Hacker News new | comments | show | ask | jobs | submit login
Strange Noise in Gravitational-Wave Data Sparks Debate (quantamagazine.org)
90 points by digital55 on July 1, 2017 | hide | past | web | favorite | 46 comments

Do they propose a mechanism for how there could be simultaneous correlations between the two detectors that aren't gravitational waves? There's few enough forces that can act through kilometers of solid earth that I'd be willing to take it as a given that they're uncorrelated, absent a signal... but then, I'm not a physicist.

Even more I'm-not-a-physicist-y speculation; could the additional correlation be some secondary gravitational wave effect hidden under the main signal, pointing to another aspect of the phenomenon?

I think the argument is that any correlation in the noise must be simply coincidental. In other words, the detection could be explained by chance rather than some true physical phenomenon measured by both detectors.

But, LIGO ran their detectors for 8 years at a lower sensitivity with no detections. Almost as soon as they turned on the upgraded detectors in September 2015, they detected the first event (GW150914) which appeared at a level that would have been undetectable previously. Also, they have made two other detections in the 1.5 years since. My guess is that if these events were due to flawed analysis, then we should have seen something during those first 8 years.

While I think it is unlikely that there is a problem here, one has to consider that upgrading the sensitivity could introduce or reveal a systematic error.

Also, it is not just a correlation that has been observed - it is a correlation that has a specific form, a rising fundamental frequency that happens to in line with what we would expect from merging black holes. I know there's an element of circularity in the argument, but what are the chances of that particular sort of signal arising from a spurious correlation?

If gravity is waves, it stands to reason that there would be lots of noise that correlated as various events in the universe reach earth.

gravity isn't waves, I think that was an overly simplistic description, but the explanation for noise does stand to reason, imho.

Gravity isn't waves, but it's a field that waves can exist in. (Similar to, the ocean isn't a wave but it can contain waves.)

“The only persons qualified to analyze this paper are in the LIGO Scientific Collaboration”

What? That's an incredibly dangerous position for a scientist of any stripe to take.

I think the end of the sentence is interesting as well... ", said Robert Wagoner, a theoretical physicist at Stanford University who is not affiliated with LIGO. “They are the only ones who have had access to the raw data.”

That kind of gives the lie to the conspiracy that many would like to believe... And before you go demanding that they release the raw data (in addition to the processed data), I hope you are aware of how many disk drives you're going to need to fit into that station wagon.

I mean yeah there's probably lots of instrument telemetry in the data, but let's give LIGO the benefit of the doubt and say that the instrumentation has consistent performance through null control regions of the timeseries and the "interesting" regions of the timeseries (bloop #1, in this case, but apparently now bloops #2, and #3 which were not analyzed by the person doing the criticism).

the 'truckloads' of data are basically all the uninteresting parts of the timeseries. Of course there is statistically going to be random events where the noises correlate at the level seen by the critique, and maybe the truckloads of data show that those are frequent enough that say 1/3 of the bloops should have that correlated noise superimposed on the 'actual signal'; and maybe that bloops 2 and 3 do not have the correlated noise if you apply the same analysis to it.

How much more 'raw data' do you need to reach a conclusions that refutes the criticism?

I've yet to find a good solution to this increasingly more prevelant sentiment in the current highly-specialized niche type of scientific advancement. I always looked up to the autodidacts of history and wanted to emulate them, but all I hear is "that's impossible these days".

It's an empirical claim, not a normative one.

it's a terrible claim. You can't claim that for all of the subdisciplines in the massive LIGO project, every single field is covered by the 'topmost expert'. Especially for a criticism about signal processing - there are so many adjacent or even distant scientific fields that are likely to have individuals with superior 'expertise', in the same class of algorithms used for the signal processing.

The claim is that only the people working in the project are qualified to criticize; that's almost the opposite. People who are working on the project are more often than not the LEAST qualified to criticize, because they're heads down in the weeds and have too much invested in the particular details and experimental choices that were made.

Moreover, for LIGO specifically, not every scientist is motivated to work for a high-risk, long-time-till-result endeavor. In fact, I'd suspect the people most likely to approach things with the right critical mindset are antiselected from working on the project.

It is interesting to see how someone not familiar with these sorts of projects interprets this stuff.

The quote has nothing to do with who is trained in signal processing in the abstract, and everything to do with who has a detailed understanding of this particular machine.

The initial paper critical of the ligo work absolutely has everything to do with signal processing in the abstract, so the quote (which refers to the criticism) must, as well.

