Jekyll2021-10-23T00:03:47-04:00https://rwdb.xyz/feed.xmlRobert Dougherty-BlissA place, perchance, to dreamRobert Dougherty-Blissrobert.w.bliss@gmail.comI hate New Jersey Nazis2021-10-22T00:00:00-04:002021-10-22T00:00:00-04:00https://rwdb.xyz/i-hate-new-jersey-nazis<p>Last Thursday, members of the Rutgers New Brunswick math department received an
email entitled <strong>Klass Act</strong> with the subject line “<strong>Nazi visiting assistant
professor of mathematics at Rutgers</strong>.” The email accuses a former Rutgers
postdoc of running an anonymous Twitter account where he posted racist, sexist,
and, well, Nazi-ish takes.</p>
<p>Having reviewed the evidence, I can say two things:</p>
<ol>
<li>
<p>It’s definitely him. Digital evidence can always be faked, but barring that
it’s him.</p>
</li>
<li>
<p>The “Nazi” label has some justification. Some tweets were just edgy
conservative opinions, but others were full-on, all-out, “the Holocaust
didn’t happen and the Aryan race is pure” type takes.</p>
</li>
</ol>
<p>As you can imagine, this has generated some controversy around the department.
Grad students are up in arms, administration is investigating, and word is
getting around to our postdoc’s former students. It seems inevitable that
something unpleasant is going to happen to our postdoc.</p>
<p>While I share the general disgust at our postdoc’s opinions, and wonder whether
he ever behaved inappropriately with students, I find myself asking: What did
he do <em>wrong</em>?</p>
<p>After all, if some punishment is going to be handed out, he must have <em>done</em>
something to deserve it. Is it because he had the wrong opinions? Or is it
because he <em>shared</em> the wrong opinions? Perhaps we should have rules about what
opinions you are allowed to share?</p>
<p>We would not be the first to try. There is a long history in this country of
banning opinions. Popular victims are folks like draft-dodgers, war-critics,
and socialists. You might enjoy <a href="https://www.popehat.com/2012/09/19/three-generations-of-a-hackneyed-apologia-for-censorship-are-enough/">Ken White’s “Three Generations of a Hackneyed
Apologia for Censorship Are
Enough.”</a>
If we take anything from this history, it’s that punishing opinions is
a dangerous game to play.</p>
<p>On the contrary, rather than do “wrong,” by hiding his opinions on an anonymous
account, our postdoc did “right.” If we had never known about his tweets, then
there would be nothing to say, and no harm done.</p>
<p>But we <em>do</em> know about his tweets. And soon <em>lots</em> of people will know about
them. The genie is out of the bottle. As a result, our postdoc is now
a headache for his employer and a potential source of discomfort for his
students. Just as he has every right to spew hate, the community should have
every right to separate themselves from him.</p>
<p>Despite his awful views, I pity our postdoc. He didn’t ask for his opinions to
be thrust upon us. He won’t even get the satisfaction of defending his views;
they are far too shameful. He probably goes out with a whimper, afraid to stand
for anything in particular.</p>Robert Dougherty-Blissrobert.w.bliss@gmail.comLast Thursday, members of the Rutgers New Brunswick math department received an email entitled Klass Act with the subject line “Nazi visiting assistant professor of mathematics at Rutgers.” The email accuses a former Rutgers postdoc of running an anonymous Twitter account where he posted racist, sexist, and, well, Nazi-ish takes.Congratulations! You are being interrogated2021-09-23T00:00:00-04:002021-09-23T00:00:00-04:00https://rwdb.xyz/congratulations-you-are-being-interrogated<p>Towards the end of 2020, I was selected for the National Security Agency’s 2021
Graduate Mathematics Program, a Summer program meant to attract young
mathematicians to the NSA. After five months of security processing and around
twenty hours of polygraph exams, I withdrew my application in disgust.</p>
<p>All NSA employees must hold “Top Secret” security clearances. This involves
a background investigation of everywhere you have lived and worked for the past
ten years—including interviews of neighbors, coworkers, friends, and
partners—followed by in-person security processing at the NSA. This screening
is comprised of two parts: A psychological evaluation and a polygraph.</p>
<p>The psychological evaluation is uncomfortable, but bearable. You are led into
a large, white office with computers hidden by privacy shields. If you didn’t
know any better, you might think you were taking an all-digital SAT. You rank
a few hundred statements about yourself from “Not true at all” to “Very true.”
The statements start benign—”I am easily irritated”—but become more unusual—”I
frequently hear voices in my head.” You then answer some fill-in-the-blank
questions, such as: “If only my father would ____,” “If only I could forget
the time that I ____.” This is followed up by a detailed interview with
a psychologist:</p>
<ul>
<li>
<p>“How much porn do you watch? What kind?”</p>
</li>
<li>
<p>“Have you ever sent naked photos of yourself to anyone?”</p>
</li>
<li>
<p>“Tell me about your mother.”</p>
</li>
</ul>
<p>Uncomfortable, but bearable. The <em>unbearable</em> part is the polygraph.</p>
<p>A polygraph is a machine which measures various physiological properties of
your body. Breathing, heart rate, sweat glands, body movement, and so on. It
consists of two tubes wrapped around your torso, a cuff around your arm, and
small metal connectors on your fingertips. The machine is run by an “examiner”
who asks questions and evaluates the polygraph’s data. In theory, if you are
lying, then your physiological measurements will be abnormal, and the examiner
can detect this.</p>
<p>In an NSA polygraph you are alone with your examiner in a small, windowless
office in a nondescript building. The examiner begins by explaining that they
are there to help you. You are both on the same team. Everything will go well
if you relax and answer every question truthfully.</p>
<p>These are lies. The polygraph exam is not an exam, or a test, or an interview.
It is an interrogation. A cleverly disguised interrogation, but an
interrogation nonetheless.</p>
<p>Your examiner asks you about forbidden things—secret foreign contacts, illegal
drugs, serious crimes—and slowly focuses in on questions where you seem
uncomfortable. Did your breath quicken when asked about classified documents?
Heart rate tick up when questioned on falsifying security paperwork? Maybe you
twitched your leg at an unfortunate time. Congratulations! You are about to be
interrogated by the NSA.</p>
<hr />
<p>Each exam is about four hours long. The testing is done in bursts of
five-to-ten minutes. Once a few tests are done, your examiner steps out to
“examine your charts.” You don’t know how long he’s gone—there are no clocks.
You do know that you are still being watched, and that the polygraph machine
you are strapped to is still recording.</p>
<p>The interrogation begins when your examiner returns to “discuss his findings.”
