NGC 891 transits 82 degrees above my southern horizon in the middle of November. Airmass 1.01. That is about as close to free as this hobby gets, and it is the only reason I believed the project was reasonable when I started it on 11 October 2023.
I thought it would take one winter.
The galaxy is not faint. SIMBAD lists V = 9.93 and an ellipse of 12.30 by 2.45 arcminutes. The Wikipedia infobox says V = 10.1 and 13.5 by 2.5 arcminutes. Take either and compute the mean surface brightness over the ellipse:
with the semi axes in arcseconds. SIMBAD's pair gives 22.26 mag/arcsec^2. Wikipedia's pair gives 22.55. Two catalogues, 0.29 apart, on an object nobody disputes. I put that first so you know what precision looks like in extended object photometry before I start quoting my own numbers to two decimals.
My zenith sky measures 20.9 on a good moonless night, meter on the same tripod hole in the lawn every time. So the mean surface of the galaxy sits about 1.4 magnitudes below my sky background, a flux ratio of , roughly 3.6 to 1 in favour of the sky. The mean is the flattering figure. Everything I actually wanted is in the halo, and the halo is much fainter than the mean.
Now the part that gets argued about in forums by people who have not written it down. Photons per pixel per second from an extended source:
is pixel pitch, is focal ratio. Aperture does not appear. Collecting area goes as , the solid angle each pixel subtends goes as , and since the diameter cancels exactly. My 100 mm f/5.5 refractor delivers the same electrons per pixel per second on a galaxy as a 300 mm f/5.5 would at the same pixel pitch.
Aperture comes back the moment you resample to the same arcseconds per pixel, because then each pixel covers a fixed patch of sky and you are collecting it with more glass, so again. Which gives the equivalence I find slightly humiliating: in the sky limited case, doubling collecting area buys exactly what doubling time buys. Signal and sky both double with area, and the ratio improves as the square root. So my 218 hours at 100 mm is about 55 hours at 200 mm, same focal ratio, same sky sampling.
I do not have a 200 mm. I have winter.
100 mm f/5.5 triplet apo, 550 mm focal length. Cooled mono CMOS, 3.8 micron pixels, 1.42 arcsec/px, field roughly 2.4 by 1.6 degrees. Filters L, R, G, B and a 7 nm H-alpha centred on 656.3 nm (Balmer alpha, the n = 3 to n = 2 transition, 656.28 nm in air). Guiding off-axis. Dither every frame, three pixels minimum.
Everything below is what survived calibration and rejection, not what I recorded.
| Season | Filter | Frames | Sub | Integration | Sensor | Median FWHM |
|---|---|---|---|---|---|---|
| 2023/24 | L | 612 | 120 s | 20.40 h | -10 C | 2.7" |
| 2023/24 | R | 96 | 300 s | 8.00 h | -10 C | 2.8" |
| 2023/24 | G | 96 | 300 s | 8.00 h | -10 C | 2.9" |
| 2023/24 | B | 108 | 300 s | 9.00 h | -10 C | 3.1" |
| 2023/24 | Ha | 60 | 600 s | 10.00 h | -10 C | 2.8" |
| 2024/25 | L | 1134 | 120 s | 37.80 h | -10 C | 2.6" |
| 2024/25 | R | 144 | 300 s | 12.00 h | -10 C | 2.7" |
| 2024/25 | G | 144 | 300 s | 12.00 h | -10 C | 2.8" |
| 2024/25 | B | 156 | 300 s | 13.00 h | -10 C | 3.0" |
| 2024/25 | Ha | 96 | 600 s | 16.00 h | -10 C | 2.9" |
| 2025/26 | L | 984 | 120 s | 32.80 h | -10 C | 2.5" |
| 2025/26 | R | 108 | 300 s | 9.00 h | -10 C | 2.6" |
| 2025/26 | G | 108 | 300 s | 9.00 h | -10 C | 2.8" |
| 2025/26 | B | 120 | 300 s | 10.00 h | -10 C | 2.9" |
| 2025/26 | Ha | 66 | 600 s | 11.00 h | -10 C | 2.8" |
| Total | 4032 | 218.00 h | 2.7" |
Recorded across three seasons: 258.00 hours. In the final stack: 218.00 hours. Sensor at minus 10 C in every single session. In early October, when ambient went above 16 C and the cooler could not hold minus 10 at a duty cycle I trust, I shut down rather than run warm and mix dark libraries. That decision cost me maybe six hours in total and I would make it again.
Usable nights: 19, then 32, then 24. Seventy five nights out of the 414 on which this galaxy was above 40 degrees for a useful stretch between astronomical dark and dawn. Eighteen percent. That is Czechia in winter and no amount of equipment fixes it.
