第 13 章
第一次测验复习:按题型整理
把前半学期的考点压成几类:维数与秩的计数、由完整解反推矩阵、真伪判断、乘积的零空间,以及行空间与零空间的交集。
Quiz 1 Review: Organized by Problem Type
The first half of the course compressed into a few problem types: dimension and rank counts, reconstructing a matrix from its complete solution, true or false, null spaces of products, and the intersection of row and null spaces.
篇内目录
- What this review covers
- Timeline
- Type 1: dimension of a span
- Type 2: counting dimensions from structure
- Type 3: reconstructing a matrix from its complete solution
- Type 4: true or false
- Why a sum of invertible matrices can be singular
- Type 5: null space of a product
- Type 6: intersection of row and null spaces
- Review points
- Check your understanding
- How this connects to the course
这次复习覆盖什么
第一次测验考的是 这一半内容,重点是矩形矩阵:零空间、左零空间、秩、四个基本子空间。矩阵可逆时这些概念都「自动成立」,一旦变成矩形,就必须逐个判断。
下面按题型整理,而不是按课堂顺序复述。
时间轴
| 时间 | 内容 |
|---|---|
| 0:09 | 复习开场:本次覆盖的范围 |
| 1:49 | 题型一:张成空间的维数 |
| 5:29 | 行化简到阶梯形 |
| 8:56 | 矩阵的具体例子 |
| 12:19 | 求秩与自由变量个数 |
| 15:44 | 题型二:由矩阵结构数维数 |
| 17:26 | 零空间很大的情形 |
| 21:06 | 平方为零的矩阵可能吗 |
| 24:39 | 反例与不可逆性 |
| 26:32 | 题型三:由完整解反推矩阵 |
| 29:59 | 乘积的零空间 |
| 33:27 | 特解加零空间的组合 |
| 36:54 | 题型四:真伪判断 |
| 40:21 | 可逆矩阵给出的答案 |
| 43:47 | 行空间在消元下保持不变 |
| 45:33 | 题型六:行空间与零空间的交集 |
题型一:张成空间的维数
题目: 是 中的三个非零向量,它们张成的子空间维数可能是多少?
答案:1、2 或 3。
- 不超过 3,因为只有三个向量;
- 不为 0,因为题目说它们非零;
- 三种都能实现:三个共线则维数 1,共面但不共线则维数 2,不共面则维数 3。
考点是「维数 ≤ 向量个数」以及「非零排除零维」。
题型二:由矩阵结构数维数
题目: 是 的行阶梯形矩阵,秩 。
- 是什么?
- 的秩是多少?
- 的秩是多少?
- 是多少?
解答:
- 且秩 3,三列全部是主元列,没有自由变量,所以 。
- 第二块是第一款的两倍,行变换会把它消掉,只剩 。秩仍为 3。
- 是 。对块做消元:用左上块消去左下块,得到 ,再消去右上块,得到 。两个 各贡献 3 个主元,秩为 6。
- 是 ,有 10 列;秩为 6,所以
这一组题只用到两条公式:,以及分块消元不改变秩。
题型三:由完整解反推矩阵
题目:已知 的一个特解是 ,零空间的一组基是 与 ,且 。求 的秩、列空间与矩阵本身。
第一步:由零空间维数得秩。
零空间维数为 2,矩阵是 ,所以
第二步:由特解得第一列。
只取 的第一列乘 2:
第三步:由零空间向量得其余列。
说明第三列为零向量。
说明
于是
验证:三列都是 的倍数,秩为 1;列空间是 的所有倍数。因此 有解当且仅当 是 的倍数。
方法论:完整解 里包含两类信息——特解给出「某一列与 的关系」,零空间基给出「列与列之间的线性关系」。两者合起来往往足以还原整个矩阵。
题型四:真伪判断
| 命题 | 判断 | 理由 |
|---|---|---|
| 方阵 的 ,则 | 真 | 方阵满秩,转置也满秩 |
| 可逆矩阵构成子空间 | 假 | 不含零矩阵;两个可逆矩阵之和可能奇异 |
| 假 | 反例 | |
| 矩阵列无关,则对任意 可解 | 真 | 列无关即满秩,方阵满秩即可逆 |
| 时行空间等于列空间 | 假 | 的行空间是 方向,列空间是 方向 |
| 与 有相同的四个子空间 | 真 | 取负不改变任何线性组合关系 |
| 四个子空间相同 是 的倍数 | 假 | 任意两个 可逆矩阵都有相同的四个子空间(全空间与零空间) |
| 交换两行,行空间与零空间不变 | 真 | 行变换保持行空间;方程组的解不变 |
| 交换两行,列空间不变 | 假 | 列空间通常改变 |
为什么「可逆矩阵之和」不一定可逆
取
这个例子仍可逆,但取
就奇异了。更简单的理由是:可逆矩阵集合不含零矩阵,而子空间必须包含零向量(对任意元素乘 0 仍在集合内),所以它根本不可能是子空间。
题型五:乘积的零空间
题目:,其中 可逆。 与 有什么关系?
