Covariant and Contravariant Tensors

Physics happens on a manifold, but we can only write it down in coordinates — and coordinates are a human choice the universe never agreed to. The whole apparatus of tensor analysis exists to guarantee that this choice leaves no fingerprints on the physics. A tensor is a collection of components labelled by indices, but what makes it a tensor is not the array of numbers: it is the fixed, multilinear rule by which those numbers rearrange when you change coordinates $x^{\mu}\to x'^{\mu}(x)$ . That rule is assembled entirely from the Jacobian matrix $\partial x'^{\mu}/\partial x^{\nu}$ and its inverse. Write a physical law as an equality of tensor components in one frame, and it is automatically an equality in every frame — which is exactly why tensors are the native grammar of relativity, continuum mechanics, and differential geometry. Within that grammar sits one fundamental split. An infinitesimal displacement $\mathrm{d}x^{\mu}$ transforms one way; the gradient $\partial_{\mu}f$ of a scalar transforms with the inverse matrix. Objects that go like $\mathrm{d}x^{\mu}$ are contravariant and wear upper indices; objects that go like $\partial_{\mu}f$ are covariant and wear lower indices. Geometrically the contravariant vectors are the arrows and velocities living in the tangent space $T_pM$ , while the covariant vectors — covectors, one-forms, gradients — live in the dual cotangent space $T_p^{*}M$ as stacks of level surfaces. The index height is not decoration. It records which transformation law applies, and the Einstein summation convention is engineered so that the only legal contraction, one upper against one lower, produces a genuine coordinate-independent scalar. The metric tensor $g_{\mu\nu}$ is the one extra piece of structure that lets us slide between the two pictures — raising and lowering indices — and it is where this article hands off to the rest of the chapter.

Key takeaways

A tensor is defined by its transformation law under a coordinate change, not by its list of components — the law is what survives when the coordinates do not.
Contravariant components carry upper indices and transform with the Jacobian $\partial x'^{\mu}/\partial x^{\nu}$ , exactly like the components of a velocity or a coordinate displacement.
Covariant components carry lower indices and transform with the inverse Jacobian $\partial x^{\nu}/\partial x'^{\mu}$ , exactly like the components of a gradient or one-form.
The two are dual: contracting one upper index against one lower index cancels the Jacobian against its inverse and yields an invariant scalar — the engine behind the summation convention.
The metric $g_{\mu\nu}$ and its inverse $g^{\mu\nu}$ convert one type into the other; in Cartesian coordinates on flat space the distinction collapses and students forget it exists.
Key idea: what makes something a tensor is its transformation law, not its array of numbers — upper indices transform with the Jacobian (contravariant), lower with the inverse (covariant), and a contracted upper–lower pair is a coordinate-invariant scalar.

The core insight

What makes something a tensor is its transformation law, not its array of numbers. Upper indices transform with the Jacobian $\partial x'^{\mu}/\partial x^{\nu}$ (contravariant), lower indices with the inverse $\partial x^{\nu}/\partial x'^{\mu}$ (covariant), and a contracted upper–lower pair gives a coordinate-invariant scalar.

Where this lives today

The covariant/contravariant bookkeeping is not a historical relic — it is running, right now, inside the most precise instruments and largest simulations physicists build. Every GPS receiver applies a relativistic clock correction derived from the Schwarzschild metric $g_{\mu\nu}$ ; without raising and lowering indices correctly to compute proper time, positioning would drift by kilometres per day. When LIGO and Virgo detect a gravitational wave, what they measure is a tiny perturbation $h_{\mu\nu}$ of the flat metric, and the entire template-matching pipeline is written in manifestly tensorial form so that no coordinate artefact masquerades as a signal. Numerical relativity — the supercomputer simulation of merging black holes that produces those templates — evolves the metric and extrinsic-curvature tensors through the covariant Einstein equations, and a single dropped index or mismatched height produces a code that violates energy conservation and crashes. The same grammar structures lattice QCD, where the gauge field carries spacetime and colour indices, and it is spreading into condensed matter, where strained graphene and other materials are modelled with an effective curved-space metric so that electrons behave as if moving on a manifold — an active frontier in which the tensor transformation law, invented for gravitation, is repurposed to engineer the electronic response of a real material. Modern computer-algebra systems now verify these index manipulations symbolically, catching the very upper/lower mismatches that derivation 3 warns against.

