Direct Sums of Subspaces

Direct sum decompositions are the backbone of eigenspace theory: when an operator is diagonalizable, the whole space splits as a direct sum of its eigenspaces, and this is exactly what makes diagonalization meaningful — coordinates along each eigenspace evolve independently under the operator. The same idea underlies the primary decomposition and Jordan form for operators that are not diagonalizable, where generalized eigenspaces replace eigenspaces but the direct-sum structure persists. In signal processing and numerical linear algebra, orthogonal direct sum decompositions ($V = U \oplus U^\perp$) are the basis of least-squares projection and the singular value decomposition, letting a signal be split into a "modeled" component and a "residual" component that do not interfere with each other. In coding theory and cryptography, error-correcting codes are frequently constructed as direct sum decompositions of a vector space over a finite field, so that information and redundancy occupy cleanly separated coordinates. In representation theory and quantum mechanics, decomposing a representation or a Hilbert space into a direct sum of irreducible pieces is precisely how physicists classify particle states and how algebraists classify group actions; the uniqueness of decomposition guaranteed by the theorem here is what makes such classifications well-defined. Finally, direct sums are the immediate stepping stone to Quotient Spaces: given any complement $W$ of $U$ in $V$, the quotient $V/U$ is naturally isomorphic to $W$, and to the Dimension Formulas for Subspaces that compute $\dim(U+W)$ in terms of $\dim U$, $\dim W$, and $\dim(U\cap W)$, specializing beautifully to $\dim U + \dim W$ exactly in the direct-sum case.

Ask the tutor: If $V = U_1 \oplus U_2 \oplus U_3$, why must $U_i \cap U_j = \{0\}$ for every pair $i \neq j$, and is that condition alone ever sufficient?

Picture $V = \mathbb{R}^2$, with $U$ the $x$-axis and $W$ the $y$-axis. Every point $(a,b)$ splits as $(a,0) + (0,b)$, and this is the only way to split it into a horizontal part plus a vertical part. That uniqueness is what a direct sum captures. Contrast this with $U$ the $x$-axis and $W$ the line $y = x$: the point $(2,2)$ can be written as $(2,2) = (0,0) + (2,2)$, or as $(2,2) = (2,0)+(0,2)$, or as $(-3,0) + (5,5) \cdot \frac15$-type combinations — infinitely many ways, since $U \cap W \ne \{0\}$ only in the axis-vs-axis case is false; here $U \cap W = \{0\}$ actually still holds since the $x$-axis meets $y=x$ only at the origin, so this sum is direct. The genuinely bad case is $U = W$: then $(2,2) = u + w$ can absorb any vector back and forth between the two copies, and uniqueness fails completely. The mental picture to keep is: a direct sum decomposition is a coordinate system. Each summand $U_i$ is one "axis" (possibly high-dimensional), and a vector's decomposition is the tuple of its coordinates along each axis. Non-directness is exactly redundancy — some nonzero vector lies simultaneously in two of the summands (or, for three or more summands, in one summand and in the sum of the others), so a piece can be shifted between two representations without changing the total.

Example 1 — Polynomials split by low-degree and a shifted complement. Let $V = \mathcal{P}_3(\mathbb{R})$, the space of real polynomials of degree at most $3$, so $\dim V = 4$ with basis $1, x, x^2, x^3$. Let

the polynomials of degree $\le 1$. Let

Note $W$ is not simply "degree $2$ and $3$ terms"; it is deliberately tilted to make the example non-trivial. Claim: $V = U \oplus W$. First, $\dim U = 2$ (basis $1,x$) and $\dim W = 2$, since $x^2 - 3x$ and $x^3+2x$ are linearly independent: no scalar multiple of $x^2-3x$ (degree $2$) can equal a nonzero multiple of $x^3+2x$ (degree $3$), and neither is a scalar multiple of the other by direct inspection of degrees.

