A velocity correlation is not fully defined until its averaging operation is stated. The same two signals can have zero covariance over a complete cycle and a negative or positive covariance over a shorter window. Subtracting a mean measured within that window changes the result again.
The short answer is to name the window, the reference mean and the quantity being estimated before interpreting a Reynolds-stress measurement. We will derive an original finite-window example, calculate its phase dependence, and test what happens when such windowed values are compared across space. The signals are fictional deterministic teaching traces, not measurements from a plasma experiment. The calculation exposes an averaging choice; it does not prove that a measured plasma flow has any particular drive.
eduKateSG · Zonal Flow
Find your next learning step
Choose the question closest to your task, or follow the original window calculation in order. All teaching chapters remain open.
ROUTE 1 · CHAPTERS 1–6
Define the averaging target
Separate the physical setting, cross-moment, local means and quarter-cycle example.
ROUTE 2 · CHAPTERS 7–13
Derive duration and phase effects
Check window placement, exact integrals and limiting durations.
ROUTE 3 · CHAPTERS 14–20
Interpret the covariance consistently
Check normalization, bounds, spatial comparison and calibration.
ROUTE 4 · CHAPTERS 21–27
Test processing and evidence
Separate dependence, numerical integration, uncertainty and an independent exercise.
ROUTE 5 · CHAPTERS 28–34
Pool, verify and continue
Record definitions, test pooling and shared noise, and check units.
Chapter contents
Define the averaging target · Chapters 1–6
1. Begin with the quantity that needs an average
2. Keep the plasma context distinct from the teaching signals
3. Define a rectangular time-window average
4. Separate a cross-moment from a covariance
Derive duration and phase effects · Chapters 7–13
7. Check the quarter-cycle result with a sign picture
8. Compare half-cycle and full-cycle windows
9. Introduce amplitude, frequency and the window centre
10. Derive the general raw cross-moment
11. Shift the quarter-cycle window without changing the signals
Interpret the covariance consistently · Chapters 14–20
14. Examine a very short window using local slopes
15. Preserve the distinction between a covariance and a correlation coefficient
16. Check a covariance bound within the same window
17. Compare window-dependent values at different radial positions
18. Do not assume every averaging operator obeys the same rules
Test processing and evidence · Chapters 21–27
21. Avoid treating more overlapping windows as independent evidence
22. Keep deterministic window sensitivity separate from sampling bias
23. Check discretisation against the analytic window result
24. Use an independently researched uncertainty source carefully
25. Relate the example to causal interpretation in plasma measurements
Pool, verify and continue · Chapters 28–34
28. Design a window comparison that changes one choice at a time
29. Record enough information to reproduce the covariance
30. Pool windowed data before or after centering deliberately
31. Check a shared-noise contribution with its assumptions visible
32. Check units before comparing a spatial gradient
CHAPTER 1 OF 34 · Define the averaging target
1. Begin with the quantity that needs an average
A Reynolds-stress contribution involves a product of fluctuating velocity components. To define a fluctuation, we first need a mean. That mean may be an ensemble average, a long-time average under suitable conditions, or a finite-window estimate. These choices are related but not interchangeable by definition.
Suppose two components are called u and v. A covariance is the average of their product after their selected means have been removed. Saying that a detector measured two velocities is therefore incomplete information for reproducing the reported covariance.
The averaging interval, weighting, treatment of gaps and reference mean all belong to the measurement description. This article isolates the finite-window choice using analytically known signals. Real plasma measurements also require geometry, calibration and a justified relation between measured quantities and the velocities entering a momentum balance.
CHAPTER 2 OF 34 · Define the averaging target
2. Keep the plasma context distinct from the teaching signals
The Zonal Flow root concerns organised flows in magnetically confined plasmas. Its Reynolds-stress discussion links a velocity covariance to momentum transport and, through the appropriate spatial balance, to possible mean-flow acceleration.
Our example uses two sinusoids to make the averaging operation exactly calculable. We do not claim that a turbulent plasma is a pair of perfectly coherent waves. A deterministic signal can nevertheless reveal a mathematical effect that a data-processing rule must handle consistently.
The distinction is important. A nonzero finite-window covariance in this example is a valid covariance of that window. It becomes misleading only if it is presented as a different target quantity without justification, such as the full-cycle covariance of the supplied signals. The calculation should help a reader specify the estimand, rather than label every short-window measurement wrong.
