A surface is stretched sideways. Its edge moves exactly as the transformation matrix predicts, but an arrow that used to be perpendicular to that edge now leans the wrong way. Multiplying every arrow by the same matrix has preserved the drawing’s organisation while losing a geometric relationship.
The gold standard of vectors here is to transform each quantity according to its job. A tangent describes movement along a surface. A surface normal describes a direction perpendicular to the surface. Those jobs impose different conditions when the transformation stretches different directions by different amounts.
The short answer is this: for an invertible linear transformation M in fixed orthonormal Euclidean coordinates, tangent vectors transform by M, while a perpendicular normal can be transformed by M inverse transpose, then normalised if a unit direction is required. Translation changes points but does not directly change these directions. The conditions matter as much as the formula.
This is an extension lesson for learners comfortable with coordinates, dot products and elementary matrices. All numerical cases are original teaching constructions. They are not a graphics-engine implementation or a physical design assessment. Begin with The Gold Standard of Vectors for the foundation, or use Learning That Transfers into Life to connect the worked checks with a broader study method.
eduKateSG · Vectors
Find your next learning step
Transform tangents and surface normals through original stretch, shear, plane and reflection examples, with perpendicular checks and clear geometric limits.
Chapter contents
Understand the question · Chapters 1–5
Work the original cases · Chapters 6–10
Test the boundaries · Chapters 11–15
Repair and transfer · Chapters 16–20
Check and continue · Chapters 21–25
A tangent vector points along a line or along a locally permitted direction on a surface. For a straight edge, subtracting its endpoint positions supplies a tangent direction. A normal is perpendicular to that direction; for a plane in three dimensions it is perpendicular to two independent tangent directions.
The distinction is relational. A normal is not identified merely by looking like an arrow that points outwards. Its dot product with the relevant tangents must be zero. That condition gives a check which remains useful when a drawing is stretched, rotated or viewed at an unfamiliar angle.
Use a two-dimensional line first. Let t = (1,1) be a tangent and n = (1,−1) a normal. The dot product is 1 − 1 = 0. Neither vector needs unit length to establish perpendicularity. The arrow’s length and its geometric role are separate issues.
Before transforming anything, write two labels: “moves along the edge” and “must remain perpendicular to the edge”. A learner who keeps these labels attached is less likely to apply a familiar matrix rule to the wrong object. The next calculation deliberately changes the geometry enough to expose that mistake.
CHAPTER 2 OF 25 · Understand the question
2. Stretch one direction and inspect the result
Let M multiply every x-coordinate by three and leave every y-coordinate unchanged. In matrix notation, M is diagonal with entries 3 and 1. The transformed tangent is Mt = (3,1).
If we also multiply the original normal by M, we obtain Mn = (3,−1). Its dot product with the transformed tangent is 3 × 3 + 1 × (−1) = 8. Since eight is not zero, this arrow is no longer perpendicular to the edge.
The component arithmetic is correct. The error lies in assuming that the normal should follow the same rule as a displacement along the edge. Non-uniform stretching changes angles, so preserving the original arrow’s component scaling is not enough to preserve its normal role.
Sketch the line through the origin in direction (3,1). A perpendicular direction is (1,−3), since 3 − 3 = 0. This gives a target that can be checked without knowing an inverse-transpose formula. The formula will be useful because it systematically produces a direction parallel to this target for more general transformations.
CHAPTER 3 OF 25 · Understand the question
3. Repair the condition before memorising the rule
The question is not “where did the old normal arrow move?” It is “what direction is perpendicular to the transformed surface?” Those questions coincide under some transformations and differ under others.
In the stretch example, a suitable new normal is (1/3,−1), which is parallel to (1,−3). Its dot product with (3,1) is one minus one, or zero. Its length differs from that of the old normal, but its perpendicular relationship is correct.
This is a first-principles repair. Identify the relationship that defines the answer, then find a transformation rule that preserves that relationship. A student can solve this small case by inspection before learning a general matrix expression.
Ask for a reconstruction check in words: “Every small movement along this transformed line has zero component in the proposed normal direction.” Because all tangent directions on the line are scalar multiples of (3,1), checking one nonzero tangent is sufficient for this two-dimensional line. A three-dimensional surface needs more care, which a later chapter develops.
