SRQ Seminar – 200181205 – Chandra
Stochastic quantisation of Yang–Mills.
(joint work with Hairer and Shen)
See also previous work of Shen on \(U (1) +\)Higgs in \(2 d\) finite lattice, arXiv 2018.
Setting and notations.
\(N > 1\). Lie group \(G =\operatorname{SU} (N)\). Lie algebra \(\mathfrak{g}=\mathfrak{s}\mathfrak{u} (N)\), skew–adjoint traceless \(N \times N\) matrices.
Introduce a \(\mathfrak{g}\)-valued connection on \(\mathbb{T}^n\) with \(n = 2, 3\):
From the connection we construct a covariant exterior derivative
with
where \([,]\) is a wedge product in the forms and a Lie bracket in the \(\mathfrak{g}\) components.
Inner product.
For \(a, b \in \mathfrak{g}\) let \(\langle a, b \rangle =\operatorname{Tr} (a^{\ast} b) = -\operatorname{Tr} (a b)\). This product makes sense ofr \(\mathfrak{g}^n\) valued functions in space or space–time by taking the \(L^2\) contraction in the spatial variables. And gives also an inner product on differential forms \(\Omega^r [\mathfrak{g}]\) which allows to defined the adjoint
In general we do not have \(\mathrm{D}_A^2 \neq 0\). A simple calculation gives
with \(F_A = \mathrm{D}_A A - \frac{1}{2} [A, A] \in \Omega^2 [\mathfrak{g}]\).
Yang–Mills measure.
where \(\mathcal{D}A\) is “Lebesgue measure” on connections. The main problem is that this measure, even formally, is invariant under an infinite dimensional group of transformations. Even after “gauge fixing” (choosing a representative for each orbit), this representative is quite rough and we do not expect to make easily sense of the non-linear expressions involved in the exponent \(\langle F_A, F_A \rangle\). So we have a problem with gauge freedom and a problem of renormalization of small scale singularities.
Gauge transformations.
Take \(g : \mathbb{T}^n \rightarrow G\) (e.g. \(C^1\)). \(g\) acts on connection as
We have covariance under this action for forms:
and same for \(\mathrm{D}^{\ast}\).
Stochastic quantisation.
Consider
on \(\mathbb{R}^M\) and a choice of \(\langle \cdot, \cdot \rangle\) on \(\mathbb{R}^M\). An equation which leaves this measure invariant is
where \(\eta\) is a \(N\) vector of white noises, defined in terms of \(\langle \cdot, \cdot \rangle\).
Yang–Milles SQ.
Using the Yang–Mills measure and out choice of inner product we obtain the equation
where \(\eta\) is a \(\mathfrak{g}^n\)–valued space–time white noise, i.e. \(\mathbb{E} [\langle \eta, f \rangle \langle \eta, g \rangle] = \langle f, g \rangle\). This will be called YM stochastic gradient flow YMSGE.
If we write \(B = g \circ A\) where \(g\) is a time–dependent gauge transformation. Then
Note that \(\mathrm{I} = 0\) if \(\partial_t g = 0\). If \(A\) satisfy the gradient flow we have
and we observe that \(g (\eta)\) is equal in distribution to \(\eta\). So \(B\) satisfy in law the same equation, provided the gauge transformation is constant in time.
Issue: we cannot make sense of the YMSGE using the theory of singular SPDEs, because the equation is not parabolic.
We have to modify the equation in order to gain parabolicity and consider
with the additional term \(- \mathrm{D}_A \mathrm{D}^{\ast}_A A\). YMSHE=Yang–Mills stochastic heat equation.
with \(i, j = 1, \ldots, N\).
If \(A\) solves YMSHE, what does \(B = g \circ A\) solves?
with
If we enforce a gauge transformation such that \((\mathrm{I}) = (\mathrm{I} \mathrm{I})\) then we recover the same dynamics (De–Turck trick). So we consider the system of equations for \((g, A)\)
then \(B = g \circ A\) satisfies
If \(g\) is adapted one can hope that \(g (\eta)\) has still the same law as \(\eta\) by using Ito integral (this requires a suitable regularization).
Goal post
We want a Markov process on connections such that the following holds
At least locally in time.
Strategy: mollify to get
\(\displaystyle \left\{ \begin{array}{l} \partial_t A_{\varepsilon} = - \mathrm{D}_{A_{\varepsilon}}^{\ast} F_{A_{\varepsilon}} - \mathrm{D}_{A_{\varepsilon}} \mathrm{D}^{\ast}_{A_{\varepsilon}} A + J_{\varepsilon} \eta\\ g_{\varepsilon}^{- 1} \partial_t g_{\varepsilon} = - \mathrm{D}_{A_{\varepsilon}}^{\ast} (g_{\varepsilon} \mathrm{d} g_{\varepsilon}^{- 1}) \end{array} \right. \) | (1) |
where \(J_{\varepsilon}\) is a non-anticipative mollifier. Thanks to this \(g_{\varepsilon}\) is adapted. Now after tranformation we have
\(\displaystyle \partial_t B_{\varepsilon} = - \mathrm{D}_{B_{\varepsilon}}^{\ast} F_{B_{\varepsilon}} - \mathrm{D}_{B_{\varepsilon}} \mathrm{D}^{\ast}_{B_{\varepsilon}} B_{\varepsilon} + g_{\varepsilon} (J_{\varepsilon} \eta) . \) | (2) |
Introduce a new equation
\(\displaystyle \partial_t \tilde{B}_{\varepsilon} = - \mathrm{D}_{\tilde{B}_{\varepsilon}}^{\ast} F_{\tilde{B}_{\varepsilon}} - \mathrm{D}_{\tilde{B}_{\varepsilon}} \mathrm{D}^{\ast}_{\tilde{B}_{\varepsilon}} \tilde{B}_{\varepsilon} + g_{\varepsilon} (J_{\varepsilon} g_{\varepsilon}^{- 1} (\eta)) . \) | (3) |
By Ito isometry eq. (3) is equal in distribution to eq. (2) for \(\varepsilon > 0\). But using the theory of regularity structures we can show that eq. (2) is equal in distribution to eq. (1) because we can see explicitly the cancellations in \(g_{\varepsilon} (J_{\varepsilon} g_{\varepsilon}^{- 1} (\eta))\) as the mollifier is removed. This comes from the BPHZ solution theory.
BPHZ theorem for regularity structures.
In this context we have an abstract equation and an approximation of the noise give a corresponding BPHZ solution where there is continuity in the law of the noise wrt. these approximations.
Renormalization.
We have to be sure that renormalization do not spoil the gauge covariance. We have to add the following counterterms
and
in order to be able to show the limits \(\varepsilon \rightarrow 0\) exists. But at this point we do not know that \(g \circ A \xequal{d} B\) in the limit. We would need \(\beta_{\varepsilon} = \alpha_{\varepsilon}\) since we have the correspondence
But we do not need so much, indeed we need only that along some subsequence of \(\varepsilon\)'s we have
where \(\theta\) can be a finite number. Then this means we just pick the wrong abstract equation and consider instead the system
If we do not have even this subsequence, then we need to play with the intensity of the noise by using the equation
and we set \(\theta_{\varepsilon} = \alpha_{\varepsilon} - \beta_{\varepsilon}\) and we can choose \(\sigma_{\varepsilon} = (| \theta_{\varepsilon} |^{1 / 2})^{- 1}\) and looking at the limit we generate a contradiction since the limit is not gauge invariant.