The physical picture

Draw a fixed arrow on a page and then lay two different grids over it. Stretch the grid — make each basis vector twice as long — and the same arrow is now spanned by half as many grid-units, so its numerical components halve. The components varied contrary to the basis: that is the etymology of "contra-variant." Now consider the family of rulers that measure those components — the dual basis, drawn as stacks of level surfaces. Those measuring surfaces stretch with the basis, so the numbers they naturally read off co-vary with it: "co-variant." The first figure makes this concrete, showing the two homes a vector-like object can live in — the tangent space of arrows and the cotangent space of level-surface stacks — and the single number their pairing produces.

Figure

A vector lives in the tangent space T_pM as an arrow; a covector lives in the dual cotangent space T_p*M as a stack of level surfaces. Their natural pairing counts pierced surfaces: ω(V) = ω_μ V^μ, one number that is the same in every coordinate frame.

The reason index height matters is that upper and lower indices are sockets and plugs: a contravariant slot can only be summed against a covariant one, and the Einstein convention's upper–lower pairing is precisely the rule that guarantees the result is coordinate-independent. Sum two upper indices with no metric in sight and you have written something that changes when you rotate your axes — physical nonsense dressed as an equation. On a flat Euclidean space in Cartesian coordinates none of this bites: the metric is the identity, an orthonormal rotation equals its own inverse-transpose, and the two component sets are numerically identical. The distinction only forces itself into view when you switch to curvilinear coordinates or curved spacetime. The second figure shows exactly that moment, using plane polar coordinates and real numbers to pull the covariant, contravariant, and physical components of a single vector apart.

Figure

Plane polar coordinates split one vector into three numbers. The coordinate basis vector e_θ has length r, so the metric is g_θθ = r². At r = 3 with angular rate V^θ = 2, the contravariant, physical, and covariant components read 2, then r·V^θ = 6, then g_θθ·V^θ = 18 — three descriptions of one arrow.

The core equations

The prototype contravariant object is the coordinate displacement, and the prototype covariant object is the gradient of a scalar; every tensor is built to imitate one or the other in each of its slots. We state the two defining transformation laws first, then the general rule, the identity that makes contraction work, and finally the metric that bridges the two worlds. A contravariant vector is anything whose components transform exactly like the coordinate differential $\mathrm{d}x^{\mu}$ — with the Jacobian of the coordinate change.

V'^{\mu}=\frac{\partial x'^{\mu}}{\partial x^{\nu}}\,V^{\nu} \qquad (1)

where: $V^{\nu}$ — components in the old coordinates $x$ ; $V'^{\mu}$ — components in the new coordinates $x'$ ; $\partial x'^{\mu}/\partial x^{\nu}$ — Jacobian matrix (summed over $\nu$ ). Reading it: each new component is a Jacobian-weighted mix of the old ones, the same mix an infinitesimal displacement or a velocity would undergo. A covariant vector — a covector or one-form — is anything whose components transform like the gradient $\partial_{\mu}f$ , with the inverse Jacobian.

\omega'_{\mu}=\frac{\partial x^{\nu}}{\partial x'^{\mu}}\,\omega_{\nu} \qquad (2)

where: $\omega_{\nu}$ — covector components in the old coordinates; $\partial x^{\nu}/\partial x'^{\mu}$ — inverse Jacobian (summed over $\nu$ ). Reading it: the lower index drags along the inverse matrix, which is why a gradient's components shrink when a velocity's would grow. A general tensor of type $(p,q)$ has $p$ upper and $q$ lower slots, and each slot transforms independently by its own copy of the appropriate matrix.

T'^{\mu_1\cdots\mu_p}{}_{\nu_1\cdots\nu_q}=\frac{\partial x'^{\mu_1}}{\partial x^{\alpha_1}}\cdots\frac{\partial x'^{\mu_p}}{\partial x^{\alpha_p}}\,\frac{\partial x^{\beta_1}}{\partial x'^{\nu_1}}\cdots\frac{\partial x^{\beta_q}}{\partial x'^{\nu_q}}\,T^{\alpha_1\cdots\alpha_p}{}_{\beta_1\cdots\beta_q} \qquad (3)

where: $p$ — number of contravariant (upper) indices; $q$ — number of covariant (lower) indices; all repeated indices are summed. Reading it: count the Jacobians ( $p$ of them) and the inverse Jacobians ( $q$ of them); that count is the tensor's type. The single algebraic fact that makes the whole scheme close is that the Jacobian and inverse Jacobian are mutually inverse matrices, expressed by the chain rule.

