the same shape everywhere

a meditation on pattern recognition, and what physics might have to say about how to live

2026-06-25 · russell jiang

i spent an evening that was supposed to be econometrics homework and ended up somewhere near the foundations of physics. it wasn’t a detour. the whole point is that it wasn’t a detour.

here is the path, roughly. it started with a question about whether a regression coefficient actually means anything, and ended with the suspicion that the universe is just doing constrained optimisation in every direction at once. nothing in between felt like a leap. each step was just the previous idea wearing different clothes. this is an attempt to write down why that happens, and why i think noticing it is the single most valuable thing you can learn to do while learning anything.

01 one number, many disguises

start with covariance. it measures how two things move together. nothing fancy, just the average of how far each variable strays from its own mean, multiplied.

scale it by the variance of x and you get the slope of a regression line. scale it instead by both standard deviations and you get the correlation coefficient, which is now politely bounded between minus one and one. take logs of both variables first and the same slope becomes an elasticity, the thing economists use to talk about how demand responds to price. log one side only and you get a semi-elasticity instead.

so the demand elasticity in an economics course, the regression slope in a statistics course, and the correlation coefficient in a probability course are not three facts. they are one fact, projected onto three different walls. the covariance is the object. everything else is a choice of lighting.

once you see this you cannot unsee it, and you start getting suspicious whenever two things in different subjects have the same shape. usually it means they are the same thing.

02 the word for it

mathematicians have a word for “the same thing wearing different clothes.” the word is isomorphism: a structure-preserving map between two objects that looks different on the surface but behaves identically underneath.

the example that made it concrete for me is embarrassingly simple. a degree-two polynomial is just three numbers, the coefficients of one, x, and x squared. so the space of those polynomials is secretly just three-dimensional space with the axes relabelled. add two polynomials, you add their coefficient vectors. scale a polynomial, you scale the vector. every operation matches. they are the same space.

this is not a cute observation. it is why you can solve a question about polynomials being linearly dependent by stacking their coefficients into a matrix and asking a computer for the null space. the polynomial problem and the vector problem are the same problem. the isomorphism is what lets you carry the tools from one world into the other without paying any toll.

and it generalises in a way that is almost unsettling: every finite-dimensional real vector space of dimension n is isomorphic to ordinary n-dimensional space. there is, in a deep sense, only one vector space of each size. polynomials, matrices, lists of numbers, solutions to certain differential equations. all the same skeleton, different skin.

03 learning as compression

here is what i think is actually going on when something “clicks.”

learning a subject in isolation means storing it as its own block of facts. ten subjects, ten blocks, no shared structure. it is expensive to hold and it does not transfer.

but if you have already internalised the shape of, say, linear independence, then the first time you meet orthogonal functions, or feedback loops, or causal graphs, you do not store a new block. you recognise an old shape and inherit all its intuition for free. the new thing costs almost nothing because you are not learning it, you are relabelling something you already own.

this is why the returns to deep foundations are not linear. each foundation you lay does not just add itself. it makes every future subject cheaper to learn, because more and more of the new material turns out to be structure you have already seen. discrete maths felt abstract and disconnected while i was doing it. then it turned out that causal inference is graph theory, that induction is the engine behind why triangular matrices are easy, that modular arithmetic is a doorway into ring theory. the foundation was not a prerequisite to get past. it was the language everything else gets written in.

the skill underneath all of this is pattern recognition, and i have started to think pattern recognition is close to what intelligence actually is. not memorising more, not even reasoning more carefully step by step, but seeing that this new unfamiliar thing is structurally identical to that old familiar thing, and inheriting the whole apparatus in one move.

04 the shape that kept reappearing

covariance was the small version of this, one object showing up in three subjects. here is the largest version i know, the one that kept pulling me further out the longer i looked at it. it starts somewhere harmless.

the shortest path between two points is a straight line. in 4-unit high school maths it shows up as a fact about complex numbers, the triangle inequality. generalise the space and it becomes the Cauchy-Schwarz inequality, which is really just the statement that a cosine is never bigger than one, dressed up for any inner product space you like. correlation being bounded by one is this same inequality applied to data. cosine similarity in a search system, the thing that decides which documents are “close,” is this same inequality again.

that much is one shape spreading, the same trick as covariance. but pull on the thread a little harder and it changes character. shortest-path is a special case of something larger: not just distance being minimised, but quantities in general settling at their extreme. and that larger idea turns out to be nearly everywhere.

light travels the path that takes the least time. a particle follows the path that makes the action stationary, which in Feynman’s telling is because it secretly explores every path at once and only the stationary one survives the interference. a chemical reaction settles where the free energy is lowest, negotiating between energy wanting to be low and entropy wanting to be high, with temperature setting the exchange rate between the two. a probability distribution, given what little you know, settles at maximum entropy, which is just being as honest as possible about your own ignorance. a regression line settles where the squared error is smallest. a neural network settles where the loss is smallest, by rolling downhill.

these are not analogies. they are, mathematically, the same move: define a quantity, then find where it is stationary. the calculus of variations is the one tool underneath all of them, and the Euler-Lagrange equation is just the infinite-dimensional version of setting a derivative to zero to find a minimum. physics, chemistry, statistics, machine learning. one verb, conjugated differently.