答案:相等。
- 若 ,则 ,所以 。
- 若 ,左乘 得 ,所以 。
因此左乘一个可逆矩阵不改变零空间。这也解释了为什么消元(左乘初等矩阵的乘积)能用来求零空间。
题型六:行空间与零空间的交集
题目:为什么向量 不可能同时是 的某一行,又属于 ?
假设它是第一行,且 。那么 的第一个分量就是
与 矛盾。
更一般地:
若 既在行空间又在零空间,则 与行空间中的每个向量都垂直;特别地,它与自己垂直,于是 ,只能是零向量。
这正是下一讲「正交」的入口:行空间与零空间不仅交集为零,而且互相垂直,维数相加为 ,因此互为正交补。
复习要点
- 维数计数:,。
- 分块消元只关心每块贡献多少个主元。
- 完整解 可以反推出矩阵的列关系。
- 子空间必须含零、对加法与数乘封闭;可逆矩阵集合不满足。
- 左乘可逆矩阵不改变零空间。
- 行空间与零空间只共享零向量。
理解检验
1. 三个非零向量在 中能张成几维?为什么?
参考答案:只能张成 1 维或 2 维。
- 上界:张成空间是 的子空间,维数不超过 2。
- 下界:三个向量都非零,张成空间至少含有一个非零方向,维数至少为 1。
- 两种都能实现:三个向量共线时维数为 1;存在两个不共线的向量时维数为 2。
关键点是:维数不超过向量个数,也不超过所在空间的维数。三个向量在 中必然相关(3 > 2),但「相关」不影响它们能张成 2 维。
2. 若 是 、秩 2, 的维数是多少?
参考答案:用公式 。这里 、,所以
从消元看:三列中恰有两个主元列、一个自由列,自由变量只有一个,零空间由对应的一个特殊解张成。(顺带:。)
3. 已知 的零空间基与一个特解,如何确定 的每一列?
参考答案:把「列」当作未知量,两类信息各给出一组方程。
设 的列为 。
- 特解给出:,即列的一个已知组合等于 。
- 零空间基向量 给出:,即列之间的一条齐次线性关系。
共有 个向量方程,而未知列有 个;再结合秩为 的条件,通常足以唯一确定各列(若信息不足以定值,则只能确定列空间与列间关系)。本讲例题中,零空间基 直接说明 , 说明 ,特解再定出 ,于是整个 被还原。
4. 为什么 时行空间通常不等于列空间?举一个对称矩阵的例子说明何时相等。
参考答案:行向量与列向量是不同方向上的对象,行空间由行张成、列空间由列张成,二者一般不同。反例:
它的行空间是 方向(第二行非零),列空间是 方向(第一列非零),两者是不同的直线。
当 对称时(),行就是列,行空间与列空间自然相等。例如
两行与两列完全相同,行空间 = 列空间 = 。更一般地,只要 与 有相同的列空间(例如 对称),两者就相等。
5. 用一句话证明行空间与零空间的交集只有零向量。
参考答案:若 同时属于行空间与 ,则 说明 与 的每一行都正交,因而与行空间的每个向量正交——特别地与自己正交;于是 ,故 。
与前后课程的关系
这一讲把第 1–12 讲压成考点。下一讲从「交集只有零」推进到「互相垂直」,并引出正交补与最小二乘的核心方程 。
What this review covers
The first quiz covers the half of the course, with emphasis on rectangular matrices: null spaces, left null spaces, rank, and the four fundamental subspaces. When a matrix is invertible these notions are automatic; when it is rectangular, each one has to be checked.
Below the material is organized by problem type rather than lecture order.
Timeline
| Time | Topic |
|---|---|
| 0:09 | Opening: what this review covers |
| 1:49 | Type 1: the dimension of a span |
| 5:29 | Elimination to echelon form |
| 8:56 | A concrete example |
| 12:19 | Computing rank and free variables |
| 15:44 | Type 2: counting dimensions from structure |
| 17:26 | When the null space is large |
| 21:06 | Can a matrix square to zero |
| 24:39 | The counterexample and invertibility |
| 26:32 | Type 3: reconstructing a matrix from its solutions |
| 29:59 | The null space of a product |
| 33:27 | Particular solution plus null space |
| 36:54 | Type 4: true or false |
| 40:21 | The answer for an invertible matrix |
| 43:47 | The row space survives row operations |
| 45:33 | Type 6: row space meets null space |
Type 1: dimension of a span
Problem. Let be nonzero vectors in . What dimensions can their span have?
Answer. 1, 2, or 3.
- At most 3, since there are only three vectors.
- Not 0, since they are nonzero.
- All three occur: collinear gives dimension 1, coplanar but not collinear gives 2, and not coplanar gives 3.
The test is simply "dimension is at most the number of vectors" plus "nonzero rules out zero".
Type 2: counting dimensions from structure
Problem. is a echelon matrix of rank .
- What is ?
- What is the rank of ?
- What is the rank of ?
- What is ?
Solutions.