CHAPTER 4 OF 25 · Understand the question
4. Derive the inverse-transpose relationship
Write the original perpendicular condition as nᵀt = 0, where the superscript T denotes transpose. Under the active transformation in our fixed Euclidean coordinate system, the tangent becomes t′ = Mt. We want a new normal n′ satisfying n′ᵀMt = 0.
Choose n′ = M⁻ᵀn, where M⁻ᵀ means the transpose of M inverse. Then n′ᵀMt = nᵀM⁻¹Mt = nᵀt = 0. The inverse cancels the transformation in exactly the place required to preserve the dot-product condition.
This establishes a valid normal direction when M is invertible. Any nonzero scalar multiple remains perpendicular, so the condition does not prescribe one compulsory length for the arrow. Normalisation is a separate step if the application requires a unit normal.
The Physically Based Rendering treatment of transformations distinguishes points, vectors and surface normals and documents this rule. The worked coordinates and diagnostic sequence in this article are original explanations of its geometric meaning, rather than copied rendering code.
For M with diagonal entries 3 and 1, M inverse has diagonal entries 1/3 and 1. Transposing a diagonal matrix leaves it unchanged. Therefore M⁻ᵀn = (1/3,−1) for n = (1,−1).
Its magnitude is √(1/9 + 1) = √10/3. Dividing by that magnitude gives the unit normal n̂′ = (1,−3)/√10. The transformed tangent has magnitude √10, so its unit direction is (3,1)/√10.
| Quantity | Components | Check against transformed tangent (3,1) |
|---|---|---|
| Original normal | (1,−1) | Dot product is 2 |
| Ordinary transformed arrow | (3,−1) | Dot product is 8 |
| Inverse-transpose normal | (1/3,−1) | Dot product is 0 |
| Unit normal | (1,−3)/√10 | Dot product is 0 |
The original normal also fails if simply left unchanged. Two plausible shortcuts therefore produce different wrong directions. The correct result is established by the perpendicular check, rather than by choosing whichever arrow appears closest to vertical in a sketch.
CHAPTER 6 OF 25 · Work the original cases
6. Explain why normalising the wrong arrow cannot repair it
Normalisation changes a nonzero vector’s magnitude while preserving its direction. If the wrong arrow has a nonzero dot product with the tangent, dividing that arrow by its length only divides the nonzero dot product by the same positive number.
In the original stretch, the wrong arrow is (3,−1), whose magnitude is √10. Its unit form has dot product 8/√10 with tangent (3,1). The value is smaller than eight, but it is still not zero.
This is a useful error diagnosis because “normalise it” can sound like a general repair whenever a normal looks wrong. Here the problem is direction, so a length adjustment cannot solve it. Transform the normal according to its defining relationship first, then normalise if needed.
For transfer, ask whether doubling an already correct normal makes it incorrect. It does not destroy perpendicularity, although an application expecting unit length would need the additional normalisation. The learner should be able to separate “perpendicular to the surface” from “has length one” in both directions.
Let M map (x,y) to (x + 2y,y). Take an original vertical tangent t = (0,1) and horizontal normal n = (1,0). Their dot product is zero.
The transformed tangent is (2,1). Ordinary transformation leaves n as (1,0), which has dot product two with the transformed tangent. Again, applying the same component rule does not preserve the normal relationship.
The inverse map sends (x,y) to (x − 2y,y). Transposing that inverse gives M⁻ᵀn = (1,−2). Its dot product with (2,1) is 2 − 2 = 0. The unit normal is (1,−2)/√5.
This example matters because a learner might mistakenly associate the correction only with unequal diagonal scaling. Shearing can also change angles. The general question is whether the transformation preserves the particular perpendicular relationships being used. A formula derived from that relationship travels more reliably than a rule attached to one familiar picture.
CHAPTER 8 OF 25 · Work the original cases
8. Recognise rotations and uniform scaling as special cases
For a rotation R, the inverse equals the transpose, so R⁻ᵀ = R. A normal and a tangent can both be rotated using the same rotation matrix while retaining their perpendicular relationship. Rotations preserve Euclidean angles and lengths.