Per pixel, over a total integration split into frames:
is object electrons per second, sky, dark current, read noise per frame. Under my sky the term swallows the rest, so
Square root of time. For equal length subframes, , so it is also the square root of the number of frames. The reason is that the signal in frames adds linearly while independent noise adds in quadrature, growing as , and the ratio of the two therefore grows as . No stacking algorithm beats that, they only fail to reach it.
| Cumulative | Integration | Frames | Relative SNR | Gain over previous |
|---|---|---|---|---|
| after winter 1 | 55.40 h | 972 | 1.00 | |
| after winter 2 | 146.20 h | 2646 | 1.62 | 62% |
| after winter 3 | 218.00 h | 4032 | 1.98 | 22% |
Seventy two hours in the third winter bought twenty two percent. Three winters bought a factor of two. Going from twenty hours to two hundred and eighteen bought a factor of 3.3. To double again from here I need 872 hours, which at my rate of usable nights is roughly nine more winters. I am 41. That is not a plan, that is an obituary.
Tilt does not average out. Stacking more frames with the same optical fault makes the fault more confident, not less. I lost 40.0 hours of luminance from the second winter because I hoped otherwise for four weeks.
The rotator had crept. One corner ran 1.4 arcsec worse in FWHM than the diagonally opposite corner, consistently, across every night of that block. I stacked it anyway for a month and told myself the field was acceptable. Then I measured it properly on a 5 by 5 grid of the frame and it was not acceptable. Deleting 40 hours took me longer to decide than to do.
The dew heater controller failed at some point around 03:00 on 14 January 2025. I found it at 05:20. Forty one luminance frames with the corrector fogging progressively, star FWHM walking from 2.6 to 6.9 arcsec across the sequence. Gone. The worse part is that I was asleep in a house eleven metres away.
December 2024: thirty one nights, zero usable. Not one. That is the month I stopped calling the schedule a schedule.
At twenty hours I had the disc, the dust lane and a soft symmetric glow around both. Nothing wrong with it. Nothing in it either.
Angular scale first, because the distance is not settled. The Wikipedia infobox gives 27.3 plus or minus 1.8 million light years, which is 8.4 plus or minus 0.6 Mpc. Mouhcine, Ibata and Rejkuba, who surveyed the outskirts from the ground, write the distance as about 10 Mpc. One kiloparsec therefore subtends arcseconds, which is 24.6" at 8.4 Mpc and 20.6" at 10 Mpc. Call it 21 to 25 arcseconds per kpc and remember that the twenty percent is real.
Three things arrived that were not there before.
The extraplanar dust. Howk and Savage imaged this galaxy with the WIYN 3.5 metre in 1997 and found hundreds of dust absorbing structures away from the midplane, tracing them out to about 1.5 kpc off the plane, so 31 to 37 arcseconds, which is 22 to 26 pixels for me. I do not resolve individual features the way a 3.5 metre does and I never will. What I have is the aggregate: the dust lane is not a line, it is a ragged fringe, the fringe is on both sides of the plane, and it is not symmetric. At twenty hours the lane had a clean edge. The clean edge was noise.
H-alpha off the plane. The extraplanar ionised gas layer here is modelled with an exponential electron scale height of about 1 kpc in the Boettcher analysis, so roughly 21 to 25 arcseconds. At ten hours of narrowband my Ha was noise with a galaxy in it. At thirty seven hours the emission is visibly not confined to the disc, and it is patchy in a way that survives every rejection test I can run.
The third thing I did not go looking for. The Mouhcine survey describes what it found around the inner galaxy like this:
The bulge and the disk of the galaxy are found to be surrounded by a previously undetected large, flat and thick cocoon-like stellar structure at vertical and radial distances of up to ~15kpc and ~40kpc respectively.
Fifteen kpc vertical is around six arcminutes. Forty kpc radial is around fifteen. They also report a stream looping out to roughly 50 kpc, which would be near nineteen arcminutes, comfortably inside my 2.4 degree field. They resolved theirs into individual stars using an 8 metre class telescope. I cannot resolve stars at 9 Mpc from a lawn. I have integrated glow.
And after the third winter there is a faint asymmetric extension off the north east end of the disc that I cannot make go away. It appears independently in each of the three seasons stacked alone. It is present in luminance and in the summed RGB. It does not follow any gradient in my flat frames, it does not sit where my dust motes sit, and it moves with the galaxy when I dither.
I am not going to claim it is their stream. I have no published surface brightness figure for that structure to check against, and that absence is exactly why this project was never plannable: I did not know how faint the target was before I started, so I could not compute the time it needed. I kept going until the residual stopped changing. That is not method, that is stubbornness with a spreadsheet attached.
If you want the arithmetic behind where my own floor sits, I wrote it out separately in @dust/limiting-magnitude-from-first-principles, including why the aperture tables everybody quotes are optimistic by about a magnitude.
Next season I am imaging a blank field thirty arcminutes away with an identical sequence, as a control. Wren counts the same rockpools at Prussia Cove year after year and publishes the boring middle years too, see @lowtide/rockpool-survey-prussia-cove, and it took me three winters to understand the point. The second survey of the same square is worth more than the first survey of a new one.
Three winters. Seventy five nights. 4032 frames. One galaxy. What I have to show is a ragged fringe on a dust lane, some Ha where it should not be, and a smudge I refuse to name. I have already started the control field.