1: \begin{abstract}

2: The paper investigates the least-squares projection method for bounded

3: linear operators, which provides a natural regularization scheme by

4: projection for many ill-posed problems. Yet, without additional assumptions,

5: the convergence of this approximation scheme cannot be guaranteed. We reveal

6: that the convergence of least-squares projection method is determined by two

7: mutually independent factors -- the \textquotedblleft kernel

8: approximability\textquotedblright\ and the \textquotedblleft offset

9: angle\textquotedblright . The kernel approximability is a necessary

10: condition of convergence described with kernel $\mathcal{N}\left( T\right) $

11: and its subspaces $\mathcal{N}\left( T\right) \cap X_{n}$, and we give

12: several equivalent characterizations for it (Theorem \ref{the1.0}). The

13: offset angle of $X_{n}$ is defined as the largest canonical angle between

14: space $T^{\ast }T(X_{n})$ and $T^{\dagger }T(X_{n})$ (which are subspaces of

15: $\mathcal{N}\left( T\right) ^{\bot }$), and it geometrically reflects the

16: rate of convergence (Theorem \ref{the1.1}). The paper also presents new

17: observations\ for the unconvergence examples of Seidman \cite[Example 3.1]%

18: {seidman1980nonconvergence} and Du \cite[Example 2.10]{du2008finite} under

19: the notions of kernel approximability and offset angle.

20: \end{abstract}