A basis promises that every vector can be written uniquely as a linear combination of chosen vectors, but uniqueness is not the same as readability. In the standard basis of $k^n$, the coordinates of $(a_1, \ldots, a_n)$ stare back at us. In a skew basis, or in a shifted polynomial basis such as $(1, X-1, (X-1)^2)$, the coefficients are no longer visible from the notation alone. The dual basis is the device that makes coordinates into functions. The guiding question is this: once a basis has been chosen, can we build linear functionals that extract the individual coordinates? If $v = a_1 e_1 + \cdots + a_n e_n$, we want a functional $e_i^*$ satisfying $e_i^*(v)=a_i$. The content of the theory is that these coordinate-reading maps exist, are unique, form a basis of the [dual space](/page/Dual%20Space), and transform contravariantly when the original basis changes. The first surprise is that the coordinate maps are not an extra Euclidean structure. They do not require an [inner product](/page/Inner%20Product), a length, or an angle. They come only from the algebra of a basis. This is why dual bases appear in linear algebra, tensor calculus, differential geometry, representation theory, and commutative algebra: whenever vectors are described by coordinates, there are dual coordinate functionals silently doing the reading. Before naming the construction, it helps to see the failure it repairs. In a nonstandard basis, the coefficient of a vector cannot be recovered by looking at a component of the ambient coordinate tuple. [example: Hidden Coordinates in a Skew Basis] Let $V=\mathbb R^2$ and take the ordered basis $\mathcal B=(e_1,e_2)$ with $e_1=(1,1)$ and $e_2=(1,-1)$. To find the $\mathcal B$-coordinates of $v=(5,1)$, write \begin{align*} (5,1)=a_1(1,1)+a_2(1,-1). \end{align*} Expanding the right-hand side gives \begin{align*} a_1(1,1)+a_2(1,-1)=(a_1+a_2,a_1-a_2). \end{align*} Thus the coefficients must satisfy \begin{align*} a_1+a_2=5 \quad \text{and} \quad a_1-a_2=1. \end{align*} Adding the two equations gives $2a_1=6$, so $a_1=3$. Substituting into $a_1+a_2=5$ gives $3+a_2=5$, hence $a_2=2$. Therefore \begin{align*} (5,1)=3(1,1)+2(1,-1), \end{align*} so the standard coordinates $(5,1)$ and the $\mathcal B$-coordinates $(3,2)$ are different. The same calculation for a general vector $(x,y)$ gives \begin{align*} (x,y)=a_1(1,1)+a_2(1,-1)=(a_1+a_2,a_1-a_2). \end{align*} Hence \begin{align*} a_1+a_2=x \quad \text{and} \quad a_1-a_2=y. \end{align*} Adding these equations gives $2a_1=x+y$, so $a_1=(x+y)/2$. Subtracting the second equation from the first gives $2a_2=x-y$, so $a_2=(x-y)/2$. The coordinate-reading functionals for the basis $\mathcal B$ are therefore \begin{align*} e_1^*(x,y)=\frac{x+y}{2} \end{align*} and \begin{align*} e_2^*(x,y)=\frac{x-y}{2}. \end{align*} Applying them to $v=(5,1)$ gives $e_1^*(5,1)=3$ and $e_2^*(5,1)=2$, so these two linear functionals recover the hidden coordinates. [/example] The example shows the whole idea in miniature. A dual basis is not obtained by reusing the original basis vectors; it is built from the coordinate extraction problem attached to that basis. ## Definition ### Coordinate Tests The construction we want is a whole family of scalar-valued tests, one test for each basis vector. Each test should return $1$ on its own basis vector and $0$ on the others, so that applying the tests to a vector reveals exactly the coefficients in its basis expansion. [definition: Dual Basis] Let $V$ be a finite-dimensional [vector space](/page/Vector%20Space) over a field $k$, and let $\mathcal B=(e_1,\ldots,e_n)$ be an ordered basis of $V$. The dual basis of $\mathcal B$ is the ordered tuple $\mathcal B^*=(e_1^*,\ldots,e_n^*)$ in $V^*$ such that \begin{align*} e_i^*(e_j)=\delta_{ij} \end{align*} for all $i,j\in\{1,\ldots,n\}$. [/definition] The