Getting close to the "why Dropbox when you can rsync" mistake (https://news.ycombinator.com/item?id=9224)
@vicapow replied to keep the Dropbox parallel alive
11,188 karma · joined February 4, 2010
@vicapow replied to keep the Dropbox parallel alive
In mathematics/applied math consider cited papers claimed to establish a certain result, but where that was not quite what was shown. Or, there is in effect no earthly way to verify that it does.
Or even: the community agrees it was shown there, but perhaps has lost intimate contact with the details — I’m thinking about things like Laplace’s CLT (published in French), or the original form of the Glivenko-Cantelli theorem (published in Italian). These citations happen a lot, and we should not pretend otherwise.
Here’s the example that crystallized that for me. “VC dimension” is a much-cited combinatorial concept/lemma. It’s typical for a very hard paper of Saharon Shelah (https://projecteuclid.org/journalArticle/Download?urlId=pjm%...) to be cited, along with an easier paper of Norbert Sauer. There are currently 800 citations of Shelah’s paper.
I read a monograph by noted mathematician David Pollard covering this work. Pollard, no stranger to doing the hard work, wrote (probably in an endnote) that Shelah’s paper was often cited, but he could not verify that it established the result at all. I was charmed by the candor.
This was the first acknowledgement I had seen that something was fishy with all those citations.
By this time, I had probably seen Shelah’s paper cited 50 times. Let’s just say that there is no way all 50 of those citing authors (now grown to 800) were working their way through a dense paper on transfinite cardinals to verify this had anything to do with VC dimension.
Of course, people were wanting to give credit. So their intentions were perhaps generous. But in no meaningful sense had they “read” this paper.
So I guess the short answer to your question is, citations serve more uses than telling readers to literally read the cited work, and by extension, should not always taken to mean that the cited work was indeed read.
The real challenges there aren't the "biggies" above, though, it's the ones in obscure journals you have to get copies of by inter-library agreements. My PhD was in applied probability and I was always happy if there were enough equations so that I could parse out the French or Russian-language explanation nearby.
I'm sure people made mistakes on their bibliographies at that time as well!
And did we all really dig up and read Metropolis, Rosenbluth, Rosenbluth, Teller, and Teller (1953)?
Edited to add: Someone made a chart! Here: https://papercopilot.com/statistics/neurips-statistics/
You can see the big bump after the book-length restriction was lifted, and the exponential rise starting ~2016.
https://loa-shared.s3.amazonaws.com/static/pdf/Didion_Malibu...
I share some of the same disappointment, especially going back and noticing disinvestment in the schools, which were one of the gems of the area.
The southern end of the central valley (San Joaquin region, whole central valley is outlined in red) is particularly hard-hit by groundwater depletion. Some of that storage does not come back, because the ground compacts after the groundwater is withdrawn.
A moment's reflection should make this clear. It's such a fundamental resource, touching everything we do. We just tend to take it for granted.
E.g.: https://en.wikipedia.org/wiki/Matilija_Dam#History
> Almost immediately after construction, the dam began silting up. The dam traps about 30% of the total sediment in the Ventura River system, depriving ocean beaches of replenishing sediment. Initially, engineers had estimated it would take 39 years for the reservoir to fill with silt, but within a few years it was clear that the siltation rate was much faster than anticipated.
There are similar sites all over the state. If you happen to live in the LA area, the Devil's Gate Dam above Pasadena is another such (but originally built for flood control, not for storage).
It's just not as easy as GP comment imagines.
(Related, and profound apologies for the fb.com link: https://www.facebook.com/DLJCSS/posts/small-quakes-do-not-pr...)
(Source: Work with Earth science domain experts in $dayjob, and am often surprised when my basic intuitive arguments turn out to be wrong.)
It's worth noting that they are mostly interested in critical phenomena in general, and earthquakes are kind of a drive-by application, treated along with fires and sand piles.
They do hint around the edges, but they don't head-on make the claim for earthquakes that small EQs materially lessen stress buildup and thereby make larger EQ's less likely.
I was looking for a credential of one of the people they interview, to see if they are really a solid earth person or more of a critical phenomena person -- their names aren't easy to find. This particular myth ("small earthquakes relieve stress") is a bit of a stinker in the solid earth community, and I think a solid earth person would be quite careful about their words as they discuss this.
