dwarfastro.com

The giant squid (the deep sea creature, not the squid nebula) spent most of a century as the animal everyone knew was down there and almost nobody could photograph. There were beaks cut from sperm whale stomachs, severed arms hauled up in nets, the occasional carcass rolled onto a beach, every bit of it proof the thing was real, and still no picture of a living one until a Japanese team filmed it in its own water in 2012. I have spent my last run of clear nights chasing a different squid, the one tangled into the Flying Bat Nebula, and I have started to feel a strange kinship with those deep sea camera crews. Mine just points the lens up instead of down.

Seventeen hours and thirty three minutes of Duo-Band data, a 35mm telescope, a Bortle 6 backyard. In the middle of the starless frame, if I let my eyes relax, a faint teal smudge sits right about where OU4 should be. I wanted it to be the Squid. I wanted it enough that I nearly wrote this as the night I finally caught it.

Then I measured it, and the measurement is the whole story.

The picture that almost had me

The first thing I made was not a photo. It was a map of where the oxygen might be. The Squid glows in OIII, which falls in the blue and green. The Flying Bat wrapped around it is hydrogen, which falls in red. So I pulled out the blue green, took away everything the red could explain, smoothed what remained, and scaled it by how much empty sky naturally flickers. Dark where there is nothing, cyan where something stands above the noise.

It looked like a discovery. Soft cyan arms, a couple of bright knots, a faint web draped across the frame that my eye read as gas before my brain could catch up. I let myself enjoy it for a while. That feeling, wanting it to be true, is the exact thing that gets a noise smudge posted as a nebula.

How I caught myself

Smoothed noise always looks like that. Blur a field of random numbers and it pools into clouds and filaments with soft edges, and the eye, which evolved to find shapes in the dark, hands you a squid. So I stopped looking and started measuring. Sigma is just how much the blank sky wobbles from pixel to pixel, and a signal’s significance is how many of those wobbles it rises above the noise. The five sigma bar I keep mentioning is the same one physics uses before it will call something a discovery.

At ten hours the oxygen excess where the Squid should be came in at 0.13 sigma above the background. At seventeen it had crept to about 0.2. A marginal detection starts near 3. A confident one is 5. Two tenths of a sigma is the noise floor with a few more hours stacked on top of it.

Then the part I had been avoiding. In this stretched eight bit image, the red and the blue green channels move together at a correlation of 0.90. The blue I had been admiring was mostly red light bleeding sideways in processing, not oxygen of its own.

Two spots did clear the bar. For an hour they were my candidate Squid cores, the proof I wanted. Then I laid the map back over the version of the frame that still had its stars.

Both of them sat dead on a star. They were not the Squid. They were the marks left where star removal pulled a bright star out and did not quite heal the hole. I had been reading scar tissue and tinted static as a creature.

What it would take

The obvious next thought is to shoot more. The trouble is the arithmetic. Faint signal climbs against noise only as fast as the square root of time, so doubling it costs four times the hours. Getting from 0.2 sigma to a confident 5 is a factor of twenty five in signal, and because the time goes as the square of that, six hundred and twenty five times the exposure. Anchored to the seventeen hours I have, that lands on the order of eleven thousand hours of clear, moonless Duo-Band data. That is closer to a prison sentence than a hobby.

Even the thirty to fifty hours I had planned would leave me near 0.3 sigma, still well inside the noise. More nights at this exact pipeline will never get there. What changes the answer is everything around the exposure.

What actually moves it

Part of what I measured is not the sky at all. It is the gear and the file. A 35mm aperture collects a fixed amount of oxygen per hour, and a dedicated scope several times that size with a real OIII filter gathers far more in the same night. That is why you will see OU4 show up in someone else’s thirty six hour capture on a bigger rig while seventeen hours here barely registers. Same object, very different light grasp. Aperture and a true narrowband filter are doing the heavy lifting there, not patience.