defining equations say that each $e_i^*$ recognises one basis vector and annihilates the others. The notation $e_i^*$ depends on the whole basis $\mathcal B$, not only on the single vector $e_i$. If $e_i$ is kept fixed but the other basis vectors change, the functional labelled $e_i^*$ may change. ### The Ambient Dual Space To make the definition fully precise, we unpack the two pieces of language it uses. First, the dual basis lives in the space of all linear scalar measurements on $V$. [definition: Dual Space] Let $V$ be a vector space over a field $k$. The dual space of $V$ is \begin{align*} V^* := \operatorname{Hom}_k(V,k), \end{align*} the vector space of all $k$-linear maps $f:V\to k$. [/definition] ### Ordering and Coordinate Notation Second, the phrase "the $i$-th coordinate" requires an ordering. A basis as a set spans the space, but a coordinate list has positions: a first vector, a second vector, and so on. Without this extra ordering data, the symbol $e_i^*$ has no well-defined index to read, so we distinguish an ordered basis from an unordered spanning list. [definition: Ordered Basis] Let $V$ be a finite-dimensional vector space over a field $k$. An ordered basis of $V$ is a tuple $\mathcal B=(e_1,\ldots,e_n)$ such that $\{e_1,\ldots,e_n\}$ is a basis of $V$. [/definition] The ordering lets us write every vector $v\in V$ in the form \begin{align*} v=\sum_{i=1}^n a_i e_i, \end{align*} where the scalar $a_i\in k$ is the $i$-th coordinate of $v$ with respect to $\mathcal B$. The next task is to turn the rule $v\mapsto a_i$ into an actual member of the dual space. [definition: Coordinate Functional] Let $V$ be a finite-dimensional vector space over a field $k$, and let $\mathcal B=(e_1,\ldots,e_n)$ be an ordered basis of $V$. For $i\in\{1,\ldots,n\}$, the $i$-th coordinate functional associated to $\mathcal B$ is the map $e_i^*:V\to k$ defined by \begin{align*} e_i^*\left(\sum_{j=1}^n a_j e_j\right)=a_i. \end{align*} [/definition] The coordinate functional definition describes the same maps by what they do to an arbitrary vector. To state their values compactly on basis vectors, we use a small piece of notation that behaves like an algebraic switch. [definition: Kronecker Delta] For indices $i,j\in\{1,\ldots,n\}$, the Kronecker delta is the scalar $\delta_{ij}$ defined by $\delta_{ij}=1$ when $i=j$, and $\delta_{ij}=0$ when $i\ne j$. [/definition] The Kronecker delta records whether a coordinate functional is being applied to its matching basis vector. This is why the compact condition $e_i^*(e_j)=\delta_{ij}$ captures the entire coordinate-extraction problem. ## Coordinates as Linear Functionals The dual basis turns the informal phrase "take coordinates" into a [linear map](/page/Linear%20Map). This matters because linear functionals can be added, scaled, composed with linear maps, and transported through change-of-basis formulas. Coordinates become part of the algebra rather than a separate bookkeeping convention. The next theorem records the coordinate expansion that motivates the whole construction. It is the statement that the dual basis really recovers a vector from its coordinate measurements. [quotetheorem:8355] This formula is often the most useful way to remember the dual basis. The vector $v$ is reconstructed by measuring it with each coordinate functional and then placing the resulting scalars back on the original basis vectors. A polynomial example makes the construction feel less tied to columns in $k^n$. The same mechanism works whenever a basis gives a coordinate system. [example: Dual Basis for Quadratic Polynomials] Let $V$ be the vector space of real polynomials of degree at most $2$, and take the ordered basis $\mathcal B=(1,X,X^2)$. Every $p\in V$ has a unique expansion \begin{align*} p(X)=a_0+a_1X+a_2X^2. \end{align*} The coordinate functionals