He will use standard techniques. He will lie (“I’m here to help you”), downplay
the seriousness of possible offenses (“Everyone has smoked a little weed”), and
apply psychological pressure (“If you don’t tell me the truth, I don’t know how
long the position will remain open”). If you do not admit something damning,
then he will suggest things that you <em>might</em> have done. (“Maybe you showed
a copy of an upcoming exam to a student of yours once?”)</p>
<p>The polygraph is similar to any police interrogation you might have seen, only
more insidious for a simple reason: It requires you to lie.</p>
<p>Let me give an example. During one of my exams, the following exchange
occurred:</p>
<blockquote>
<p><strong>Examiner</strong>: Over the past few hours we’ve been together, you’ve seemed like
a fairly honest person. I want to verify that by asking you some more
questions.</p>
<p><strong>Robert</strong>: Sounds good.</p>
<p><strong>E</strong>: At the agency, we think of trustworthiness as a scale from 0 to 10, with
0 being the least trustworthy and 10 being the most trustworthy. Would you
say you’re a 0 or a 10?</p>
<p><strong>R</strong>: I suppose I’m pretty close to a 10.</p>
<p><strong>E</strong>: [Pauses] You’re pretty close to a 10? What does that mean? Are you a 9? So
one out of ten times we tell you a secret, you’ll sell it to the Russians?</p>
<p><strong>R</strong>: OK, when you put it that way, I suppose I’m a 10.</p>
<p><strong>E</strong>: Excellent. So, during the exam, I will ask you the following: Have you ever
lied to cover up a mistake?</p>
<p><strong>R</strong>: Surely I ha-</p>
</blockquote>
<p>At this point, my interrogator slammed a notepad onto the table and held the
point of his pen over it.</p>
<blockquote>
<p><strong>E</strong>: Someone who lied to cover up a mistake? That kind of person wouldn’t work
here. I’ll ask you again, and if I have to write anything down on this
paper, we’re going to have problems. Have you ever lied to cover up
a mistake?</p>
</blockquote>
<p>My examiner (Mike, or Mark, or Adam, or whatever) didn’t want the
truth—everyone has lied to cover up a mistake—he wanted me to lie.</p>
<p>Why? Because deceit is necessary for the polygraph.</p>
<p>Given a reading from his machine, how does the examiner know if it indicates
lying? Your blood pressure might be higher, your sweat glands might have
activated, but how do you know that indicates lying?</p>
<p>One way to be sure is to have an example of a “control” lie. If the examiner
could know for sure that you lied when answering one question, then he would
have a “clean” example to compare with other questions. Thus, the examiner
needs you to lie at some point. This leads to absurd exchanges like my example.</p>
<p>This is a stupid idea. I immediately recognized what the examiner was trying to
do, which gave me a choice: Respond truthfully or lie as instructed. Once
I decided, how should I ignore that we are now playing a game? That the
examiner is manipulating me? How do I trust anything from their mouth? How do
I trust <em>anything</em> that <em>anyone</em> at the NSA tells me?</p>
<p>It was at this moment the bottom fell out for me. <em>This man is an idiot. I’m
being watched on camera by other idiots. After my exam, all the idiots will get
together and have one, big idiot meeting about how stupid they think I am. We
are all lying to each other. I can’t trust anyone in this building, and they
don’t trust me.</em></p>
<p>I was tired, frustrated, and anxious. Neither I nor the examiners were being
truthful; I started to say what I thought they wanted to hear. Anything to make
it stop.</p>
<p>When the exam finally did stop, I was dumped into a lobby with around eighty
other applicants. No one talked about their exams, no one shared any looks, and
especially no one stood up and said, “Hey, was that fucking <em>bullshit</em> or
what?” We sat there in a maddening silence, together but isolated.</p>
<hr />
<p>To be fair to the polygraphy fans out there, I should say that I’m not an
expert. (The agency is clear that researching polygraphs makes you look very
suspicious.) But, since leaving the NSA applicant pool I have read up on
polygraphs, and the <em>actual</em> experts seem unimpressed.</p>
<p>Dr. Drew Richardson, an FBI Laboratory Scientist who specialized in polygraphy,
summarized the field in his 1997 testimony to congress:</p>
<blockquote>
<ol>
<li>[Polygraphy] is completely without any theoretical foundation and has absolutely no validity.</li>
<li>[Anyone] can be taught to beat this type of polygraph exam in a few minutes.</li>
</ol>
</blockquote>
<p>The government itself admits that polygraphs are not good enough to deny
security clearances. <a href="https://www.dni.gov/files/NCSC/documents/Regulations/SEAD-4-Adjudicative-Guidelines-U.pdf">Security Executive Agent
Directive-4</a>,
the standing policy on security clearances, states that you cannot be denied
a clearance just for failing a polygraph. You have to <em>confess</em> something
disqualifying. Hence the examiner.</p>
<p>The examiner is not there to run a machine; he is there to produce
a confession. <a href="https://books.google.com/books?id=_5AeDSz_dQ4C&pg=PA90&lpg=PA90#v=onepage&q&f=false">According to a report on government
secrecy</a>,
“over 95% of the information the NSA develops on individuals who do not meet
federal security guidelines is derived via [voluntary admissions from] the
polygraph process.” The polygraph is essentially the NSA’s only way to reject
applicants.</p>
<p>The examiner certainly isn’t there to catch spies. Aldrich Ames, the most
famous convicted spy in American history, passed two polygraph exams while
spying for the Soviet Union. <a href="http://fas.org/irp/congress/1994_rpt/ssci_ames.htm">In a letter to the Federation of American
Scientists</a>, Ames wrote:</p>
<blockquote>
<p>[The polygraph’s] most obvious use is as a coercive aid to interrogators,
lying somewhere on the scale between the rubber truncheon and the diploma on
the wall behind the interrogator’s desk. [It] has done little more than create
confusion, ambiguity and mistakes.</p>
</blockquote>
<hr />
<p>In total, I was polygraphed five times. That equals about twenty hours of exams
spread over three months. Add travel and waiting and it was much more time. If
I had any courage at all I would have quit after the first session. <em>Who are
they to ask what kind of porn I watch? To suggest that I lie and cheat and
steal at my workplace?</em></p>
<p>I could never sleep in the hotel the night before. My work suffered from the
time spent and the stress incurred. During every exam, I considered ripping off
the damn tubes and wires attached to me and storming out.</p>
<p>Yet I returned. Not one, not twice, but four times! Why?</p>
<p>I returned because the NSA has an alluring mystique. For all the heated
discourse over the Snowden revelations, most people can still only guess what
the NSA actually does. Every page of <a href="https://en.wikipedia.org/wiki/The_Puzzle_Palace"><em>The Puzzle
Palace</em></a> feels like
a real-life spy story, because <em>it is</em>. You can’t drive by the agency’s vast
compounds without wondering: What do they do in there?</p>
<p>The NSA offered an opportunity to do big things with mathematics, much more
important that whatever I would do on the outside. Are they morally dubious
things? Certainly. But I love my country enough to do morally dubious things
for it. Or at least, I thought I did.</p>
<p>I knew that the NSA inflicted a kind of Orwellian nightmare on the rest of the
world—we all know that thanks to Snowden—but I assumed that I would be shielded
from it. Even as an applicant, I felt that I was only ticking formalities.
I was in the club! I should have been above the fray and beyond the reach of
bureaucratic paranoia.</p>
<p>How could I forget what Orwell himself taught us? In <em>1984</em>, Syme is the member
of The Party tasked with turning the common language (“Oldspeak”) into a new,
simplified language with fewer, less-expressive words (“Newspeak”). Syme is the
ultimate insider. He understands the game and he believes in his work in the
full, cynical sense. Orwell summarizes his fate as such:</p>
<blockquote>
<p>One of these days, thought Winston with sudden deep conviction, Syme will be
vaporized. He is too intelligent. He sees too clearly and speaks too plainly.
The Party does not like such people. One day he will disappear. It is written
in his face.</p>
</blockquote>
<p>Insiders, outsiders, they are all the same. No one is free from The Party.
I was a fool to think otherwise.</p>Robert Dougherty-Blissrobert.w.bliss@gmail.comTowards the end of 2020, I was selected for the National Security Agency’s 2021 Graduate Mathematics Program, a Summer program meant to attract young mathematicians to the NSA. After five months of security processing and around twenty hours of polygraph exams, I withdrew my application in disgust.The second-rate mind of an expositor2021-08-15T00:00:00-04:002021-08-15T00:00:00-04:00https://rwdb.xyz/the-second-rate-mind-of-an-expositor<p><em>Exposition, criticism, appreciation, is work for second-rate minds.</em></p>
<p>At least, so says Godfrey Harold Hardy, the godfather of stuffy mathematicians
everywhere. To Hardy, the intellectual world is divided into two camps: Those
who <em>create</em>, and those who <em>commentate</em>. The creators are artists, scientists,
and politicians. The commentators are art-<em>critics</em>, expositors, and bloggers.