05 the word that gets it wrong

it is tempting, once you have noticed all this, to draw a tidy life lesson from it. nature takes the path of least action, so maybe we should too. stop straining. take the path of least resistance. go where the current carries you and let things happen.

i believed that for about an hour, and then i realised it is a misreading, and the misreading is instructive.

least action does not mean laziness. the word “least” is doing something more specific than it sounds. the path a particle takes is not the one that costs the least effort. it is the one where the action is stationary, the path where nudging slightly to either side does not make things any better or worse. and in Feynman’s picture it only gets selected because every other path was secretly explored too, and this is the one that survived once all the alternatives interfered and cancelled. the straight line looks effortless. but it was chosen out of an enormous invisible space of paths not taken.

so the thing that looks like surrender is actually the opposite. it is not the absence of forces. it is their balance.

06 find your equilibrium

the better word, the one that survives contact with the actual mathematics, is equilibrium.

think about the free energy again. there is an equation for it, and it is one of those rare equations that says something true about more than what it was written for:

ΔH is enthalpy, the system’s pull toward low energy, the part of nature that wants to settle and rest and let go of heat. ΔS is entropy, the pull toward disorder, the part that wants to spread out and explore and never sit still. and T is temperature, which sets how much the second pull matters against the first. a system settles where ΔG is at its lowest, which is to say it settles not at lowest energy and not at highest disorder but at the exact point where the two stop fighting.

that is the whole push and pull of a life in one line. the part of you that wants to rest, set against the part of you that wants to move, with the conditions you happen to be living under deciding how much each one wins. and the resting point is not either extreme. it is the negotiated middle, the place where the tension resolves. nothing about it is drift.

equilibrium in general is like this. it is not the place where no forces act on you. it is the place where the forces cancel. and the difference between those two things is the whole difference between a life lived well and a life merely allowed to happen.

the pure go-with-the-flow philosophy has a quiet trap in it. let it happen can slide, without you noticing, into never choosing anything, calling your avoidance peace. but equilibrium does not let you off that easily. to find the balance point you have to actually move through the space. you have to try directions, feel where the tension pulls, lean into things and notice what pushes back, and settle only where the pushing resolves. you cannot find a stationary point without first exploring the paths around it. the stillness is real, but it is earned, and it sits at the end of a search rather than at the start of a surrender.

and your temperature changes. the conditions you are in shift where your balance point sits, the same way heating a reaction moves where it settles. the equilibrium you find at nineteen is not the one you will find at forty, and that is not failure, it is just the terms of the negotiation changing. the work is not to find the resting place once. it is to keep finding it as the temperature moves.

so when something in your life feels effortless, it is worth asking which kind of effortless it is. the empty kind, where you have stopped moving because moving is hard. or the balanced kind, where you have moved enough, in enough directions, that nothing is pulling at you anymore. they look identical from the outside. they are opposites from the inside.

07 why it matters

there is a version of going through life the way you can go through a degree: collecting tricks. a method for this kind of problem, a method for that kind, each one its own isolated thing. it works, mostly. but a person who only has tricks is stuck the moment a tool breaks or a situation arrives that none of their methods quite fit, because they were never holding the structure underneath, only the surface. the whole of part one was about the alternative: that underneath the tricks there are shapes, and if you learn to see the shapes you stop carrying a pile of separate methods and start carrying one way of seeing that travels.

what surprised me is that this turned out to be true past the edge of the maths. the same habit of looking for the shape underneath, the one that let me see covariance and elasticity as one object, polynomials and vectors as one space, is the habit that let me see least action and free energy and a settled life as one object too. i went looking for whether a regression coefficient means anything, and found a thread running from covariance through isomorphism through the calculus of variations all the way out to the principle that nature settles wherever its forces come into balance. it is all connected, not in a vague everything-is-one way but in a precise, provable, same-skeleton way. and then, almost by accident, the same shape had something to say about how to live, which is far more than i went looking for. you do not get that second kind of seeing without practising the first. learning to recognise structure in the abstract is what makes you able to recognise it in your own life, where the stakes are higher and the labels are hidden.

the still path is the balanced one, not the empty one. find where the forces cancel. that is the equilibrium, and it is the one thing in all of this you actually have to go looking for yourself ■

a dialogue distillate: my thoughts, drafted with claude.