associated to $\mathcal B$ are therefore \begin{align*} e_1^*(p)=a_0,\quad e_2^*(p)=a_1,\quad e_3^*(p)=a_2. \end{align*} For example, on the basis vectors themselves these satisfy \begin{align*} e_1^*(1)=1,\quad e_1^*(X)=0,\quad e_1^*(X^2)=0, \end{align*} and similarly each $e_i^*$ reads the coefficient of the corresponding basis element. Now take the shifted ordered basis $\mathcal C=(1,X-1,(X-1)^2)$. Its dual basis reads coefficients in expansions of the form \begin{align*} p(X)=b_0+b_1(X-1)+b_2(X-1)^2. \end{align*} For $p(X)=X^2$, use $X=(X-1)+1$. Squaring gives \begin{align*} X^2=((X-1)+1)^2. \end{align*} Expanding the square gives \begin{align*} ((X-1)+1)^2=(X-1)^2+2(X-1)+1. \end{align*} Reordering in the basis $\mathcal C$ gives \begin{align*} X^2=1+2(X-1)+(X-1)^2. \end{align*} Thus the $\mathcal C$-coordinates of $X^2$ are $(1,2,1)$. If $\mathcal C^*=(c_1^*,c_2^*,c_3^*)$, then \begin{align*} c_1^*(X^2)=1,\quad c_2^*(X^2)=2,\quad c_3^*(X^2)=1. \end{align*} So the first shifted dual functional sends $X^2$ to $1$, whereas the first standard dual functional sends $X^2$ to $0$; the coefficient being read depends on the chosen basis. [/example] The example warns against a common mistake: a dual basis is not an absolute list of coefficient maps on a vector space. It is tied to a chosen basis. Changing the basis changes what the word "coefficient" means. ## Existence and Uniqueness The definition of a dual basis says what the desired functionals should do on the basis vectors. The next question is whether those prescriptions determine actual linear maps on the whole space. This is where the universal role of a basis enters: specifying a linear map on basis vectors determines it everywhere. [quotetheorem:414] Existence says that the coordinate-reading maps really are linear functionals. Uniqueness says there is no hidden choice: once the ordered basis is fixed, the dual basis is forced. This is the dual version of coordinate expansion. A vector is determined by its coordinates; a linear functional is determined by its values on a basis. The dual basis therefore gives a coordinate system for the dual space: instead of measuring a vector by its coefficients, it measures a functional by the values it assigns to the chosen basis vectors. The formula also gives a dimension count. If $\dim_k V=n$, then $\dim_k V^*=n$. The equality of dimensions is not the same as a chosen identification of $V$ with $V^*$; the dual basis construction depends on the chosen basis. [example: Row Functionals and the Standard Dual Basis] Let $V=k^n$ with standard ordered basis $\mathcal E=(e_1,\ldots,e_n)$, where $e_i$ has $1$ in the $i$-th position and $0$ in every other position. For $x=(x_1,\ldots,x_n)\in k^n$, the standard basis expansion is \begin{align*} x=x_1e_1+\cdots+x_ne_n. \end{align*} If $f:k^n\to k$ is linear, then applying $f$ and using linearity gives \begin{align*} f(x)=f(x_1e_1+\cdots+x_ne_n)=x_1f(e_1)+\cdots+x_nf(e_n). \end{align*} Set $a_i=f(e_i)$ for each $i$. Then \begin{align*} f(x_1,\ldots,x_n)=a_1x_1+\cdots+a_nx_n. \end{align*} The scalars $a_i$ are forced, because evaluating at $e_i$ gives \begin{align*} f(e_i)=a_i. \end{align*} The standard dual basis consists of the coordinate projection maps \begin{align*} e_i^*(x_1,\ldots,x_n)=x_i. \end{align*} Therefore, for every $x=(x_1,\ldots,x_n)$, \begin{align*} (a_1e_1^*+\cdots+a_ne_n^*)(x)=a_1e_1^*(x)+\cdots+a_ne_n^*(x)=a_1x_1+\cdots+a_nx_n=f(x). \end{align*} Hence the row vector $(a_1,\ldots,a_n)$ represents the functional $a_1e_1^*+\cdots+a_ne_n^*$ in the standard dual basis. [/example] In matrix language, vectors are often columns and covectors are rows. The dual basis explains why rows multiply columns to produce scalars: a row is a linear functional written in the dual basis. ## Change of Basis A basis is rarely permanent. Computations