\frac{\partial x'^{\mu}}{\partial x^{\nu}}\,\frac{\partial x^{\nu}}{\partial x'^{\rho}}=\delta^{\mu}{}_{\rho} \qquad (4)

where: $\delta^{\mu}{}_{\rho}$ — Kronecker delta, the components of the identity map and itself a genuine $(1,1)$ -tensor. Reading it: wherever an upper transformation meets its matching lower transformation, they annihilate into a delta. That identity is exactly what makes a contracted upper–lower pair a coordinate-independent scalar — the deepest reason index height is worth tracking.

S=\omega_{\mu}V^{\mu}=\omega'_{\mu}V'^{\mu} \qquad (5)

where: $S$ — the invariant scalar formed by pairing a covector with a vector. Reading it: the value $S$ is the same real number in every coordinate system, because substituting (1) and (2) collapses the two Jacobians via (4). Finally, the metric tensor and its inverse are the isomorphism that lets a single geometric vector be described either way — lowering an index turns a contravariant component into a covariant one, and raising it undoes that.

V_{\mu}=g_{\mu\nu}V^{\nu},\qquad V^{\mu}=g^{\mu\nu}V_{\nu},\qquad g^{\mu\alpha}g_{\alpha\nu}=\delta^{\mu}{}_{\nu} \qquad (6)

where: $g_{\mu\nu}$ — covariant metric; $g^{\mu\nu}$ — its matrix inverse. Reading it: the metric is the dictionary between tangent and cotangent spaces, and only when $g_{\mu\nu}=\delta_{\mu\nu}$ do the two component sets read the same.

Derivation AWhy basis vectors are covariant while the components are contravariant

The goal is to earn the terminology rather than memorize it: to show, from the chain rule alone, that a coordinate basis vector transforms with the inverse Jacobian while the components of any fixed vector transform with the Jacobian, so that the two vary in exactly opposite ways. This is the concrete meaning of "contra-variant" — the numbers run contrary to the basis — and it is the reason upper and lower indices are assigned as they are. We treat the coordinate basis vector as the partial-derivative operator $\mathbf{e}_{\mu}=\partial/\partial x^{\mu}$ , transform it, then demand that the physical vector itself be frame-independent and read off what the components must do.

Step 1. Write the new basis vector as a derivative operator and apply the chain rule: $\mathbf{e}'_{\mu}=\partial/\partial x'^{\mu}=(\partial x^{\nu}/\partial x'^{\mu})\,\partial/\partial x^{\nu}=(\partial x^{\nu}/\partial x'^{\mu})\,\mathbf{e}_{\nu}$ . Why: the chain rule forces the inverse Jacobian $\partial x^{\nu}/\partial x'^{\mu}$ to appear, so the basis vectors themselves transform covariantly, like lower-index objects.

Step 2. Insist that the geometric vector is one and the same object in both frames, $V=V'^{\mu}\mathbf{e}'_{\mu}=V^{\nu}\mathbf{e}_{\nu}$ . Why: this is the whole point of the construction — the arrow does not care which grid we drew, only its description does, so the two expansions must be equal.

Step 3. Substitute Step 1 into Step 2: $V'^{\mu}(\partial x^{\nu}/\partial x'^{\mu})\mathbf{e}_{\nu}=V^{\nu}\mathbf{e}_{\nu}$ , match the coefficients of $\mathbf{e}_{\nu}$ , and invert with (4) to get $V'^{\mu}=(\partial x'^{\mu}/\partial x^{\nu})\,V^{\nu}$ . Why: the components pick up the Jacobian — the inverse of what the basis picked up — so components and basis vary oppositely, which is precisely why the components are called contravariant and wear upper indices.

Go deeper · Derivation B · 1 minThe contraction of an upper with a lower index is a coordinate invariant

Derivation BThe contraction of an upper with a lower index is a coordinate invariant

The goal is to prove the single claim that justifies the entire upper–lower bookkeeping: that the pairing $\omega_{\mu}V^{\mu}$ evaluates to the same number in every coordinate system, so that any equation built by contracting matched indices is automatically frame-independent. Everything the Einstein convention promises rests on this one cancellation. We substitute the two transformation laws, collect the Jacobian against the inverse Jacobian, and use the chain-rule identity (4) to reduce the pair to a Kronecker delta.