- With three columns and rank 3, every column is a pivot column. No free variables, so .
- The second block is twice the first and is eliminated. The rank remains 3.
- is . Block elimination gives ; each contributes 3 pivots, so the rank is 6.
- is with 10 columns and rank 6:
Only two facts are used: and block elimination preserves rank.
Type 3: reconstructing a matrix from its complete solution
Problem. For , one particular solution is , a basis of the null space is and , and . Find the rank, the column space, and itself.
Step 1: rank from the null-space dimension.
The null space has dimension 2 and the matrix is , so
Step 2: the first column from the particular solution.
is twice the first column:
Step 3: the remaining columns from the null-space vectors.
forces the third column to be zero.
gives
Therefore
Check: all three columns are multiples of , so the rank is 1 and the column space is the line through . The system is solvable exactly when is a multiple of .
Method. A complete solution carries two kinds of information: the particular solution relates one column to , and the null-space basis records linear relations among columns. Together they often determine the whole matrix.
Type 4: true or false
| Statement | Verdict | Reason |
|---|---|---|
| Square with implies | True | Full rank is preserved by transpose |
| The invertible matrices form a subspace | False | The zero matrix is missing; sums can be singular |
| implies | False | Counterexample |
| An matrix with independent columns is solvable for every | True | Independent columns in a square matrix mean invertible |
| When , the row space equals the column space | False | For the row space is the direction and the column space the direction |
| and have the same four subspaces | True | Negation preserves every linear relation |
| Same four subspaces implies is a multiple of | False | Any two invertible matrices have the same four subspaces |
| Swapping two rows leaves the row and null spaces unchanged | True | Row operations preserve the row space; solutions are unchanged |
| Swapping two rows leaves the column space unchanged | False | The column space generally changes |
Why a sum of invertible matrices can be singular
Take
which is singular. A simpler argument: the set of invertible matrices does not contain the zero matrix, and a subspace must contain zero, so the set cannot be a subspace at all.
Type 5: null space of a product
Problem. Let with invertible. How do and compare?
Answer. They are equal.
- If , then , so .
- If , multiply on the left by to get , so .
Left multiplication by an invertible matrix does not change the null space. This is why elimination, which is left multiplication by a product of elementary matrices, can be used to find null spaces.
Type 6: intersection of row and null spaces
Problem. Why can the vector never be both a row of and a member of ?
Suppose it were the first row and . Then the first component of would be
contradicting .
More generally,
If lies in both, it is orthogonal to every vector in the row space, in particular to itself. Then , so .
This is the doorway to the next lecture: the row space and null space are not only disjoint, they are perpendicular, with dimensions adding to . They are orthogonal complements.
Review points
- Dimension counts: and .
- Block elimination: count pivots contributed by each block.
- The complete solution can reveal the columns of .
- A subspace must contain zero and be closed under addition and scaling; invertible matrices satisfy neither.
- Left multiplication by an invertible matrix preserves the null space.
- The row space and null space share only the zero vector.
Check your understanding
1. What dimensions can three nonzero vectors span in , and why?
Reference answer: Only dimension 1 or dimension 2.
- Upper bound: the span is a subspace of , so its dimension is at most 2.
- Lower bound: all three vectors are nonzero, so the span contains a nonzero direction and has dimension at least 1.
- Both occur: collinear vectors span a line (dimension 1); two non-collinear vectors among them span the plane (dimension 2).
The key point: the dimension is bounded both by the number of vectors and by the ambient space. Three vectors in are necessarily dependent (), but dependence does not prevent them from spanning dimension 2.
2. If is with rank 2, what is ?
Reference answer: Use . Here and , so
In elimination terms: among the three columns exactly two are pivots and one is free, so there is a single free variable and the null space is spanned by one special solution. (For reference, .)
3. Given a null-space basis and one particular solution, how do you recover each column of ?
Reference answer: Treat the columns as unknowns; the two kinds of information supply equations.
Let the columns be .
- The particular solution gives : one known combination of the columns equals .
- Each null-space basis vector gives : a homogeneous linear relation among the columns.
That is vector equations for unknown columns; together with the rank condition they usually determine the columns uniquely (if not, only the column space and the relations are determined). In the lecture's example, the null vector forces , forces , and the particular solution then fixes , recovering all of .
4. Why is the row space usually different from the column space when ? When are they equal?
Reference answer: Rows and columns are objects in different directions; the row space is spanned by the rows and the column space by the columns, and in general they differ. Counterexample:
Its row space is the direction (second row), while its column space is the direction (first column).
They coincide when is symmetric (), since then rows and columns are the same vectors. For example
has identical rows and columns, so both spaces are . More generally, they coincide whenever has the same column space as .
5. Prove in one sentence that the row space and null space intersect only at zero.
Reference answer: If lies in both, then makes orthogonal to every row, hence to every vector of the row space — in particular to itself, so and therefore .
How this connects to the course
This lecture compresses Lectures 1–12 into exam technique. The next moves from "the intersection is only zero" to "they are perpendicular", introducing orthogonal complements and the central least-squares equation .