent baseline inputs with 3 ps standard deviation each. The closure standard deviation is about 5.20 ps; the shared AB produces a covariance of 9 ps² between the first two. The difference of their closures is −6 ps with standard deviation 6 ps because four unshared edge observations survive. A conclusion drawn from treating both large-looking values as independent evidence would misstate what was measured.
Finally remove AD from the record. CABC still exists; CABD and CACD do not. CBCD still exists, and another possible loop can be built from the available edges but adds no third independent direction. A quality record should identify the missing edge, keep the available loops, and avoid pretending the six-edge comparison was reproduced.
Worked comparison
| Observation | Valid inference | Not established |
|---|---|---|
| Small signed loop sum | Tested loop is consistent at its uncertainty | Stations are correctly calibrated |
| Nonzero aligned loop sum | Some nonclosing contribution or modelling mismatch survives | Which baseline caused it |
| Shared AB in two loops | The two loop errors may be correlated | Two independent confirmations |
| No AD measurement | Some checks are unavailable | AD had zero error |
Questions and practical checks
Why is this called a closure delay?
It is the signed sum of measured baseline delays around a closed loop. In a triangle that uses AB, BC and CA, each station enters twice with opposite signs.
What if my software stores AC rather than CA?
Use minus AC in the clockwise expression. Confirm the underlying sign convention and the applicable time reference before combining the values.
Can source structure be present when the closure is zero?
Yes. Some structure contributions can cancel in a particular geometry or be represented as station differences. Use different geometry and independent source evidence to test the explanation.
Are all triangles from a four-station scan independent?
No. Four triangles built from six consistently directed baselines satisfy one exact linear relation. Shared observations can also make different triangle estimates statistically correlated.
Can I use a closure to correct a station clock?
Closure cancels a station clock term under the matching assumptions. To estimate or validate it, retain suitable baseline information or use independent clock and calibration observations.
Continue with the connected subjects
Build the VLBI observation
VLBI geometric delays: project baselines and see why clocks can be confounded
VLBI source structure: derive why the group-delay phase slope can change
VLBI source switching: check whether time alignment is supported
Understand network limits
Residual redundancy: see which errors an observation model cannot expose
Free networks: distinguish relative consistency from a fixed datum
Normal-equation combination: recognise common information before merging estimates
Sources and scope
Xu and colleagues: closure observables, common epochs and a radio-source structure case
Anderson and Xu: what geodetic VLBI closure statistics reveal and can miss
Xu and colleagues: closure-guided source-structure reweighting and baseline orientation
Global Geodetic Observing System: the purpose and measurement basis of VLBI
The numerical cases in this guide are original fictional teaching examples, not measurements from a real site. This is an educational explanation, not a field safety instruction, engineering approval or an official certification. For a real project, use qualified practitioners, current product documentation and site-specific evidence. Sources were checked on 4 November 2026.