I was reading the treatment of the large $N$ limit of the Non-Linear Sigma Model (NLSM) in Peskin & Schroeder, Sec. 13.3, and I noticed something strange in the evaluation of the path-integral by steepest descent.
P&S are evaluating the NLSM euclidean partition function: $$Z_E=\intop \mathcal D n \,\delta [n^2 -1]\exp [-N\intop \text d ^d x \frac{1}{2\tau_0}(\partial _\mu n)^2],$$ where $n$ is an $N$-component vector subject to the constraint $n^2=1$, represented here by a functional $\delta$, and $\tau_0$ is the bare 't Hooft coupling, which is held fixed in the $N\to \infty$ limit.
Next, they get rid of $\delta [n^2-1]$ by integrating-in an auxiliary field $\alpha$: $$ Z_E=\intop \mathcal D n \mathcal D \alpha \,\exp [-N\intop \text d ^d x \lbrace \frac{1}{2\tau_0}(\partial _\mu n)^2-\frac{i}{2\tau_0}\alpha(n^2-1)\rbrace ].$$
The "$i$" in front of $\alpha(n^2-1)$ correctly reproduces a $\delta$-function, as also Ron Maimon remarked [here][1].
However, as they go on in doing the large $N$ limit by the steepest descent method (I omit for short the passages, see Eqs. 13.114 through 13.115 in P&S), they arrive to the following equation for the saddle point in the $\alpha$ integration: $$\langle x\vert \frac{1}{-\partial ^2 +i\alpha}\vert x \rangle = \frac{1}{\tau_0}.$$ They argue that such equation requires $\alpha = -im^2$, where $m^2$ is a positive real number given by: $$\frac{1}{2\pi} \ln \frac{\Lambda}{m} = \frac{1}{\tau_0}$$
($\Lambda$ is a cut-off).
Now I am confused. The original $\alpha$ was supposedly a real scalar field, but the steepest descent method requires it to be pure imaginary.
So my **question**: does this mean that
1. The auxiliary field $\alpha$ was really imaginary from the beginning, or
2. The auxiliary field is real, but the expansion is done around a non-physical value of $\alpha$.
3. Neither of these.
Either possibilities (1) or (2) sound equally phony to me, so any suggestion is appreciated.
[1]: https://physics.stackexchange.com/questions/28700/auxiliary-field-and-loop-expansion