The creators possess a <em>first-rate</em> mind, sharp enough to forge connections
heretofore unseen. The commentators merely have <em>second-rate</em> minds, just
capable enough to <em>understand</em> what the greats do.</p>
<p>Like any respectable gate-keeper, Hardy makes sure that his preferred activity
is placed in the most prestigious category. Professional mathematicians are not
only creators, but creators of the <em>highest order.</em> “Immortality may be a silly
word,” says Hardy, “but probably a mathematician has the best chance of
whatever it may mean.”</p>
<p>Hardy is onto something here. Mathematical immortality <em>is</em> alluring. Finding
“the truth” or being a swell teacher is nice, but the real game is showing
everyone, for all time, just how great you really were. Why publish papers? Why
prepare talks? Why bash your head against difficult problems? Because if you
are successful, then you get to put a new bullet point on your CV, and if you
get enough bullet points in the right places, then you get respect, possibly
forever.</p>
<p>As Hardy puts it, “a man’s first duty, a young man’s at any rate, is to be
ambitious.” I am in <em>total agreement</em> with Hardy on this point. It is
refreshing to read such a blunt statement of purpose. Another excellent example
of this honesty is Richard Hamming’s <a href="https://www.cs.virginia.edu/~robins/YouAndYourResearch.pdf">“You and Your
Research”</a>, wherein
Hamming states that a scientist ought to desire <em>great</em> discoveries and the
prestige that they bring. Sure enough, both Hardy and Hamming produced their
share of amazing research.</p>
<p>And yet, for all the wonderful work that Hardy’s <em>first-rate</em> mind produced,
being a moderately well-known mathematician does not merit even a <em>footnote</em> in
any account of human history. He is less known to the man on the street, and
has made a smaller impact on their life, than the town barber. Occasionally
a mathematician stumbles into <em>true immortality</em>, like Euler, Newton, and
Euclid, but even then no one but mathematicians has more than a passing
knowledge of their <em>names</em>, much less what they <em>did</em>. Hardy himself wouldn’t
rank in the top fifty English historical figures, even of just the 20th
century. He would be lucky to be an answer on <em>Jeopardy!</em>.</p>
<p>In contrast, who are the most famous mathematical people alive today? They are
the mathematics <em>communicators</em>. People like <a href="https://www.singingbanana.com/">James
Grime</a> and <a href="https://standupmaths.com/">Matt
Parker</a>, both featured in Brady Haran’s excellent
<a href="https://www.numberphile.com/"><em>Numberphile</em></a> series. They are the mathematical
<em>spokespeople</em>, like <a href="https://www.poshenloh.com/">Po-Shen Loh</a>, who calls
himself an “educator and coach” <em>before</em> “academic and researcher.” They are
pop-mathematics authors like <a href="http://eugeniacheng.com/">Eugenia Cheng</a>, who
writes <em>New York Times</em> bestsellers. They are textbook authors like James
Stewart, who wrote one of the most popular calculus texts of the past twenty
years. Even my <a href="https://sites.math.rutgers.edu/~zeilberg/">famous advisor</a> is
<em>at least</em> as well known in the mathematics community for his <em>opinions</em> as he
is for his <em>research</em>. Expositors may have “second-rate minds,” but they also
have first-rate impact.</p>
<p>Lost in his pontificating, Hardy forgot that <em>no one cares</em> what mathematicians
do. Even other mathematicians <em>barely</em> care. Of all the academic disciplines,
mathematics has got to be one of the most obscure and meaningless to the
average person. Therefore, the greatest impact a mathematically-inclined person
might have on the world is to <em>explain</em> mathematics to the public, or to work
in some area that <em>uses</em> mathematics.</p>
<p>There is no greater demonstration of Hardy’s absolute disconnect than the
following passage, where he describes how a person might defend what
non-mathematical they do:</p>
<blockquote>
<p>“I do what I do because it is the one and only thing that I can do at all
well. I am a lawyer, or a stockbroker, or a professional cricketer, because
I have some real talent for that particular job. I am a lawyer because I have
a fluent tongue, and am interested in legal subtleties; I am a stockbroker
because my judgment of the markets is quick and sound; I am a professional
cricketer because I can bat unusually well. <strong>I agree that it might be better
to be a poet or a mathematician, but unfortunately I have no talent for such
pursuits.</strong>”</p>
</blockquote>
<p>Let’s do as Hardy suggests and pick a famous, contemporary athlete. Who is
better remembered by the world, Hardy or Babe Ruth?</p>
<p>I have great professional admiration for Hardy. The mathematical community
surely owes him a debt of gratitude. However, I must remind his spirit that,
outside of this small cult of strange people called “mathematicians,” he will
soon be forgotten. The tiny chance that he has to be remembered by the world is
not for his research, but for his reluctantly-written <em>Apology</em>.</p>
<p>Fortunately we do not need to pick favorites. The aim of research is not new
facts, but <em>interesting stories</em>. In this sense, expositors and researchers are
two sides of the same coin: A mine is of little use to the world without
a refinery to process the minerals. Together, researchers and expositors
produce stories for experts <em>and</em> non-experts, which lets everyone share in the
joy of human ingenuity. Amen.</p>Robert Dougherty-Blissrobert.w.bliss@gmail.comExposition, criticism, appreciation, is work for second-rate minds.COVID Brings Out the Sun for Stormy Petrels Women’s Volleyball2021-06-23T00:00:00-04:002021-06-23T00:00:00-04:00https://rwdb.xyz/covid-brings-out-the-sun-for-stormy-petrels-women-s-volleyball<p>The past year has been strange for sports at all levels. Shortened seasons,
cancelled games, and a surplus of “Out (COVID)” on injury reports. It felt like
everyone, from the NBA to your local rec league, just <em>barely</em> got through
their seasons. This disruption produced some unlikely outcomes, one of which
I’d like to explore here.</p>
<p>I went to my first Oglethorpe women’s volleyball game on September 2, 2017. The
Petrel’s were taking on Illinois Tech in the second game of the 2017 Stormy
Petrel Challenge. They fell to an early 0-2 deficit, and the third set seemed
lost when they found themselves down 17-22. Jacoby Sims scored three quick
kills to force an Illinois timeout at 20-22, and the Petrels rallied to a 26-24
victory. They won the next two sets by 5 and 3 points, respectively, completing
the 3-2 comeback. The game was fast, explosive, and tense. I was hooked.
I couldn’t wait to see the rest of the season.</p>
<p>After the Illinois Tech game, the Petrels went on a 13 game losing streak. They
ended the season 4-21 overall, and 2-12 in conference play. Needless to say,
that first game set my expectations a little too high.</p>
<p>You see, the Stormy Petrels are a <em>terrible</em> women’s volleyball team. I mean
<em>atrocious</em>. They have not had a winning season in the past ten years. In those
ten years, they went 16-127 in conference play, and never won more than
3 conference games in a single season. Whether we bring in new coaches, new
players, or new staff, it’s always the same. The Petrels are the bottom of the
barrel of the Southern Athletic Association, and they have been for a long
time.</p>
<p>And yet, I just can’t quit them. Even though they’re awful, even though no one
goes to the games, and even though I don’t even live in Georgia anymore,
I still hold out hope that maybe, just maybe, <em>this</em> is the year they turn it
around. Well, finally, <em>this</em> year came to pass, and the Stormy Petrels were
not a <em>complete</em> joke.</p>
<table>
<thead>
<tr>
<th>Season</th>
<th>Oglethorpe Conference record</th>
</tr>
</thead>
<tbody>
<tr>
<td><strong>2020-2021</strong></td>
<td><strong>6-5</strong></td>
</tr>
<tr>
<td>2019-2020</td>
<td>0-16</td>
</tr>
<tr>
<td>2018-2019</td>
<td>2-12</td>
</tr>
<tr>
<td>2017-2018</td>
<td>2-12</td>
</tr>
<tr>
<td>2016-2017</td>
<td>3-11</td>
</tr>
<tr>
<td>2015-2016</td>
<td>0-14</td>
</tr>
<tr>
<td>2014-2015</td>
<td>3-11</td>
</tr>
<tr>
<td>2013-2014</td>
<td>0-14</td>
</tr>
<tr>
<td>2012-2013</td>
<td>2-12</td>
</tr>
<tr>
<td>2011-2012</td>
<td>3-11</td>
</tr>
<tr>
<td>2010-2011</td>
<td>1-14</td>
</tr>
</tbody>
</table>
<p>In a shortened, conference-only season, second-year coach Anna Braun led the
Stormy Petrels to a <strong>6-5</strong> finish. You may look at that and say, “Robert,
that’s only 55%. It took a shortened season, a COVID-depleted conference, and
over ten years years to <em>barely</em> win half of their games?” And while, yes, you
would be correct, this is the Petrel’s best percentage, first winning season,
and best conference ranking since joining the SAA. For Oglethorpe, 6-5 is as
good as winning the title.</p>
<p>To explain how amazing this is, I want to provide some context. First we need
to take a detour into the world of analytics.</p>
<h1 id="introducing-elo">Introducing: <em>Elo</em></h1>
<p><a href="https://en.wikipedia.org/wiki/Elo_rating_system"><em>Elo</em></a> is a numerical system
used to rate the strength of teams. Arpad Elo first introduced it to rank the
strength of chess players, but it is easily adaptable to any head-to-head
competition. Here’s how it works.</p>
<p>An “average” team has an Elo score of 1500. Stronger teams have higher scores,
and weaker teams have lower scores. Your score increases if you win a game, and
decreases if you lose a game. Two teams with equal scores have even odds of
winning, having a score of 70 points higher gives you about a 60% chance of
winning, and having a score of 150 points higher gives you about a 70% chance
of winning.</p>
<table>
<thead>
<tr>
<th>Elo difference</th>
<th>Win probability for stronger team</th>
</tr>
</thead>
<tbody>
<tr>
<td>0</td>
<td>50%</td>
</tr>
<tr>
<td>75</td>
<td>60.5%</td>
</tr>
<tr>
<td>150</td>
<td>70%</td>
</tr>
<tr>
<td>300</td>
<td>85%</td>
</tr>
<tr>
<td>600</td>
<td>97%</td>
</tr>
</tbody>
</table>
<p>Like I said, you get a higher Elo score by winning games, and a lower score by
losing games. How much your score changes depends on the following ideas:</p>
<ol>
<li>
<p>A strong team beating a weak team gets a small score bump. (We expect strong
teams to beat weak teams.)</p>
</li>
<li>
<p>A strong team beating a strong team gets a moderate score bump. (We learn
about the strong teams, but not their relation to a third, weaker team.)</p>
</li>
<li>
<p>A weak team beating a strong team gets a <em>big</em> score bump. (This is the most
surprising situation.)</p>
</li>
</ol>
<p>The “base change” is 40 Elo points. That’s the most you could increase or
decrease after a game. You get a proportion of that that depends on your odds
of winning or losing the game. If you win, then you get 40 times the
probability that you would lose. If you lose, then you lose 40 times the
probability that you would win. For the other side, Elo is a zero-sum game: The
