The heart of linear algebra isn't "crunching matrices" — it's an epistemological claim: an object is invariant, but the basis (set of coordinate axes) you describe it in decides how complex it looks. The same vector has messy coordinates in one basis and a single component in another. A change of basis is a change of perspective, and the right one makes a tangled problem decouple on its own.
Non-trivial points: ① Eigenvectors are the directions a transformation only stretches, never rotates — a system's natural axes. Find the eigenbasis and a coupled transformation becomes independent scalings of each component (diagonalization). Finding eigenvectors is asking "in what directions does this system keep itself?" ② Seemingly unrelated techniques — Fourier transforms, principal component analysis (PCA), word embeddings — are all the same move: re-express the data in a basis where the problem becomes linear and separable. ③ The whole deep-learning revolution is one sentence: automatically learn a basis where entangled data becomes linearly separable. "Representation" just means "choice of basis."
The Fourier transform: a messy sound wave is a writhing curve in the time domain, nearly intractable; switch to the frequency basis (sinusoids as axes) and the same signal becomes a few clean vertical lines — denoising, compression, and tuning turn trivial. The signal didn't change; a change of basis made the hard problem vanish.
LLM word embeddings: words live in a high-dimensional space, and "king − man + woman ≈ queen" works because the learned basis aligns semantic axes like "gender" and "royalty" into independent coordinates. Likewise a stuck product decision — if you keep weighing "cost vs experience" on entangled axes, try another basis (say "reversibility vs learning value"), and the dilemma often decouples into two separately decidable problems.
Topology studies the properties that survive continuous deformation (stretching, bending — but no tearing, no gluing). A handled coffee mug can be continuously squeezed into a donut; to topology they are the same object, because both have exactly one hole. The power: once you stop fussing over distance, angle, and size — all mutable — and instead ask "what can't change no matter how I deform this," you've grasped the system's essential skeleton.
Non-trivial points: ① Topological invariants (connectedness, number of holes, boundaries) jump only when something tears — so they're robust to noise and perturbation: stable under any wiggle, and when they change, it's a qualitative leap. This is the mathematical language of phase transitions, identity persistence, and critical points. ② "Can you get from A to B" is purely topological, independent of distance — that's why it's more fundamental than metric. ③ A Buddhist resonance: a river, a "self" — endlessly deforming, yet seemingly carrying some invariant (the continuing stream) through it all; topology also allows the other answer — perhaps there is no invariant, only the deformation itself (non-self). The discipline is to first separate the mutable details from the skeleton that breaks only when torn.
The Seven Bridges of Königsberg: can you walk every bridge exactly once? Euler's insight was to erase the shapes and distances of islands and bridges, keeping only the graph of "what connects to what." Once reduced to pure connectivity, the answer is immediate (impossible). This was the founding moment of topological thinking: the key to the problem is first throwing away the irrelevant metric detail.
In distributed systems, a "network partition" is essentially a topological event: connectivity doesn't degrade gradually — at some instant it breaks discontinuously, forcing the CAP tradeoff (you can keep consistency or availability, not both). In parenting: a child's core temperament is a topological invariant — it gets continuously stretched and shaped by growth, but you can't "tear" it off by force. The wise move isn't to fight the invariant but to deform along with it — channel the same energy toward a different outlet.
Category theory proposes a radical shift: don't study an object's internal construction; study the arrows (maps) between objects and how those arrows compose. A thing's essence lies not in its material but in how it relates to and chains with other things. This is exactly the mathematical origin of software engineering's "program to the interface, not the implementation" — what a module is matters less than its input/output types and whether it composes cleanly with others.
Non-trivial points: ① Composability is the whole game: if A→B and B→C reliably chain into A→C, you can stack arbitrarily complex systems without losing control. This is the mathematics of modularity — Lego is Lego only because its interfaces compose. ② Isomorphism is the rigorous form of analogy: when an invertible, composition-preserving map exists between two structures, they are mathematically the same thing — every theorem proven in one domain transfers intact to the other. Recognizing "my scheduling problem is isomorphic to graph coloring" means inheriting that domain's entire solved body of results for free. ③ A functor is a structure-preserving translation — a "reliable analogy" that maps the whole web of relations across seamlessly.
Function composition and "interface over implementation": the Unix pipe cat | grep | sort is category theory's spirit made live — each command is an arrow (input stream → output stream); you don't care how it's written internally, only that the types connect, so you can chain them into powerful pipelines at will. The entire Unix philosophy is one line: make small programs composable.
Designing an AI agent pipeline: each agent is an arrow (input type → output type), and whether the system runs depends on whether one's output type cleanly connects to the next's input — designing agents is designing composable morphisms. And when you notice your "team task-scheduling problem" is isomorphic to graph coloring, you can import that field's mature algorithms wholesale, saving the cost of thinking from scratch. Spotting an isomorphism is the sharpest move in cross-disciplinary transfer.
High-dimensional data almost always curls up on a low-dimensional "manifold": it looks like hundreds of variables, yet only a handful of directions truly vary. A core act of intelligence is finding those few axes that matter and compressing away the rest — that is compression, and it is abstraction. And "similarity" hides an often-missed trap: without specifying a space and a metric, "similar" is meaningless. Two things sit close together in one projection and opposite in another.
Non-trivial points: ① Good dimensionality reduction keeps the signal and drops the noise — the same coin as Card 1's "choosing a basis" (e.g. PCA). ② The curse of dimensionality: in high dimensions, volume concentrates near the shell and "the distance between any two points tends to equalize" — so "nearest neighbor" loses meaning, and intuitions that work in low dimensions fail en masse. Counterintuitive, but unavoidable in high-dimensional data. ③ For BigCat: embedding vectors, semantic retrieval, vector databases all "define similarity in some learned low-dimensional space" — retrieval quality equals reduction quality. ④ But never forget: reduction is lossy — a map, not the territory itself. Compressing a person into three labels is efficient; the cost is that the dimension you compressed away is often exactly the one you most needed to see this time.
Any world map: projecting a 3D earth onto a 2D sheet must distort something — the Mercator projection, to preserve angles, draws high-latitude Greenland far larger than it really is. Every reduction preserves one thing while sacrificing another; the question is always "which dimension should I preserve this time?"
RAG (retrieval-augmented generation) and semantic search: the system compresses both documents and the query into one embedding space and retrieves relevant chunks by cosine similarity — recall is only as good as whether this reduction preserved the few semantic axes that actually determine relevance. Everyday decisions work the same: evaluating a candidate or plan, you mentally reduce it to two or three dimensions and score it; the real risk is that you happened to project away the one dimension that mattered most. Just asking "did I compress away an important axis?" often rescues a bad judgment.