Hold a mug in front of a wall.
From one angle, the shadow looks like a rectangle.
Rotate the mug and the handle suddenly appears.
The mug did not become more real.
The projection direction changed which part of reality could survive the flattening.
Projection direction is the rule that decides which distinctions become visible, which become small and which collapse into the same place.
This is the first pillar beneath How Lossy Works | Projection. The master owns the general bargain of reducing a richer world to a smaller representation. This article owns orientation: why changing the viewpoint, basis or axis can change the story even when the source remains fixed.
Quick Read
A projection maps a source onto a lower-dimensional representation. The result depends not only on the source but also on the direction or basis used for projection. A camera angle can hide depth. A map orientation can foreground one route. PCA chooses axes of maximal variance rather than human semantic importance. A two-dimensional plot can therefore be truthful about one relationship while hiding another. Responsible analysis compares plausible projections and asks whether the claim survives a reasonable change of viewpoint.
source → choose view/basis → project → visible structure + collapsed structure → interpretation → alternate view → compare → bounded claim
The Object Is Not the View
A cube projected head-on can look like a square.
Rotate it and multiple faces become visible.
The source object has one geometry.
The receiver can experience many shadows of that geometry.
Projection direction decides which dimensions align with the viewing surface and which point partly or entirely away from it.
Foreshortening Is Loss Made Visible
A long object pointing toward the camera can look short.
Its real length has not disappeared physically.
It has been compressed into depth, a dimension the flat image does not directly preserve.
Foreshortening is a simple everyday demonstration that projection error can depend on orientation.
Occlusion Is Directional Too
Two objects can overlap from one viewpoint and separate from another.
The first view may make one object disappear behind another.
Nothing was deleted from the world.
The projection direction caused two rays to land on the same part of the image.
Map Orientation Changes Intuition
North-up feels natural because it is conventional.
Rotate the map so a traveller’s forward direction points upward and the same street network can become easier to navigate.
Neither orientation changes the street geometry.
It changes the receiver’s relation to the representation.
The Axis Is a Question
Plot income against age.
Now plot income against education.
Same people.
Different projection.
Choosing axes is already a decision about which relationships deserve visible space.
Coordinate Choice Can Reveal or Hide Structure
A spiral can look complicated in Cartesian coordinates and simple in a coordinate system aligned to angle and radius.
A periodic process can become clearer when represented by phase rather than raw clock time.
The source has not changed.
The representation has been rotated conceptually toward a structure we care about.
PCA Makes Direction Selection Explicit
Principal Component Analysis chooses orthogonal directions that capture as much variance as possible in descending order. Current scikit-learn documentation describes PCA as linear dimensionality reduction using singular value decomposition to project centred data into a lower-dimensional space, with principal axes corresponding to directions of high variance.
This is powerful because the axes are data-derived rather than chosen manually.
But it also means the objective is variance, not automatically interpretability, fairness, rarity or decision importance.
Maximum Variance Is Not Maximum Meaning
A rare medical condition may occupy a low-variance direction.
A safety-critical fault may be statistically uncommon.
A minority language may contribute little total variance to a global dataset.
Discarding low-variance directions can therefore remove something socially or operationally important.
Variance tells us how much the data move along a direction. It does not tell us what the direction is worth.
Scaling Before PCA Changes the Direction
One variable is measured in dollars.
Another is measured on a 1–5 scale.
If raw magnitudes differ greatly, the high-variance numeric scale can dominate the principal directions.
Standardising features can produce a different projection because the upstream geometry has changed.
Normalisation owns the broader transformation problem. Projection Direction owns the consequence: a different geometry produces different directions worth keeping.
Whitening Changes the Downstream Geometry Again
PCA whitening rescales retained components to unit variance. Scikit-learn notes that whitening removes relative variance-scale information while sometimes improving compatibility with downstream estimators.
The representation gains one property and gives up another.
This is classic lossy design: preserve what the next job needs, not everything the source contained.
Supervised Projection Can Choose a Different Direction Entirely
An unsupervised projection asks what structure exists in X.
A supervised representation may ask what direction best separates a target Y.
The chosen shadow changes because the job changed.
Task information acts like a new lamp shining on the same source.
Projection Direction Is Receiver-Dependent
A doctor wants risk-relevant structure.
A teacher wants misconception structure.
A logistics planner wants travel-time structure.
A designer wants perceptual similarity.
The same high-dimensional source can support different useful projections because the receiver is solving a different problem.
A Camera Teaches the Same Lesson
A portrait shot from slightly below can increase apparent dominance.
A wide-angle lens close to the face changes proportion.
A side profile removes one eye and reveals nose contour.
Photography does not merely record an object; viewpoint selects a projection of it.
This is why the same person can look “different” without any deceptive editing.
Cross-Sections Are Directional Projections of Structure
Cut through a geological formation north–south.
Now cut east–west.
Each section reveals different relationships.
A section can be truthful and still miss a fault or layer that lies mostly outside that cutting plane.
Medical Imaging Depends on View Selection
Axial.
Coronal.
Sagittal.
Different planes answer different anatomical questions.
The safest reading often combines multiple views rather than treating one slice as the complete object.
Multiple Views Reduce Hidden-Dimension Risk
If a structure appears in one projection but disappears in another, ask why.