CHAPTER 3 OF 34 · Define the averaging target
3. Define a rectangular time-window average
For a window beginning at t0 and lasting T, define the average of a function f as the integral of f from t0 to t0+T divided by T. We use uniform weighting within the interval and no weighting outside it.
Write the resulting average as meanT(f). Define mu=meanT(u) and mv=meanT(v). The centered window covariance is CT=meanT[(u−mu)(v−mv)]. Expanding the product and using the fact that mu and mv are constants within this particular integral gives CT=meanT(uv)−mu mv.
This identity holds exactly for the defined window. It does not require random signals, stationarity or a large number of samples. Those additional concepts become relevant when interpreting CT as an estimate of a population or long-time covariance. Keeping the exact algebra separate from that statistical interpretation makes the later checks simpler.
CHAPTER 4 OF 34 · Define the averaging target
4. Separate a cross-moment from a covariance
The raw cross-moment is meanT(uv). The covariance subtracts the product of the two window means. These quantities coincide if at least one of the two means is zero, but they need not coincide in a short interval.
A pair of signals with zero mean over a complete cycle can have nonzero means in a fraction of that cycle. Using their full-cycle means and using their local window means therefore define different products to average.
An instrument or analysis script should state which operation it performs. A label such as correlation can conceal a raw product, a centered covariance, a lagged correlation or a normalized correlation coefficient. Reproducing the calculation requires its actual definition. In the examples below, every reported centered covariance explicitly uses the two means calculated within the same stated window.
CHAPTER 5 OF 34 · Define the averaging target
5. Choose signals with a known complete-cycle reference
Let u(t)=cos t and v(t)=sin t. The angle t is a dimensionless phase coordinate for this first example. Equivalently, it represents angular frequency times physical time with unit amplitudes.
Over a complete interval of length 2π, both signals have zero mean and the average of their product is zero. Their complete-cycle covariance is consequently zero, independent of the starting phase. The two signals are in quadrature; the positive and negative product contributions cancel over the complete cycle.
This reference is known from the supplied functions, rather than estimated from noisy data. We can therefore ask exactly how a finite-window operation differs from it. No sampling error is needed to produce the difference. Later sections introduce physical amplitudes, angular frequency and a shifted starting phase without changing the underlying logic.
CHAPTER 6 OF 34 · Define the averaging target
6. Work a quarter-cycle window directly
Take the window from zero to π/2. The mean of cos t is [sin(π/2)−sin 0]/(π/2)=2/π. The mean of sin t is also 2/π.
The raw cross-moment is the integral of cos t sin t divided by π/2. Its integral over this interval is one-half, so the raw cross-moment is 1/π. The centered covariance is therefore CT=1/π−4/π², approximately−0.086975.
The raw cross-moment is positive, while the centered covariance is negative. There is no sign contradiction: the second quantity subtracts the positive product of the two local means. Within this quarter cycle, one signal decreases while the other increases. Their departures from their respective local means tend to have opposite signs, producing the negative centered covariance.
CHAPTER 7 OF 34 · Derive duration and phase effects
7. Check the quarter-cycle result with a sign picture
At the beginning of the window, u is above its window mean and v is below its window mean. Near the end, u is below its mean and v is above its mean. Those opposite departures explain the dominant negative covariance contributions.
There is a small central part where both departures have the same sign, so it would be inaccurate to claim that every instantaneous centered product is negative. The integral determines the final result. The sign picture is a diagnostic explanation, not a substitute for computing the integral.
This distinction is useful when interpreting plots. A scatter of two components can reveal a local increasing or decreasing relationship while the uncentered product tells a different story about their offsets. Specify the center before assigning physical meaning to a positive or negative product average.
CHAPTER 8 OF 34 · Derive duration and phase effects
8. Compare half-cycle and full-cycle windows
For the window from zero to π, meanT(u)=0 and meanT(v)=2/π. The raw cross-moment is zero, so the centered covariance is zero too. For zero to 2π, both means and the cross-moment vanish.
The half-cycle zero is specific to this alignment. It is not a theorem that every half-cycle window produces zero covariance. The window includes equal positive and negative cos t sin t contributions and has a zero mean for cos t under the chosen starting phase.
