Life’s Loose Tolerances

Does Evan Know What Hes Talking About?
No, I just read the three books at the bottom. Numbers are ±3 orders of magnitude.

Maybe I would’ve been a real biochem head if my first exposure to it had been the work of David S. Goodsell. Every bio reading and lecture slide I had ever seen depicted proteins as magic blobs, somehow animated into doing all that delicate continuous work by starbursts labeled "ATP", WordArt arrows as the animating force of creation. Goodsell shows the exquisite lil’ mechanics of it all, proteins whirring, yanking, squeezing, linking, walking, frantically jiggling folded chains that make themselves and you.

Goodsell and Peter Hoffmann also revealed something that has never left my mind since: in life "slip" is not only tolerated, but essential. Living things do not "optimize"; they spend to prevent catastrophic outcomes like DNA damage, and let the rest of its machinery be kinda shoddy stuff that works on average and doesn’t down take the whole system if it screws up. (That’s not to say proteins aren’t good at their jobs; most are absurdly efficient, with some motor proteins approaching 88% of energy-in converted to work-out. No machine you can see or touch or manufacture has ever approached that.)

In living cells only a few essential processes are fastidious, and a few ATP-dependent proteins "tightly coupled" where spending some energy means getting some good work each time. The rest of the thousands of proteins come out of the factory a little glitchy, take a few steps back and only on average more steps forward, and through all that tolerate and make use of ten trillion collisions with water molecules per second. Life thrives in chaos, sees what sticks, is very precise in a few domains and catch-as-catch-can for the rest.

An example of a bit of no-slip paired to mostly high-slip: the "central dogma" of biochem, DNA ▶ RNA ▶ the proteins that do the work of sustaining a cell. DNA mutations are neat in the fullness of time, but the next generation damn well better work like its parent or there’d be nothing fighting entropy today. So DNA replication is low-frequency, metabolically expensive and has a lot of checks. The cascade of enzymes that copy DNA make around one error per billion nucleotides, "like copying a thousand books and making only one mistake." Compare that to the high-frequency downstream dirty work of turning information into Life: protein transcription and translation.

In the first step of making proteins, enzymes read DNA and make messenger RNA. These enzymes average one error per 10,000 nucleotides. That means the DNA ▶ mRNA process has one hundred thousand times as many fuckups as DNA ▶ DNA, which means about half of all RNA strands have mistakes. And it just gets sloppier: ribosomes put one wrong amino acid in the chain for every 10,000 it has to link, so at the end of all this one out of every 24 proteins has a misplaced amino acid.

All those errors are, tautologically since I can write this down and you can read it, very okay. While a DNA screwup can kill a cell outright or permanently weaken its descendants, errors in RNA transcription and protein translation are non-lethal. Even good RNA is short-lived and disposable, so bad strands don’t make a lot of proteins or might not even work at all in the ribosome. And all proteins have a certain, mostly-brief working life. Malformed proteins get diced up and recycled pretty fast, so they don’t have a lot of time to do damage. Finally the information flows only one way: when you forget the eggs and make a nasty cake, the recipe doesn’t update to keep your error. Same deal here. A bad protein doesn’t reach into the nucleus and change the instructions to make it.

I think that’s so cool! Life is a few bits of careful work coupled to an insane number of non-lethal, low-cost mistakes. I’m not kidding when I say that is lovely advice.

The Machinery of Life, David S. Goodsell
Our Molecular Nature: The Body’s Motors, Machines and Messages, David S. Goodsell
Life’s Ratchet: How Molecular Machines Extract Order from Chaos, Peter M. Hoffmann

Sentinel Sees Something

Does Evan Know What Hes Talking About?
Charlie does, I just dick around with radar because it looks neat.

Always show love to passive radiometers, but sometimes you gotta go active: satellite radar can see earthstuff at night, through clouds and even peek under dry sand (something something dielectric, ask a scientist). Applications? Glad you didn't ask: you can use satellite radar to monitor illegal fishingfind oil spills, watch ice floes crunch, and spy on whatever. But radar data is harder to interpret and process than what you'd get from Landsat. You have to deal with polarizations and different sensor modes to even get the data, then just eyeballing it shows you a bunch of speckly nonsense, and getting useful knowledge out of it? Buddy.