Step 1. Transform both factors into the primed frame and multiply: $\omega'_{\mu}V'^{\mu}=\big(\tfrac{\partial x^{\alpha}}{\partial x'^{\mu}}\omega_{\alpha}\big)\big(\tfrac{\partial x'^{\mu}}{\partial x^{\beta}}V^{\beta}\big)$ . Why: writing both primed objects in terms of unprimed ones exposes an inverse Jacobian and a Jacobian sharing the summed index $\mu$ , which is the setup for cancellation.

Step 2. Reassociate the derivatives and apply (4): $\tfrac{\partial x^{\alpha}}{\partial x'^{\mu}}\tfrac{\partial x'^{\mu}}{\partial x^{\beta}}=\delta^{\alpha}{}_{\beta}$ , so the product becomes $\delta^{\alpha}{}_{\beta}\,\omega_{\alpha}V^{\beta}$ . Why: the two matrices are mutual inverses, and their product over the shared upper–lower index is exactly the identity — the geometric fact that a slot and its dual annihilate.

Step 3. Collapse the delta: $\delta^{\alpha}{}_{\beta}\,\omega_{\alpha}V^{\beta}=\omega_{\beta}V^{\beta}=\omega_{\mu}V^{\mu}$ , identical to the unprimed pairing. Why: the primed and unprimed contractions are the same real number, so $\omega_{\mu}V^{\mu}$ is a true scalar — this is why physical laws written as contracted-index equalities hold in all frames.

Go deeper · Derivation C · 1 minThe mixed Kronecker delta is a genuine tensor but the all-lower delta is not

Derivation CThe mixed Kronecker delta is a genuine tensor but the all-lower delta is not

The goal is to show that $\delta^{\mu}{}_{\nu}$ has numerically identical components in every coordinate frame and really is a $(1,1)$ -tensor, whereas the symbol $\delta_{\mu\nu}$ with both indices down is frame-dependent and is not a tensor at all — it merely coincides with the metric in special coordinates. This is the sharpest possible illustration that index position carries physical content, not typographical whim. We push $\delta^{\alpha}{}_{\beta}$ through the $(1,1)$ transformation law, then attempt the same for $\delta_{\alpha\beta}$ and watch it fail.

Step 1. Apply the mixed law: $\delta'^{\mu}{}_{\nu}=\tfrac{\partial x'^{\mu}}{\partial x^{\alpha}}\tfrac{\partial x^{\beta}}{\partial x'^{\nu}}\delta^{\alpha}{}_{\beta}=\tfrac{\partial x'^{\mu}}{\partial x^{\alpha}}\tfrac{\partial x^{\alpha}}{\partial x'^{\nu}}$ , since the delta sets $\beta=\alpha$ . Why: the built-in delta contracts the two transformation matrices onto the same index, which is the arrangement (4) knows how to simplify.

Step 2. Use the chain-rule identity (4): $\tfrac{\partial x'^{\mu}}{\partial x^{\alpha}}\tfrac{\partial x^{\alpha}}{\partial x'^{\nu}}=\delta^{\mu}{}_{\nu}$ , so $\delta'^{\mu}{}_{\nu}=\delta^{\mu}{}_{\nu}$ . Why: the components come out identical in the new frame, which is the definition of a numerically invariant tensor — the mixed delta is the unique isotropic $(1,1)$ -tensor.

Step 3. Now try both indices down: $\delta'_{\mu\nu}=\tfrac{\partial x^{\alpha}}{\partial x'^{\mu}}\tfrac{\partial x^{\beta}}{\partial x'^{\nu}}\delta_{\alpha\beta}$ , a product of two inverse Jacobians that generally does not simplify to $\delta_{\mu\nu}$ . Why: with no upper index to cancel against, nothing collapses — so $\delta_{\mu\nu}$ changes form under coordinate change and is not a tensor; the object that legitimately lowers an index is the metric $g_{\mu\nu}$ , not a frozen delta. *The mixed delta $\delta^{\mu}{}_{\nu}$ is frame-invariant, and lowering it requires the metric, giving $g_{\mu\nu}$ rather than a coordinate-independent $\delta_{\mu\nu}$ .*

Limiting cases

The whole covariant/contravariant distinction is a story about the metric and the coordinate map; when either is trivial, the distinction weakens or vanishes.