loser’s score goes down the exact amount the winner’s goes up.</p>
<p>Let’s see some examples.</p>
<p>If an average team (Elo 1500) met another average team (Elo 1500), we’d predict
that both teams have a 50% chance to win. The winner had a 50% of losing, so
they would get 20 points (Elo 1520) and the loser would fall 20 points (Elo
1480).</p>
<p>If an average team (Elo 1500) met an above average team (Elo 1600), we’d
predict that the above average team has about a 64% chance to win. If the
stronger team won, then they would only get 36% of 40, or 14 points (Elo 1614),
while the average team would only fall 14 points (Elo 1496). If, on the other
hand, the <em>average</em> team won, then they would get 64% of 40, or 26 points (Elo
1526), while the stronger team would fall 26 points (Elo 1574).</p>
<p>With this description of Elo out of the way, let’s see the history of the SAA
in a graph.</p>
<p><img src="/images/elo.png" alt="Elo scores" /></p>
<p>Using data compiled from the 2012-2013 season to the 2019-2020 season, I have
computed the Elo scores of every women’s volleyball team in the SAA and plotted
them above. (I left out Sewanee because they didn’t play in the 2020-2021
season.) The graph shows a clear divide into three groups:</p>
<ol>
<li>
<p><strong>The pack leaders,</strong> Berry, Birmingham-Southern, and Hendrix. These teams
have always been at least average, and have been moderately to very strong
in recent years.</p>
</li>
<li>
<p><strong>The middling runts,</strong> Rhodes, Centre, and Millsaps. These teams were
decent a while ago, but have been in a sharp decline recently. They are
firmly below average.</p>
</li>
<li>
<p><strong>The Stormy Petrels,</strong> the dregs of the SAA.</p>
</li>
</ol>
<p>Now that we have some context for how bad the Stormy Petrels are, 6-5 doesn’t
sound so bad, does it? How does Oglethorpe’s Elo look after this miraculous
winning season?</p>
<p><img src="/images/elo2.png" alt="Elo redux" /></p>
<p>Wow! Oglethorpe’s Elo <em>skyrocketed</em> after this barely-winning year. The wins
came from Centre (x2), Hendrix, Rhodes, and Millsaps (x2). The losses came from
Birmingham-Southern (x3), Rhodes, and Hendrix. Except for Rhodes, which nearly
cancels out, the losses were to teams so overpowered they barely lowered
Oglethorpe’s rating. The wins, since Oglethorpe had such a meagre rating,
produced much more.</p>
<h1 id="a-postseason-shocker-simulating-with-elo">A postseason shocker: simulating with Elo</h1>
<p>The <em>real</em> shocker is that 6-4 was enough to put the Stormy Petrels into the
SAA playoffs with their first top-4 finish <em>ever</em>. (Only the top-4 teams played
this year.) We can quantify how surprising this was by <em>simulating</em> the
2020-2021 season.</p>
<p>Once we know the Elo score of two teams, we can simulate a matchup between them
by flipping a biased coin. If Elo says the stronger team has a 75% chance to
win, then we flip a coin that comes up heads 75% of the time, and tails 25% of
the time. If we see heads, then the stronger team “won,” and otherwise they
“lost.” Proceeding in this way, we can simulate an entire season of games as
long as we know the Elo scores of the teams involved to a sufficiently high
accuracy.</p>
<p>So, I did that. I simulated the 2020-2021 season 10,000 times and figured out
how likely it was that each team made the playoffs. Oglethorpe had around a 20%
chance, which is probably their best chance at finishing top-4 in a <em>long</em>
time. The three top-dogs (Berry, Birmingham-Southern, Hendrix) all made the
playoffs, so Oglethorpe had to leap over the middle-of-the-pack teams (Centre,
Rhodes, Millsaps) to get fourth place.</p>
<table>
<thead>
<tr>
<th>Team</th>
<th>Probability of top-4 SAA finish (bold = actual)</th>
</tr>
</thead>
<tbody>
<tr>
<td><strong>Berry</strong></td>
<td><strong>97.92%</strong></td>
</tr>
<tr>
<td><strong>Birmingham-Southern</strong></td>
<td><strong>96.78%</strong></td>
</tr>
<tr>
<td><strong>Hendrix</strong></td>
<td><strong>85.73%</strong></td>
</tr>
<tr>
<td>Millsaps</td>
<td>53.29%</td>
</tr>
<tr>
<td>Centre</td>
<td>26.08%</td>
</tr>
<tr>
<td>Rhodes</td>
<td>21.09%</td>
</tr>
<tr>
<td><strong>Oglethorpe</strong></td>
<td><strong>19.11%</strong></td>
</tr>
</tbody>
</table>
<p>A 20% event happens, on average, once every five times, so it isn’t so crazy
that Oglethorpe would make the top-4. Of course, going any further than the
first round would probably require beating a big-dog team, and accordingly the
probability that Oglethorpe would win the SAA title was exceedingly small.</p>
<table>
<thead>
<tr>
<th>Team</th>
<th>Probability of winning SAA title (bold = actual)</th>
</tr>
</thead>
<tbody>
<tr>
<td>Berry</td>
<td>38.73%</td>
</tr>
<tr>
<td><strong>Birmingham-Southern</strong></td>
<td><strong>34.09%</strong></td>
</tr>
<tr>
<td>Hendrix</td>
<td>18.22%</td>
</tr>
<tr>
<td>Millsaps</td>
<td>4.17%</td>
</tr>
<tr>
<td>Centre</td>
<td>2.72%</td>
</tr>
<tr>
<td>Rhodes</td>
<td>1.31%</td>
</tr>
<tr>
<td>Oglethorpe</td>
<td>0.76%</td>
</tr>
</tbody>
</table>
<h1 id="whats-next-for-the-stormy-petrels">What’s next for the Stormy Petrels?</h1>
<p>Hot off their highest finish in a decade, where do the Stormy Petrels go? Will
the new coach lead us to glory? Will she crash and burn how all the rest have?
We don’t know for sure, but all the machinery I’ve laid out here will enable us
to make predictions about the next season. As soon as the schedule is released
I’ll be churning those out.</p>
<p>My mind tells me that the likeliest outcome is for Oglethorpe to have
a terrible year, like we always do. My heart tells me that we’re going to win
it all, baby.</p>Robert Dougherty-Blissrobert.w.bliss@gmail.comThe past year has been strange for sports at all levels. Shortened seasons, cancelled games, and a surplus of “Out (COVID)” on injury reports. It felt like everyone, from the NBA to your local rec league, just barely got through their seasons. This disruption produced some unlikely outcomes, one of which I’d like to explore here.Admission to Candidacy2021-05-18T00:00:00-04:002021-05-18T00:00:00-04:00https://rwdb.xyz/admission-to-candidacy<p>On Monday, 2021 May 17, I passed my oral qualifying exam at Rutgers University.
This means that I have <em>advanced to candidacy</em>. In other words, I have cleared
the necessary hurdles to be a “real” PhD student, and am “ready” to be a
researcher.</p>
<p>Of course, this implicitly means that I was <em>incapable</em> of doing research
before passing my exams. <a href="/publications/gcf">None</a> of
<a href="publications/az-recurrences">my</a> previous <a href="/publications/dyck">publications</a>
were <a href="/publications/beukers">“real”</a><sup id="fnref:quality" role="doc-noteref"><a href="#fn:quality" class="footnote" rel="footnote">1</a></sup>.</p>
<p>Thankfully, now that I am a “PhD candidate” as opposed to a mere “graduate
student,” the tides will turn. Indeed, with some luck, both the <em>quality</em> and
<em>quantity</em> of my work will improve because I am not <em>wasting my time on
formalities.</em></p>
<p>The purpose of a PhD program is to produce competent researchers. The <em>Big Lie</em>
about qualifying exams is that they support this purpose. In truth, qualifying
exams are mostly <em>redundant</em>. They measure things that the department already
knows. They exist <em>only</em> to place students into high-risk, low-reward
situations for long periods of time, which detracts from their opportunities to
do research.</p>
<p>I propose a modification to the qualifying exam structure, but let’s first see
how they work at Rutgers now.</p>
<p><strong>Written Qualifying Exam.</strong> This is a sequence of three 2.5 hour-long exams on
the three courses you take in your first semester. In each, you are asked five
random questions, and have to answer three correctly (on average) to pass. You
have three attempts, which will take you a full year to exhaust. The written
qualifying exam supposedly measures your technical ability in “basic”
mathematical subjects.</p>
<p>“But wait,” I hear you ask. “The exams are on the courses you took in your
first semester? Didn’t you get grades in those courses? Don’t those already
measure your ability?”</p>
<p>I am sorry to tell you that the current program mostly disregards your
coursework. “Your grades don’t matter” is a common refrain among graduate
students because it is true. Slave over every homework assignment? Go to every
office hour? Get perfect marks on every midterm? Congratulations! You have
achieved nothing. <em>One</em> exam, completed under threat of expulsion, is the
department’s <em>ultimate measure</em> of technical mastery.</p>
<p>Of course, your grade <em>is</em> a measure of your ability. You spent an entire
semester completing homework assignments, taking exams, and interacting with
your professors. The department would <em>already know</em> who puts in work and who
is “technically competent” if it cared to look.</p>
<p>I see two possibilities:</p>
<ol>
<li>
<p>The department is out-of-touch with student assessment or does not trust
professors to evaluate students.</p>
</li>
<li>
<p>The department wants to haze you.</p>
</li>
</ol>
<p>Which do you think it is?</p>
<p><strong>Oral Qualifying Exam.</strong> This is an 80-120 minute long meeting with
a committee of four faculty members. You will be questioned on a syllabus that
you wrote in consultation with your committee. In theory you have two attempts,
but it is exceedingly rare to fail. Supposedly this exam ensures that you have
not “overspecialized” before beginning research.</p>
<p>Unlike the written qualifying exam, there is no “standard” oral exam. Your
experience is highly dependent on your area and advisor. This gives it some
merit: Your syllabus is likely relevant to your proposed research, and the exam
is graded <em>subjectively</em> by your committee members. However, this subjectivity
makes the exam, again, redundant.</p>
<p>Suppose that your committee is confident that you will pass. Perhaps you’ve
been meeting with them regularly to discuss your syllabus. What, in this case,
is the purpose of the exam? Don’t they already know your knowledge? Suppose
instead that your committee is confident that you will <em>not</em> pass. Then why on
Earth are you taking the exam? In both cases, the exam merely makes you
demonstrate what the committee <em>already knows</em> in a high-pressure situation.</p>
<p>The only situation where the exam has <em>true</em> merit is when the committee does
not know if you are ready, and they are honestly willing to fail you. I posit
that this situation is <em>nearly nonexistent</em>. If it were a common occurrence,
then there would be a significant number of failed oral qualifying exams. There
are not (I think).</p>
<p>The benefit of the oral exam process is having four faculty members sign off on
your skills. This has nothing to do with the <em>exam</em> itself, but rather what
your committee thinks about you personally. Passing demonstrates that you can
convince four people to say that they like you.</p>
<p><strong>The Problem.</strong> Qualifying exams <em>take time that you could be doing research</em>.
