第 21 章

特征值与特征向量:矩阵作用下方向不变的向量

从投影、置换、旋转与三角矩阵四类例子出发,用 det(A−λI)=0 求特征值,并说明实特征值、复特征值与特征向量短缺三种可能。

Eigenvalues and Eigenvectors: Directions the Matrix Leaves Alone

Four examples — projection, permutation, rotation, triangular — build intuition; then det(A−λI)=0 finds eigenvalues, and three failure modes appear: complex values, repeats, and too few eigenvectors.

首发于 2019.09.25中 / EN
系统课程linear-algebraeigenvalueseigenvectorsdeterminant
语言
字号
目录
篇内目录

特征向量在问什么

矩阵是作用在向量上的函数:输入 xx,输出 AxAx。大多数向量的方向会被改变。少数特殊向量的方向不变,只是长度被缩放——这些就是特征向量:

Ax=λxAx = \lambda x

λ\lambda 称为特征值。它可以是负数或零,甚至可以是复数;「方向不变」允许反向。

注意这个方程和 Ax=bAx=b 完全不同:λ\lambda 与 xx 都是未知量,而且乘在一起。消元法不能直接用,需要换一个思路。

零特征值的含义

若 λ=0\lambda=0,则 Ax=0Ax=0。所以零特征值存在,等价于 AA 奇异,对应的特征向量就是零空间中的向量。这样一来,「行列式为零」与「有零特征值」是同一件事。

时间轴

时间内容
0:07特征值与特征向量的总体介绍
3:10AxAx 与 xx 同向意味着什么
6:43判断一个向量是否为特征向量
9:43投影矩阵:零空间与列空间方向
12:50置换与反射矩阵的例子
15:57λ=−1\lambda=-1 的方向
19:15求法:解 det⁡(A−λI)=0\det(A-\lambda I)=0
22:25特征向量构成整条直线
25:322×2 的特征多项式
31:34平移 A+3IA+3I 的特征值加 3
34:40迹与行列式给出特征值的和与积
37:49旋转 90∘90^\circ 的复特征值
41:02复特征值成共轭对
47:05重根与退化情形
50:11退化矩阵的例子

四个例子,先建立直觉

投影矩阵

设 PP 把向量投影到一个平面。

  • 平面内的向量投影后不变:Px=xPx=x,所以 λ=1\lambda=1,整个平面都是特征向量。
  • 垂直于平面的向量投影为零:Px=0Px=0,所以 λ=0\lambda=0,对应的特征向量是平面的法向量。

投影矩阵的特征值只有 11 和 00,与「投影两次等于投影一次」完全吻合:P2=PP^2=P 迫使 λ2=λ\lambda^2=\lambda。

置换矩阵

A=[0110]A=\begin{bmatrix}0&1\\1&0\end{bmatrix}

它交换两个分量。

  • x=(1,1)x=(1,1):交换后不变,λ=1\lambda=1。
  • x=(−1,1)x=(-1,1):交换后变成 (1,−1)=−x(1,-1)=-x,λ=−1\lambda=-1。

两个特征向量垂直。这个矩阵是对称的,后面会看到这不是巧合。

旋转矩阵

逆时针旋转 90∘90^\circ:

Q=[0−110]Q=\begin{bmatrix}0&-1\\1&0\end{bmatrix}

从几何上看,任何非零向量旋转 90∘90^\circ 后都不可能保持方向,所以应当「没有实特征向量」。代数上:

det⁡(Q−λI)=∣−λ−11−λ∣=λ2+1\det(Q-\lambda I) = \begin{vmatrix}-\lambda&-1\\1&-\lambda\end{vmatrix} = \lambda^2+1

解得 λ=±i\lambda=\pm i。实矩阵可以有复特征值,而且成共轭对出现。QT=−QQ^T=-Q,是反对称矩阵,这类矩阵的特征值纯虚。

三角矩阵

A=[3103]A=\begin{bmatrix}3&1\\0&3\end{bmatrix}

特征值可以直接从对角线读出:λ1=λ2=3\lambda_1=\lambda_2=3。但求特征向量时:

A−3I=[0100]A-3I=\begin{bmatrix}0&1\\0&0\end{bmatrix}

零空间只有一维,基是 (1,0)(1,0)。只有一个线性无关的特征向量,无法张成整个平面。这是重复特征值带来的「特征向量短缺」。

求法:先求 λ,再求 x

关键变形

把 Ax=λxAx=\lambda x 移项:

