Past exam of the mathematics course of the University of Cambridge / 2012 / iii / Paper 57 / 4

Use the paper's mode normalization, effectively $M_p=1$, and set $C=ϵ+1-c_s^2$, $K=k_1+k_2+k_3$, $s_2=k_1{k}_2+k_1{k}_3+k_2{k}_3$, $s_3=k_1{k}_2{k}_3$. Write $⟨{\zeta }_1{\zeta }_2{\zeta }_3{⟩}_c=(2\pi {)}^3{\delta }^3(\sum {\mathbf{k}}_i)B$. There are two source consistency issues: the supplied interaction picture expansion makes an external-before-vertex Wick contraction contain $u_k({\tau }^{\prime })$, with positive exponential $e^{ik{c}_s{\tau }^{\prime }}$; it converges at $-\infty (1-i0)$, whereas the PDF prints $-\infty (1+i\epsilon )$. Also the stated positive interaction has the opposite overall bispectrum sign from the printed target. The following calculation identifies both issues directly.

For the stated mode expansion, the ordered contraction isA spatial derivative at the vertex supplies $i{p}_j$; the two differentiated legs consequently give $-{\mathbf{k}}_i\cdot {\mathbf{k}}_j$. There are two Wick contractions for each choice of the undifferentiated external leg. The extra $a({\tau }^{\prime })$ in the stipulated in-in formalism integral makes the vertex coefficient $a^2ϵC/c_s^2$, using $a^2=1/(H^2{\tau }^{\prime 2})$. This is consistent if the named $H_{\mathrm{int}}$ is the cosmic-time generator expressed as a function of $\tau$, since $dt=ad\tau$; if it were already the conformal-time generator, that extra $a$ would have to be removed.

For the spatial-gradient cubic curvature bispectrum, define the late-time scalar gradient vertex integralWith $\omega =c_{s}K$, the first two terms of the product obey $({\tau }^{\prime -2}-i\omega {\tau }^{\prime -1})e^{i\omega {\tau }^{\prime }}=-{\partial }_{{\tau }^{\prime }}(e^{i\omega {\tau }^{\prime }}/{\tau }^{\prime })$. The other two use ${\int }_{-\infty }^0{e}^{i\omega {\tau }^{\prime }}d{\tau }^{\prime }=-i/\omega$ and ${\int }_{-\infty }^0{\tau }^{\prime }{e}^{i\omega {\tau }^{\prime }}d{\tau }^{\prime }=1/{\omega }^2$, with the same vacuum prescription. The endpoint divergence $-1/{\tau }_f$ is real, whileIn particular the finite $-K$ term comes from the endpoint of the total derivative and must not be discarded along with the real divergence.

Let $S={\mathbf{k}}_1\cdot {\mathbf{k}}_2+{\mathbf{k}}_1\cdot {\mathbf{k}}_3+{\mathbf{k}}_2\cdot {\mathbf{k}}_3$. The ordered vertex expectation, after stripping the momentum delta, is $Z=-2S{H}^{4}CI/(64{ϵ}^2{c}_s^5{k}_1^3{k}_2^3{k}_3^3)$. Since $\mathrm{Re}(-2iZ)=2\mathrm{Im}Z$, the literal positive Hamiltonian givesThe printed positive-sign target is obtained by replacing the displayed interaction Hamiltonian by its negative, with the convergent contour above. This is also the expected sign if a positive cubic $\zeta (\partial \zeta {)}^2$ coefficient was intended as a Lagrangian term, by the interaction Hamiltonian relation proved in question 3(a). Under that explicit repair,Restoring the overall $(2\pi {)}^3{\delta }^3(\sum {\mathbf{k}}_i)$ reproduces all three cyclic terms in the requested expression. An equilateral bispectrum configuration is a concrete sign check: for $k_i=k$, $S=-3k^2/2$, $Q=-17k/9$, so $SQ=17k^3/6>0$. The literal interaction gives $B=-17H^{4}C/(96{ϵ}^2{c}_s^4{k}^6)$, while the printed target is positive. This cannot be repaired by merely changing the overall definition of $\zeta$, because its field expansion and interaction were specified together. The contour typo and Hamiltonian sign are therefore documented source repairs, rather than silently altered contractions.

Factor the momentum dependence intoWrite $s=+1$ for the intended target, obtained with the Hamiltonian sign repair in part (i), and $s=-1$ for the literal stated Hamiltonian. The same sign choice applies to both limits and cancels in their amplitude ratio. The primordial bispectrum contribution is $s{H}^4(ϵ+1-c_s^2)\mathcal{G}/(16{ϵ}^2{c}_s^4)$. HenceAt fixed $H,ϵ$ and momentum triangle, the ratio to the unit-sound-speed amplitude isKeeping only the leading slow-roll order makes this $100/c_s^4$. To compare the two pieces of this same interaction at a fixed $c_s$, their ratio is instead $(1-c_s^2)/ϵ\simeq 100(1-c_s^2)$. Thus the noncanonical piece overtakes the slow-roll piece once $1-c_s^2\gg 0.01$, while both have the common raw $c_s^{-4}$ enhancement from the interaction and its external mode normalization. For instance $c_s^2=0.1$ gives an exact ratio $9100$ to the canonical limit with these fixed parameters.

A raw primordial bispectrum amplitude is not the same as normalized primordial non-Gaussianity: here $P_{\zeta }(k)=H^2/(4ϵc_s{k}^3)$, so the ratio $B/P_{\zeta }^2$ scales as $(ϵ+1-c_s^2)/c_s^2$ at fixed shape. The small-sound-speed enhancement in normalized non-Gaussianity is therefore $c_s^{-2}$. These are tree-level weak-interaction predictions; approaching arbitrarily small $c_s$ eventually requires checking the perturbative regime.