Is $b$ in Support Vector Regression constant or not?

In Support Vector Regression, the following dual optimization problem is solved with respect to $\alpha_i$ and $\alpha_i^*$:

\begin{aligned} & \text{maximize} && \begin{cases} -\frac{1}{2}\sum_{i,j = 1}^{j}(\alpha_j - \alpha_j^*)x_i,x_j \\ -\epsilon \sum_{i = 1}^{l}(\alpha_i + \alpha_i^*) + \sum_{i = 1}^{l} y_i (\alpha_i - \alpha_i^*) \end{cases} \\ & \text{subject to} && \begin{cases} \sum_{1}^{l} (\alpha_i - \alpha_i^*) = 0\\ \alpha_i, \alpha_i^* \in [0, C] \end{cases} \end{aligned} \label{dual_lin_opt}

In case of a linear SVR, the final form of the approximated function is

$$f(x) = \sum_{i = 1}^{l}(\alpha_i - \alpha_i^*)x_ix + b. \label{sv_exp_lin}$$

I want to understand how $b$ is derived in this context. As the problem of SVR is usually stated, it seems that $b$ is a constant term but in papers that explain the derivation, $b$ seems to be different for every observation. For instance, in Smola (2003) $b$ is

$$\max\{-\epsilon + y_i - \omega x_i | \alpha_i < C \text{ or } \alpha_i^* > 0 \} \le b \le \min\{-\epsilon + y_i - \omega x_i | \alpha_i > 0 \text{ or } \alpha_i^* < C\}$$

And it is stated that the inquelities become equalities for some $\alpha_i$ or $\alpha_i^*$ $\in (0, C)$. But that means $b$ differs for every observation because it depends on the index $i$ of the observation. However, I have never seen the formula for SVR with $b_i$. So, is $b$ constant or not?

$b$ is defined as being a single constant for the model. If it needed to be defined for every training point, then you'd need some separate way to pick a $b$ for the test point.
The bounds you gave for $b$ are $$\max_i\{-\epsilon + y_i - \omega x_i | \alpha_i < C \text{ or } \alpha_i^* > 0 \} \le b \le \min_i\{-\epsilon + y_i - \omega x_i | \alpha_i > 0 \text{ or } \alpha_i^* < C\} ,$$ otherwise what would the max and min be over? Since both bounds are tight for certain $i$, that constrains $b$ to be a single value.
• I didn't recognize that it is the maximum and minimum over all $i$s. I thought it corresponds to the $\alpha_i$ or $\alpha_i^*$ that min./max. the value of the set. Can you recommend a paper or book that goes into more detail of the derivation of $b$ and proves your final statement? – random_guy Oct 17 '16 at 11:16
• Yeah, $b$ fitting is done after the $\alpha$s are fit. I believe Scholkopf and Smola's book Learning with Kernels has more details here, but my copy is currently across an ocean and I can't easily check. If you follow citations from the tutorial you linked you might find a derivation; alternatively, depending on your C++ skills the libsvm source is reasonably understandable if you just want to check that a good implementation of the algorithm does it that way. – Dougal Oct 17 '16 at 11:28