diff --git a/docs/5.Logistic回归.md b/docs/5.Logistic回归.md
index 703e1e68..c60fa829 100644
--- a/docs/5.Logistic回归.md
+++ b/docs/5.Logistic回归.md
@@ -245,6 +245,10 @@ def gradAscent(dataMatIn, classLabels):
     return array(weights)
 ```
 
+大家看到这儿可能会有一些疑惑，就是，我们在迭代中更新我们的回归系数，后边的部分是怎么计算出来的？为什么会是 alpha * dataMatrix.transpose() * error ?因为这就是我们所求的梯度，也就是对 f(w) 对 w 求一阶导数。具体推导如下:
+
+![f(w)对w求一阶导数](../images/5.Logistic/LR_21.png)
+
 > 测试算法: 使用 Logistic 回归进行分类
 
 ```python
diff --git a/images/5.Logistic/LR_21.png b/images/5.Logistic/LR_21.png
new file mode 100644
index 00000000..94e410fa
Binary files /dev/null and b/images/5.Logistic/LR_21.png differ