Projection¶
Core Idea¶
Map a richer source onto a lower-dimensional target along a chosen direction, collapsing the dimensions perpendicular to the target while preserving those parallel — the direction determines what gets lost.
How would you explain it like I'm…
Shadow on the Wall
When you hold your hand in front of a flashlight, it makes a flat shadow on the wall. The shadow keeps some things, like the shape of your fingers, but loses others, like how thick your hand is. A projection is making that flatter picture, where you keep some of the thing and throw the rest away.
Flattening With Leftovers
A projection is squashing something with lots of detail down onto a simpler version, along a chosen direction, keeping what lines up with your target and tossing the rest. A shadow does this: a 3D hand becomes a 2D outline, and the direction of the light decides what gets flattened. A neat clue that it's a true projection is that doing it twice gives the same answer as doing it once: a shadow of a shadow is the same shadow. It's useful when the part you keep holds the important stuff and the part you throw away was just noise, but misleading when you accidentally throw away the part that mattered. So the thrown-away part deserves a name too: it reminds you that you have a picture from one direction, not the whole thing.
Collapse Along a Direction
A Projection is the structural move of mapping a richer, higher-dimensional object onto a lower-dimensional or constrained representation along a chosen direction, deliberately collapsing the dimensions perpendicular to the target while keeping those parallel to it. Four parts define it: a *source* with more degrees of freedom than will be kept; a *target* onto which the source is mapped; a *projection direction* (sometimes explicit, sometimes hidden in the choice of target) that determines what gets lost; and a *residual* — the orthogonal complement, everything the projection threw away. Its algebraic signature is *idempotence*: applying the projection a second time gives the same result, which is what distinguishes it from an arbitrary lossy map. A projection is informative when the target captures the load-bearing variation and the residual is just noise; it's misleading when the target was a poor choice and the residual carried the real signal. Naming the residual is what turns 'I have the picture' into the honest 'I have *a* picture, from one direction.'
A Projection is the structural move of mapping a higher-dimensional or richer object onto a lower-dimensional or constrained representation along a chosen direction or onto a chosen target, deliberately collapsing the dimensions perpendicular to the target while preserving those parallel to it. Four commitments define it: a source with more degrees of freedom than will be retained; a target subspace, surface, or constrained representation onto which the source is mapped; a projection direction — sometimes explicit, sometimes implicit in the choice of target — that determines what gets lost; and a structural identity, idempotence, whereby applying the projection a second time gives the same result, the algebraic signature distinguishing a projection from an arbitrary lossy map. The skeleton has four parts: a source space with structure to be reduced; a target; a projection direction (the equivalence relation collapsing source into target); and a residual, the orthogonal complement of what was kept. The move is informative when the target captures the load-bearing variation and the residual is unimportant or noise, and misleading when the target was a poor choice and the residual carried the signal — so the residual is a first-class object, not an afterthought; naming it turns 'I have the picture' into the honest 'I have *a* picture, from one direction.' Projection is the structural complement of representation: it is the *act* of producing a particular representation by deciding what to drop. A solid's shadow is a projection (3D source, 2D target, direction set by the light); a flat map is a projection (curved surface to flat plane); a summary statistic is a projection (high-dimensional reality to one scalar); an executive summary is a projection (long argument to short statement, direction set by what the reader needs). The idempotence signature — projecting twice equals projecting once — is what makes 'summary of a summary equals the summary' hold when the directions agree, and its failure is a diagnostic that what looked like a clean projection was really a richer, drifting transformation.
Broad Use¶
- Linear algebra: Least-squares regression projects the response onto the column space; PCA projects onto the leading eigendirections.
- Cartography: A spherical Earth maps to a flat plane via a projection that preserves some properties and necessarily distorts others.
- Graphics: A 3D scene maps to a 2D image via a camera projection, with depth as the lost residual.
- Databases: The relational select operation literally drops columns not in the target.
- Statistics: Dimensionality reduction and sufficient statistics retain the inferentially relevant variation and discard the rest.
- Optimization: Projection-onto-convex-sets solves constrained problems by repeatedly projecting onto the feasible set.
- Reporting: A status report projects complex work onto a short summary chosen to answer the audience's question.
Clarity¶
Exposes that every reduced representation is the output of a projection choice, making the direction and the discarded residual first-class objects a consumer must interrogate.
Manages Complexity¶
Compresses a high-dimensional object into a few decision-relevant dimensions, and engineers comparability by aligning two sources onto a shared target.
Abstract Reasoning¶
Idempotence — projecting twice equals projecting once — is the honesty check distinguishing a clean projection from a drifting transformation, and a metric that keeps moving under re-summarization fails it.
Knowledge Transfer¶
- PCA → any reduction: "Preserve the leading directions the reader needs, reserve the residual for footnotes."
- Map projections → dashboards: "No projection preserves everything" — a metric tuned for trend distorts magnitude, and the choice is which property is load-bearing.
- Projected-gradient → constrained settings: "Move, then project back to feasible" recurs as clipping in ML and snap-to-spec in engineering.
Example¶
A satellite's twelve-channel image is projected onto a single vegetation-index direction carrying the crop-stress signal; the other ten channels become residual, and a different decision (fire risk) demands a different projection of the same image.
Relationships to Other Primes¶
Parents (1) — more general patterns this builds on
- Projection is a kind of, typical Abstraction — The file: 'Projection is the precise, geometric, idempotent SPECIAL CASE within the broader family of reductions that abstraction names.' Projection drops the dimensions perpendicular to a chosen target (vs abstraction's free-form essence-extraction).
Children (1) — more specific cases that build on this
- Perspective is a kind of, typical Projection — The file: projection is the substrate-neutral mechanical act;
perspectiveis the viewpoint-laden version 'wearing projection's neutral clothing' — perspective = projection whose DIRECTION encodes interests. Projection is the more-general structural parent of the (canon) perspective. ADDITIVE parent edge; canon perspective is the art-depth one, so confidence medium and owner verifies the sense matches.
Path to root: Projection → Abstraction
Not to Be Confused With¶
- Projection is not Perspective because projection is the mechanical act along a given direction, whereas perspective carries a viewpoint-laden direction encoding interests that must be argued.
- Projection is not Abstraction because projection drops dimensions perpendicular to a target and is idempotent, whereas abstraction discards detail to keep an essence with no such signature.
- Projection is not Representation because projection is the operation of deciding what to drop, whereas representation is the resulting reduced form.