Possibilities include:
- the structure is genuinely anisotropic;
- the first view aligned with it unusually well;
- the second view collapsed it;
- the apparent structure belonged partly to the projection rule.
View diversity becomes a hostile test.
Projection Direction Can Create Apparent Separation
Two clouds overlap in three dimensions.
Viewed from one direction they appear separate.
Viewed from another they overlap.
The first plot did not necessarily invent the difference.
It selected a direction in which the difference became visible.
And It Can Hide Real Separation
Two groups separated mostly along a discarded dimension can collapse onto the same region after projection.
The second pillar, Projection Collisions, owns this many-to-one ambiguity.
Projection Direction Can Also Hide Continuity
A curved manifold can look like separated islands when flattened poorly.
A continuous trajectory can fold onto itself.
Conversely, distinct structures can overlap.
Continuity is not guaranteed to survive arbitrary views.
The Viewpoint Is Part of the Evidence Receipt
When a conclusion depends on a projection, record:
- source variables;
- scaling;
- projection method;
- axes or components;
- parameters;
- random seed where relevant;
- which dimensions were retained;
- which claim the view is being used to support.
Without that receipt, the picture can outlive the assumptions that produced it.
Projection Is a Controlled Loss Budget
Every direction preserves something better.
Every direction hides something else.
The important question is whether the hidden structure matters for the receiver’s job.
The fourth pillar, Projection Objectives, owns that preservation choice directly.
A Better Projection-Direction Model
source geometry → receiver goal → scaling/coordinate choice → direction/basis → projection → visible structure → hidden structure → alternate views → stability check → bounded interpretation
A 28-Lens Projection Direction Audit
- Source: what higher-dimensional object is being viewed?
- Receiver: who needs the projection?
- Goal: what question should the view answer?
- Coordinates: how is the source represented upstream?
- Scaling: were variables normalised or standardised?
- Direction: what viewing direction is used?
- Basis: are axes original or learned?
- Objective: why were these axes selected?
- Variance: is high variance being privileged?
- Semantics: do retained axes map to meaningful concepts?
- Depth: what disappears into the omitted dimension?
- Occlusion: which objects can hide others?
- Foreshortening: what appears artificially small?
- Separation: which groups become more distinct?
- Overlap: which distinctions collapse?
- Continuity: could a path fold or break visually?
- Orientation: would rotation change receiver intuition?
- Alternate view: what happens from another reasonable direction?
- Method: would another projection family show the same structure?
- Scale level: what dimensions were already aggregated?
- Feature loss: was information removed before projection?
- Receiver bias: what visual pattern is especially persuasive?
- Traceability: can the axes and parameters be reconstructed?
- Stability: does the claim survive small changes?
- Collision: could different source states share this shadow?
- Return path: can the claim be checked in the richer source?
- Limit: what should the projection not be used to claim?
- Decision: is the preserved structure sufficient for the actual job?
Laboratory 1: Project a Cube
Use a transparent cube or box. Photograph it from three directions. List which edges overlap and which faces disappear in each image.
Laboratory 2: Rotate the Dataset
Create a simple three-dimensional point cloud with two groups separated mainly along one axis. Compare several two-dimensional views. Which view reveals the separation and which hides it?
Laboratory 3: Axis Choice as Question Choice
Take a small student dataset and plot three different pairs of variables. Write a different legitimate question each plot can answer and one conclusion each plot cannot support.
For Primary Readers
Shine a torch at one toy from several sides. Ask why the toy stays the same while the shadow changes.
For Secondary Readers
Separate source object from viewing rule. Explain how a change of axis can reveal one pattern while hiding another.
For Advanced Readers
Model projection direction as a basis-selection problem. A low-dimensional representation preserves components aligned with the chosen subspace and suppresses components orthogonal or poorly aligned to it; semantic adequacy therefore depends on the relationship between subspace objective and downstream task.
Common Misconceptions
- “If two projections disagree, one must be wrong.” Different views can preserve different true relationships.
- “PCA finds the most important dimensions.” PCA finds high-variance directions under its preprocessing and metric, not universal importance.
- “Rotation is cosmetic.” In projection, direction can determine which information survives.
- “A clean 2D separation proves the source groups are completely separate.” Omitted dimensions can contain overlap or continuity.
- “One good projection is enough.” Important claims deserve alternate-view and richer-source checks.
Research Corridor
- scikit-learn — PCA — principal axes, explained variance and whitening.
- scikit-learn — t-SNE — a contrasting nonlinear projection whose result depends on optimisation and parameters.
- eduKateSG — Normalisation.
- eduKateSG — Projection Master.
Frequently Asked Questions
What is projection direction?
It is the orientation, basis or subspace relative to which a richer object is mapped into a lower-dimensional representation.
Why does changing projection direction change the result?
Because different source components align differently with the retained dimensions. Some become prominent, some shrink and some collapse into omitted dimensions.
How should an analyst choose a projection?
Choose according to the structure required by the task, record the preprocessing and objective, and test whether important conclusions survive reasonable alternate views.
Final Thought: The Lamp Is Part of the Picture
We often stare at the shadow and forget the light that made it.
A projection becomes intellectually honest when we treat direction not as invisible setup, but as part of the claim.
PROJECTION · FOUR PILLAR LEGS
Return to Projection, or continue through Projection Collisions, Projection Stability and Projection Objectives. Return to the Information & Representation Hub.