If you invented a new sort of internal combustion engines and I disputed whether it could work based on abstract ignition theory, my claims are still implicitly dependent on my understanding your design and the actual physical device, regardless of what I concentrate on in my abstract.

Please reread. Those other experts are apparently Fourier analyzing processed data to "remove the signal". Whether they are doing it correctly or not, any real check will require looking at the underlying data, which is what the non-LIGO associated scientist said. Misquoting someone is not a good way to start an argument.

no. LIGO released the non-processed data, and the other scientists processed it separately without using (presumably high-bandpass) filters that the LIGO scientists used.

page 3:


I think you are misinterpreting the quote. I think you believe it means:

"The only persons who have the right critical mindset to assess this paper are in the LIGO Scientific Collaboration"

I believe instead it is simply saying:

"The paper has no obvious flaws that anyone unfamiliar with LIGO could point out. Therefore the best people who could assess it are the LIGO team"

"The paper has no obvious flaws that anyone unfamiliar with LIGO could point out. Therefore the best people who could assess it are the LIGO team"

What's the point of peer review, then? Why bother publishing detailed experimental reports and providing access to data if the team gives it the two thumbs up and no one on the team sees any flaws. Their word should be good enough.

I'm not sure I understand what you are saying. What is the point in peer review of what?

The fact that one team is more qualified to analyse a paper that another team has produced has nothing to do with whether or not that paper should be independently reviewed.

If you want to look at the data and run their code w/o installing any software, check out their ready to view/run Jupyter notebooks:


1. click on a notebook to see the html rendering 2. or clone+run to execute/edit the code (need to create login)

code links:

original criticism: http://www.nbi.ku.dk/gravitational-waves/correlations.html

response: https://github.com/spxiwh/response_to_1706_04191/blob/master...

definitely click through the quanta article's links.. fascinating reads all the way down

It is fascinating, and I learned that one of the events was first detected by an unmodeled search for strange events in the signals:

"Finally, LIGO runs 'unmodelled' searches, which do not search for specific signals, but instead look for any coherent non-Gaussian behaviour in the observatories. These searches actually were the first to find GW150914, and did so with remarkably consistent parameters to the modelled searches..."


I have done lots of signal processing. You cant just subtract out a noisy signal and complain the residual is correlated, there will be bits remaining. If they want to show correlated noise, why cant they use any other part of the signal?

The correlated remainder is only in the vicinity of the signal event.

The 2nd team is arguing the presence of residual correlations after the signal is removed means the total correlation isn't well described by a gravity wave signature.

This is experimental data, so there's clearly going to be noise and there's clearly going to be complexities in the detector response to the event. The question boils down to whether the residuals are significant enough and distinct enough from expected detector response behaviors to call the detection into question or suggest something unexpected is contained inside the detected signal.

It all boils down to what they subtract. It is said that they subtract the wave signal. Is this signal known for sure or very predictable ? It also depends on how they subtract it. Is it in the fequency domain or the space doman ? I didn't read the arXiv article and don't plan too. Just commenting and asking.

Think of it as a matched filter tuned to the impulse response of a gravity wave. If that filter matches the observed signal then it's a detection, if it doesn't then it's not. The debate is over the significance of the inevitable remainder after subtracting the model of the signal. Is it an expected amount of remainder or an unexpectedly large amount of remainder.

If any correlation in the residual signals would be enough to doubt the discovery then I don't understand why they didn't simply combined the signals into one by minimizing the residual correlation.

Your suggestion to just minimize the residual correlations would unfortunately boil down to "analyze the data in the way that best hides any problems in the data" which isn't what anyone wants.

There are specific signal shapes/characteristics that would be producable by a gravitational wave. Anything that doesn't match that shape/doesn't have those characteristics by definition isn't from a gravity wave. Ignoring those characteristics and just saying "if it's correlated it's from a gravity wave" would be a highly incorrect way to analyze the data, which is why none of the scientists in the debate would advocate doing that.

I'm not advocating that any correlated signal should be considered a gravitational wave. I'm just saying that, if residual correlation is such a problem, then it would make more sense to create a combined signal by minimizing the residual correlation.

This discussion would then simply be about how well the combined signal matches a gravitational wave, which seems a lot more sensible.

I suspect they did that in part, but it's also critical to keep the signals distinct precisely because you expect the noise to be uncorrelated between the two signals so you're introducing an additional square root of two gaussian noise factor in the combination.

This sort of thing is much more about statistics and error propagation than one encounters in more traditional signal processing.

