Here the argument is a more complex. There is an easy part, when the low frequency has a higher modulation localization. Indeed, using Bernstein, (?) and (?), we estimate as follows:
\begin{split}
\left\lvert \int_{ \mathbb{R}^{ 1+3 } } Q_{ \succeq k }u v \bar w
dx dt \right\rvert
&\lesssim \lVert Q_{ \succeq k } u \rVert_{ L^{ 2 }_{ t,x } }
\lVert \widetilde P_{ k } (v \bar w) \rVert_{ L_{ t,x }^{ 2 } }
\lesssim 2^{ -\frac{k}{2} } \lVert u \rVert _{ \dot X^{ 0,\frac{1}{2}
,\infty } }
\cdot 2^{ \frac{k}{2} } \lVert v \bar w \rVert _{ L_{ t }^{ 2 } L_{ x }^{
\frac{3}{2} } }\\
&\lesssim \lVert u \rVert _{ V_{ \Delta }^{ 2 } }
\lVert v \rVert _{ L_{ t }^{ 4 } L_{ x }^{ 3 } }
\lVert w \rVert _{ L_{ t }^{ 4 } L_{ x }^{ 3 } }
\lesssim \lVert u \rVert_{ V_{ \Delta }^{ 2 } }
\lVert v \rVert_{ V_{ \Delta }^{ 2 } }
\lVert w \rVert_{ V_{ \Delta }^{ 2 } }.
\end{split}
For modulations 2k \preceq j \preceq k we change the strategy. Fix such a j. We notice that
\int_{ \mathbb{R}^{ 1+3 } } Q_{ j }u \cdot v \bar w d xd t
= \sum_{q,q' \in \mathcal{C}_k} \int Q_{ j }u \cdot v_q \bar w_{q'} d xd t,
where the sum is restricted to the pairs with d(q,q') \lesssim 2^{k}; here we use the notation v_q = \chi_q(D) v for the component of v with the property that \hat v is localized in q, and similarly for w_{q'}. Then we fix such a pair (q,q'). We note that also have
\int Q_{ j }u \cdot v_q \bar w_{q'} d xd t = \sum_{I,I' \in \mathcal{I}_{j}} \int Q_{ j }u \cdot v_{q,I} \bar w_{q',I'},
where v_{q,I} = \chi_I(D_t) v_{q} is the component of v_q with the property that \hat v_q(\xi,\tau) has \tau localization in I. Since Q_j u is localized in the region |\tau| \lesssim 2^j, it suffices to restrict the summation over pairs of intervals with the property d(I,I') \lesssim 2^{j}. We remark that, for a free solution (e^{-it\Delta} f)_q=e^{-it\Delta} f_q, while the time frequency localization translates into (e^{-it\Delta} f_q)_I= e^{-it\Delta} \chi_I(\xi^2) f_{q}; essentially this produces a localization of \xi in S, the strip characterized by \xi^2 \in I. This is a subset of the parallelepiped of size C2^j \times C2^k \times C2^k, where the size of C2^j comes in the direction of \xi_0, where \xi_0 is the center of q.
Lemma. i) Assume that j \leq k \leq 0 and S is parallelepiped centered at \xi_0, |\xi_0| \approx 1 and of dimensions \approx 2^{ j }\times 2^{ k } \times 2^{ k }. If \hat f is supported in S, then
\begin{aligned}
\|e^{-it\Delta}f\|_{L_t^pL_x^q}
\lesssim
2^{\frac{2k+j}2 \left(1-\frac{2}{q}\right) - \frac{2k}{p}}
\|f\|_{L_x^2}
\end{aligned}
whenever 2<p\leq\infty, \quad 2\leq q\leq\infty,\quad \frac1p+\frac1q\leq\frac12. In particular, \|e^{-it\Delta}f\|_{L_{t,x}^4} \lesssim 2^{\frac{j}4}\|f\|_{L_x^2}.
ii) Assume u \in U^4_\Delta has similar localization properties. Then
\| u \|_{L^4_{t,x}} \lesssim 2^\frac{j}4 \|u\|_{U^4_\Delta}.
Assuming the above lemma, we conclude the proof of (5.2). For fixed pair (q,q') and I,I' we have
\begin{aligned}
\left\lvert \int_{ \mathbb{R}^{ 1+3 } } Q_{ j }u \cdot v_{q,I} \bar w_{q',I'} d xd t \right\rvert
&\lesssim \lVert Q_{ j }u \rVert_{ L_{t,x}^{2 }}
\lVert v_{q,I} \rVert_{ L_{ t,x }^{ 4 } } \lVert w_{q',I'} \rVert_{ L_{ t,x } ^{ 4 }(T) }\\
& \lesssim 2^{-\frac{j}2} \|u\|_{V^2_\Delta} 2^\frac{j}4 \|v_{q,I}\|_{U^4_\Delta} 2^\frac{j}4 \|w_{q',I'}\|_{U^4_\Delta} \\
& \lesssim \|u\|_{V^2_\Delta} \|v_{q,I}\|_{U^4_\Delta} \|w_{q',I'}\|_{U^4_\Delta}.
\end{aligned}
Then we estimate the global term using Lemma (3.4) as follows
\begin{aligned}
\left\lvert \int_{ \mathbb{R}^{ 1+3 } } Q_{ j }u \cdot v \bar w d xd t \right\rvert
& \lesssim \|u\|_{V^2_\Delta} \left(\sum_q \sum_I \|v_{q,I}\|^2_{U^4_\Delta} \right)^\frac12 \left( \sum_{q'} \sum_{I'} \|w_{q',I'}\|^2_{U^4_\Delta} \right)^\frac12 \\
& \lesssim \|u\|_{V^2_\Delta} \langle j \rangle^2 \|v\|_{V^2_\Delta} \|w\|_{V^2_\Delta}.
\end{aligned}
Strictly speaking, the result in Lemma (3.4), does not apply verbatim above, and requires some slight adjustments. First, the localization of \chi_q(D_x) \chi_I(D_t) is not entirely in spatial variables, as required by Lemma (3.4). However, recalling that we work with Q_{\preceq 2k} v and Q_{\preceq 2k} w, then, just as we did for free solutions, this translates into a parallelepiped localization solely on the spatial variables. Second, the statement Lemma (3.4) is made in terms of isotropic objects (cubes to be more precise), while here the localization is made on parallelepipeds. This is again not a problem, since the argument Lemma (3.4) provides the argument in this setup as well, and we made sure we picked up the correct factor \langle j \rangle since 2^j is the smallest scale that appears in the parallelepiped.
The summation over the range 2k \preceq j \preceq k of the above estimate, combined with (Proof), gives (5.2).
Thus we are left with establishing the claim in Lemma ?, which we prove below.