For a transformation M = sR with nonzero uniform scale s, the normal rule gives M⁻ᵀn = (1/s)Rn, while the ordinary arrow rule gives Mn = sRn. These results are parallel because their ratio is s², which is positive. After unit normalisation they give the same direction.
This explains why a faulty implementation can appear correct on objects that are only rotated or uniformly scaled. Those test cases remove the mechanism that exposes the mistake. Passing them is useful but insufficient evidence for general non-uniform transformations.
A good test set includes a rotation, uniform scaling, non-uniform scaling and shear. Predict which cases should allow the ordinary direction rule to agree after normalisation. Then check the prediction with dot products rather than relying on a rendered object’s appearance.
An affine map can be written F(p) = Mp + b, where b is a translation. For two points p and q, the displacement between their images is F(q) − F(p) = M(q − p). The translation cancels.
Therefore tangent directions follow M, not Mp + b. Adding b to a normal or tangent treats a direction as though it were a point. This is a different error from using M instead of M⁻ᵀ for a normal, and the two errors should be diagnosed separately.
For an original check, translate every point on the line y = x by b = (4,−2) without stretching or rotating it. The line moves, but tangent (1,1) and normal (1,−1) remain valid directions. The line’s position changed; its orientation did not.
The earlier affine point-weight guide explains why points and displacements respond differently to a changed origin or translation. Use that route if the learner is still adding a position shift to every component list.
CHAPTER 10 OF 25 · Work the original cases
10. Transform a plane equation and its offset
A plane equation has the form n·p = c. Its normal n and scalar offset c work together. Under p′ = Mp + b with invertible M, substitute p = M⁻¹(p′ − b) to obtain n′·p′ = c + n′·b, where n′ = M⁻ᵀn.
Take the original plane x + 2y + 2z = 6. Let M scale x by two, y by one and z by one half, then translate by b = (3,−1,2). The transformed normal is (1/2,2,4).
The new right-hand side is 6 + (1/2) × 3 + 2 × (−1) + 4 × 2 = 13.5. Thus the transformed plane is 0.5x′ + 2y′ + 4z′ = 13.5. Omitting the offset update would place a parallel plane in the wrong location.
Check a known point. Original p = (6,0,0) lies on the first plane. Its image is (15,−1,2), and 0.5 × 15 − 2 + 8 = 13.5. This checks both direction and location, which a normal-direction test alone cannot do.
CHAPTER 11 OF 25 · Test the boundaries
11. Require two independent tangents for a surface
In three dimensions, being perpendicular to one tangent leaves a whole plane of possible normal candidates. A surface normal must be perpendicular to the entire tangent plane, so two independent tangent directions are needed to establish its line of direction.
Take t1 = (1,1,0) and t2 = (0,0,1). Their cross product is n = (1,−1,0). Under diagonal scaling M = diag(3,1,2), the transformed tangents are t1′ = (3,1,0) and t2′ = (0,0,2).
The inverse-transpose normal is n′ = (1/3,−1,0). Its dot products with both transformed tangents are zero. Checking only t2′ would be too weak: many wrong arrows in the xy-plane are also perpendicular to the z-direction.
Ask the learner to supply a deliberately wrong candidate that passes one check but fails the other. For example, (1,0,0) is perpendicular to t2′ but has dot product three with t1′. A failed counterexample teaches why the second independent tangent is necessary.
CHAPTER 12 OF 25 · Test the boundaries
12. Recompute a normal from transformed edges
For an explicit triangle, another route is to transform its vertices, subtract transformed positions to form two edges, and take their cross product. This constructs a normal directly from the transformed geometry.
In the preceding example, t1′ × t2′ = (2,−6,0), parallel to the inverse-transpose result (1/3,−1,0). Their magnitudes differ by six, the determinant of M. The difference in scale is expected when a cross product represents oriented area as well as direction.
For invertible M, the identity (Mu) × (Mv) = det(M) M⁻ᵀ(u × v) connects the two methods. It also reveals an orientation issue when the determinant is negative. Perpendicular direction, unit length and orientation must not be silently merged.