Enough reading, let's check out a radar image:

San Francisco, CA as seen from the ESA's Sentinel 1A satellite, which carries a C-band synthetic aperture radar. This thing emits microwaves, listens for the echoes, and draws you a picture. Why's it look so different than a Landsat scene? And why's it so pixelly and fuzzy? Unlike passive sensors which get to collect reflectance data using The Best Emitter We Got (the Sun), active sensors have to hurl their own energy and see what comes back; they can only emit so much, so you'll get less image detail and miss stuff that's blocked from the sensor's view by steep terrain (see the scalloped hills in the upper left? Layover.) Also microwaves have longer wavelengths than what Landsat picks up. This can be useful in seeing through vegetation (transparent to the longer-wavelength energy emitted by Sentinel) but can create "what am I looking at" moments like the triple Golden Gate Bridge.

Fun fact: the bright cross on the ship is probably from a radar reflector, which helps the vessel show up on ship tracking radars. Being seen is a good thing when moving a mountain of goods or crude in a vehicle that takes a few miles to even change course.

This radar stuff is complicated! But you, me, and everyone we know are in luck: the wonderful 👏✨👏 Charlie Loyd 👏✨👏 showed me a quick way to process radar data just because I asked. And through his sacrifice you shall know too.

We're going to run Charlie's principal component analysis (PCA) python script on Sentinel 1A radar images. PCA takes different images as inputs (say, bands 3, 4, 5, 7 on a Landsat scene or two different radar polarizations of the same area) and tries to pick out what matters; if the analysis detects redundant information across the different input images, it throws it out, leaving only the juicy bits (bytes?) that account for the variation between them. Say you asked a program to analyze a sample of 1,000 cars. You'd be miffed if it reported "This is big: every car has four wheels!" You care about differences and deviations, the spice of life, right? That's what PCA's for. Don't cite me, I've done this like twice.

Charlie on PCA:

Think of an n-band image as a dataset with n dimensions. In other words, each pixel is a point in n-D space. For most human-important scenes you get something consistent: the first principal component will be the sorta-mostly-mean of the bands (materials that are bright in one part of the visible spectrum are usually bright in others). The second principal component will usually be basically vegetation v. not vegetation (close to NDVI, a standard index). The third – at least over cropland – tends to be wet v. dry surface features. This is the basis for a standard transform called the Tasseled Cap, which is a set of coefficients first taken from PCA of fields. PCA is also really interesting for, say, archeology, because in the furthest-down correlations you sometimes find important things like obscure combinations of plant species that like minerals leached out of ancient walls. (You could also, less wholesomely, use it for oil exploration.) PCA lets you surface stuff you’d never find if you tried.

There are a dozen channels, so it gets really weird as you go down the PCs. Some of them seem to be water vapor content, cloud height – it’s very hard to tell.

So let's grab some radar data and PCA it up. Sentinel 1A has been up since 2014, has a twin in Sentinel 1B and a beams down lots of free synthetic aperture radar data.

  1. Head to the Sentinel data hub at

  2. Sign up for a free account in the top right.

  3. Drag a box around a small area you like (try to find somewhere with interesting topography, which looks super neat under radar. But don't take my word for it).

  4. Put this string in the search field: S1A_IW_GRDH_1SDV*
    That'll ask the server for Sentinel 1A data, in the mode "interferometric wide swath," processed to the "Level 1 Ground Range Detected" product. 

    • Sentinel 1A's radar data comes in a few flavors:
      IW - interferometric wide swath - 250km swath, 5x20m spatial resolution
      EW - extra-wide swath - 400km swath, 20x40m spatial resolution
      SM - strip map - 80km swath, 5x5m spatial resolution

      If you want to get higher-resolution, narrower-swath maps over a smaller area, try this query over Chicago or Houston: *_S1_GRDH*
      If you see little color previews, you're set.