(A−λI)x=0(A-\lambda I)x=0

要存在非零的 xx,矩阵 A−λIA-\lambda I 必须奇异,于是

det⁡(A−λI)=0\det(A-\lambda I)=0

这称为特征方程。它是关于 λ\lambda 的 nn 次多项式,有 nn 个根(计重数)。

二阶的便捷形式

对 2×22\times2 矩阵,

det⁡(A−λI)=λ2−(tr⁡A)λ+det⁡A\det(A-\lambda I)=\lambda^2-(\operatorname{tr}A)\lambda+\det A

其中迹是主对角线之和。因此:

λ1+λ2=tr⁡A,λ1λ2=det⁡A\lambda_1+\lambda_2=\operatorname{tr}A, \qquad \lambda_1\lambda_2=\det A

这两个关系对任意 nn 也成立:所有特征值之和等于迹,之积等于行列式。

完整例子

A=[3113]A=\begin{bmatrix}3&1\\1&3\end{bmatrix}

先算特征多项式:

det⁡(A−λI)=∣3−λ113−λ∣=(3−λ)2−1=λ2−6λ+8\det(A-\lambda I) = \begin{vmatrix}3-\lambda&1\\1&3-\lambda\end{vmatrix} = (3-\lambda)^2-1 = \lambda^2-6\lambda+8

因式分解:

(λ−4)(λ−2)=0(\lambda-4)(\lambda-2)=0

所以 λ1=4\lambda_1=4,λ2=2\lambda_2=2。验证:和 4+2=64+2=6 等于迹,积 88 等于行列式。

再求特征向量。

λ=4\lambda=4:

A−4I=[−111−1]A-4I=\begin{bmatrix}-1&1\\1&-1\end{bmatrix}

零空间由 (1,1)(1,1) 张成。

λ=2\lambda=2:

A−2I=[1111]A-2I=\begin{bmatrix}1&1\\1&1\end{bmatrix}

零空间由 (−1,1)(-1,1) 张成(也可取 (1,−1)(1,-1))。

两个特征向量垂直。矩阵是对称的,特征值都是实数。

三个结构性事实

加 cIcI 只平移特征值

若 Ax=λxAx=\lambda x,则

(A+cI)x=Ax+cx=(λ+c)x(A+cI)x=Ax+cx=(\lambda+c)x

所以特征向量完全不变,特征值整体加 cc。课堂上正是这样从置换矩阵 [0110]\begin{bmatrix}0&1\\1&0\end{bmatrix} 得到 [3113]\begin{bmatrix}3&1\\1&3\end{bmatrix}:特征值从 1,−11,-1 变成 4,24,2,特征向量仍是 (1,1)(1,1) 与 (−1,1)(-1,1)。

但 A+BA+B 没有这种性质

一个常见错误是:若 Ax=λxAx=\lambda x、By=αyBy=\alpha y,就断言 A+BA+B 的特征值是 λ+α\lambda+\alpha。这不成立,因为 xx 一般不是 BB 的特征向量。−x-x 只有在 BB 是 II 的倍数时才碰巧成立。

同理,ABAB 的特征值一般也不是特征值的乘积。

迹与行列式总是对的

无论矩阵是否对称、是否重复,下面的恒等式都成立:

∑iλi=tr⁡A,∏iλi=det⁡A\sum_i \lambda_i = \operatorname{tr}A, \qquad \prod_i \lambda_i = \det A

这也是判断计算是否出错的第一道检查。

三种「坏情况」

情况例子后果
复特征值旋转 90∘90^\circ没有实特征方向;特征值成共轭对
重复特征值[3103]\begin{bmatrix}3&1\\0&3\end{bmatrix}可能只有一维特征空间
特征向量不足同上无法对角化,下一讲处理