The above link does not answer that question. It is relating stress release to "fault strength", or the maximum shear stress that can be withstood by the fault. There is an incidental relationship with depth that plays a role.
The video linked nearby (on criticality) also does not address the question at issue.
I'm only replying because I work adjacent to this area, and my understanding is that the idea that small EQ's release stress is a myth. Here [1] is another link, listed as #1 in the "Myths" category. And you can dig up quotes from none other than Lucy Jones [2] saying that this is a myth.
I don't work directly in this area, so I'm not willing to say absolutely no. But I'd really like to see a head-on reference supporting the claim that it's not a myth.
I feel like I’m being a bit curmudgeonly, but I don’t read them much any more.
See: https://fitelson.org/PrSAT/, and the linked paper: https://fitelson.org/pm.pdf
The paper starts off slow, but have patience to read up to section 4, Applications, which is kind of surprising.
But looking into the specifics again after all these years [1], I read:
"The N.S.A. wanted to clamp a lid on as much of the affair as it could. Within days, the agency’s National Computer Security Center, where the elder Morris worked, asked Purdue University to remove from its computers information about the internal workings of the virus."
and that CERT at CMU was one response to the incident [2].
So there is a whiff of the incident being steered away from public prosecution and towards setting up security institutions.
Robert Morris did get a felony conviction, three years probation, and a $10K fine. As for hn users, aside from pg, Cliff Stoll has a minor role in the story.
[1] https://archive.nytimes.com/www.nytimes.com/times-insider/20...
The model was that NASA did stuff that was pathfinding, typically in response to science objectives, and that commercial applications would follow. By design, it’s not mass production.
This works for Earth science stuff like land surface monitoring, methane monitoring, land subsidence, groundwater monitoring, sea level rise, etc. NASA developed these remote sensing technologies that have made it into commercial applications.
So there is a synergy between NASA science and commercial space. It does not have to be either/or.
>> What surprised me was that initial budget request, which basically said, we, America, are never going to launch another space telescope. We're going to turn off 95% of the ones we have in orbit. We are getting out of that business, we don't want to ask those questions anymore.
So this hits on a few key points. It’s not just that this budget request is tossing out perfectly good technology maturation plans for getting the next large space telescope built (https://science.nasa.gov/astrophysics/programs/habitable-wor...), among other goals.
It’s also (see the second sentence) that the budget request will result in de-orbiting perfectly-functioning operational missions like OCO-2 (https://ocov2.jpl.nasa.gov/), and deactivating perfectly functional instruments onboard ISS that are returning data continuously right now. It’s a multi-billion dollar self-own. There’s no sense in it. (https://www.planetary.org/charts/fy-2026-active-mission-canc...).
For many of these missions, having a long-term continuous dataset is super-valuable —- obviously so for a CO2 monitoring mission, or missions monitoring land surface temperature, vegetation/forests, etc. They are built, launched, and returning data. It’s all gravy at this point.
As nearby commenters note, this has nothing to do with cost savings. It’s more like a mix of pure spite, owning some libs in Maryland and California, and an object lesson in who the boss is.
Suppose you emulate a forward model y = F(x), by choosing a design X = {x1, ..., xN}, and making a training set T = {(x1, y1), ..., (xN, yN)}.
With T, you train an emulator G. You want to know how good y0hat = G(x0) is compared to y0 = F(x).
If there is a stochastic element to the forward model F, there will be noise in all of the y's, including in the training set, but also including y0! (Hopefully your noise has expectation 0.)
(This would be the case for a forward model that uses any kind of Monte Carlo under the hood.)
In this case, because the trained G(x0) is averaging over (say) all the nearby x's, you can see variance reduction in y0hat compared to y0. This, for example, would apply in a very direct way to G's that are kernel methods.
I have observed this in real emulation problems. If you're pushing for high accuracy, it's not even rare to see.
More speculatively, one can imagine settings in which (deterministic) model error, when averaged out over nearby training samples in computing y0hat, can be smaller than the single-point model error affecting y0. (For example, there are some errors in a deterministic lookup table buried in the forward model, and averaging nearby runs of F causes the errors to decrease.)
I have seen this claim credibly made, but verifying it is hard -- the minute you find the model error that explains this[*], the model will be fixed and the problem will go away.