This provides an independent geometric check for supplied triangle data. It does not mean that every normal in a graphics model must be replaced by a face normal. A model may use a separately defined smooth shading normal; its intended role should be preserved and validated rather than overwritten without understanding the representation.
CHAPTER 13 OF 25 · Test the boundaries
13. Handle reflections without losing orientation
Let M reflect the x-coordinate, so M = diag(−1,1,1). Its determinant is negative. Take original edges u = (1,0,0), v = (0,1,0), with ordered cross product u × v = (0,0,1).
The transformed edges are (−1,0,0) and (0,1,0), whose ordered cross product is (0,0,−1). The inverse-transpose direction applied to the original normal remains (0,0,1). Both arrows are perpendicular to the transformed plane, but they point in opposite orientations.
The cross-product identity accounts for the difference through det(M) = −1. Which orientation is wanted depends on the object’s definition: preserved normal orientation, vertex winding, outward-facing convention and the surrounding representation may need explicit coordination.
Do not repair every negative determinant by blindly reversing a normal. First identify the orientation contract. The mathematical lesson is that an unoriented normal line contains two opposite directions, while an oriented surface chooses one of them according to additional information.
CHAPTER 14 OF 25 · Test the boundaries
14. Identify the limit of a singular transformation
If M is singular, M inverse does not exist, so the inverse-transpose formula is unavailable. This does not automatically mean every transformed surface lacks a normal. The geometry must be inspected.
For example, M = diag(1,1,0) leaves a non-degenerate triangle in the xy-plane non-degenerate. Its transformed edges still have cross product in the z-direction. A normal can be computed from those edges even though the global transformation has no inverse.
The same map can collapse a differently oriented triangle to a line. Then its two transformed edges are dependent and their cross product is zero. A zero vector cannot be normalised into a unique unit surface normal. The missing dimension is a geometric issue, not a small numerical inconvenience.
Never replace a failed inverse with an arbitrary small denominator and describe the answer as the uniquely correct normal. A numerical policy may be needed in software, but it must reflect the model’s intended treatment of degenerate geometry. The learning exercise should make the lost information visible.
CHAPTER 15 OF 25 · Test the boundaries
15. Keep active transformation and coordinate change separate
This guide actively transforms a shape while keeping the original Euclidean axes fixed. A passive change of basis describes the same geometry using different coordinate building blocks. Those are distinct operations even when both use matrices.
The familiar component dot product nᵀt assumes the orthonormal coordinate setting stated at the start. In a non-orthonormal basis, the Euclidean inner product needs the appropriate metric representation. Treating every new coordinate list as though its axes remained perpendicular and equally scaled can produce a different error.
A learner should therefore state whether the object changed, the coordinate description changed, or both changed. The phrase “transform the vector” is too broad to carry those distinctions by itself.
The Vectors subject root develops bases and warns against calculating Euclidean length from non-orthonormal coefficients using an unjustified sum of squares. Return there before extending the normal formula into a different coordinate convention. More notation is useful only when it preserves the geometric meaning.
CHAPTER 16 OF 25 · Repair and transfer
16. Distinguish interpolation from normal transformation
An interpolated attribute and a transformed surface normal raise different questions. Perspective interpolation asks which original surface weights correspond to a screen location. Normal transformation asks which direction remains perpendicular after a geometric transformation.
The earlier perspective-interpolation guide addresses the first question. Applying inverse-transpose correction does not replace the interpolation rule, and perspective-correct interpolation does not automatically fix a wrongly transformed normal.
Even after component interpolation, a resulting normal may need normalisation if its intended use requires unit length. Interpolating two unit vectors does not generally yield another unit vector. For instance, averaging (1,0,0) and (0,1,0) gives (0.5,0.5,0), whose magnitude is √0.5.
The operations should be named separately in an explanation: transform the geometric data, interpolate according to the declared representation, and normalise a valid nonzero result where needed. A polished image can conceal several compensating mistakes, so tests should isolate each mathematical contract.
CHAPTER 17 OF 25 · Repair and transfer
17. Use a small illumination calculation as a diagnostic
Suppose a deliberately simplified diffuse factor is max(0,n̂·l̂), with both normal and light directions expressed in the same frame and normalised. This is a mathematical illustration, not a full material or lighting model.