对称矩阵是「好情况」的代表:特征值全为实,特征向量两两正交。反对称矩阵是另一个极端:特征值纯虚。一般矩阵介于两者之间。

核心回顾

  • Ax=λxAx=\lambda x:矩阵作用后方向不变的向量。
  • 零特征值   ⟺  \iff 矩阵奇异。
  • 求法:解 det⁡(A−λI)=0\det(A-\lambda I)=0 得 λ\lambda,再解 (A−λI)x=0(A-\lambda I)x=0 得 xx。
  • 二阶:λ2−tr⁡(A)λ+det⁡A=0\lambda^2-\operatorname{tr}(A)\lambda+\det A=0。
  • 特征值之和等于迹,之积等于行列式。
  • 加 cIcI 平移特征值、保持特征向量;A+BA+B 一般没有这种关系。
  • 实矩阵可能有复特征值;重复特征值可能导致特征向量不足。

常见误区

  1. 把 Ax=λxAx=\lambda x 当成普通方程组用消元解。两个未知量相乘,必须先消去 xx。
  2. 认为特征向量唯一。特征向量张成一个空间,任何非零倍数都有效。
  3. 忘记零特征值的几何含义。它就是零空间。
  4. 对 A+BA+B 直接相加特征值。只有 B=cIB=cI 时才成立。
  5. 以为实矩阵的特征值一定是实数。旋转矩阵给出反例。

理解检验

1. 投影矩阵的特征值为什么只能是 0 和 1?

参考答案:投影满足 P2=PP^2=P。设 Px=λxPx=\lambda x,x≠0x\ne0。两边再作用一次 PP:

P2x=P(λx)=λPx=λ2x.P^2x=P(\lambda x)=\lambda Px=\lambda^2 x .

但 P2x=Px=λxP^2x=Px=\lambda x,所以 λ2x=λx\lambda^2x=\lambda x。由于 x≠0x\ne0,得 λ2=λ\lambda^2=\lambda,即 λ(λ−1)=0\lambda(\lambda-1)=0,故 λ=0\lambda=0 或 λ=1\lambda=1。几何上,投影到平面内的向量不动(λ=1\lambda=1),法向向量被压成零(λ=0\lambda=0)。

2. 用迹与行列式验证 [3113]\begin{bmatrix}3&1\\1&3\end{bmatrix} 的特征值。

参考答案:迹 =3+3=6=3+3=6,行列式 =9−1=8=9-1=8。特征多项式为

λ2−(tr⁡A)λ+det⁡A=λ2−6λ+8=(λ−2)(λ−4),\lambda^2-(\operatorname{tr}A)\lambda+\det A=\lambda^2-6\lambda+8=(\lambda-2)(\lambda-4),

所以 λ=2,4\lambda=2,4。检验:和 2+4=62+4=6 等于迹,积 2⋅4=82\cdot4=8 等于行列式,与正文一致。

3. 旋转 90∘90^\circ 的矩阵为什么没有实特征向量?

参考答案:几何上,QxQ\mathbf{x} 与 x\mathbf{x} 垂直且长度不变,一个非零向量不可能等于自己的实倍数,所以不存在实特征向量。代数上,

det⁡(Q−λI)=∣−λ−11−λ∣=λ2+1,\det(Q-\lambda I)=\begin{vmatrix}-\lambda&-1\\1&-\lambda\end{vmatrix}=\lambda^2+1,

方程 λ2+1=0\lambda^2+1=0 在实数中没有解,只有 λ=±i\lambda=\pm i。QQ 是反对称矩阵(QT=−QQ^T=-Q),其特征值必为纯虚数或零。

4. 一个 3×33\times3 矩阵有特征值 2,2,52,2,5,它的迹与行列式各是多少?

参考答案:迹等于特征值之和:2+2+5=92+2+5=9。行列式等于特征值之积(计重数):2⋅2⋅5=202\cdot2\cdot5=20。这两个恒等式对任意矩阵都成立,是检查计算的第一道关口。

5. 为什么 [3103]\begin{bmatrix}3&1\\0&3\end{bmatrix} 只有一维特征空间?

参考答案:特征值是 λ=3\lambda=3(二重)。求特征向量解 (A−3I)x=0(A-3I)x=0:

A−3I=[0100],A-3I=\begin{bmatrix}0&1\\0&0\end{bmatrix},

方程只有 x2=0x_2=0,x1x_1 任意,所以零空间是 span⁡{(1,0)}\operatorname{span}\{(1,0)\},维数为 1。特征值的代数重数是 2,但几何重数只有 1,缺少一个线性无关的特征向量,因此这个矩阵不能对角化。

与前后课程的关系

这一讲定义特征值与特征向量,并给出求法。下一讲讨论对角化:当矩阵有足够多的线性无关特征向量时,可以写成 A=SΛS−1A=S\Lambda S^{-1},从而快速计算幂与理解长期行为。重复特征值导致的短缺,将由 Jordan 形式处理。

What an eigenvector asks

A matrix is a function on vectors: input xx, output AxAx. Most vectors change direction. A few keep their direction and only change length — those are the eigenvectors:

Ax=λxAx = \lambda x

The scalar λ\lambda is the eigenvalue. It may be negative or zero, or even complex; "same direction" allows reversal.