Case	Result	Meaning
Cartesian, Euclidean metric	$g_{\mu\nu}=\delta_{\mu\nu}\Rightarrow V^{\mu}=V_{\mu}$	raising and lowering do nothing numerically — the elementary-physics limit where one kind of index suffices
Orthogonal curvilinear coordinates	$g_{\mu\nu}=\mathrm{diag}(h_1^2,\dots,h_n^2)$	diagonal but non-unit metric: $V_{\mu}=h_{\mu}^2V^{\mu}$ (no sum), so the two component sets differ by scale factors — the origin of "physical components"
Linear (affine) change	$x'^{\mu}=A^{\mu}{}_{\nu}x^{\nu}$	the Jacobian is the constant matrix $A$ ; contravariant components go with $A$ , covariant with $(A^{-1})^{\mathsf T}$ — special relativity uses this with $A$ a Lorentz matrix
Orthogonal transformation	$(A^{-1})^{\mathsf T}=A$	for rotations the two matrices coincide, so "Cartesian tensors" never need to distinguish index height

Common trapThinking the components are the vector, and equating mismatched indices

The single most damaging habit is to treat the array $V^{\mu}$ as "the vector" and then write things like $V^{\mu}=V_{\mu}$ or $A^{\mu}=B_{\mu}$ . A vector is a frame-independent geometric object; $V^{\mu}$ and $V_{\mu}$ are two different arrays of numbers describing it, tied together only through the metric by $V_{\mu}=g_{\mu\nu}V^{\nu}$ . An equation whose free indices sit at different heights on the two sides is not coordinate-invariant — the two sides transform with opposite Jacobians and cannot stay equal after a coordinate change. In Cartesian coordinates on flat space the two arrays happen to hold the same numbers, which is exactly what lulls students into ignoring the rule until polar coordinates or curved spacetime punish them. The discipline is simple and non-negotiable: every free index must match in height on both sides, and every summed index must appear once up and once down.

Worked examples

1 · One vector, three components in polar coordinates (fully worked)

A particle at radius $r=3$ has contravariant angular component $V^{\theta}=2\ \mathrm{s}^{-1}$ (an angular rate). Find the covariant component $V_{\theta}$ and the physical (orthonormal-frame) component, and confirm they are three descriptions of one vector.

Step 1. Read the metric off the line element. Plane polar coordinates have $\mathrm{d}s^2=\mathrm{d}r^2+r^2\,\mathrm{d}\theta^2$ , so $g_{rr}=1$ and $g_{\theta\theta}=r^2$ . At $r=3$ this gives $g_{\theta\theta}=9$ , and the inverse is $g^{\theta\theta}=1/9$ . The angular metric component is $g_{\theta\theta}=9$ .

Step 2. Lower the index with (6): $V_{\theta}=g_{\theta\theta}V^{\theta}=9\times 2=18$ . The covariant component is numerically nine times the contravariant one — the two are emphatically not the same array, unlike in Cartesian coordinates. The covariant component is $V_{\theta}=18$ .

Step 3. Recover the physical component using the scale factor $h_{\theta}=\sqrt{g_{\theta\theta}}=r=3$ : $v_{\theta}=h_{\theta}V^{\theta}=3\times 2=6$ , which is exactly the familiar $r\dot\theta$ of elementary mechanics. So one geometric vector slot is described by three different numbers — $V^{\theta}=2$ , $v_{\theta}=6$ , $V_{\theta}=18$ — related by successive factors of $r$ . The three components are 2 (contravariant), 6 (physical), and 18 (covariant).

2 · Checking invariance under a linear coordinate change (fully worked)

Take the 2D linear change with Jacobian $A=\begin{pmatrix}2&0\\1&1\end{pmatrix}$ , a vector $V^{\nu}=(1,3)$ , and a covector $\omega_{\nu}=(4,2)$ . Transform each and verify the contraction $\omega_{\mu}V^{\mu}$ is unchanged.

Step 1. Transform the vector with the Jacobian, per (1): $V'^{\mu}=A^{\mu}{}_{\nu}V^{\nu}$ , so $V'^{1}=2(1)+0(3)=2$ and $V'^{2}=1(1)+1(3)=4$ . The transformed vector is $V'^{\mu}=(2,4)$ .