The other limit is self inflicted, and it is the more interesting one. The image I measured is an eight bit export out of the DWARF’s processing app, already stretched for display. Eight bits leaves each channel only 256 levels, the stretch has thrown away the link between pixel value and collected light, and the channels have been blended, which is exactly why red and blue tracked so tightly. That is the worst kind of file to do faint signal measurement on. The honest version of this measurement lives in the linear stack, before any stretch, at full precision. Run through Siril on that linear data, with the oxygen pulled straight off the green and blue sensor pixels and the gradient removed cleanly, the same seventeen hours should give a lower noise floor and a tighter, more trustworthy limit. It will not conjure a detection that is not there, only measure the absence more honestly, which is the right thing to ask of it.

That is the next post. I am going to take this exact seventeen hour data set back into Siril, do the whole thing on the linear master, and report what the floor really looks like once the vendor stretch is out of the way. This one is the before. The DWARF 3 and its app pack a remarkable amount into their size, and they also quietly cap what you can pull from the faintest targets. That is worth knowing before you spend seventeen hours finding the edge of it.

So no catch tonight. The Squid is out there, glowing faintly in a color my small scope can barely reach, and seventeen hours was never quite going to land it. The deep sea crews waited a hundred years and a lot of bad footage for one clean frame. I can wait a few more nights, with a better process and, one day, a bigger lens.

If you want the background on OU4 itself, what it is and why it hides so well, that is on a separate page.

Clear Skies.

---

A note on how I ran the numbers

I should be straight about one thing, because on this blog the method matters more than the mystique. I did not hand write the analysis. I gave the two exports, the starless frame and the one with stars, to an AI assistant, Claude Opus 4.8, and asked it to run the significance test: the channel correlation, the sigma measurement, the overlay on the stars. It did the arithmetic I would have done slowly and probably worse. The data is mine, the imaging is mine, and the conclusion that there is nothing here yet is mine. The model only counted, and counting is the part people are worst at when they already want a particular answer. The exact prompt is below, so you can run the same check on your own frames with the same model.

Give it both images and ask for the numbers, not the verdict you are hoping for. Here is the prompt to try:

"I have two exports of the same deep sky frame at identical scale and registration, one starless and one with stars. The target is a faint OIII nebula superimposed on brighter Ha. Do this and show the numbers.

Load both as RGB float arrays. Treat R as the Ha proxy and (G + B) / 2 as the OIII proxy. Report the Pearson correlation between R and the OIII proxy over the starless frame. A value near 0.9 means the blue is mostly Ha crossing channels rather than independent oxygen.

Remove the Ha correlated component. Fit the OIII proxy as a cubic function of R by least squares, evaluate it, and subtract, so the residual is the OIII that R cannot predict. The cubic absorbs the nonlinear display stretch.

Convolve the residual with a Gaussian of sigma about 40 pixels so the test responds to structure larger than single pixel noise. Estimate the sky from four corner boxes, each roughly 12 percent of width by 14 percent of height, inset from the edge. Compute the mean m and standard deviation s of those samples. This s is the noise level of the blank sky.

Define significance as z = (residual - m) / s. In plain terms, sigma is the standard deviation of the empty sky, and z is the number of standard deviations a pixel rises above the background. Zero a 70 pixel border to drop smoothing artifacts. Report the peak z, the count of pixels above z = 3, and the area fraction above z = 3.

Threshold z at 3 and overlay the survivors on the starry frame. For each surviving blob, state whether its centroid sits on a star. Blobs centered on stars are star removal residuals, not nebula.

Estimate the diffuse OIII significance inside a region I will mark, or near the frame center if I do not. Then, since shot noise limited signal to noise scales as the square root of integration time, compute the hours to reach a target significance as hours_now * (target / current_sigma) ^ 2, using my current integration time. Report milestones for z = 0.4, 1, 3, and 5.

Judge the result on this scale, which follows the physics convention where five sigma is the standard for a claimed detection. Below 3 sigma is no detection, 3 to 5 is marginal, 5 or above is confident. Do not infer signal the statistics do not support."

---

Frequently Asked Questions