A broad response smooths the light
Prerequisites: the two geometry chapters, weighted averages and elementary summation. Let C(t) denote the variable part of the central continuum, after subtracting a constant reference level. Let Ψ(τ) describe the line’s response per unit delay, normalized so its integral is one. A linear, time-independent response gives
The integral adds all earlier inputs with their response weights. With cloud delays τj and normalized weights wj, it becomes L(t)=∑jwjC(t−τj). A sharp flare is repeated over the delay distribution. Fast fluctuations can wash out because one cloud is responding to a bright phase while another is responding to a dim one.
For a constant input, a normalized response returns that same constant. No cloud responds to a future input. Those two limits are useful checks of a numerical calculation.
What a correlation peak measures
Consider the ideal cross-correlation RCL(s)=∫C(t)L(t+s)dt. Substituting the response equation and changing the integration order gives
Here RCC is the continuum’s autocorrelation: how similar it is to itself after a shift. The observed correlation mixes the response with the time structure of the driver. A broad, asymmetric or poorly sampled fluctuation can therefore shift a correlation peak away from the response mean.
The Lab reports both quantities for its saved synthetic driver. Its cross-correlation peak is a maximum on the stated finite lag grid. Its mean delay is the weighted average of cloud delays. Changing inclination can reshape the correlation even when the symmetric ring’s mean delay stays fixed.
Real observations arrive at uneven times. An interpolated cross-correlation, or ICCF, estimates one curve between neighbouring observations and compares it with the other. It should not extrapolate outside the observed interval. Long gaps remain a limitation even when interpolation draws a continuous line through them. The method guide makes the overlap rule and lag range explicit.
Estimate a virial product
Let σv be the response-weighted standard deviation of the recession velocities, in metres per second. A chosen delay estimate τ gives
The dimensionless factor f describes geometry, motion and the adopted line-width definition. A width from the mean spectrum need not equal the width of its variable part, often measured through an rms spectrum. The convention must accompany the number.
For a uniform thin circular ring, ⟨sin2ϕ⟩=1/2. Its mean velocity is zero and its mean delay is r/c, giving
This is an analytic reference for that ring and that width definition. A finite annulus with equal response per radius has a different radial average, so the same formula must not be applied automatically.
Worked example. For the twenty-light-day ring with M=5.0×107M⊙ and i=30∘, the dispersion is about 1.27×103kms−1. The virial product is 6.25×106M⊙, and the ring factor is eight. These are calculated teaching values. Turning the same ring to sixty degrees makes the virial product 1.875×107M⊙; the actual mass is still 5.0×107M⊙.
How much does the lag depend on the data?
Measurement-error perturbations ask what happens if measured fluxes move within their quoted errors. Resampling observing epochs asks how strongly the particular set of nights determines the result. A flux-randomization/random-subset-selection procedure combines these operations. Repeating it produces a distribution of lag estimates, conditional on the data and analysis choices.
The spread of that distribution is not a complete accounting of model uncertainty. A different continuum proxy, a seasonal gap or an evolving response can affect the answer in another way. Test the lag window and inspect the full correlation before quoting only its maximum.
Exercises
- A measured line dispersion is ten percent too large. With lag unchanged, by what factor does the virial product change?
- A ring is almost face-on. Why can the mass become difficult to infer from the line width?
- An ICCF peak lies on the largest allowed lag. What should be checked before reporting it?
- A fifty-day observing season is used to estimate a forty-day lag. Why is the effective paired record much shorter than fifty days?
Answers and diagnostic reasoningShow the answer
- The product increases by 1.12=1.21, or twenty-one percent.
- The projected width approaches zero. Small width errors and uncertainty in inclination then strongly affect the geometric correction.
- Expand or justify the window, count overlapping pairs and inspect gaps. A boundary maximum is not an interior, localized peak.
- Shifting one curve by forty days leaves only about ten days of possible overlap, before accounting for the actual sampling. A large correlation can depend on very few shared features.
Sources
Peterson and Horne, arXiv:astro-ph/0407538v1, Sections 1–2, for response and virial-factor assumptions.
Pancoast, Brewer and Treu, arXiv:1407.2941v2, Sections 2 and 4, for direct modelling and comparisons between inferred mean delays and cross-correlation summaries.
AGN Watch NGC 5548 optical archive, the real-data input used by the next method guide.
Try the prediction
- Predict whether the correlation peak must equal the response-weighted delay.
- At fixed radius, compare several inclinations and read both delay summaries.
- Explain the differences using the synthetic driver and the delay distribution; keep this simulation separate from the observed tutorial.