diff --git a/tests/Make.inc b/tests/Make.inc
index 4367ad5d..0c8332a7 100644
--- a/tests/Make.inc
+++ b/tests/Make.inc
@@ -1,5 +1,5 @@
 
-bin_PROGRAMS = Test_GaugeAction Test_cayley_cg Test_cayley_coarsen_support Test_cayley_even_odd Test_cayley_ldop_cr Test_cf_coarsen_support Test_cf_cr_unprec Test_cheby Test_contfrac_cg Test_contfrac_even_odd Test_cshift Test_cshift_red_black Test_dwf_cg_prec Test_dwf_cg_schur Test_dwf_cg_unprec Test_dwf_cr_unprec Test_dwf_even_odd Test_dwf_force Test_dwf_fpgcr Test_dwf_hdcr Test_gamma Test_hmc_WilsonFermionGauge Test_hmc_WilsonGauge Test_lie_generators Test_main Test_multishift_sqrt Test_nersc_io Test_quenched_update Test_remez Test_rng Test_rng_fixed Test_simd Test_stencil Test_wilson_cg_prec Test_wilson_cg_schur Test_wilson_cg_unprec Test_wilson_cr_unprec Test_wilson_even_odd Test_wilson_force Test_wilson_force_phiMdagMphi Test_wilson_force_phiMphi
+bin_PROGRAMS = Test_GaugeAction Test_cayley_cg Test_cayley_coarsen_support Test_cayley_even_odd Test_cayley_ldop_cr Test_cf_coarsen_support Test_cf_cr_unprec Test_cheby Test_contfrac_cg Test_contfrac_even_odd Test_contfrac_force Test_cshift Test_cshift_red_black Test_dwf_cg_prec Test_dwf_cg_schur Test_dwf_cg_unprec Test_dwf_cr_unprec Test_dwf_even_odd Test_dwf_force Test_dwf_fpgcr Test_dwf_hdcr Test_gamma Test_hmc_WilsonFermionGauge Test_hmc_WilsonGauge Test_lie_generators Test_main Test_multishift_sqrt Test_nersc_io Test_partfrac_force Test_quenched_update Test_remez Test_rng Test_rng_fixed Test_simd Test_stencil Test_wilson_cg_prec Test_wilson_cg_schur Test_wilson_cg_unprec Test_wilson_cr_unprec Test_wilson_even_odd Test_wilson_force Test_wilson_force_phiMdagMphi Test_wilson_force_phiMphi
 
 
 Test_GaugeAction_SOURCES=Test_GaugeAction.cc
@@ -42,6 +42,10 @@ Test_contfrac_even_odd_SOURCES=Test_contfrac_even_odd.cc
 Test_contfrac_even_odd_LDADD=-lGrid
 
 
+Test_contfrac_force_SOURCES=Test_contfrac_force.cc
+Test_contfrac_force_LDADD=-lGrid
+
+
 Test_cshift_SOURCES=Test_cshift.cc
 Test_cshift_LDADD=-lGrid
 
@@ -110,6 +114,10 @@ Test_nersc_io_SOURCES=Test_nersc_io.cc
 Test_nersc_io_LDADD=-lGrid
 
 
+Test_partfrac_force_SOURCES=Test_partfrac_force.cc
+Test_partfrac_force_LDADD=-lGrid
+
+
 Test_quenched_update_SOURCES=Test_quenched_update.cc
 Test_quenched_update_LDADD=-lGrid
 
diff --git a/tests/Test_contfrac_force.cc b/tests/Test_contfrac_force.cc
new file mode 100644
index 00000000..a39d596d
--- /dev/null
+++ b/tests/Test_contfrac_force.cc
@@ -0,0 +1,126 @@
+#include <Grid.h>
+
+using namespace std;
+using namespace Grid;
+using namespace Grid::QCD;
+
+#define parallel_for PARALLEL_FOR_LOOP for
+
+int main (int argc, char ** argv)
+{
+  Grid_init(&argc,&argv);
+
+  std::vector<int> latt_size   = GridDefaultLatt();