Step 2. Transform the covector with the inverse Jacobian, per (2): $\omega'_{\mu}=(A^{-1})^{\nu}{}_{\mu}\omega_{\nu}$ . Here $A^{-1}=\tfrac12\begin{pmatrix}1&0\\-1&2\end{pmatrix}$ , so $\omega'_{1}=\tfrac12(1)(4)+\tfrac12(-1)(2)=2-1=1$ and $\omega'_{2}=0(4)+1(2)=2$ . The transformed covector is $\omega'_{\mu}=(1,2)$ .

Step 3. Contract in both frames. Original: $\omega_{\nu}V^{\nu}=4(1)+2(3)=10$ . Primed: $\omega'_{\mu}V'^{\mu}=1(2)+2(4)=10$ . The Jacobian on the vector cancelled the inverse Jacobian on the covector exactly as derivation 2 promised. The contraction is $10$ in both frames — a coordinate invariant.

Physics of the portrait

Covariant and contravariant tensors are the vocabulary of one specific intellectual achievement: the "absolute differential calculus" built in Padua in the closing decades of the nineteenth century, a machinery so far ahead of its applications that it sat nearly unused for a generation until relativity came looking for it.

Origins · 1 minGregorio Ricci-Curbastro and how this was discovered

Gregorio Ricci-Curbastro

1853–1925 · founder of the absolute differential calculus, the tensor calculus of general relativity

Gregorio Ricci-Curbastro spent almost his entire career as professor of mathematical physics at the University of Padua, and his great work was to weld scattered nineteenth-century results — Riemann's quadratic differential forms, Christoffel's 1869 study of how such forms transform, Beltrami's differential parameters, Lipschitz's invariants — into a single, systematic calculus of quantities defined by their transformation law. Between roughly 1884 and 1894 he developed what he called the absolute differential calculus: a method for writing geometric and physical relations so that they hold in every coordinate system, with objects sorted by how many indices transform "co" and how many "contra." The terminology itself predates him — James Joseph Sylvester had coined "covariant" and "contravariant" around 1851 in the algebra of invariants — but Ricci gave the ideas their differential-geometric home and their calculational power. In 1900 he and his brilliant student Tullio Levi-Civita published the definitive account, Méthodes de calcul différentiel absolu et leurs applications, in Mathematische Annalen. For fifteen years it was admired by a handful of geometers and used by almost no one; Ricci's own colleagues thought the formalism ponderous. Then Einstein, guided by his friend Marcel Grossmann, discovered that general covariance demanded exactly this apparatus, and by 1915–1916 the Ricci calculus had become the indispensable language of gravitation. The curvature contraction now universally called the Ricci tensor, sitting at the heart of Einstein's field equations, carries his name. Ricci lived just long enough to see the calculus he built in near-obscurity become one of the central tools of twentieth-century physics.

History

1851

— Sylvester coins "covariant" and "contravariant" in the algebra of invariants

1869

— Christoffel studies how quadratic differential forms transform, producing the Christoffel symbols

1900

— Ricci-Curbastro and Levi-Civita publish the absolute differential calculus

1908

— Minkowski recasts special relativity in four-vector, upper-and-lower-index form

1916

— Einstein founds general relativity on tensor calculus and introduces the summation convention

Going further

References. Stone & Goldbart, Mathematics for Physics, ch. on Tensors and Differential Geometry · Arfken & Weber, Mathematical Methods for Physicists, ch. on Tensor Analysis · Nakahara, Geometry, Topology and Physics, ch. 5 (manifolds and tensors) · Byron & Fuller, Mathematics of Classical and Quantum Physics, ch. on tensors · Hassani, Mathematical Physics, ch. on tensor analysis · Riley, Hobson & Bence, Mathematical Methods for Physics and Engineering · Schutz, A First Course in General Relativity, ch. 3 · Misner, Thorne & Wheeler, Gravitation, ch. 3. Key checkpoints for the results in this article: $\partial_{\mu}f$ transforms with the inverse Jacobian; the chain-rule identity $\partial x'^{\mu}/\partial x^{\nu}\,\partial x^{\nu}/\partial x'^{\rho}=\delta^{\mu}{}_{\rho}$ holds; the contraction $\omega_{\mu}V^{\mu}$ is invariant, confirmed numerically in worked example 2 as $10=10$ under the Jacobian $A=\big(\begin{smallmatrix}2&0\\1&1\end{smallmatrix}\big)$ ; and the polar metric $g_{\theta\theta}=r^2$ gives $V_{\theta}=18$ and physical component $6$ at $r=3,\ V^{\theta}=2$ , with scale factors $h_r=1,\ h_\theta=r$ .