[*] E.g., with a plot of y0hat overlaid on y0, and the people who maintain the forward model say "do you have y0 and y0hat labeled correctly?"
He is thinking about a random choice among the 20 edges branching out from each vertex.
The CLT gives a result about a recentered and rescaled version of the sum of iid variates. CLT does not give a result about the sum itself, and the article is invoking such a result in the “files” and “lakes” examples.
I’m aware that it can appear that CLT does say something about the sum itself. The normal distribution of the recentered/rescaled sum can be translated into a distribution pertaining to the sum itself, due to the closure of Normals under linear transformation. But the limiting arguments don’t work any more.
What I mean by that statement: in the CLT, the errors of the distributional approximation go to zero as N gets large. For the sum, of course the error will not go to zero - the sum itself is diverging as N grows, and so is its distribution. (The point of centering and rescaling is to establish a non-diverging limit distribution.)
So for instance, the third central moment of the Gaussian is zero. But the third central moment of a sum of N iid exponentials will diverge quickly with N (it’s a gamma with shape parameter N). This third-moment divergence will happen for any base distribution with non-zero skew.
The above points out another fact about the CLT: it does not say anything about the tails of the limit distribution. Just about the core. So CLT does not help with large deviations or very low-probability events. This is another reason the post is mistaken, which you can see in the “files” example where it talks about the upper tail of the sum. The CLT does not apply there.
I passed through the highly-irreproducible eras described in the section you link, and that you summarize in your last paragraph. There was so much different FP hardware, and so many different niche compilers, that my takeaway became “you can’t rely on reproducibility across any hardware/os version/compiler/library difference”.
There are still issues with libraries and compilers, summarized farther down (https://www.open-std.org/jtc1/sc22/wg21/docs/papers/2025/p37...). And these can be observed in the wild when changing compilers on the same underlying hardware.
But your point is that irreproducibility at the level of interrupts or processor scheduling is not a thing on contemporary mainstream hardware. That’s important and I hadn’t realized that.
Indeed!
The GRACE measurement of mass change is one of the more revolutionary advances in Earth science remote sensing in the last few decades. It has provided a unique and completely novel view of groundwater mass change. Grace is the main reason we know so much about the massive groundwater loss in the Oglala aquifer in the US Midwest, in the Central Basin in California, and in northern India. Water well data exists but it is very sparse and idiosyncratic.
It’s also our main window into mass losses in ice sheets in high latitudes (Greenland, Antarctica). We have radar altimetry data from Antarctica, but because of glacial rebound and other effects, it’s not easy to translate height changes into mass changes. Grace measures mass change directly.
Several authors of the cited study are on the science team. It is a JPL instrument.
The original Grace pair used radio to measure separation and velocity, while the follow-up Grace-FO uses a laser. I assume the small wavelength of the laser provides a more accurate measurement. It’s possible that Grace-FO has a slightly higher spatial resolution (I’ve worked with Grace but not Grace-FO); the horizontal resolution of Grace is about 100km or about 1 degree.
From an inference perspective the measurement is very interesting. They pool about a month’s worth of observations of the distance and velocity of a pair of satellites, and do a Bayesian inversion to obtain a parameterized gravitational potential for that month. The map from gravitational potential to observation is known analytically, so it’s readily possible to get a spatial covariance for the gravitational potential, as well as the point estimate.
They do give an example of a Chernoff (exponential) bound for a sum of iid random variables. The bound of course has an exponential form - they just don’t call it a large deviation. So it’s a bit of a missed opportunity, oven that the name is in the chapter title.
These bounds come up all over the place in CS, but especially lately in learning theory.
Similarly, we are bad at estimating small proportions ("easily shave 2%"). What is being claimed in the parentheses here is that there's a probability distribution of "how much costs are shaved" and that we can estimate where the bulk of its support is.
But we're not really good at making such estimates. Maybe there is some probability mass around 2%, but the bulk is around 0.5%. It seems like that's a small difference (just 1.5%!) but it's a factor of 4 in terms of savings.
So now we have a large number (annual spend), multiplied by a very uncertain number (cost shave, with poor experimental support), leading to a very uncertain outcome in terms of savings.
And it can be that, in reality, the costs of changing service turn out to overwhelm this outcome.