This equation is not Ax=bAx=b: λ\lambda and xx are both unknown, and they multiply. Elimination does not apply directly, so a different idea is needed.

What a zero eigenvalue means

If λ=0\lambda=0, then Ax=0Ax=0. A zero eigenvalue therefore exists exactly when AA is singular, and its eigenvectors are the null space. "Determinant zero" and "a zero eigenvalue" are the same statement.

Timeline

TimeTopic
0:07The big picture: eigenvalues and eigenvectors
3:10What it means for AxAx to be parallel to xx
6:43Testing whether a vector is an eigenvector
9:43Projection: nullspace and column-space directions
12:50Permutation and reflection examples
15:57A direction with λ=−1\lambda=-1
19:15The method: solve det⁡(A−λI)=0\det(A-\lambda I)=0
22:25Eigenvectors form whole lines
25:32The 2×2 characteristic polynomial
31:34Shifting by 3I3I adds 3 to the eigenvalues
34:40Trace and determinant give the sum and product
37:49A 90∘90^\circ rotation has complex eigenvalues
41:02Complex eigenvalues come in conjugate pairs
47:05Repeated eigenvalues and degenerate cases
50:11An example of a degenerate matrix

Four examples for intuition

Projection matrix

Let PP project onto a plane.

  • A vector already in the plane is unchanged: Px=xPx=x, so λ=1\lambda=1, and the whole plane consists of eigenvectors.
  • A vector perpendicular to the plane projects to zero: Px=0Px=0, so λ=0\lambda=0, and the normal direction supplies eigenvectors.

The eigenvalues are only 11 and 00, consistent with P2=PP^2=P, which forces λ2=λ\lambda^2=\lambda.

Permutation matrix

A=[0110]A=\begin{bmatrix}0&1\\1&0\end{bmatrix}

It swaps the two components.

  • x=(1,1)x=(1,1): unchanged, so λ=1\lambda=1.
  • x=(−1,1)x=(-1,1): becomes (1,−1)=−x(1,-1)=-x, so λ=−1\lambda=-1.

The two eigenvectors are perpendicular. The matrix is symmetric, and that is not a coincidence.

Rotation matrix

Rotate counterclockwise by 90∘90^\circ:

Q=[0−110]Q=\begin{bmatrix}0&-1\\1&0\end{bmatrix}

Geometrically, no nonzero vector can keep its direction after a quarter turn, so there should be no real eigenvector. Algebraically,

det⁡(Q−λI)=∣−λ−11−λ∣=λ2+1\det(Q-\lambda I) = \begin{vmatrix}-\lambda&-1\\1&-\lambda\end{vmatrix} = \lambda^2+1

so λ=±i\lambda=\pm i. A real matrix can have complex eigenvalues, and they occur in conjugate pairs. Here QT=−QQ^T=-Q: the matrix is antisymmetric, and such matrices have purely imaginary eigenvalues.

Triangular matrix

A=[3103]A=\begin{bmatrix}3&1\\0&3\end{bmatrix}

The eigenvalues can be read directly off the diagonal: λ1=λ2=3\lambda_1=\lambda_2=3. But for eigenvectors,

A−3I=[0100]A-3I=\begin{bmatrix}0&1\\0&0\end{bmatrix}

whose null space is only one-dimensional, spanned by (1,0)(1,0). There is only one independent eigenvector, not enough to span the plane. A repeated eigenvalue can create a shortage.

The method: find λ\lambda first, then xx

The key rearrangement

Move everything to one side:

(A−λI)x=0(A-\lambda I)x=0

For a nonzero xx to exist, A−λIA-\lambda I must be singular, so

det⁡(A−λI)=0\det(A-\lambda I)=0

This is the characteristic equation, a polynomial of degree nn in λ\lambda with nn roots counted with multiplicity.