+  std::vector<int> simd_layout = GridDefaultSimd(Nd,vComplex::Nsimd());
+  std::vector<int> mpi_layout  = GridDefaultMpi();
+
+  const int Ls=9;
+
+  GridCartesian         * UGrid   = SpaceTimeGrid::makeFourDimGrid(GridDefaultLatt(), GridDefaultSimd(Nd,vComplex::Nsimd()),GridDefaultMpi());
+  GridRedBlackCartesian * UrbGrid = SpaceTimeGrid::makeFourDimRedBlackGrid(UGrid);
+  GridCartesian         * FGrid   = SpaceTimeGrid::makeFiveDimGrid(Ls,UGrid);
+  GridRedBlackCartesian * FrbGrid = SpaceTimeGrid::makeFiveDimRedBlackGrid(Ls,UGrid);
+
+  std::vector<int> seeds4({1,2,3,4});
+  std::vector<int> seeds5({5,6,7,8});
+  GridParallelRNG          RNG5(FGrid);  RNG5.SeedFixedIntegers(seeds5);
+  GridParallelRNG          RNG4(UGrid);  RNG4.SeedFixedIntegers(seeds4);
+
+  int threads = GridThread::GetThreads();
+  std::cout<<GridLogMessage << "Grid is setup to use "<<threads<<" threads"<<std::endl;
+
+  LatticeFermion phi        (FGrid); gaussian(RNG5,phi);
+  LatticeFermion Mphi       (FGrid); 
+  LatticeFermion MphiPrime  (FGrid); 
+
+  LatticeGaugeField U(UGrid);
+
+  SU3::HotConfiguration(RNG4,U);
+  
+  ////////////////////////////////////
+  // Unmodified matrix element
+  ////////////////////////////////////
+  RealD mass=0.01; 
+  RealD M5=1.8; 
+  OverlapWilsonContFracTanhFermion Dcf(U,*FGrid,*FrbGrid,*UGrid,*UrbGrid,mass,M5,1.0);
+  Dcf.M   (phi,Mphi);
+
+  ComplexD S    = innerProduct(Mphi,Mphi); // pdag MdagM p
+
+  // get the deriv of phidag MdagM phi with respect to "U"
+  LatticeGaugeField UdSdU(UGrid);
+  LatticeGaugeField tmp(UGrid);
+  
+
+  Dcf.MDeriv(tmp , Mphi,  phi,DaggerNo );  UdSdU=tmp;
+  Dcf.MDeriv(tmp , phi,  Mphi,DaggerYes ); UdSdU=(UdSdU+tmp);  
+  
+  LatticeFermion Ftmp      (FGrid);
+
+  ////////////////////////////////////
+  // Modify the gauge field a little 
+  ////////////////////////////////////
+  RealD dt = 0.0001;
+
+  LatticeColourMatrix mommu(UGrid); 
+  LatticeColourMatrix forcemu(UGrid); 
+  LatticeGaugeField mom(UGrid); 
+  LatticeGaugeField Uprime(UGrid); 
+
+  for(int mu=0;mu<Nd;mu++){
+
+    SU3::GaussianLieAlgebraMatrix(RNG4, mommu); // Traceless antihermitian momentum; gaussian in lie alg
+
+    PokeIndex<LorentzIndex>(mom,mommu,mu);
+
+    // fourth order exponential approx
+    parallel_for(auto i=mom.begin();i<mom.end();i++){
+      Uprime[i](mu) =
+	  U[i](mu)
+	+ mom[i](mu)*U[i](mu)*dt 
+	+ mom[i](mu) *mom[i](mu) *U[i](mu)*(dt*dt/2.0)
+	+ mom[i](mu) *mom[i](mu) *mom[i](mu) *U[i](mu)*(dt*dt*dt/6.0)
+	+ mom[i](mu) *mom[i](mu) *mom[i](mu) *mom[i](mu) *U[i](mu)*(dt*dt*dt*dt/24.0)
+	+ mom[i](mu) *mom[i](mu) *mom[i](mu) *mom[i](mu) *mom[i](mu) *U[i](mu)*(dt*dt*dt*dt*dt/120.0)