Prerequisites. Tensor Algebra and the Tensor Product · Manifolds and Tangent Spaces · Linear Algebra (dual spaces, change of basis). Leads to. The Metric Tensor · Raising and Lowering Indices · Tensor Contraction and the Trace · The Covariant Derivative.

Ask the tutor: Ask me to walk you through raising and lowering a specific tensor's indices with a metric of your choice, or to prove any object obeys — or fails — the tensor transformation law.

If a tensor is "just" its components, why insist it is defined by a transformation law?

Because the same array of numbers can be a tensor or not, depending on how it responds to a coordinate change. A tensor is an equivalence class: one geometric object, with a different component list in each frame, all tied together by the law in (3). The law is what guarantees that an equation true in one frame is true in all frames — the property that makes tensors worth using. If you only ever work in one coordinate system, you can pretend a tensor is its array; the moment you change coordinates, only the objects obeying the transformation law keep their physical meaning, which is why the definition is placed on the law and not the numbers.

What is the real difference between a vector and a covector — do we need both?

A vector lives in the tangent space $T_pM$ ; think of it as an arrow, a velocity, a displacement. A covector lives in the dual space $T_p^{*}M$ ; think of it as a machine that eats a vector and returns a number — a gradient, a stack of level surfaces. They are genuinely different animals: they transform with opposite matrices, and there is no coordinate-free way to turn one into the other without extra structure. That extra structure is the metric. So on a bare manifold you truly need both, and the pairing $\omega(V)=\omega_{\mu}V^{\mu}$ is the only frame-independent thing you can build from one of each. The metric merely lets you pretend a vector is a covector by supplying the missing dictionary.

Why does the Einstein summation convention only sum an upper index against a lower one?

Because that is the only pairing that produces a coordinate-independent result. When you sum an upper index (which carries a Jacobian) against a lower one (which carries the inverse Jacobian), the two matrices cancel into a Kronecker delta, as derivation 2 shows, and the sum is a true scalar or lower-rank tensor. Summing two upper indices leaves two uncancelled Jacobians, so the result changes when you change coordinates — it is not a tensor and usually not physically meaningful. The convention encodes this rule visually: a repeated index appearing once up and once down is a legal, invariant contraction; a repeated index at the same height is a warning sign that something is wrong or that a metric has been silently used.

In Cartesian coordinates V^μ and V_μ are identical — so why bother with the distinction?

Because "identical in Cartesian coordinates" is a coincidence of the Euclidean metric being the identity, not a general truth. The instant you switch to polar, cylindrical, or spherical coordinates, the metric acquires factors like $r^2$ , and $V_{\theta}=r^2V^{\theta}$ — the two arrays diverge, as worked example 1 shows with the numbers 2, 6, and 18. In curved spacetime there is no coordinate system in which they coincide everywhere. Ignoring the distinction is fine for elementary mechanics on flat space, but it produces wrong equations of motion in curvilinear coordinates and is simply undefined in general relativity. The safe habit is to track index height always, so that nothing breaks when the metric stops being trivial.

Is the metric tensor part of the definition of covariant and contravariant, or something extra?

Something extra. The covariant and contravariant transformation laws (1) and (2) are defined purely from the coordinate change — they need no metric, and they exist on any smooth manifold. What the metric adds is a bridge between the two types: without it, a vector and a covector are separate objects that can only be paired, never equated. The metric $g_{\mu\nu}$ supplies the isomorphism $V_{\mu}=g_{\mu\nu}V^{\nu}$ that lets you represent one geometric vector in either language and raise or lower indices at will. So the distinction is prior and more fundamental; the metric is the additional geometric structure that makes the two representations interchangeable.

How does this connect to the covariant derivative that comes next?

Ordinary partial derivatives of tensor components do not themselves form a tensor: differentiating $V^{\mu}$ mixes in derivatives of the basis vectors, spoiling the clean transformation law. The covariant derivative $\nabla_{\mu}$ repairs this by adding Christoffel-symbol correction terms that exactly cancel the offending pieces, so that $\nabla_{\mu}V^{\nu}$ is a genuine $(1,1)$ -tensor. Everything in this article — knowing which slots are up, which are down, and how each must transform — is the prerequisite for defining that correction, because the Christoffel symbols are precisely the objects that make the derivative respect index height. Raising and lowering indices, the metric, and the covariant derivative are the immediate sequels, and they all take the transformation law established here as given.