often become easier after changing coordinates, diagonalising an operator, or choosing a basis adapted to a subspace. The dual basis then changes in the opposite direction from the original basis. This opposite movement is the source of the word contravariant in tensor calculus. We use the standard coordinate notation $[x]_{\mathcal B}$ and change-of-basis matrix $P_{\mathcal C\leftarrow\mathcal B}$: if $\mathcal B$ and $\mathcal C$ are ordered bases of $V$, then \begin{align*} [x]_{\mathcal C}=P_{\mathcal C\leftarrow\mathcal B}[x]_{\mathcal B}. \end{align*} The next theorem states the inverse-transpose rule. It is the computational heart of dual bases under coordinate change. [quotetheorem:416] The inverse appears because a functional must give the same scalar after the vector coordinates have changed. The transpose appears because functionals act like row vectors against coordinate columns. A concrete Euclidean example shows why the dual basis is not generally the same as the original geometric vectors. Orthogonality can hide this issue in the standard basis, but a skew basis exposes it. [example: Dual Basis in a Non-Orthonormal Basis] Let $V=\mathbb R^2$ and let $\mathcal B=(e_1,e_2)$ with $e_1=(1,0)$ and $e_2=(1,1)$. To compute the dual basis, write an arbitrary vector $v=(x,y)$ in the form \begin{align*} (x,y)=a_1(1,0)+a_2(1,1). \end{align*} Expanding the right-hand side coordinate by coordinate gives \begin{align*} a_1(1,0)+a_2(1,1)=(a_1,0)+(a_2,a_2)=(a_1+a_2,a_2). \end{align*} Therefore the coefficients satisfy \begin{align*} a_1+a_2=x \quad \text{and} \quad a_2=y. \end{align*} The second equation gives $a_2=y$. Substituting this into the first equation gives \begin{align*} a_1+y=x. \end{align*} Subtracting $y$ from both sides gives \begin{align*} a_1=x-y. \end{align*} Thus \begin{align*} (x,y)=(x-y)(1,0)+y(1,1). \end{align*} The coordinate functionals for the basis $\mathcal B$ are therefore \begin{align*} e_1^*(x,y)=x-y \end{align*} and \begin{align*} e_2^*(x,y)=y. \end{align*} They satisfy the required dual-basis values: \begin{align*} e_1^*(1,0)=1-0=1,\quad e_1^*(1,1)=1-1=0 \end{align*} and \begin{align*} e_2^*(1,0)=0,\quad e_2^*(1,1)=1. \end{align*} As row vectors acting on standard coordinate pairs, these functionals are represented by $(1,-1)$ and $(0,1)$, because \begin{align*} (1,-1)\cdot(x,y)=1\cdot x+(-1)\cdot y=x-y \end{align*} and \begin{align*} (0,1)\cdot(x,y)=0\cdot x+1\cdot y=y. \end{align*} The dual basis is therefore not the original list of geometric vectors $(1,0),(1,1)$; it is the list of coordinate-reading covectors $(1,-1),(0,1)$ in the standard row-vector representation. [/example] The example shows a useful distinction. In an [inner product space](/page/Inner%20Product%20Space), a functional may be represented by a vector through the inner product, but that representing vector is not usually the same as the original basis vector. ## Pairings and Natural Maps The expression $e_i^*(v)$ is an instance of a general pairing between the dual space and the original space. This pairing is the algebraic operation behind matrix multiplication, evaluation of differential forms on tangent vectors, and the contraction of tensors. [definition: Evaluation Pairing] Let $V$ be a vector space over a field $k$. The evaluation pairing between $V^*$ and $V$ is the map $\operatorname{ev}:V^*\times V\to k$ defined by \begin{align*} \operatorname{ev}(f,v)=f(v). \end{align*} [/definition] The evaluation pairing is bilinear: it is linear in the functional and in the vector separately. Dual bases are exactly the pairs of bases for which the evaluation matrix is the identity matrix. To use duality for subspaces, the next construction records which functionals vanish on a given subspace. [definition: Annihilator] Let $V$ be a vector space over a field $k$, and let $W\subset V$ be a subspace. The annihilator of $W$ is \begin{align*} W^\circ := \{f\in V^*: f(w)=0 \text{ for all } w\in W\}. \end{align*} [/definition] Annihilators connect dual bases with systems of linear equations. If a subspace is spanned by some basis vectors, then the dual basis functionals corresponding to the complementary basis vectors cut it out. [example: A Coordinate Subspace and Its Annihilator] Let $V=k^4$ with standard basis $\mathcal E=(e_1,e_2,e_3,e_4)$, and let $W=\operatorname{span}(e_1,e_3)$. Every vector in $W$ has the form \begin{align*} ae_1+be_3=a(1,0,0,0)+b(0,0,1,0)=(a,0,b,0). \end{align*} Thus $v=(x_1,x_2,x_3,x_4)$ lies in $W$ exactly when $x_2=0$ and $x_4=0$. We compute the functionals that vanish on all of $W$. The standard dual basis satisfies \begin{align*} e_i^*(x_1,x_2,x_3,x_4)=x_i. \end{align*} So $e_2^*$ and $e_4^*$ vanish on $W$, because for every $(a,0,b,0)\in W$, \begin{align*} e_2^*(a,0,b,0)=0 \end{align*} and \begin{align*} e_4^*(a,0,b,0)=0. \end{align*} Hence every linear combination $\alpha e_2^*+\beta e_4^*$ lies in $W^\circ$. Conversely, let $f\in W^\circ$. Since $\mathcal E^*=(e_1^*,e_2^*,e_3^*,e_4^*)$ is the standard dual basis of $V^*$, write \begin{align*} f=\lambda_1e_1^*+\lambda_2e_2^*+\lambda_3e_3^*+\lambda_4e_4^*. \end{align*} Because $e_1\in W$ and $f$ vanishes on $W$, \begin{align*} 0=f(e_1)=\lambda_1e_1^*(e_1)+\lambda_2e_2^*(e_1)+\lambda_3e_3^*(e_1)+\lambda_4e_4^*(e_1)=\lambda_1. \end{align*} Because $e_3\in W$ as well, \begin{align*} 0=f(e_3)=\lambda_1e_1^*(e_3)+\lambda_2e_2^*(e_3)+\lambda_3e_3^*(e_3)+\lambda_4e_4^*(e_3)=\lambda_3. \end{align*} Therefore \begin{align*} f=\lambda_2e_2^*+\lambda_4e_4^*. \end{align*} This proves \begin{align*} W^\circ=\operatorname{span}(e_2^*,e_4^*)\subset V^*. \end{align*} The dual basis records the two coordinate equations $x_2=0$ and $x_4=0$ that define this coordinate subspace. [/example] This is a useful model for more general subspaces: linear equations are elements of the dual, and a choice of basis lets us express those equations in coordinates. The next natural question is what happens if we dualise once more and let vectors act on functionals by evaluation. [definition: Bidual Map] Let $V$ be a vector space over a field $k$. The bidual map is the linear map $\iota_V:V\to V^{**}$ defined by \begin{align*} \iota_V(v)(f)=f(v) \end{align*} for $v\in V$ and $f\in V^*$. [/definition] Unlike the identification of $V$ with $V^*$ coming from a chosen basis, the bidual map is canonical: it does not require selecting coordinates. Finite-dimensionality determines when this canonical map reaches all of $V^{**}$. [quotetheorem:8356] This theorem separates two ideas that are often conflated. The map $V\to V^{**}$ is natural, while an isomorphism $V\to V^*$ usually requires extra data such as a basis or a nondegenerate [bilinear form](/page/Bilinear%20Form). ## Computations and Applications Dual bases are computational tools. They let us compute coefficients, represent linear maps by matrices, express tensors, and translate between geometric and algebraic descriptions. A linear map $T:V\to W$ is determined by its action on a basis of $V$. The dual basis of $W$ then extracts the coordinates of each image vector, producing the entries of the matrix of $T$. Concretely, if $\mathcal B=(u_1,\ldots,u_m)$ is a basis of $U$ and $\mathcal C=(v_1,\ldots,v_n)$ is a basis of $V$ with dual basis $\mathcal C^*=(v_1^*,\ldots,v_n^*)$, then the matrix of $T:U\to V$ in these bases has entries \begin{align*} A_{ij}=v_i^*(T(u_j)). \end{align*} This is the coordinate-extraction form of the [matrix representation bijection](/theorems/382) between linear maps and matrices. This formula explains why rows of a matrix are covectors: the $i$-th row records the $i$-th coordinate functional of the output applied to the images