These three windows are a useful test set for a calculation. The quarter-cycle centered result is negative, while the two listed longer-window results vanish. A routine that assumes the supplied signals have zero mean in every interval will incorrectly report the quarter-cycle raw value as its centered covariance.
| Window in phase coordinates | Mean of u | Mean of v | Raw cross-moment | Centered covariance |
|---|---|---|---|---|
| Zero to π/2 | 2/π | 2/π | 1/π | 1/π−4/π² |
| Zero to π | Zero | 2/π | Zero | Zero |
| Zero to 2π | Zero | Zero | Zero | Zero |
CHAPTER 9 OF 34 · Derive duration and phase effects
9. Introduce amplitude, frequency and the window centre
Now let u(t)=A cos(ωt+φ) and v(t)=B sin(ωt+φ), with positive amplitudes A and B and angular frequency ω. The window lasts T in physical time and is centred at tc.
Define θ=ωtc+φ as the phase at the window centre and h=ωT/2 as its half-width in phase. Introduce sinc h=sin h/h, with the continuous value one at h=0.
A centered integration gives mu=A cos θ sinc h and mv=B sin θ sinc h. The expressions show separately how the window location and duration matter. θ sets which part of the cycle is observed, while h controls how much cancellation occurs across the interval. A duration in seconds alone is not enough to compare windows when their signal frequencies differ.
CHAPTER 10 OF 34 · Derive duration and phase effects
10. Derive the general raw cross-moment
Using cos a sin a=sin(2a)/2, the raw product is AB sin(2ωt+2φ)/2. Averaging symmetrically around tc gives meanT(uv)=AB sin(2θ) sinc(2h)/2.
The product of the window means is AB sin(2θ)[sinc h]²/2. Subtracting yields CT=AB sin(2θ){sinc(2h)−[sinc h]²}/2.
This is the complete centered rectangular-window result for these supplied quadrature signals. It separates an amplitude factor AB, a window-centre phase factor sin(2θ), and a duration factor involving h. The full-cycle covariance remains zero, yet finite-window covariance can have either sign. A useful implementation check is to recover the earlier quarter-cycle result by substituting h=π/4 and θ=π/4 with A=B=1.
CHAPTER 11 OF 34 · Derive duration and phase effects
11. Shift the quarter-cycle window without changing the signals
Keep a quarter-cycle duration, so h=π/4. A window beginning at zero has centre phase θ=π/4 and the negative covariance already calculated. Shift the window by a quarter cycle: its centre becomes θ=3π/4, and sin(2θ) changes sign.
The centered covariance now equals the positive value 4/π²−1/π. The signals, amplitudes and full-cycle reference have not changed. The selected phase interval has changed.
A sequence of short-window covariance values can therefore oscillate even when the underlying periodic traces remain exactly unchanged. Such an oscillation is not automatically evidence that a physical transport mechanism alternates its full-cycle sign. It can be a consequence of a moving observation window. Whether that local-window behaviour is itself the intended physical observable must be stated in the analysis.
CHAPTER 12 OF 34 · Derive duration and phase effects
12. Check a shifted half-cycle window
For a half-cycle duration, h=π/2. Then sinc(2h)=0 and [sinc h]²=4/π². The general result becomes CT=−2AB sin(2θ)/π².
A half-cycle window centred at θ=π/2 gives zero. A half-cycle window centred at θ=π/4 gives −2AB/π², approximately−0.202642AB. This second window begins at phase−π/4 and ends at phase3π/4.
The example explains why a duration-only rule is insufficient. Calling an interval half a period does not establish which components of the local mean subtraction will cancel. A reproducible processing description includes its start or centre and how that position is chosen relative to the data. Phase locking, event alignment and an arbitrary moving window can produce different summaries of the same record.
CHAPTER 13 OF 34 · Derive duration and phase effects
13. Test the complete-cycle and integer-cycle limits
If the window contains an integer number of complete cycles, h is an integer multiple of π. Both sinc h and sinc(2h) vanish, and the centered covariance is zero for every centre phase.
This exact cancellation depends on the supplied periodic functions, their common frequency and the exact cycle count. A window described approximately as one period can retain a residual. Signals with changing frequency, several frequencies, trends or nonperiodic fluctuations need their own averaging analysis.
The limit is nevertheless a powerful check for this teaching model. A numerical integration that gives a substantial nonzero covariance for a complete cycle should prompt review of endpoint handling, units, sampling weights and the mean subtraction. It does not require inventing a new physical correlation to explain a discrepancy in an exactly known test signal.