If you demonstrate mastery of your first-semester courses, then you should
spend your second semester and the subsequent Summer meeting with potential
advisors. Doing anything else is a <em>waste of time</em>, as far as producing
researchers goes.</p>
<p><strong>The Solution.</strong></p>
<p>Here is a proposal:</p>
<ul>
<li>Offer a written qualifying exam to students as soon as they arrive.
<ul>
<li>Pass? Skip relevant first semester required courses.</li>
<li>Fail? Take relevant first semester required courses.</li>
</ul>
</li>
<li>First semester required courses.
<ul>
<li>A? Continue in the program.</li>
<li>Less than an A? You must pass a written qualifying exam by the end of
your first Summer.</li>
</ul>
</li>
<li>Oral exams are replaced with a requirement that four faculty members agree
that you know your syllabus sufficiently well. It is up to them to decide if
a formal exam is necessary.</li>
</ul>
<p>I hasten to say that I am not griping about Rutgers <em>in particular</em>. I am very
fond of Rutgers, its math department, and all of the people in it.</p>
<p>In fact, nothing I’ve said seems specific to Rutgers. I assume that this silly
exam process happens at most mathematics departments. Perhaps the situation is
similar in other STEM areas, and perhaps even in the humanities. All the more
reason to evaluate the evaluators.</p>
<div class="footnotes" role="doc-endnotes">
<ol>
<li id="fn:quality" role="doc-endnote">
<p>I don’t think that my outlined contributions are particularly
<em>good</em>, but they <em>are</em> real. <a href="#fnref:quality" class="reversefootnote" role="doc-backlink">↩</a></p>
</li>
</ol>
</div>Robert Dougherty-Blissrobert.w.bliss@gmail.comOn Monday, 2021 May 17, I passed my oral qualifying exam at Rutgers University. This means that I have advanced to candidacy. In other words, I have cleared the necessary hurdles to be a “real” PhD student, and am “ready” to be a researcher.Birthday Dominoes2021-03-06T00:00:00-05:002021-03-06T00:00:00-05:00https://rwdb.xyz/birthday-dominoes<p>(<em>This post is dedicated to the most important Pisces birthday I know: EK.</em>)</p>
<p><a href="https://en.wikipedia.org/wiki/Dominoes"><em>Dominoes</em></a> is a well-known game that
no one actually knows how to play. A much more accessible game is <em>tiling
dominoes</em>: I give you a grid, and you tell me if you can cover the whole thing
with dominoes.</p>
<p>For example, look at this 2x2 grid:</p>
<p><img src="/images/2x2-dom.jpg" alt="2x2" /></p>
<p>Easy! What about this 4x4 grid?</p>
<p><img src="/images/4x4-dom.jpg" alt="4x4" /></p>
<p>Also easy! What about a 3x3 grid?</p>
<p><img src="/images/3x3-dom.jpg" alt="3x3" /></p>
<p>It can’t be done! Any grid you can cover with dominoes has to have an even
number of squares, but a 3x3 grid has 9. We lose, through no fault of our own.</p>
<p>This pretty much solves grids. You can tile them with dominoes if and only if
they have an even number of squares. But this gives us two natural questions to
ask:</p>
<ol>
<li>
<p>What if we used something other than dominoes?</p>
</li>
<li>
<p>What if we used something other than grids?</p>
</li>
</ol>
<h2 id="something-other-than-dominoes">Something other than dominoes</h2>
<p>Dominoes are pieces with two blocks “snapped” together. What if we used more
than two blocks? This these exist, of course, and we call them <em>triominoes</em>,
<em>quadominoes</em>, <em>pentominoes</em>, and in general, <em>polyominoes</em>.</p>
<p>There is only “one” domino—two blocks stuck at their ends—but there are
<em>three</em> triominoes!</p>
<p><img src="/images/triominoes.jpg" alt="triominoes" /></p>
<p>That pesky 3x3 grid that we couldn’t tile with dominoes is a cinch with
triominoes:</p>
<p><img src="/images/3x3-filled.jpg" alt="3x3-filled" /></p>
<p>The most famous polyomino is the
<a href="https://en.wikipedia.org/wiki/Pentomino"><em>pentomino</em></a>, which has five blocks
stuck together. These are commonly used for fun in brain teasers, if you’re
into that sort of thing.</p>
<h2 id="something-other-than-grids">Something other than grids</h2>
<p>The shape of the board is just as important as the number of spaces. For
example, look at this “T” with four spaces:</p>
<p><img src="/images/t.jpg" alt="t" /></p>
<p>Even though there are an even number of spaces, we can’t tile this with
dominoes! Here’s a grid with 9 spaces that we cannot tile with triominoes:</p>
<p><img src="/images/weird-9.jpg" alt="weird 9" /></p>
<p>In general, a board with $n$ spaces is only guaranteed to have a tiling of
monomioes (pieces with 1 block) and an $n$-omio (a piece with $n$ blocks), both
of which are kind of cheating. Board layout plays a big role.</p>
<h2 id="the-heart-board">The heart board</h2>
<p>I propose that we think about tiling <em>the heart board</em>, something I invented
for just this occasion. Here are the first four heart boards:</p>
<p><img src="/images/filled.png" alt="the heart boards" /></p>
<p>It’s a bit hard to make out the “grid” in these pictures, so here are the first
two drawn by hand (kinda):</p>
<p><img src="/images/hand-hearts.jpg" alt="first two hearts" /></p>
<p>The number of blocks in the first four hearts is 10, 43, 96, and 169,
respectively. The hearts follow a pattern that generalizes to arbitrary sizes.
The $n$th heart board is the set of all $(x, y)$ such that</p>
<ul>
<li>$0 \leq x < 2n$</li>
<li>$0 \leq y < 4n$</li>
<li>$x \leq y$</li>
<li>$y \leq 5n - x$</li>
<li>$y \leq x + 3n$,</li>
</ul>
<p>and also the reflection of these points about the $y$-axis.</p>
<p>The $n$th heart board has exactly</p>
\[10 n^2 + 3n - 3\]
<p>spaces in it.</p>
<p>There are lots of questions to ask about this board, but let’s settle for just
one: When can the heart board be tiled by dominoes?</p>
<p>The first heart board cannot be tiled with dominoes. That’s easy enough to see
by hand because it only has 10 blocks:</p>
<p><img src="/images/hand-1-try.jpg" alt="trying to color the first heart" /></p>
<p>The second heart board is much bigger at 43 blocks, but this is an odd size so
no tiling by dominoes could exist. In fact, the general size $10 n^2 + 3n - 3$
is odd when $n$ is even, so only the “odd” heart boards could hope to be tiled
by dominoes anyway.</p>
<p>So what about the third heart block, the one in the bottom left of the
computer-generated image above? Can it be tiled using dominoes? It turns out
that <em>no</em>, it cannot be. How do I know this? My computer told me. In fact,
using the <a href="https://github.com/jwg4/polyomino">polyomino library</a>, my computer
told me more:</p>
<table>
<thead>
<tr>
<th style="text-align: right">Heart board number</th>
<th>Domino tiling?</th>
</tr>
</thead>
<tbody>
<tr>
<td style="text-align: right">1</td>
<td>no</td>
</tr>
<tr>
<td style="text-align: right">3</td>
<td>no</td>
</tr>
<tr>
<td style="text-align: right">5</td>
<td>no</td>
</tr>
<tr>
<td style="text-align: right">7</td>
<td>no</td>
</tr>
<tr>
<td style="text-align: right">9</td>
<td>no</td>
</tr>
<tr>
<td style="text-align: right">11</td>
<td>no</td>
</tr>
<tr>
<td style="text-align: right">13</td>
<td>no</td>
</tr>
<tr>
<td style="text-align: right">15</td>
<td>no</td>
</tr>
<tr>
<td style="text-align: right">17</td>
<td>no</td>
</tr>
<tr>
<td style="text-align: right">19</td>
<td>no</td>
</tr>
</tbody>
</table>
<p>It seems like no heart boards can be tiled by dominoes. Is this true?