Check $U \cap W = \{0\}$. Suppose $p \in U \cap W$. Since $p \in U$, $p = a_0 + a_1 x$ for some $a_0, a_1$. Since $p \in W$, $p = s(x^2 - 3x) + t(x^3+2x)$ for some $s,t \in \mathbb{R}$, i.e.

Equating this with $a_0 + a_1 x$ (matching coefficients of $1, x, x^2, x^3$, which is valid since $1,x,x^2,x^3$ is a basis of $V$):

From $t=0, s=0$ we get $a_1 = 0$ as well, and $a_0 = 0$. Hence $p = 0$. So $U \cap W = \{0\}$.

Check $U + W = V$. We show $1, x, x^2-3x, x^3+2x$ span $V$ and, since $\dim V = 4$ and we have $4$ vectors, it suffices to check they are linearly independent (in a finite-dimensional space, a spanning set of size $\dim V$ is automatically a basis — this is a standard consequence of the dimension theorem for spanning/independent sets). Suppose

Collecting by degree:

Matching coefficients against $1,x,x^2,x^3$ (a basis, so all coefficients must vanish): $c_4 = 0$, $c_3 = 0$, then $c_2 - 3(0)+2(0) = c_2 = 0$, and $c_1 = 0$. So all $c_i = 0$: the four vectors are linearly independent, hence a basis, hence $U + W = \operatorname{span}(1,x,x^2-3x,x^3+2x) = V$. By the corollary, $V = U \oplus W$. As an explicit instance, decompose $p(x) = 3 + 5x - 2x^2 + 7x^3$. We want $p = u + w$ with $u = a_0+a_1x \in U$ and $w = s(x^2-3x)+t(x^3+2x) \in W$. Matching coefficients as above: $t = 7$ (from $x^3$), $s = -2$ (from $x^2$), $a_0 = 3$ (from constant term), and from the $x$-coefficient, $a_1 + (2t - 3s) = 5$, i.e., $a_1 + (14 + 6) = 5$, so $a_1 = 5 - 20 = -15$. Thus

and one checks $u + w = 3 - 15x + 7x^3 - 2x^2 + 20x = 3 + 5x - 2x^2+7x^3 = p(x)$. ✓. This decomposition is unique by directness, matching the theorem's guarantee.

Example 2 — Three subspaces of $\mathbb{R}^3$: pairwise trivial intersection is not enough. This example simultaneously illustrates directness of three lines and sets up the counterexample below. In $\mathbb{R}^3$, let

Each $U_i$ is $1$-dimensional. We test condition (c) of the theorem directly, for each $i$. For $i=1$: is $U_1 \cap (U_2+U_3) = \{0\}$? A general element of $U_2 + U_3$ is $b(0,1,0) + c(1,1,1) = (c, b+c, c)$. For this to lie in $U_1 = \operatorname{span}(1,0,0)$ we need the second and third coordinates to vanish: $b + c = 0$ and $c = 0$, forcing $b=c=0$, giving the zero vector. So $U_1 \cap (U_2+U_3) = \{0\}$. For $i=2$: is $U_2 \cap (U_1+U_3) = \{0\}$? Elements of $U_1+U_3$: $a(1,0,0)+c(1,1,1) = (a+c, c, c)$. Lying in $U_2$ requires first and third coordinates zero: $a + c = 0$, $c=0$, so $a=c=0$: zero vector again. Trivial intersection holds. For $i=3$: is $U_3 \cap (U_1+U_2) = \{0\}$? Elements of $U_1+U_2$ are $(a,b,0)$; elements of $U_3$ are $(c,c,c)$. Equality forces $c=0$ (third coordinate), hence $a=b=0$ too. Trivial intersection holds. All three conditions in (c) hold, so by the theorem $U_1 + U_2 + U_3$ is direct: $U_1 \oplus U_2 \oplus U_3$. Indeed $\dim(U_1+U_2+U_3) = 1+1+1=3 = \dim \mathbb{R}^3$, consistent with the sum being all of $\mathbb{R}^3$ and direct.