The reversed R-kernel with its bottom term removed
Let \(x\le y\) and \(t\le y\). For comparable pairs write \[ \overline R_{x,z}=\operatorname{reflect}_{d(x,z)}R_{x,z}; \] set this value to zero outside Bruhat order. Then \[ \sum_{x<z\le y}\overline R_{x,z}R_{z,t} =\delta_{x,t}-R_{x,t}. \] Here \(\delta_{x,t}\) is one if \(x=t\) and zero otherwise.
Proof
Step 1. Extend the kernel sum to the common lower ideal
The reversed R-kernel identity gives \[ \sum_{z\le t}\overline R_{x,z}R_{z,t}=\delta_{x,t}. \] Since \(t\le y\), every \(z\le t\) also belongs to \(L(y)\). For an extra index \(z\le y\) with \(z\not\le t\), the factor \(R_{z,t}\) is zero.
Thus adding those zero terms gives \[ \sum_{z\le y}\overline R_{x,z}R_{z,t}=\delta_{x,t}. \] This places every value of \(t\le y\) on the same finite index set.
Step 2. Restrict to indices above x
If \(x\not\le z\), then \(R_{x,z}=0\), and its reflected polynomial is zero. Therefore the terms outside \(x\le z\) vanish.
The full lower-ideal sum is consequently the sum over \(x\le z\le y\).
Step 3. Remove the bottom contribution
Because \(x\le y\), the index \(z=x\) occurs. Its value is \[ \overline R_{x,x}R_{x,t} =\operatorname{reflect}_0(1)R_{x,t} =R_{x,t}. \] Subtracting it from Step 1 leaves exactly the sum over \(x<z\le y\), proving the formula.
If \(t=x\), its right side is \(1-R_{x,x}=0\). If \(x<t\), it is \(-R_{x,t}\). If \(x\not\le t\), it is zero by support. These cases account for all terms when the identity is used inside a sum over \(t\in L(y)\).
References
The local source declarations below give the construction and proof.
Lean formalization
OAI.KLInvariance.sum_proper_eq_sub: OrdinaryR.lean:278.
OAI.KLInvariance.CanonicalP.reversed_kernel_on_lower: CanonicalKL.lean:133.
OAI.KLInvariance.CanonicalP.reversed_kernel_on_proper: CanonicalKL.lean:149.