Amazingly, yes! Not a single heart board can be tiled by dominoes.</p>
<p>The proof of this fact relies on a very simple observation: If you color the
squares of an odd heart board in an alternating fashion, then then there are
exactly two more squares of one color than the other. For example, look at the
first heart board:</p>
<p><img src="/images/hand-color.jpg" alt="filled first heart" /></p>
<p>Every domino you put down must cover exactly one of each color. In the above
picture, every domino covers one red and one blue tile. Once you place four
dominoes, there are no blue tiles left, but two red tiles. We can’t cover those
pieces with dominoes! (This is what happened in our attempted tiling above.)</p>
<p>This pattern persists for every odd heart board. Here is a plot of the first
four hearts again, now colored in this alternating way:</p>
<p><img src="/images/filled-alt.png" alt="filled, alternating colored hearts" /></p>
<p>The first and third heart boards above (left column) have exactly two more red
squares than blue squares. The second and fourth (right column) actually have
a <em>bigger</em> difference in the number of squares, but we already knew that they
couldn’t be colored with dominoes.</p>
<p>This pattern is somewhat tricky to prove, but once you know the idea it’s just
calculations. See <a href="https://math.stackexchange.com/questions/4051519/can-you-tile-a-heart-with-dominoes">my math.SE
question</a>
for details.</p>
<p>We’ve left lots of questions on the table that would be easy to answer. For
example, what’s the proportion of squares that are red versus blue in the even
heart boards? A harder project: Let $L_n$ be a set of lines which intersect the
square $[-n, n] \times [0, n]$ and each other. When is the region “inside”
$L_n$ tileable with dominoes?</p>
<p>I don’t know the answer to any of these offhand, but they sound fun. I hope
that this delivery on my promise of math art taught everyone something new.
Here’s to much more in the future!</p>
<p><img src="/images/filled-alt-big.png" alt="big hearts" /></p>Robert Dougherty-Blissrobert.w.bliss@gmail.com(This post is dedicated to the most important Pisces birthday I know: EK.)The Central Limit Theorem from 10,000 feet2020-12-10T00:00:00-05:002020-12-10T00:00:00-05:00https://rwdb.xyz/the-central-limit-theorem-from-10-000-feet<p>The Central Limit Theorem (CLT) states, roughly, that averaging large samples
produces an approximately normal distribution. More formally, it says (in
a special case) that if $\{X_k\}$ is an iid sequence of random variables with
mean $0$ and variance $1$, then $n^{-1/2} \sum_{k = 1}^n X_k$ converges weakly
to the standard normal distribution.</p>
<p>Every undergrad in the world has probably heard “and proving the central limit
is beyond the scope of this course.” Well, here I am to show a very
<em>high-level</em> proof. Sort of. I won’t give all the technical details, because
I don’t think they matter. What’s amazing about the theorem is that it follows
from relatively “simple” asymptotics.</p>
<p>So, here’s the CLT from 10,000 feet.</p>
<p>“It turns out”<sup id="fnref:levy" role="doc-noteref"><a href="#fn:levy" class="footnote" rel="footnote">1</a></sup> that weak convergence of distributions is more-or-less
equivalent to pointwise convergence of characteristic functions. If</p>
\[L(t) = \lim_n E[e^{itX_n}]\]
<p>exists everywhere and is continuous at $0$, then $L(t)$ is a characteristic
function of some random variable $X$, and the $X_n$ converge weakly $X$. For
us, this means that $n^{-1/2} \sum_{k = 1}^n X_k$ would converge weakly to
a standard normal if we could show that</p>
\[\lim_n E[e^{it n^{-1/2} \sum_{k = 1}^n X_k}] = e^{-t^2 / 2}.\]
<p>Suppose that the $X_k$ are distributed as $X$. Independence tells us that the
characteristic function of our sum is “nice”:</p>
\[E[e^{it n^{-1/2} \sum_{k = 1}^n X_k}] = E[e^{it n^{-1/2} X}]^n.\]
<p>At this point we would love to say, “the characteristic function of $X
/ \sqrt{n}$ is (insert thing related to the characteristic function of $X$),”
but this is not true. Characteristic functions do not behave nice under scalar
multiples. There are no nice answers here. We have to do some approximations.</p>
<p>Let’s use the following asymptotic expansion of the exponential function:</p>
\[e^{it n^{-1/2} X} = 1 + \frac{it X}{\sqrt{n}} - \frac{t^2 X^2}{n} + O(n^{-3/2}).\]
<p>Taking expectations yields</p>
\[E[e^{it n^{-1/2} X}] = 1 - \frac{t^2}{2n} + O(n^{-3 / 2}),\]
<p>so the characteristic function of $n^{-1/2} \sum_{k = 1}^n X_k$ is</p>
\[(1 - \frac{t^2}{2n} + O(n^{-3 / 2}))^n.\]
<p>If we take a log and apply its asymptotic expansion, then we get</p>
\[n(-\frac{t^2}{2n} + O(n^{-3 / 2})) = -\frac{t^2}{2} + O(n^{-1/2}).\]
<p>Thus, the characteristic function of $n^{-1/2} \sum_{k = 1}^n X_k$ is</p>
\[-\frac{t^2}{2} + O(n^{-1/2}),\]
<p>which means that it goes to $t^2 / 2$ as $n \to \infty$, which is the (log)
characteristic function of the standard normal. Done!</p>
<hr />
<p>Let’s look at some steps in a little more detail. Say, 5,000 feet.</p>
<p>Other than the business about characteristic functions, the only part we
cheated was the asymptotics. I wrote</p>
\[e^{it n^{-1/2} X} = 1 + \frac{it X}{\sqrt{n}} - \frac{t^2 X^2}{n} + O(n^{-3/2}),\]
<p>dropping the implicit $X^3$ factor in the big-Oh term. This is true if we
regard $X$ as just some constant, but as soon as we take expectations in the
next step, we’re getting into fishy territory. The $X$’s that we dropped may
have been important. To really examine this, we need to know how that big-Oh
term behaves with the $X$’s back in it.</p>
<p>It turns out that the exponential function has nice remainder properties for
certain arguments. If we write</p>
\[R_n(x) = e^{ix} - \sum_{k = 0}^n \frac{(ix)^k}{k!},\]
<p>then</p>
\[|R_n(x)| \leq \min(2 |x|^n / n!,\ |x|^{n + 1} / (n + 1)!).\]
<p>This is not so hard to prove: Note that $R_n’(x) = i R_{n - 1}(x)$, so that</p>
\[R_n(x) = i \int_0^x R_{n - 1}(t)\ dt.\]
<p>Since $R_0(x) = e^{ix} - 1$ is bounded in absolute value by $\min(2, |x|)$,
induction proves the claim.</p>
<p>Putting all this together, we get</p>
\[e^{i t X} = 1 + it X - \frac{t^2 X^2}{2} + R_2(t X).\]
<p>Expectation will kill the linear term ($E[X] = 0$), and the remainder satisfies</p>
\[|R_2(t X)| \leq \min( t^2 |X|^2, |t|^3 |X|^3 / 6),\]
<p>or</p>
\[|R_2(t X)| \leq t^2 \min( X^2, |t| |X|^3 / 6).\]
<p>The minimum is bounded by the integrable $X^2$, and also goes to $0$ as $t \to
0$. Thus</p>
\[E[R_2(t X)] = o(t^2),\]
<p>meaning that dividing by $t^2$ produces something that tends to $0$ as $t \to
0$. Letting $t \mapsto t / \sqrt{n}$ gives</p>
\[E[e^{it n^{-1/2} X}] = 1 - \frac{t^2}{2n} + o(t^2 / n),\]
<p>so the full characteristic function is</p>
\[(1 - \frac{t^2}{2n} + o(t^2 / n))^n.\]
<p>Taking a log and doing the right simplifications yields a log-characteristic
function of</p>
\[-\frac{t^2}{2} + n o(t^2 / n) \to -\frac{t^2}{2}\]
<p>for all $t$. <em>Now</em> we’re done.</p>
<hr />
<p>So, back to the 10,000 foot view. What did we do?</p>
<ol>
<li>
<p>Note that the characteristic functions had a <em>kind of</em> nice form, but needed
some asymptotic analysis.</p>
</li>
<li>
<p>Do asymptotic analysis assuming that the random variables are unimportant
with respect to $n$, then go back and check this if we need to.</p>
</li>
</ol>
<p>With this in mind, let’s try to prove something else: <em>Poisson distributions
with large means are approximately normal</em>.</p>
<p>More formally, let’s prove this: Let $X_n$ be a sequence of Poisson random
variables with mean $\lambda_n$ where $\lambda_n \to \infty$. I want to show
that</p>
\[G_n = \frac{X_n - \lambda_n}{\sqrt{\lambda_n}}\]
<p>converges weakly to a standard normal distribution.</p>
<p>Fortunately, here the characteristic functions are easy to compute. Let
$\phi_X(t) = E[e^{itX}]$ be the characteristic function of a random variable
$X$. Then,</p>
\[\phi_{G_n}(t) = E[e^{itG_n}] = e^{-it\sqrt{\lambda_n}} \phi_{X_n}(t / \sqrt{\lambda_n}).\]
<p>The characteristic function for a Poisson random variable is well-known:</p>
\[\phi_{X_n}(x) = e^{\lambda_n(e^{ix} - 1)}.\]
<p>So, with $x = t / \sqrt{\lambda_n}$, we get</p>
\[\begin{align*}
\log \phi_{G_n}(t) &= -it \sqrt{\lambda_n} + \log \phi_{X_n}(x) \\
&= -it \sqrt{\lambda_n} + \lambda_n (e^{ix} - 1).