For the diagonal stretch example, the correct unit normal is (1,−3)/√10. Let the light direction be (0,−1). Their dot product is 3/√10, approximately 0.9487. The incorrectly transformed unit arrow (3,−1)/√10 gives 1/√10, approximately 0.3162.
Both outputs lie between zero and one and may look plausible in an isolated image. A range check therefore cannot establish that the normal is geometrically correct. The tangent-perpendicular test is more direct evidence about the transformation itself.
This diagnostic also requires a common coordinate frame. A correct normal compared with a light direction from an incompatible frame can produce a wrong result. Use the existing ray-tracing and reflection lesson to connect vector relationships with a wider geometric application after the normal calculation is secure.
If a learner writes a component list correctly but cannot say whether it is a point, tangent or normal, repair the object distinction first. Give a line and ask for one point on it, one movement along it and one perpendicular direction.
If those roles are clear but matrix multiplication fails, work a diagonal scale before introducing a shear. If multiplication is secure but the learner keeps using M for every direction, return to the defining dot-product condition and let the failed value eight remain visible.
If the only difficulty is unit normalisation, use a correct normal whose length is easy to check. Do not introduce a new transformation at the same moment. Separate the direction repair from the length repair so the learner knows what changed.
A useful teaching sequence is a labelled sketch, one failed calculation, a perpendicular-condition repair, then a general expression. The learner should explain each transition before receiving a larger collection of exercises. Keep the diagram available until its relationships can be reconstructed from the symbols alone.
Task A: t = (2,1), n = (1,−2), and M = diag(2,3). Find the transformed tangent and a correct normal. Answer: t′ = (4,3), and M⁻ᵀn = (1/2,−2/3), parallel to (3,−4). Their dot product is zero.
Task B: what would ordinary transformation give for n? Answer: (2,−6). Its dot product with (4,3) is 8 − 18 = −10, so it fails. Normalising it cannot make that dot product zero.
Task C: translate the entire construction by (7,−5). Does the tangent or normal direction change? Answer: no. Point positions change, while the translation cancels in endpoint differences. A plane or line equation still needs its offset updated.
Task D: a transformed triangle’s two edge vectors are (1,2,0) and (2,4,0). Can their cross product supply a unit face normal? Answer: no. The edges are dependent and their cross product is zero. State the degeneracy rather than dividing by its zero magnitude.
On the next study day, remove the chapter title and present a line with a tangent and normal under a new stretch. Ask the learner to choose the transformation rule before showing a worked example. Method selection is part of the learning.
Then change the task from a line to a plane and require two independent tangent checks. Next add a translation and ask separately about direction and plane location. Finally introduce a reflection and ask whether perpendicularity alone selects an orientation.
A learner who can calculate but cannot explain should revisit one small counterexample. A learner who can explain but makes arithmetic errors needs cleaner component organisation and substitution checks. These are different repair pathways.
For a parent or tutor, the useful question is, “Which relationship must still be true after the change?” It invites the learner to use a defining property rather than search for a formula by visual resemblance. Successful transfer is the ability to preserve that property when the familiar diagram is replaced.
Are normals vectors? In Euclidean geometric representation they can be drawn and stored as vector components, but their transformation rule follows their perpendicular relationship to a surface. Calling them arrows does not settle how that role should transform.
Why not always use inverse transpose for every vector? A displacement along the transformed surface must follow the geometry through M. Applying the normal rule to it generally gives the wrong displacement. The two objects serve different jobs.
Does the formula require a unit original normal? No. A nonzero normal of any length supplies the perpendicular direction. Normalise the transformed result when unit length is part of the intended use.
Can a singular transform be handled by the same inverse formula? No. Inspect the transformed geometry and its dimension. Some triangles retain a well-defined normal from transformed edges; others collapse.
Will a correct normal guarantee a correct image? No. Geometry, visibility, interpolation, coordinate frames, materials and lighting can introduce separate errors. The normal check establishes one specific relationship within that larger system.
CHAPTER 22 OF 25 · Check and continue
22. Recover the same normal from an implicit curve
A second route to the normal can test the matrix reasoning independently. Start with the unit circle x² + y² = 1. At a point on the circle, a normal direction is supplied by the gradient (2x,2y), which is perpendicular to the local tangent.