CHAPTER 14 OF 34 · Interpret the covariance consistently
14. Examine a very short window using local slopes
For small h, sinc(2h)−[sinc h]² is approximately−h²/3. The covariance is therefore approximately−AB sin(2θ)ω²T²/24.
The same result follows from local linear approximations. For two signals with slopes u’ and v’ at the window centre, their centered covariance over a short uniform window is approximately u’v’T²/12. Here u’=−Aω sin θ and v’=Bω cos θ, giving the expression above.
This independent derivation provides both a check and an interpretation: the short-window covariance reflects the local slopes of smooth deterministic traces after their local offsets are removed. Its approach to zero as T² does not mean the two full signals disappear. The window is shrinking around an almost constant pair of values, and centering removes those local constants.
CHAPTER 15 OF 34 · Interpret the covariance consistently
15. Preserve the distinction between a covariance and a correlation coefficient
A normalized correlation coefficient divides CT by the product of the two centered root-mean-square fluctuations. It is dimensionless. CT itself has units of the product of the two velocity components.
For a sufficiently short window around a phase where both slopes are nonzero, both centered variances also shrink with T². The normalized coefficient can approach plus or minus one even while the covariance approaches zero. The local traces become nearly linearly related over that small interval.
A strong normalized coefficient therefore does not by itself establish a large momentum flux. The fluctuation magnitudes matter. Near a point where one slope vanishes, the short-window approximation needs higher-order terms and should not be used blindly. When reporting a Reynolds-stress-related measurement, retain the dimensional covariance and the averaging definition alongside any normalized correlation plot.
CHAPTER 16 OF 34 · Interpret the covariance consistently
16. Check a covariance bound within the same window
For a given averaging window, the magnitude of CT cannot exceed the product of the centered root-mean-square values of u and v. This is the covariance form of the Cauchy–Schwarz inequality.
The word centered is essential. Using complete-cycle root-mean-square values to normalize a locally centered covariance mixes two different definitions. Such a ratio may be a chosen diagnostic, but it should not be described automatically as the standard window correlation coefficient.
A practical check computes both window variances with the same weights and mean treatment as the covariance. A substantial bound violation suggests inconsistent centering, weights, units or arithmetic. Passing the bound establishes internal consistency; it does not prove that the window represents an appropriate long-time average or the specific velocity components required by a plasma momentum model.
CHAPTER 17 OF 34 · Interpret the covariance consistently
17. Compare window-dependent values at different radial positions
Suppose the supplied phase φ varies with radial position x while A, B, ω and T remain fixed. Then θ depends on x. Let K=sinc(2h)−[sinc h]². The window covariance is ABK sin[2θ(x)]/2.
Its spatial derivative is ABK cos[2θ(x)]dθ/dx. The derivative need not vanish even though the complete-cycle covariance is zero at every position. A fixed clock window samples different phase portions of the signals at different locations.
This is a comparison between two averaging targets. It should not be inserted uncritically as a claimed false force in an otherwise unchanged mean-flow equation. A momentum balance must be derived consistently for its chosen averaging operation. The example shows why identifying that operation is necessary before comparing a measured covariance gradient with a theoretical long-time or ensemble-averaged stress.
CHAPTER 18 OF 34 · Interpret the covariance consistently
18. Do not assume every averaging operator obeys the same rules
An ideal Reynolds average used in a derivation is typically supplied with properties that simplify mean and fluctuating terms. A sliding finite-time filter does not automatically behave as the same projection. Filtering a filtered trace can change it again, and a locally filtered mean can retain time dependence.
The algebra for CT inside a fixed window is exact. The further step of using a sequence of those values in a reduced dynamical equation needs its own treatment of means, derivatives and any retained filter terms.
This is a mathematical scope condition rather than an instruction to avoid finite-window diagnostics. Such diagnostics can be highly useful. Their interpretation becomes clearer when the theory and the processing use compatible definitions. Matching a symbol called stress in two documents is not sufficient if one denotes a local filtered covariance and the other a differently averaged transport term.
CHAPTER 19 OF 34 · Interpret the covariance consistently
19. Keep baseline choices distinct from calibration offsets
Adding a constant offset a to u and b to v leaves the locally centered covariance unchanged. The new window means increase by a and b, and those constants cancel in the centered departures.