      You can also hit the hamburger to the left of the search bar to narrow the results by date.

  5. If you see thumbnails with color previews, you're on the right track. These are ~800mb compressed files that include two tiffs, one for each polarization.

  6. Find at least two scenes from different dates, with color preview thumbnails and the same footprint, then hit the arrow icon that appears on mouseover to download the files. This will be slow.

  7. The .tiffs will be in the "measurement" folder after you decompress. They're too big and speckly to run through PCA as-is; you'll shrink them to cut down on artifacts. Good old gdalwarp can drop these giant files down to 2500px wide using the cubic resampling method. You can use another resample method, but make sure it's smooth!

gdalwarp -of GTiff -r cubic -ts 2500 0 input.tif output.tif

I did this with four images, two polarizations from two dates over coastal Bulgaria. Thus:

8. Merge your resized files into a single multi-band tiff. You can use either or QGIS > Raster > Miscellaneous > Merge.
9. It's time for PCA: make sure you have the rasterio, sklearn, and numpy python packages, download Charlie's PCA script and feed it the multi-band tiff you just made. This will yield a new multiband tiff, with each band corresponding to a principal component.

$ python your_stacked_file.tif your_output_file.tif

10. I'll let the real C-band explain what you get.

Now you have a tiff whose bands are the principal components of the inputs. Principal component (PC) 1 is basically going to be “brightness”, or something close to an average of all the inputs. So it makes sense to use L*a*b* color, assigning the first three channels in that order. You can use Photoshop or, say, imagemagick. To get the color channels distributed well, you can equalize them and then lower their contrast.

I throw out the fourth PC, because it’s mostly noise from what I’ve seen. One of the nice things about PCA is that it tends to “comb” noise into the later PCs, giving you cleaner signal up front. But in practice, band 4 does seem to have some plausible features in among the static. Mixing it into band 3 might give you something good, or then again it might not.

The disadvantage of PCA, of course, is that there’s no defined interpretation of the PCs. If PC 2, say, seems to give you wet v. dry, or forest v. bare soil, that’s great, but there’s no theoretical or physical rigor to that interpretation, only statistics. Which is still great for a lot of purposes, but it’s not the same as, like, NDVI or whatever.

Something I’d like to try is throwing it much deeper stacks – like 10 observations instead of 2. You get as many bands out of PCA as you put in, so you should be able to pick out a lot more. Of course 10 observations means you’re looking at a longish time period, and there will be intrinsic change, which might throw things off a bit – PCA will start to produce dimensions that are “about” time more than about surface signal. Which is not necessarily bad!

Because you’re looking at a bunch of multiscale-heterogenous surfaces for which every pixel is presumably mixed, you expect your histograms to tend generally to converge on bell curves. (Massive generalization, broken by water among other things.) This is useful for L*a*b* because the a* and b* channels are pretty sensitive – anything far off the midpoint is really saturated. In general, you expect an L*a*b* representation of an arbitrary multiscale-heterogenous scene (like a landscape) to have a* and b* channels that approximate tight (small σ) bell curves. This is what equalizing and then lowering contrast is supposed to get you. But it can get thrown off by the big flat nodata areas, among other things. (Come to think of it, it wouldn’t be hard to mask them out and do the adjustments in skimage … hm.) Just try to get the peaks of the a* and b* bells on the midpoints as a starting point.

–Charlie Loyd, your favorite and mine


11. Open your output tiff in Photoshop (it'll be in RGB and look terrible but it's just not using most of the range). Also create a new document with the same dimensions, but in the "Lab Color" color mode. (Hint: select all > copy first before creating a new document will automatically populate dimensions in your new document window.)

12. Go into just the red channel on your output > select all > copy > paste into the Lightness channel in your fresh document. Repeat for the others; blue channel > a channel, green > b channel. Add a levels or curves adjustment layer and hit "auto" for each channel or futz away with the sliders. 