+	+ mom[i](mu) *mom[i](mu) *mom[i](mu) *mom[i](mu) *mom[i](mu) *mom[i](mu) *U[i](mu)*(dt*dt*dt*dt*dt*dt/720.0)
+	;
+    }
+
+  }
+  
+  Dcf.ImportGauge(Uprime);
+  Dcf.M          (phi,MphiPrime);
+
+  ComplexD Sprime    = innerProduct(MphiPrime   ,MphiPrime);
+
+  //////////////////////////////////////////////
+  // Use derivative to estimate dS
+  //////////////////////////////////////////////
+
+
+  LatticeComplex dS(UGrid); dS = zero;
+  for(int mu=0;mu<Nd;mu++){
+    mommu   = PeekIndex<LorentzIndex>(UdSdU,mu);
+    mommu=Ta(mommu)*2.0;
+    PokeIndex<LorentzIndex>(UdSdU,mommu,mu);
+  }
+
+  for(int mu=0;mu<Nd;mu++){
+    forcemu = PeekIndex<LorentzIndex>(UdSdU,mu);
+    mommu   = PeekIndex<LorentzIndex>(mom,mu);
+
+    // Update PF action density
+    dS = dS+trace(mommu*forcemu)*dt;
+
+  }
+
+  Complex dSpred    = sum(dS);
+
+  std::cout << GridLogMessage << " S      "<<S<<std::endl;
+  std::cout << GridLogMessage << " Sprime "<<Sprime<<std::endl;
+  std::cout << GridLogMessage << "dS      "<<Sprime-S<<std::endl;
+  std::cout << GridLogMessage << "predict dS    "<< dSpred <<std::endl;
+
+  std::cout<< GridLogMessage << "Done" <<std::endl;
+  Grid_finalize();
+}
diff --git a/tests/Test_partfrac_force.cc b/tests/Test_partfrac_force.cc
new file mode 100644
index 00000000..05925e55
--- /dev/null
+++ b/tests/Test_partfrac_force.cc
@@ -0,0 +1,126 @@
+#include <Grid.h>
+
+using namespace std;
+using namespace Grid;
+using namespace Grid::QCD;
+
+#define parallel_for PARALLEL_FOR_LOOP for
+
+int main (int argc, char ** argv)
+{
+  Grid_init(&argc,&argv);
+
+  std::vector<int> latt_size   = GridDefaultLatt();
+  std::vector<int> simd_layout = GridDefaultSimd(Nd,vComplex::Nsimd());
+  std::vector<int> mpi_layout  = GridDefaultMpi();
+
+  const int Ls=9;
+
+  GridCartesian         * UGrid   = SpaceTimeGrid::makeFourDimGrid(GridDefaultLatt(), GridDefaultSimd(Nd,vComplex::Nsimd()),GridDefaultMpi());
+  GridRedBlackCartesian * UrbGrid = SpaceTimeGrid::makeFourDimRedBlackGrid(UGrid);
+  GridCartesian         * FGrid   = SpaceTimeGrid::makeFiveDimGrid(Ls,UGrid);
+  GridRedBlackCartesian * FrbGrid = SpaceTimeGrid::makeFiveDimRedBlackGrid(Ls,UGrid);
+
+  std::vector<int> seeds4({1,2,3,4});
+  std::vector<int> seeds5({5,6,7,8});
+  GridParallelRNG          RNG5(FGrid);  RNG5.SeedFixedIntegers(seeds5);
+  GridParallelRNG          RNG4(UGrid);  RNG4.SeedFixedIntegers(seeds4);
+
+  int threads = GridThread::GetThreads();
+  std::cout<<GridLogMessage << "Grid is setup to use "<<threads<<" threads"<<std::endl;
+
+  LatticeFermion phi        (FGrid); gaussian(RNG5,phi);
+  LatticeFermion Mphi       (FGrid); 
+  LatticeFermion MphiPrime  (FGrid); 
+
+  LatticeGaugeField U(UGrid);
+
+  SU3::HotConfiguration(RNG4,U);
+  