The convenient 2×22 \times 2 form

For a 2×22\times2 matrix,

det⁡(A−λI)=λ2−(tr⁡A)λ+det⁡A\det(A-\lambda I)=\lambda^2-(\operatorname{tr}A)\lambda+\det A

Hence

λ1+λ2=tr⁡A,λ1λ2=det⁡A\lambda_1+\lambda_2=\operatorname{tr}A, \qquad \lambda_1\lambda_2=\det A

Both identities hold for every nn: the eigenvalues sum to the trace and multiply to the determinant.

A full example

A=[3113]A=\begin{bmatrix}3&1\\1&3\end{bmatrix}

The characteristic polynomial:

det⁡(A−λI)=∣3−λ113−λ∣=(3−λ)2−1=λ2−6λ+8\det(A-\lambda I) = \begin{vmatrix}3-\lambda&1\\1&3-\lambda\end{vmatrix} = (3-\lambda)^2-1 = \lambda^2-6\lambda+8

Factor:

(λ−4)(λ−2)=0(\lambda-4)(\lambda-2)=0

So λ1=4\lambda_1=4 and λ2=2\lambda_2=2. Check: the sum 4+2=64+2=6 is the trace, and the product 88 is the determinant.

Now the eigenvectors.

For λ=4\lambda=4:

A−4I=[−111−1]A-4I=\begin{bmatrix}-1&1\\1&-1\end{bmatrix}

Its null space is spanned by (1,1)(1,1).

For λ=2\lambda=2:

A−2I=[1111]A-2I=\begin{bmatrix}1&1\\1&1\end{bmatrix}

Its null space is spanned by (−1,1)(-1,1) (or (1,−1)(1,-1)).

The two eigenvectors are perpendicular. The matrix is symmetric, and both eigenvalues are real.

Three structural facts

Adding cIcI shifts eigenvalues

If Ax=λxAx=\lambda x, then

(A+cI)x=Ax+cx=(λ+c)x(A+cI)x=Ax+cx=(\lambda+c)x

The eigenvectors are unchanged and every eigenvalue increases by cc. In the lecture, adding 3I3I to the permutation matrix [0110]\begin{bmatrix}0&1\\1&0\end{bmatrix} produces [3113]\begin{bmatrix}3&1\\1&3\end{bmatrix}: the eigenvalues move from 1,−11,-1 to 4,24,2, while the eigenvectors stay (1,1)(1,1) and (−1,1)(-1,1).

But A+BA+B does not work that way

A common mistake: given Ax=λxAx=\lambda x and By=αyBy=\alpha y, conclude that A+BA+B has eigenvalue λ+α\lambda+\alpha. This is false, because xx is generally not an eigenvector of BB. It works only when BB is a multiple of II.

Likewise, the eigenvalues of ABAB are generally not the products of the eigenvalues.

The trace and determinant identities always hold

Whatever the matrix, symmetric or not, repeated eigenvalues or not:

∑iλi=tr⁡A,∏iλi=det⁡A\sum_i \lambda_i = \operatorname{tr}A, \qquad \prod_i \lambda_i = \det A

These are the first check on any eigenvalue computation.

Three "bad cases"

CaseExampleConsequence
Complex eigenvalues90∘90^\circ rotationNo real eigendirection; values come in conjugate pairs
Repeated eigenvalues[3103]\begin{bmatrix}3&1\\0&3\end{bmatrix}The eigenspace may be one-dimensional
Too few eigenvectorsSame exampleCannot be diagonalized; the next lecture addresses it

Symmetric matrices are the good case: real eigenvalues and orthogonal eigenvectors. Antisymmetric matrices are the opposite extreme: purely imaginary eigenvalues. General matrices sit in between.

Core review

  • Ax=λxAx=\lambda x: vectors whose direction survives the matrix.
  • A zero eigenvalue is equivalent to singularity.
  • Method: solve det⁡(A−λI)=0\det(A-\lambda I)=0 for λ\lambda, then (A−λI)x=0(A-\lambda I)x=0 for xx.
  • For 2×22\times2: λ2−tr⁡(A)λ+det⁡A=0\lambda^2-\operatorname{tr}(A)\lambda+\det A=0.
  • Eigenvalues sum to the trace and multiply to the determinant.
  • Adding cIcI shifts eigenvalues and preserves eigenvectors; A+BA+B generally does not.
  • Real matrices can have complex eigenvalues; repeated eigenvalues can leave too few eigenvectors.