13. And here's what I got: the PCA picked out forests in light blue, agricultural fields in screaming pink, impervious surfaces in yellow, and a whole lot of other hues I couldn't decode. PCA stripped away the common information and left what was distinct between the 4 input bands, and the colors just reflect what the analysis thought should be lumped together. PCA is trying to tell me something, but it’s beyond my ken.

Man that's small. Click through already.

Man that's small. Click through already.

14. The end! No moral. Remember that your inputs are just numbers to the PCA script; it doesn't know anything about remote sensing. It simply picks out anomalies in your images, and it's your job to figure out what they mean. It's in your hands now. Your burden. I'm sorry.

"The mill is there no longer; the wind is still there."

Does Evan Know What Hes Talking About?
No, I really liked a college class about the French Revolution. “Twelve Who Ruled” remains a favorite nutty true story.

The real lasting accomplishment of the Russian Revolution is the thrall in which it holds college-age men, but I was never that much of a wiener. Now the French Revolution, there's a dustup worth venerating. I loved learning about that crazy-ass decade, a completely insane upheaval that also made too much sense. No wonder the rest of the continent was spooked enough to gin up some armies and stamp on those embers.


A passage of Les Misérables nicely sums up the period's complete nut-assery. Victor Hugo can be a windbag but it's still a beautiful story, starting 15 years after the revolution's end and climaxing during another fit of Republican fury, the June Rebellion. My favorite part comes early in the novel when a Bishop, who we're told is the noblest of dudes, heads out of town to administer last rites to a dying Republican. He's a patient man but it is a testy visit. You'd be too if your vocation was "priest" and your host recently retired from a distinguished career as "priest murderer."

A sample of the exchange:


Republican, speaking of the French Revolution: "Alas! The work was incomplete, I admit: we demolished the ancient regime in deeds; we were not able to suppress it entirely in ideas. To destroy abuses is not sufficient; customs must be modified. The mill is there no longer; the wind is still there."

Bishop: "You have demolished. It may be of use to demolish, but I distrust a demolition complicated with wrath."

Republican: "Right has its wrath, Bishop; and the wrath of right is an element of progress. In any case, and in spite of whatever may be said, the French Revolution is the most important step of the human race since the advent of Christ. Incomplete, it may be, but sublime. It set free all the unknown social quantities; it softened spirits, it calmed, appeased, enlightened; it caused the waves of civilization to flow over the earth. It was a good thing. The French Revolution is the consecration of humanity."

When the bandannas come off and the barricades come down, every revolutionary has to answer to the junta colonel if they're unlucky, the grandkid if they're lucky, or the journalist if they're lonely: was it worth it? To this Republican, a member of the Convention no less, of course it was. If the revolutionaries were to be blamed for anything, it was being too ahead of the curve. Which, after discounting the cruelty and bloodshed, seems fair. It really was a little too much a little too fast. The whole exercise was doomed from the first conversation in the Vendée, when the new state finally got around to telling the provinces about the complete inversion of the known order that had been going down:

Farmer: "What's up, guy? Nice sash."

Prefect: "Thanks. Listen up: the king is dead, your priest is in jail, we renamed all the months, women are holding muskets, and France is trying to repel like four separate invasions. As to that last bit, we're going to need your food for eating and your sons for fighting."

Farmer: "Wow. I'm not into any of that. Even one of those would super suck, so all of those things together are making me real mad."

Prefect: "Tant pis, hayseed. Now make with the vittles and the bodies or I'm coming back with a noose."

Farmer: "Damn."


So of course it didn't stick. But it did uncork a whole mess of cool new ideas. All republics owe those fiery Frenchmen a debt of gratitude, not for their choppy-chop governance style but for laying the bloody foundations of pluralism and representative government. Now go read Twelve Who Ruled, I'll wait.

Voir Direst

The metal detector bottleneck makes for a big assey line, but effects a pleasing visual for the tardy. All those heights and colors and ages and clothes, wearing the same blank indignation. Economic segregation being what it was, and is, and will be, the last place to really mix is when we're against the sticky gears of New York municipal bureaucracy.

And it turns out My Big Fat Greek Wedding, screened to keep us awake, really is universal. That alone was worth the burned day.