Choose the original point p = (1/√2,1/√2). Its normal direction is parallel to (1,1). Apply the horizontal stretch M = diag(3,1). The new point is p′ = (3/√2,1/√2), and the transformed circle is the ellipse x′²/9 + y′² = 1.
The gradient of the ellipse’s defining expression is (2x′/9,2y′). At p′ this becomes (2/(3√2),√2), parallel to (1/3,1). Inverse-transpose transformation of the original normal direction (1,1) gives exactly that direction.
Ordinary transformation would give (3,1), which points along the stretched position arrow rather than the ellipse’s normal at that point. For a circle the radial direction and normal coincide, but a non-uniform stretch does not preserve that coincidence. This is another way an attractive visual shortcut can fail.
The example is an extension for learners who know derivatives. It is not needed to understand the earlier dot-product argument. Its value is that a different representation of the same geometry produces a consistent result. Independent representations provide stronger checking than repeating the same matrix multiplication twice.
To transfer the idea, stretch the vertical direction instead and predict which normal component is reduced by the inverse rule. Then compare the transformed implicit equation. Keep the point, curve and normal attached to the same transformation; mixing an old point with a new curve creates an unrelated calculation.
CHAPTER 23 OF 25 · Check and continue
23. Compose transformations in the correct order
Suppose M acts first and N acts second. The total active transformation is NM. A tangent follows that order: t′ = NMt. The associated normal follows M⁻ᵀ first and N⁻ᵀ second, giving n′ = N⁻ᵀM⁻ᵀn.
This agrees with (NM)⁻ᵀ. The inverse reverses multiplication order and the transpose reverses it again. Writing the intermediate geometric stages is often easier to check than trying to remember a verbal rule about reversed order.
Use the earlier tangent t = (1,1) and normal n = (1,−1). Let M = diag(3,1), then let N shear (x,y) to (x + 2y,y). The tangent first becomes (3,1), then (5,1).
The normal first becomes (1/3,−1). Applying N⁻ᵀ gives (1/3,−5/3). Its dot product with (5,1) is 5/3 − 5/3 = 0. The sequential calculation preserves the defining relationship through both stages.
If the normal operations are applied in the wrong order, the result is (1/3,−3), whose dot product with (5,1) is −4/3. The error can remain hidden in a long chain of individually correct matrix multiplications if the order is not attached to the physical stages.
A practical worksheet therefore includes a row for each stage: current tangent, current normal and their dot product. If the perpendicular condition first fails at stage two, inspect the stage-two rule and ordering before recomputing the entire scene. This localises the first unstable move and makes the correction explainable.
The component calculations assume that the axes describe compatible spatial directions and a common Euclidean length convention. A list containing length, temperature and a model score is a data vector, but it does not automatically define a physical tangent or surface normal under this geometry.
Likewise, a matrix used to rescale numerical features in a statistical model may have a different purpose from an active stretch of a surface. Before carrying the inverse-transpose rule into a new application, identify the surface, its tangent space and the inner product that defines perpendicularity.
This is a useful final boundary check. The rule is powerful because it follows a specific geometric relationship. Treating it as a general instruction for every column of numbers would remove the very meaning that made the derivation valid. A learner should be able to name both a new example where the rule applies and a component list for which the surface-normal question has not even been defined.
CHAPTER 25 OF 25 · Check and continue
25. Report the result and choose the next guide
A reproducible answer names the original surface directions, the active transformation, coordinate convention, normal rule, normalisation requirement and orientation convention. It includes dot-product checks against the transformed tangent space.
For the main case, write: “The stretch maps tangent (1,1) to (3,1). Ordinary transformation of normal (1,−1) gives a non-perpendicular arrow. Inverse-transpose transformation gives (1/3,−1), whose dot product with the new tangent is zero; its unit form is (1,−3)/√10.”
For foundations, return to Vectors. For the adjacent representation question, read perspective interpolation. For the technical reference, use Physically Based Rendering: applying transformations.
For wider learning, continue through The Gold Standard of Education and How X Works. The next useful test is a new transformation for which the learner can state the preserved relationship before calculating.