The raw cross-moment does change. It gains terms involving the offsets and the original local means, as well as ab. A raw-product calculation can therefore be much more sensitive to a baseline offset than a correctly centered covariance.
This useful invariance applies to constants within the chosen window. A drifting offset is a time-dependent signal and may contribute to the covariance after centering. Deciding whether it is instrument drift or a physical low-frequency component requires independent evidence. Mean subtraction does not, by itself, identify the physical origin of a component it removes or retains.
CHAPTER 20 OF 34 · Interpret the covariance consistently
20. Check multiplicative calibration separately
If the measured components are guu and gvv, with constant gains gu and gv, their centered covariance is gu gv CT. Additive constants cancel after local centering, but multiplicative calibration errors remain.
A normalized correlation coefficient can hide this amplitude dependence because positive gains cancel between covariance and the centered standard deviations. The dimensional momentum-flux interpretation still depends on calibrated component magnitudes and signs.
This distinction helps explain why a visually stable correlation coefficient need not establish a stable or accurately calibrated stress. Record the component calibration, units, orientation and any sign convention before interpreting the magnitude. If a component axis is reversed, the covariance sign reverses too. A sign is informative only after both the physical orientation and the processing definition have been fixed.
CHAPTER 21 OF 34 · Test processing and evidence
21. Avoid treating more overlapping windows as independent evidence
Neighbouring moving windows often share much of the same data. Their resulting covariance values are then related through shared observations, even before any physical temporal correlation in the signals is considered.
Plotting many overlapping windows can give a smooth time-resolved description. It does not automatically supply that many independent measurements for an uncertainty calculation. The number of displayed points and the effective independent information are different quantities.
For the deterministic sinusoids here, the moving-window curve is completely specified by the phase and duration formula. In an experimental record, uncertainty also depends on noise, temporal correlation, sampling and the estimator. Increasing the count of overlapping plot points cannot be assumed to reduce uncertainty by the square root of that displayed count. The statistical argument needs to match the dependence in the actual data.
CHAPTER 22 OF 34 · Test processing and evidence
22. Keep deterministic window sensitivity separate from sampling bias
The quarter-cycle result was obtained from exact integrals of known functions. It exists without finite sample noise. Calling it a biased estimate is meaningful only after specifying the target it is intended to estimate and a statistical sampling model.
Similarly, replacing a sample covariance denominator N by N−1 addresses a particular finite-sample correction under suitable assumptions. It does not generally transform the covariance of a selected quarter cycle into the complete-cycle covariance of these deterministic traces.
The first task is to identify the quantity being summarised. The second is to choose a sampling estimator and uncertainty treatment for that quantity. Mixing these tasks can lead an analyst to apply a familiar statistical correction to a discrepancy caused by window alignment or a different reference mean. An exact known-signal test is useful because it separates those issues cleanly.
CHAPTER 23 OF 34 · Test processing and evidence
23. Check discretisation against the analytic window result
To approximate the defined uniform integral, a numerical routine must use sample times and weights consistent with that integral. Counting both endpoints as full-weight samples in a periodic interval can alter a simple average because the endpoint values are repeated in phase.
A useful verification compares progressively finer quadrature with the exact CT formula. The discrepancy should shrink in the expected manner for the chosen integration scheme. This is a check of numerical approximation, separate from whether the selected duration and centre answer the intended physical question.
Do not use a successful integration test to claim that every experimental processing issue is solved. Missing samples, irregular spacing, filtering and timestamp uncertainty need their own treatment. The analytic example gives a controlled reference against which those implementation choices can be examined rather than concealed in a final covariance number.
CHAPTER 24 OF 34 · Test processing and evidence
24. Use an independently researched uncertainty source carefully
Sciacchitano and Wieneke’s 2016 primary study, PIV uncertainty propagation, derives uncertainty expressions for velocity-derived quantities including Reynolds stresses. Its institutional abstract distinguishes finite-sample uncertainty and the effective number of independent samples.
That study concerns particle image velocimetry and should not be described as a validation of a particular plasma probe or the fictional signals in this article. It provides methodological context for keeping estimator uncertainty and measurement dependence explicit. The exact sinusoidal window calculations here are independently constructed.