Common misconceptions

  1. Treating Ax=λxAx=\lambda x as an ordinary system to eliminate. Two unknowns multiply; eliminate xx first.
  2. Expecting a unique eigenvector. Eigenvectors fill a space; any nonzero multiple works.
  3. Missing the geometry of a zero eigenvalue. It is the null space.
  4. Adding eigenvalues of AA and BB. Only valid when one matrix is a multiple of II.
  5. Assuming a real matrix has real eigenvalues. The rotation matrix is the counterexample.

Check your understanding

1. Why can a projection matrix only have eigenvalues 0 and 1?

Reference answer: A projection satisfies P2=PP^2=P. Suppose Px=λxPx=\lambda x with x≠0x\ne0. Apply PP again:

P2x=P(λx)=λPx=λ2x.P^2x=P(\lambda x)=\lambda Px=\lambda^2 x .

But P2x=Px=λxP^2x=Px=\lambda x, so λ2x=λx\lambda^2x=\lambda x. Since x≠0x\ne0, λ2=λ\lambda^2=\lambda, i.e. λ(λ−1)=0\lambda(\lambda-1)=0, hence λ=0\lambda=0 or λ=1\lambda=1. Geometrically, vectors in the target plane stay put (λ=1\lambda=1) and the normal direction collapses to zero (λ=0\lambda=0).

2. Use the trace and determinant to verify the eigenvalues of [3113]\begin{bmatrix}3&1\\1&3\end{bmatrix}.

Reference answer: Trace =3+3=6=3+3=6 and determinant =9−1=8=9-1=8. The characteristic polynomial is

λ2−(tr⁡A)λ+det⁡A=λ2−6λ+8=(λ−2)(λ−4),\lambda^2-(\operatorname{tr}A)\lambda+\det A=\lambda^2-6\lambda+8=(\lambda-2)(\lambda-4),

so λ=2,4\lambda=2,4. Check: the sum 2+4=62+4=6 equals the trace and the product 2⋅4=82\cdot4=8 equals the determinant, matching the text.

3. Why does a 90∘90^\circ rotation have no real eigenvector?

Reference answer: Geometrically, QxQ\mathbf{x} is perpendicular to x\mathbf{x} and of the same length, so a nonzero vector can never be a real multiple of its image. Algebraically,

det⁡(Q−λI)=∣−λ−11−λ∣=λ2+1,\det(Q-\lambda I)=\begin{vmatrix}-\lambda&-1\\1&-\lambda\end{vmatrix}=\lambda^2+1,

and λ2+1=0\lambda^2+1=0 has no real solution, only λ=±i\lambda=\pm i. Since QQ is antisymmetric (QT=−QQ^T=-Q), its eigenvalues must be purely imaginary or zero.

4. A 3×33\times3 matrix has eigenvalues 2,2,52,2,5. What are its trace and determinant?

Reference answer: Trace equals the sum of the eigenvalues: 2+2+5=92+2+5=9. Determinant equals their product counted with multiplicity: 2⋅2⋅5=202\cdot2\cdot5=20. Both identities hold for every matrix and are the first check when computing.

5. Why does [3103]\begin{bmatrix}3&1\\0&3\end{bmatrix} have only a one-dimensional eigenspace?

Reference answer: The only eigenvalue is λ=3\lambda=3 (double). Solve (A−3I)x=0(A-3I)x=0:

A−3I=[0100],A-3I=\begin{bmatrix}0&1\\0&0\end{bmatrix},

whose only equation is x2=0x_2=0 with x1x_1 free, so the nullspace is span⁡{(1,0)}\operatorname{span}\{(1,0)\}, of dimension 1. The algebraic multiplicity is 2 but the geometric multiplicity is 1; one independent eigenvector is missing, so the matrix is not diagonalizable.

How this connects

This lecture defines eigenvalues and eigenvectors and shows how to compute them. The next discusses diagonalization: when there are enough independent eigenvectors, A=SΛS−1A=S\Lambda S^{-1} makes powers and long-term behavior easy to read. The shortage caused by repeated eigenvalues leads to Jordan form.

来源