The inversion sequence records the sign change of a word
Let \(\omega=(i_1,\ldots,i_n)\) be any word, let \(w=w(\omega)\), and let \(c_\omega(t)\) count the occurrences of \(t\) in its left inversion sequence. For \(t\in W\) and \(\varepsilon\in\mathbb Z/2\mathbb Z\), \[ A_\omega(w^{-1}tw,\varepsilon) =\bigl(t,\varepsilon+c_\omega(t)\bmod2\bigr). \]
Proof
Step 1. The empty word
For the empty word, \(w=1\), the action is the identity, and the inversion sequence is empty. Both sides are \((t,\varepsilon)\).
Step 2. Append one letter
Suppose the formula holds for \(\omega\), and append a simple generator \(s=s_i\). The new word product is \(ws\). Its input label is \[ (ws)^{-1}t(ws)=s\,w^{-1}tw\,s. \] Applying \(A_i\) conjugates this label to \(w^{-1}tw\) and adds one to the bit exactly when \[ s\,w^{-1}tw\,s=s. \] Multiplying this equality on both sides by \(s\) shows that it is equivalent to \(w^{-1}tw=s\), or \(t=wsw^{-1}\).
Step 3. Apply the induction hypothesis
The remaining action is \(A_\omega\). The induction hypothesis, with its initial bit increased by the indicator from Step 2, gives the output \[ \left(t,\varepsilon+\mathbf1_{t=wsw^{-1}}+c_\omega(t)\bmod2\right). \]
The inversion sequence of the appended word is the old sequence followed by \(wsw^{-1}\). Its multiplicity of \(t\) is therefore \[ c_{\omega i}(t)=c_\omega(t)+\mathbf1_{t=wsw^{-1}}. \] This identifies the output with the required formula for \(\omega i\). Every finite word is obtained by successive right appends, so induction proves the theorem.
References
The local declarations below give the construction and its word calculations.
Lean formalization
OAI.KLInvariance.CoxeterSupport.wordAction_record: Coxeter.lean:60.