of the input basis vectors. [example: Computing a Matrix Entry by a Dual Functional] Let $T:\mathbb R^2\to\mathbb R^2$ be defined by $T(x,y)=(2x+y,x-3y)$. Use the domain basis $\mathcal B=(v_1,v_2)$ with $v_1=(1,1)$ and $v_2=(1,-1)$, and use the standard codomain basis $\mathcal E=(e_1,e_2)$. The dual basis of $\mathcal E$ consists of the coordinate projections $e_1^*(u_1,u_2)=u_1$ and $e_2^*(u_1,u_2)=u_2$, so each matrix entry is obtained by applying the appropriate coordinate projection to $T(v_j)$. For the first domain basis vector, \begin{align*} T(v_1)=T(1,1)=(2\cdot 1+1,1-3\cdot 1)=(3,-2). \end{align*} Therefore \begin{align*} e_1^*(T(v_1))=e_1^*(3,-2)=3 \end{align*} and \begin{align*} e_2^*(T(v_1))=e_2^*(3,-2)=-2. \end{align*} Thus the first column of the representing matrix is $(3,-2)$. For the second domain basis vector, \begin{align*} T(v_2)=T(1,-1)=(2\cdot 1+(-1),1-3(-1))=(1,4). \end{align*} Therefore \begin{align*} e_1^*(T(v_2))=e_1^*(1,4)=1 \end{align*} and \begin{align*} e_2^*(T(v_2))=e_2^*(1,4)=4. \end{align*} Thus the second column is $(1,4)$. Hence the matrix of $T$ from $\mathcal B$ to $\mathcal E$ has first column $(3,-2)$ and second column $(1,4)$, equivalently first row $(3,1)$ and second row $(-2,4)$. [/example] The formula extends beyond ordinary matrix computations. In differential geometry, coordinate vector fields have dual coordinate one-forms. The same Kronecker-delta relation appears, with tangent vectors replacing abstract basis vectors. [example: Coordinate One-Forms as a Dual Basis] Let $M$ be a smooth $n$-manifold, and let $(U,\varphi)$ be a chart with coordinate functions $(x_1,\ldots,x_n)$. At a point $p\in U$, write $\varphi(p)=(t_1,\ldots,t_n)$. The coordinate tangent vector $\partial_{x_j}|_p$ acts on a smooth function $f$ by differentiating $f\circ\varphi^{-1}$ in the $j$-th coordinate at $\varphi(p)$, and the one-form $dx_i|_p$ acts by \begin{align*} dx_i|_p(v)=v(x_i). \end{align*} Therefore \begin{align*} dx_i|_p(\partial_{x_j}|_p)=(\partial_{x_j}|_p)(x_i). \end{align*} Since $x_i\circ\varphi^{-1}(t_1,\ldots,t_n)=t_i$, differentiating in the $j$-th coordinate gives \begin{align*} (\partial_{x_j}|_p)(x_i)=\frac{\partial t_i}{\partial t_j}. \end{align*} If $i=j$, this derivative is $1$; if $i\ne j$, the coordinate $t_i$ is independent of $t_j$, so the derivative is $0$. Hence \begin{align*} dx_i|_p(\partial_{x_j}|_p)=\delta_{ij}. \end{align*} Thus $dx_1|_p,\ldots,dx_n|_p$ are dual to $\partial_{x_1}|_p,\ldots,\partial_{x_n}|_p$. Consequently, if $\omega$ is a one-form on $U$, then at each $p\in U$ its coefficients in this dual basis are obtained by evaluation on the coordinate tangent vectors: \begin{align*} \omega_i(p)=\omega_p(\partial_{x_i}|_p). \end{align*} So \begin{align*} \omega_p=\sum_{i=1}^n \omega_i(p)\,dx_i|_p. \end{align*} Equivalently, on the chart one writes \begin{align*} \omega=\sum_{i=1}^n \omega_i\,dx_i. \end{align*} This is the local coordinate version of the dual-basis construction. [/example] When both a vector and a functional have been expanded in matching bases, there is still a compatibility question: how do the coordinates of the covector combine with the coordinates of the vector to produce the scalar $f(v)$? The answer must respect the dual-basis pairing $e_i^*(e_j)=\delta_{ij}$, so all cross-terms vanish and only matching coordinates contribute. [quotetheorem:8357] The theorem is the abstract form of row-times-column multiplication. It is also the seed of tensor contraction: a covariant slot consumes a vector slot and returns a scalar. ## Common Pitfalls and Boundary Cases The dual basis is simple once the basis is fixed, but several mistakes recur because different identifications collapse in familiar Euclidean examples. A frequent mistake is to identify $e_i$ with $e_i^*$. These objects live in different