\end{align*}\]
<p>We can do some asymptotics here, since $e^{ix} - 1 = ix - x^2 / 2 + O(x^3)$. This gives</p>
\[\begin{align*}
\log \phi_{G_n}(t) &= -it \sqrt{\lambda_n} + it \sqrt{\lambda_n} - \frac{t^2}{2} + O(t^2 / \sqrt{\lambda_n}) \\
&= -\frac{t^2}{2} + O(t^2 / \sqrt{\lambda_n}).
\end{align*}\]
<p>Letting $n \to \infty$ shows that $G_n$ converges weakly to a standard normal.</p>
<div class="footnotes" role="doc-endnotes">
<ol>
<li id="fn:levy" role="doc-endnote">
<p>Here’s where many of the technical details are hidden. There is a lot
of functional analysis hidden in the following statement. I don’t
understand it, you don’t understand it, so let’s just drop it and be on our
way. <a href="#fnref:levy" class="reversefootnote" role="doc-backlink">↩</a></p>
</li>
</ol>
</div>Robert Dougherty-Blissrobert.w.bliss@gmail.comThe Central Limit Theorem (CLT) states, roughly, that averaging large samples produces an approximately normal distribution. More formally, it says (in a special case) that if $\{X_k\}$ is an iid sequence of random variables with mean $0$ and variance $1$, then $n^{-1/2} \sum_{k = 1}^n X_k$ converges weakly to the standard normal distribution.Martingales and Pólya’s Urn2020-11-15T00:00:00-05:002020-11-15T00:00:00-05:00https://rwdb.xyz/martingales-and-p%C3%B3lya-s-urn<p>Consider an urn which begins with two balls, one white and one black. At each
time $n = 1, 2, \dots$, a ball is chosen uniformly at random and duplicated in
the urn. What happens in the long run?</p>
<p>This is an exercise in Williams’ <em>Probability with Martingales</em>, but I think it
is instructive to see how martingales can be “found” naturally when you look
for them, <em>a lá</em> the method of summation factors taught in <em>Concrete
Mathematics</em>.</p>
<p>Let $B_n$ be the number of black balls chosen by time $n$, so that $B_0 = 0$
and $B_1$ is uniformly distributed on ${0, 1}$. The proportion of balls that
are black after time $n$ is</p>
\[M_n = \frac{B_n + 1}{n + 2}.\]
<p>I have suggestively called this “$M_n$” for “Martingale.” Indeed, $M_n$ is
a martingale adapted to the sequence of sigma algebras defined by $F_n
= \sigma(A_1, \dots, A_n)$, where $A_n = [\text{a black ball was chosen at time
$n$}]$. It is not hard to check that this is true, but it seems lucky to me
that <em>exactly</em> what we want to study happens to be a martingale. What if we
were not so lucky? What if we had to <em>define</em> a martingale from scratch?</p>
<p>Let’s look for some motivation from our current problem. To check that $M_n$ is
a martingale, we will need to compute $E_n[B_{n + 1}] = E[B_{n + 1} | F_n]$, so
we may as well do that now:</p>
\[\begin{align*}
E_n[B_{n + 1}] &= B_n + E_n[[\text{black at time $n$}]] \\
&= B_n + \frac{B_n + 1}{n + 1} \\
&= (1 + 1 / (n + 1)) B_n + \frac{1}{n + 1}.
\end{align*}\]
<p>Okay, we now know that $E_n[B_{n + 1}]$ is related to $B_n$ by some type of
equation. Can we see how to define a martingale from here? <em>Yes!</em></p>
<p>In general, suppose that the sequence of random variables $B_n$ satisfies</p>
\[E_n[B_{n + 1}] = a_n B_n + b_n\]
<p>for some constant sequences $a_n$ and $b_n$. Let’s try to <em>make</em> a martingale
by defining $M_n = c_n B_n + d_n$ for some to-be-determined constant sequences
$c_n$ and $d_n$. To make $M_n$ a martingale, we need</p>
\[E_n[M_{n + 1}] = c_{n + 1} (a_n B_n + b_n) + d_{n + 1} = c_n B_n + d_n = M_n.\]
<p>It would suffice to choose $c_n$ such that</p>
\[c_{n + 1} a_n = c_n\]
<p>and $d_n$ such that</p>
\[c_{n + 1} b_n + d_{n + 1} = d_n.\]
<p>But these conditions are easy to satisfy! They give, say,</p>
\[c_{n + 1} = c_1 \prod_{k = 1}^n a_k^{-1}\]
<p>and</p>
\[d_{n + 1} = d_1 - \sum_{k = 1}^n c_{k + 1} b_k.\]
<p>In the case of our exercise, we have $a_n = 1 + 1 / (n + 1)$ and $b_n = 1 / (n + 1)$. If we solve the implied recurrences (being thankful that the sum which occurs in $d_{n + 1}$ telescopes), then we get precisely</p>
\[c_n = d_n = \frac{1}{n + 2},\]
<p>so the $M_n$ we would define is</p>
\[M_n = \frac{1}{n + 2} B_n + \frac{1}{n + 2} = \frac{B_n + 1}{n + 2}.\]
<p>Magic!</p>
<hr />
<p>Now that we have this martingale, what does it get us? Well, immediately it
tells us that</p>
\[E[M_n] = E[M_0] = \frac{1}{2},\]
<p>so we should always “expect” there to be a pretty even mix of balls in the urn.</p>
<p>What else? Well, this is a nonnegative (bounded!) martingale, so it has a limit
almost surely. Call that limit $M_\infty = \lim M_n$. What does this limit look
like?</p>
<p>Let’s take a moment generating function approach. Perhaps we can work out
$E[\exp(M_n z)]$, and then take a limit to get $E[\exp(M_\infty z)]$.</p>
<p>Exercise: If $B_n$ is as defined before (the number of black balls chosen by
time $n$), then $B_n$ is uniformly distributed on ${1, 2, \dots, n + 1}$.
(Hint: Work out a recurrence using the law of total probability.)</p>
<p>Using the above exercise, we can write</p>
\[E[\exp(M_n z)] = \sum_{k = 1}^{n + 1} \exp\left( \frac{(k + 1) z}{n + 2} \right) \frac{1}{n + 1}.\]
<p>This is just a geometric sum, a task fit for a computer. Maple readily spits
out</p>
\[E[\exp(M_n z)] = \frac{1}{n + 1} \frac{e^{(n + 4) z / (n + 2)} - e^{2z / (n + 2)}}{e^{z / (n + 2)} - 1}.\]
<p>Maple can also take the limit for us (but it is not so hard by hand in this
case) to show that</p>
\[\lim_n E[\exp(M_n z)] = \frac{e^z - 1}{z}.\]
<p>Well! This is intensely interesting. The function on the right is the moment
generating function for the <em>uniform distribution on $(0, 1)$!</em> Could this be
true? Could the limit of a discrete process be a continuous one? Look at these
graphs:</p>
<p><img src="/images/polya.png" alt="Urn histogram" />
<img src="/images/polya-1.png" alt="Urn paths" /></p>
<p>Why, it <em>must</em> be true! The graphs confirm it!</p>
<p>To assuage the hardened hearts of doubters, we should justify that</p>
\[\lim_n E[\exp(M_n z)] = E[\exp(M_\infty z)].\]
<p>This is not hard:</p>
\[|\exp(M_n z)| = \exp(M_n \operatorname{Re} z) \leq e\]
<p>for $|z| < 1$, say. So we can apply the dominated convergence theorem to get
our moment generating function</p>
\[E[\exp(M_\infty z)] = \frac{e^z - 1}{z},\]
<p>which shows once and for all that $M_\infty$ is in fact uniform on $(0, 1)$.