+  ////////////////////////////////////
+  // Unmodified matrix element
+  ////////////////////////////////////
+  RealD mass=0.01; 
+  RealD M5=1.8; 
+  OverlapWilsonPartialFractionTanhFermion Dpf(U,*FGrid,*FrbGrid,*UGrid,*UrbGrid,mass,M5,1.0);
+  Dpf.M   (phi,Mphi);
+
+  ComplexD S    = innerProduct(Mphi,Mphi); // pdag MdagM p
+
+  // get the deriv of phidag MdagM phi with respect to "U"
+  LatticeGaugeField UdSdU(UGrid);
+  LatticeGaugeField tmp(UGrid);
+  
+
+  Dpf.MDeriv(tmp , Mphi,  phi,DaggerNo );  UdSdU=tmp;
+  Dpf.MDeriv(tmp , phi,  Mphi,DaggerYes ); UdSdU=(UdSdU+tmp);  
+  
+  LatticeFermion Ftmp      (FGrid);
+
+  ////////////////////////////////////
+  // Modify the gauge field a little 
+  ////////////////////////////////////
+  RealD dt = 0.0001;
+
+  LatticeColourMatrix mommu(UGrid); 
+  LatticeColourMatrix forcemu(UGrid); 
+  LatticeGaugeField mom(UGrid); 
+  LatticeGaugeField Uprime(UGrid); 
+
+  for(int mu=0;mu<Nd;mu++){
+
+    SU3::GaussianLieAlgebraMatrix(RNG4, mommu); // Traceless antihermitian momentum; gaussian in lie alg
+
+    PokeIndex<LorentzIndex>(mom,mommu,mu);
+
+    // fourth order exponential approx
+    parallel_for(auto i=mom.begin();i<mom.end();i++){
+      Uprime[i](mu) =
+	  U[i](mu)
+	+ mom[i](mu)*U[i](mu)*dt 
+	+ mom[i](mu) *mom[i](mu) *U[i](mu)*(dt*dt/2.0)
+	+ mom[i](mu) *mom[i](mu) *mom[i](mu) *U[i](mu)*(dt*dt*dt/6.0)
+	+ mom[i](mu) *mom[i](mu) *mom[i](mu) *mom[i](mu) *U[i](mu)*(dt*dt*dt*dt/24.0)
+	+ mom[i](mu) *mom[i](mu) *mom[i](mu) *mom[i](mu) *mom[i](mu) *U[i](mu)*(dt*dt*dt*dt*dt/120.0)
+	+ mom[i](mu) *mom[i](mu) *mom[i](mu) *mom[i](mu) *mom[i](mu) *mom[i](mu) *U[i](mu)*(dt*dt*dt*dt*dt*dt/720.0)
+	;
+    }
+
+  }
+  
+  Dpf.ImportGauge(Uprime);
+  Dpf.M          (phi,MphiPrime);
+
+  ComplexD Sprime    = innerProduct(MphiPrime   ,MphiPrime);
+
+  //////////////////////////////////////////////
+  // Use derivative to estimate dS
+  //////////////////////////////////////////////
+
+
+  LatticeComplex dS(UGrid); dS = zero;
+  for(int mu=0;mu<Nd;mu++){
+    mommu   = PeekIndex<LorentzIndex>(UdSdU,mu);
+    mommu=Ta(mommu)*2.0;
+    PokeIndex<LorentzIndex>(UdSdU,mommu,mu);
+  }
+
+  for(int mu=0;mu<Nd;mu++){
+    forcemu = PeekIndex<LorentzIndex>(UdSdU,mu);
+    mommu   = PeekIndex<LorentzIndex>(mom,mu);
+
+    // Update PF action density
+    dS = dS+trace(mommu*forcemu)*dt;
+
+  }
+
+  Complex dSpred    = sum(dS);
+
+  std::cout << GridLogMessage << " S      "<<S<<std::endl;
+  std::cout << GridLogMessage << " Sprime "<<Sprime<<std::endl;
+  std::cout << GridLogMessage << "dS      "<<Sprime-S<<std::endl;
+  std::cout << GridLogMessage << "predict dS    "<< dSpred <<std::endl;
+
+  std::cout<< GridLogMessage << "Done" <<std::endl;
+  Grid_finalize();
+}