Is it really noise or just more signal?

Cosmic background radiation is noise but it can be mapped out into a pattern. Could the correlation be a view of a type of cosmic background gravitational waves?

The pattern is in a different dimension than the noise (time vs spatial). A pattern by definition isn't noise.

Well, except:

> The main claim of Jackson’s team is that there appears to be correlated noise in the detectors at the time of the gravitational-wave signal. This might mean that, at worst, the gravitational-wave signal might not have been a true signal at all, but just louder noise.

If that's an accurate characterization, "louder noise" does seem a lot like signal.

Couldn't the "correlated noise" just be a superposition of numerous small gravitational-wave signals?

Not by definition. There is such a thing as structured noise.

Obviously the simplest solution is to have a third detector, located far away from the other two. I'm surprised the article didn't even mention that.

Let's wait and see if their article gets through peer review in a similarly respected journal to the three published detection papers'.

where there's smoke...


Scientists keep saying and repeating how correlation doesn't mean anything to the point it's become a joke

Now they are dabbling about a spurious correlation in a small section of the data? For real?

No, really, if anyone would be trying to prove anything based on that correlation alone they would have been laughed and dismissed by the majority of the scientific community

But suddenly when it's to reject a result then it's important?

> points out that Jackson’s team could have misused a common data-processing technique called the Fourier transform... The error, Harry writes, has to do with the technical assumption that the input data signal be “cyclical,”

Correct. Not to mention the sample is too short (and the even is short as well and there's noise and the sampling rate is not too big).

If you play with FFT on these conditions you start seeing things that "aren't there"

Scientists don't say correlation doesn't mean anything. What they do say is that correlation doesn't imply causation.

If the basic model underlying the original LIGO finding would predict no correlation between the recordings, then a significant amount of correlation should lead scientists to revise the model, and reexamine the findings. Very much standard scientific practice.

There are several possible sources of correlation on the "impulse response" of the LIGO detectors (the whole detector, from building construction to the electronics and detectors in itself) so it is not surprising that might happen. Or maybe it's something in the source of the signal

But the signal is still stronger than the noise

Besides the "poking the data in weird ways until it stops meaning anything" issue

No. The word "Correlation" as used in the article to discuss correlations in the data generated by two remote detectors has a very specific technical meaning. It's not simply one of the words in the trope "correlation does not imply causation."

Correlation as used here and put simply means that when the signal coming out of one the detectors increases, the signal coming out of the other detector (which is physically separated from the first detector by a very large distance) also increases, and when the signal coming out of the first detector decreases the signal coming out of the second detector also decreases.

The existence of that mathematical correlation between those two signals coming out of those two remote detectors absolutely implies there is a causation to the correlation (yes, mathematical correlation can imply a causation, even while it alone is not able to indicate what that causation is).

The question at hand is not whether there is a correlation or whether there is a causation. The question is what is the causation for the correlation and whether that causation is a gravity wave or something else.

The odds are extremely good that that causation is a gravity wave, but it's a big enough experiment and a big enough result that it's worth looking very closely at the analyses that argue in favor of it being a gravity wave (and those that argue against it).

... what? This is precisely the correlation meant in the aphorism you mentioned and is also the only sense of the word correlation of which I am aware of. Signals can be similar (correlated) purely by chance and this cannot be taken to imply a causative link between them.

If a correlation is consistent, some sort of causation is quite likely. There are of cause spurious correlations but in physics experiments we can mostly get rid of them by repeating the experiment several times, unlike in economics for example.

The more important point, however, is that a correlation between A and B does not imply that A causes B or B causes A. It could also be that unknown factor C causes A and B.

Agreed, though repeated measurements weren't mentioned so I thought it was important to emphasize that finding a correlation in one measurement cannot be taken to imply causation, which sounded like what OP was saying.

Your point is a very good one, though. If we view causation through a causal graph, such that vertices are variables and direct causal links are arrows (directed edges), a statement "correlation doesn't imply causation" can be taken to mean several things:

  * From correlation, we cannot determine the direction of the arrow.
  * From correlation, we cannot claim there *is* a (direct) arrow between two vertices.
  * From correlation, we cannot claim there is even a path from one vertex to the other.

Indeed. I think it is also important to note that the unknown factor can be something very abstract. For example if you measure the oscillations of two pendulums with the same properties, they will also be correlated, but simply because they are driven by the same underlying physics...

Guidelines | FAQ | Support | API | Security | Lists | Bookmarklet | Legal | Apply to YC | Contact