The source record was checked on 25 October 2026. Use specialist plasma diagnostic methods for an actual experiment and retain their geometric and calibration assumptions. The general lesson transferable across measurement settings is narrower: a statistical summary needs both a defined averaging target and an uncertainty analysis appropriate to how the observations were obtained.
CHAPTER 25 OF 34 · Test processing and evidence
25. Relate the example to causal interpretation in plasma measurements
Alonso and colleagues’ primary study of zonal-flow-like structures in the TJ-II stellarator examines possible drive and damping mechanisms. Its abstract qualifies the causal conclusion despite a compatible turbulence-driven acceleration scale, noting the possible relevance of local Reynolds-stress measurements.
This supports caution in moving from a measured correlation to an established causal momentum balance. Our window example adds a different, explicitly mathematical question: which covariance has the processing actually calculated? It does not reinterpret the study’s data or assert that its result was caused by the window effect derived here.
A sound comparison combines appropriate diagnostics, averaging definitions, spatial information and competing terms. Demonstrating that a correlation can provide the right scale is useful evidence, but it is not by itself a complete account of the flow’s evolution. Keep that distinction visible when transferring the model to another experiment.
CHAPTER 26 OF 34 · Test processing and evidence
26. Try an independent changed-frequency exercise
Let u(t)=2 cos(3t) metres per second and v(t)=5 sin(3t) metres per second, with t in seconds. Calculate their centered covariance over the interval from zero to π/6 seconds. Identify the fraction of a complete cycle contained in that interval.
Then repeat the calculation after adding π/2 to both signal phases while preserving the same clock interval. Predict whether the magnitude, sign or complete-cycle reference changes. Finally calculate the covariance over a complete period and explain why dividing the short-window data by a different sample-count denominator cannot supply that result automatically.
Before substituting numbers, record A, B, ω, T, h and θ. This separates physical time from phase and makes unit errors visible. The resulting covariance must have units of square metres per square second, rather than metres per second or a dimensionless correlation coefficient.
CHAPTER 27 OF 34 · Test processing and evidence
27. Check the changed-frequency solution
The angular frequency is three radians per second, so the complete period is 2π/3 seconds. The interval π/6 seconds is one-quarter of that period. Thus h=π/4 and θ=π/4 for the initial phases.
The centered covariance is AB(1/π−4/π²)=10(1/π−4/π²), approximately−0.86975 square metres per square second. Adding π/2 to both phases changes θ to3π/4 and reverses the sign, giving approximately+0.86975 in the same units.
The complete-cycle covariance remains zero in either case. The difference is an exact consequence of duration, phase alignment and local mean subtraction. It is not a finite-sample correction problem in this constructed example. The solution also shows why a stated window length in seconds must be compared with the relevant signal frequencies before its phase coverage is understood.
CHAPTER 28 OF 34 · Pool, verify and continue
28. Design a window comparison that changes one choice at a time
A useful comparison holds the underlying trace and calibration fixed while changing only duration, centre or mean definition. Report those changes separately. Otherwise a changed covariance could reflect several altered processing choices at once.
For example, first compare locally centered and raw-product values within the same interval. Then keep centering fixed and shift the interval. Finally change its duration while preserving a stated centre. Each step tests a different dependence exposed by the analytic formula.
In measured data, compare the resulting summaries with their uncertainty and the relevant physical time scales. A window long enough to reduce one sampling issue may also wash out a time-dependent feature of interest. The appropriate duration follows the scientific question and the behaviour of the data; there is no universal window length supplied by this sinusoidal example.
CHAPTER 29 OF 34 · Pool, verify and continue
29. Record enough information to reproduce the covariance
Record the two measured or reconstructed components, their orientation and units, the timestamps and any prefiltering. State the start or centre, duration, weighting, mean definition and handling of missing data. Distinguish dimensional covariance from any normalized or lagged quantity.
If the covariance will enter a momentum balance, name that balance’s averaging assumptions and its spatial geometry. State which uncertainty sources are propagated and whether overlapping windows or correlated observations affect the effective information.
A compact report can still contain these essentials. The aim is to let another reader reproduce the number and assess whether it matches the theoretical quantity being compared. A long list of covariance values without the averaging definition is less informative than a clearly specified estimate with an honest statement of its limits.