Wow!</p>
<hr />
<p>I’ve been thinking a lot about martingales for a probability class I’m taking
this semester. I still don’t quite <em>get</em> them, but I’m trying to do more and
more exercises to understand why we care about them. I <em>have</em> seen some good
examples so far, but I’m just now getting to writing down some of my favorites.</p>Robert Dougherty-Blissrobert.w.bliss@gmail.comConsider an urn which begins with two balls, one white and one black. At each time $n = 1, 2, \dots$, a ball is chosen uniformly at random and duplicated in the urn. What happens in the long run?An experimental approach to the drunk passenger problem2020-07-24T00:00:00-04:002020-07-24T00:00:00-04:00https://rwdb.xyz/the-drunk-passenger-problem<p>An apparently common question in
<a href="https://en.wikipedia.org/wiki/Quantitative_analysis_(finance)">quant</a>
interviews is as follows:</p>
<blockquote>
<p>Suppose that 100 passengers are boarding an airplane, all with assigned
seats. The first passenger is drunk, and sits in a random seat. Each following
passenger sits in their assigned seat, if it is available, and in a random seat
otherwise. What is the probability that the last passenger sits in their
assigned seat?</p>
</blockquote>
<p>I just wanted to quickly write up what I thought were four nifty proofs,
including one “experimental.”</p>
<p>First, you have to observe the fundamental recurrence. Let $p(n)$ be the
probability that the last passenger sits in the correct seat given that there
are $n$ passengers. It’s easy to check that $p(1) = 1$ and $p(2) = p(3) = 1/2$.</p>
<p>Suppose that there are $n$ passengers. If the drunk sits in his seat, then the
probability of success is $1$. If the drunk sits in the $k$th person’s seat
with $k > 1$, then the problem has been reduced to a drunk passenger with
a total of $n - k + 1$ passengers. (After the drunk sits, passengers $2$
through $k - 1$ sit, then passenger $k$ is so irate that they drink themselves
into a stupor.) Therefore,</p>
\[p(n) = \frac{1}{n} \sum_{1 \leq k < n} p(k).\]
<p>We now have a few ways to go about this.</p>
<h2 id="guess-and-check">Guess-and-check</h2>
<p>It’s easy to see that $p(2) = p(3) = 1 / 2$, so maybe the sequence is just
constant after $n = 2$? In principle we would want to compute a few more terms,
but without a computer this gets hard to do in your head. We can jump straight
to induction:</p>
\[\sum_{1 \leq k < n + 1} p(k) = 1 + \frac{n - 1}{2} = \frac{n + 1}{2}.\]
<p>Dividing by $n + 1$ gives $1 / 2$, so $p(n)$ is constant beyond $n \geq 2$.</p>
<h2 id="smarter-recurrences">Smarter recurrences</h2>
<p>There’s a standard trick to get rid of summations in recurrences, and in this
case it works exceptionally well. We can replace $n$ with $n + 1$ in the
fundamental recurrence, giving the following two equations:</p>
\[\begin{align*}
p(n) &= \frac{1}{n} \sum_{1 \leq k < n} p(k) \\
p(n + 1) &= \frac{1}{n + 1} \sum_{1 \leq k < n + 1} p(k).
\end{align*}\]
<p>Having the denominators on the right-hand side is a little awkward, so let’s
multiply them to the other side, then subtract the two equations:</p>
\[(n + 1)p(n + 1) - n p(n) = p(n).\]
<p>Simplifying yields</p>
\[(n + 1)(p(n + 1) - p(n)) = 0,\]
<p>and since $n$ is positive, we see that $p(n)$ is constant for $n \geq 2$.</p>
<h2 id="true-guess-and-check"><em>True</em> guess-and-check</h2>
<p>The previous method works so well because $p(n + 1) = p(n)$ falls out, and the
first guess-and-check method works because the induction is easy. However,
I claim that the <em>true</em> guess and check method is even easier!</p>
<p>We know, <em>a priori</em>, that the sequence $p(n)$ defined by $p(2) = 1 / 2$ and</p>
\[p(n) = \frac{1}{n} \left( 1 + \sum_{2 \leq k < n} p(k) \right)\]
<p>satisfies a linear recurrence with polynomial coefficients. (We just showed how
to prove such a thing in the previous method. Even without getting supremely
lucky, the result still holds.) Therefore, since $p(n) = 1 / 2$ for the first
dozen terms or so (exercise!), it follows<sup id="fnref:follows" role="doc-noteref"><a href="#fn:follows" class="footnote" rel="footnote">1</a></sup> that $p(n) = 1 / 2$ for all
$n \geq 2$.</p>
<h2 id="generating-functions">Generating functions</h2>
<p>This is an almost completely different way to do it.</p>
<p>Let $P(z) = \sum_{k \geq 2} p(k) z^k$. The fundamental recurrence can be
rewritten as</p>
\[(n + 1) p(n) = 1 + \sum_{2 \leq k \leq n} p(k)\]
<p>for $n \geq 2$, and it is an easy application of well-known
generatingfunctionology that this implies</p>
\[zP'(z) + P(z) = \frac{z^2}{1 - z} + \frac{P(z)}{1 - z}.\]
<p>This is a linear differential equation, with general solution given by</p>
\[P(z) = \frac{c + z^2 / 2}{1 - z}\]
<p>for some constant $c$. It is easy to check that $P’‘(0) = 1$, and this gives $c
= 0$. Thus</p>
\[P(z) = \frac{z^2}{2(1 - z)} = \frac{z^2}{2} + \frac{z^3}{2} + \frac{z^4}{2}
+ \cdots.\]
<p>We should note that this approach is almost entirely mechanical. Computers can
derive the differential equation and then solve it, almost unaided by humans.</p>
<h2 id="conclusions">Conclusions</h2>
<p>I don’t know what this problem has to do with the stock market, but it sure is
fun!</p>
<div class="footnotes" role="doc-endnotes">
<ol>
<li id="fn:follows" role="doc-endnote">
<p>To be completely rigorous, we need to say, “and the order of the
recurrence is no larger than $K$,” and then check $K$ terms, but at a glance we
know that $K$ is certainly no more than ten, or whatever. <a href="#fnref:follows" class="reversefootnote" role="doc-backlink">↩</a></p>
</li>
</ol>
</div>Robert Dougherty-Blissrobert.w.bliss@gmail.comAn apparently common question in quant interviews is as follows:A small binomial sum2020-06-24T00:00:00-04:002020-06-24T00:00:00-04:00https://rwdb.xyz/a-small-binomial-sum<p>I recently came across an interesting binomial sum in <a href="https://math.stackexchange.com/questions/3729998">this Math.SE
question</a>. As usual,
Zeilberger’s algorithm makes short work of it, but the “human” approach reveals
a small miracle.</p>
<p>We want to evaluate</p>
\[R_n(z) = \sum_k (-1)^k {2k \choose k} {k \choose n - k} z^k.\]
<p>Our first step is to use the identity ${2k \choose k} = (-4)^k {-1/2 \choose
k}$ and write</p>
\[R_n(z) = \sum_k {-1/2 \choose k} {k \choose n - k} (4z)^k.\]
<p>There might be a nice, direct way to evaluate this sum, but I don’t know it.
Instead, let’s apply the <em>snake oil method</em>. Let $R(x) = \sum_{n \geq 0} R_n(z)
x^k$. By expanding $R_n(z)$ and interchanging the order of summation, it is
easy to show that</p>
\[R(x) = \frac{1}{\sqrt{1 + 4zx(1 + x)}}.\]
<p>The radicand in the denominator is a quadratic polynomial in $x$ with
discriminant $16z(z - 1)$. The square root disappears precisely when $z = 0$ or
$z = 1$! If $z = 0$ the sum is trivially $1$, and if $z = 1$ we get</p>
\[R(x) = \frac{1}{1 + 2x},\]
<p>so</p>
\[R_n(1) = \sum_k 4^k {-1/2 \choose k} {k \choose n - k} = (-2)^n.\]
<p>In other words, the sum given has a nice, exponential solution in <em>only</em> the
case the Math.SE question asked. Remarkable!</p>
<h2 id="reflections">Reflections</h2>
<p>Zeilberger’s produces a recurrence for $R_n(z)$ straight away. Running
<code class="language-plaintext highlighter-rouge">ZeilbergerRecurrence(f, n, k, R, 0..n) assuming n::posint</code> where $f$ is the
summand of $R_n(z)$ yields</p>
\[(n + 2) R_{n + 2}(z) + 2z(2n + 3) R_{n + 1}(z) + 4z(n + 1) R_n(z) = 0.\]
<p>So, in general, $R_n(z)$ is holonomic in $n$ and needs no more than three terms
in its defining recurrence. Of course, if $z = 1$, then <code class="language-plaintext highlighter-rouge">ZeilbergerRecurrence</code>
instead yields</p>
\[F(n + 1) + 2 F(n) = 0,\]
<p>so it sometimes simplifies. Maple fails to solve the general recurrence when
$z$ is symbolic.</p>
<p>If we were good enough with complex functions, then we could work out
asymptotics, but in any event this is another avenue to explore.</p>Robert Dougherty-Blissrobert.w.bliss@gmail.comI recently came across an interesting binomial sum in this Math.SE question. As usual, Zeilberger’s algorithm makes short work of it, but the “human” approach reveals a small miracle.