vector spaces: $e_i\in V$, while $e_i^*\in V^*$. In $\mathbb R^n$ with the standard inner product and standard basis, the distinction is hidden by a convenient identification; outside that setting, it must be kept visible. [remark: The Dual Basis Depends on the Whole Basis] The notation $e_i^*$ should not be read as a construction from the single vector $e_i$ alone. It depends on the entire ordered basis $(e_1,\ldots,e_n)$, because the condition $e_i^*(e_j)=0$ for $j\ne i$ refers to all the other basis vectors. [/remark] Another mistake is to expect the same finite-dimensional behaviour in arbitrary infinite-dimensional spaces. The algebraic dual of an infinite-dimensional vector space is usually much larger than the span of coordinate functionals attached to a basis, and topological duals introduce additional continuity constraints. [remark: Infinite-Dimensional Warning] For an infinite-dimensional vector space with a Hamel basis, each basis vector still has an associated coordinate functional. However, the full algebraic dual contains many functionals whose behaviour is not captured by finite coordinate lists. In normed spaces, the continuous dual is a different object again, because linear functionals must also be bounded. [/remark] The finite-dimensional theory is therefore the clean setting in which the dual basis is both a coordinate tool and a basis of the whole dual space. Infinite-dimensional duality keeps the same evaluation idea, but the structural consequences require more care. ## Beyond and Connected Topics Dual bases are one of the first places where linear algebra becomes functorial. A linear map $T:V\to W$ induces a pullback map $T^*:W^*\to V^*$ by composition, sending a functional on $W$ to a functional on $V$. This reversal of arrows is the abstract version of the inverse-transpose rule for [change of basis](/page/Change%20Of%20Basis). The construction is also the entry point to tensor algebra. Once $V^*$ is available, tensors can be built from tensor products of copies of $V$ and $V^*$. The dual basis then supplies the coordinate tensors used in multilinear algebra, differential forms, and tensor fields. In differential geometry, dual bases appear as coordinate one-forms $dx_i$, dual to coordinate vector fields $\partial_{x_i}$. This is the language behind cotangent spaces, differential forms, Riemannian metrics, and integration on manifolds. The page [Cambridge III Differential Geometry](/page/Cambridge%20III%20Differential%20Geometry) develops this direction at course level. In Lie theory, dual spaces occur in representations, coadjoint actions, weights, roots, and invariant bilinear forms. A basis of a [Lie algebra](/page/Lie%20Algebra) gives a dual basis of its linear functionals, but the additional bracket structure creates richer geometry. The page [Lie Algebras I: Foundations](/page/Lie%20Algebras%20I%3A%20Foundations) is a natural continuation. In commutative algebra, dual modules are subtler than dual vector spaces because bases may not exist and Hom functors need not preserve exactness. Free modules retain a version of the dual basis story, while projective modules generalise the parts that survive without a global basis. See [Cambridge III Commutative Algebra](/page/Cambridge%20III%20Commutative%20Algebra) for the module-theoretic setting. ## References Androma, [Cambridge IB Linear Algebra](/page/Cambridge%20IB%20Linear%20Algebra). Androma, [Lie Algebras I: Foundations](/page/Lie%20Algebras%20I%3A%20Foundations). Androma, [Cambridge III Commutative Algebra](/page/Cambridge%20III%20Commutative%20Algebra). Androma, [Cambridge III Differential Geometry](/page/Cambridge%20III%20Differential%20Geometry). Sheldon Axler, *Linear Algebra Done Right* (2015). Kenneth Hoffman and Ray Kunze, *Linear Algebra* (1971). Serge Lang, *Linear Algebra* (1987).