CHAPTER 30 OF 34 · Pool, verify and continue
30. Pool windowed data before or after centering deliberately
Suppose several disjoint windows are combined with weights wj that sum to one, proportional to the represented observation duration. Let their separate means be muj and mvj and their centered covariances Cj. The covariance of the pooled observations equals the weighted sum of Cj plus the weighted covariance of the separate means: sum(wj Cj)+sum[wj(muj−mu)(mvj−mv)], where mu and mv are the pooled means.
This identity follows by expanding each observation around its own window mean and then around the pooled mean. The within-window departures have zero mean within their respective windows, so the cross terms vanish. Averaging already-centered covariance values omits the between-window contribution.
For a concrete example, pool the two equally long intervals [0,π/2] and [π,3π/2] from the original unit signals. This is a deliberately noncontiguous selected record, not a complete half-cycle interval. Both individual centered covariances equal1/π−4/π². Their mean pairs are (2/π,2/π) and (−2/π,−2/π), so the pooled means vanish. The between-window contribution is4/π². The pooled covariance is therefore1/π, positive despite both individual centered covariances being negative.
The calculation explains why window centering followed by averaging is a different operation from pooling and centering once. Neither operation should be substituted for the other without stating the target. It is particularly relevant when records are grouped around events or when some intervals are excluded.
CHAPTER 31 OF 34 · Pool, verify and continue
31. Check a shared-noise contribution with its assumptions visible
Let the recorded components be uobs=u+η and vobs=v+η, where the same additive component η enters both channels. The centered covariance expands into CT plus the window variance of η and the two cross-covariances involving η and the physical signals.
If those two cross-covariances are zero in a supplied teaching case, the shared component adds its nonnegative window variance to the measured covariance. With unit original signals in the quarter-cycle window and a stipulated η variance0.1, the observed covariance becomes approximately0.013025 rather than−0.086975. A sign change can therefore occur without changing the supplied physical traces.
The zero cross-covariances are explicit conditions in this calculation. They are not guaranteed merely by calling something noise, particularly in a finite record. For distinct channel noises, their own cross-covariance replaces the shared variance term and can have different behaviour. Assess the measurement model before treating a correlation as entirely physical.
This example also separates two issues. Window sensitivity remains present in exact noiseless traces. Shared-channel contamination is an additional measurement effect. A longer window might change both contributions, but its effect cannot be assumed to remove either one without analysing the signals, dependence and calibration. A practical check examines reference measurements or independent diagnostics capable of distinguishing shared instrument components from the physical components of interest.
CHAPTER 32 OF 34 · Pool, verify and continue
32. Check units before comparing a spatial gradient
For physical velocity components, A and B each have units of metres per second. The sinc and phase factors are dimensionless, so CT has units of square metres per square second. Dividing a covariance difference by a radial separation in metres gives an acceleration-scale quantity in metres per square second.
That dimensional match is necessary but insufficient for assigning a force contribution. The component directions, density convention, averaging operation and omitted terms still need to match the selected balance. A numerical gradient formed from two differently defined window statistics can carry the right units while answering the wrong physical question. Use units as an early consistency check, then examine those definitions before interpreting the result.
Is a short-window covariance always wrong?
No. It correctly describes the supplied window when calculated consistently. The error is interpreting it as another averaging target without showing why that interpretation is justified.
Can locally centered covariance have the opposite sign from the raw product average?
Yes. The quarter-cycle example has a positive raw cross-moment and a negative centered covariance. The product of the two nonzero window means accounts for the difference.
Does zero full-cycle covariance mean every local window is uncorrelated?
No. Contributions can cancel over a complete cycle while individual intervals have nonzero covariance. Window location and duration determine which contributions remain.
Does a large correlation coefficient establish a large Reynolds stress?
No. A normalized coefficient also depends on the fluctuation scales removed by normalization. A dimensional stress-related interpretation needs calibrated component amplitudes and the appropriate averaging definition.
CHAPTER 34 OF 34 · Pool, verify and continue
34. Continue through the relevant learning routes
For the physical setting, return to the Zonal Flow subject guide. For a complementary spatial balance, read why a Reynolds-stress gradient can accelerate a flow without net momentum input. That article supplies a restricted momentum and energy model; this article explains why its averaging assumptions must be matched when constructing a covariance from data.
Connect the reasoning through The Gold Standard of Education and the How X Works hub. A useful final explanation names the target average, reconstructs the local means, checks phase and duration, and states what additional evidence supports the physical interpretation.