Update matrix-timing-tests timings [skip ci]

.gitea/workflows/Merge-Checker.yaml

@@ -3,6 +3,8 @@ name: Merge-Checker
 on:
   pull_request:
     branches: ["**"]
+    paths-ignore:
+        - 'unit-tests/timing-results/**'
 
 jobs:
   build_and_test:
@@ -36,67 +38,46 @@ jobs:
               echo "Warning: $test_exec not found or not executable"
             fi
           done
-      - name: Run matrix-timing-tests
+
+      - name: Run matrix-timing-tests with per-test timing output and save results
         run: |
           mkdir -p unit-tests/timing-results
           if [ -x build/unit-tests/matrix-timing-tests ]; then
             echo "Running matrix-timing-tests with timing"
             /usr/bin/time -v build/unit-tests/matrix-timing-tests -d yes &> unit-tests/timing-results/matrix-timing-tests.txt
-            cat unit-tests/timing-results/matrix-timing-tests.txt
           else
             echo "matrix-timing-tests executable not found or not executable"
             exit 1
           fi
-
-      - name: Compare timing results
-        id: check_diff
+      - name: Commit and push timing results
+        if: github.event.pull_request.head.repo.full_name == github.repository
         run: |
-          git show origin/${{ github.event.pull_request.head.ref }}:unit-tests/timing-results/matrix-timing-tests.txt > old.txt || echo "" > old.txt
-          cp unit-tests/timing-results/matrix-timing-tests.txt new.txt
+          git config --global user.name "ci-bot"
+          git config --global user.email "ci-bot@local"
 
-          echo "Comparing timing results for changes ≥ 0.1s (ignoring 'Timing Tests' lines)..."
+          BRANCH_NAME="${{ github.event.pull_request.head.ref }}"
 
-          changed=0
+          echo "Checking if last commit was a timing update"
+          LAST_COMMIT_MSG=$(git log -1 --pretty=%B)
 
-          awk -v changed_ref=/tmp/timings_changed.flag '
-            BEGIN {
-              change_threshold = 0.1
-            }
-            FILENAME == "old.txt" && /^[0-9]+\.[0-9]+ s: / {
-              label = substr($0, index($0, ":") + 2)
-              if (label != "Timing Tests") {
-                label_times[label] = $1
-              }
-            }
-            FILENAME == "new.txt" && /^[0-9]+\.[0-9]+ s: / {
-              new_time = $1
-              label = substr($0, index($0, ":") + 2)
-              if (label == "Timing Tests") next
-
-              old_time = label_times[label]
-              delta = new_time - old_time
-              if (delta < 0) delta = -delta
-
-              if (old_time != "" && delta >= change_threshold) {
-                printf "⚠️  %.3f s → %.3f s: %s (Δ=%.3f s)\n", old_time, new_time, label, delta
-                system("touch " changed_ref)
-              } else if (old_time == "") {
-                printf "🆕 New timing entry: %.3f s: %s\n", new_time, label
-                system("touch " changed_ref)
-              }
-            }
-            END {
-              if (!system("test -f " changed_ref)) {
-                exit 0
-              } else {
-                print "✅ Timings haven’t changed significantly (Δ < 0.1s)."
-                exit 0
-              }
-            }
-          ' old.txt new.txt
-
-          if [ -f /tmp/timings_changed.flag ]; then
-            echo "timings_changed=true" >> $GITHUB_OUTPUT
+          if echo "$LAST_COMMIT_MSG" | grep -q "Update matrix-timing-tests timings"; then
+            echo "Last commit was a timing update, skipping commit."
+            exit 0
           else
-            echo "timings_changed=false" >> $GITHUB_OUTPUT
-          fi
+            echo "Last commit name was: {$LAST_COMMIT_MSG}"
+          fi
+
+          echo "Checking out source branch $BRANCH_NAME"
+          git stash
+          git fetch origin "$BRANCH_NAME"
+          git checkout -B "$BRANCH_NAME" "origin/$BRANCH_NAME"
+          git stash pop
+
+          git add unit-tests/timing-results/matrix-timing-tests.txt
+
+          if git diff --quiet --cached; then
+            echo "No changes to commit"
+          else
+            git commit -m "Update matrix-timing-tests timings [skip ci]"
+            git push origin "$BRANCH_NAME"
+          fi

.vscode/settings.json

@@ -75,9 +75,5 @@
     },
     "clangd.enable": true,
     "C_Cpp.dimInactiveRegions": false,
-    "editor.defaultFormatter": "xaver.clang-format",
-    "clangd.inactiveRegions.useBackgroundHighlight": false,
-    "clangd.arguments": [
-        "--compile-commands-dir=${workspaceFolder}/build"
-    ],
+    "editor.defaultFormatter": "xaver.clang-format"
 }

CMakeLists.txt

@@ -4,11 +4,9 @@ project(Vector3D)
 add_subdirectory(src)
 add_subdirectory(unit-tests)
 
-set(CMAKE_CXX_STANDARD 17)
+set(CMAKE_CXX_STANDARD 11)
 
-add_compile_options(-Wall -Wextra -Wpedantic)
-add_compile_options (-fdiagnostics-color=always)
-set(CMAKE_COLOR_DIAGNOSTICS ON)
+add_compile_options(-fdiagnostics-color=always -Wall -Wextra -Wpedantic)
 
 include(FetchContent)
  

README.md

@@ -1,12 +1,3 @@
 # Introduction
 This matrix math library is focused on embedded development and avoids any heap memory allocation unless you explicitly ask for it.
-It uses templates to pre-allocate matrices on the stack.
-
-# Building
-1. Initialize the repositiory with the command:
-```bash
-cmake -S . -B build -G Ninja
-```
-
-2. Go into the build folder and run `ninja`
-3. That's it. You can test out the build by running `./unit-tests/matrix-tests`
+It uses templates to pre-allocate matrices on the stack.

src/CMakeLists.txt

@@ -41,40 +41,6 @@ target_link_libraries(vector-3d
     PRIVATE
 )
 
-# SVD
-add_library(svd 
-    STATIC
-    SVD.cpp
-)
-
-target_link_libraries(svd
-    PUBLIC
-    vector-3d-intf
-    PRIVATE
-)
-
-set_target_properties(svd
-    PROPERTIES
-    LINKER_LANGUAGE CXX
-)
-
-# QR (eigenvalues/eigenvectors via implicit shifted QR iteration)
-add_library(qr 
-    STATIC
-    QR.cpp
-)
-
-target_link_libraries(qr
-    PUBLIC
-    vector-3d-intf
-    PRIVATE
-)
-
-set_target_properties(qr
-    PROPERTIES
-    LINKER_LANGUAGE CXX
-)
-
 # Matrix
 add_library(matrix 
     STATIC
@@ -85,8 +51,6 @@ target_link_libraries(matrix
     PUBLIC
     vector-3d-intf
     PRIVATE
-    svd
-    qr
 )
 
 set_target_properties(matrix

src/Matrix.cpp

@@ -1,39 +1,3 @@
-// This #ifndef section makes clangd happy so that it can properly do type hints
-// in this file
-#ifndef MATRIX_H_
-#define MATRIX_H_
-#include "Matrix.hpp"
-#endif
-
-// Forward-declare QR::EigenQR so the Matrix::EigenQR implementation below can
-// call it even when Matrix.cpp is pulled in through QR.hpp's own include chain
-// (QR.cpp -> QR.hpp -> Matrix.hpp -> Matrix.cpp), where the QR namespace has
-// not been declared yet at this point. If we are not already inside that
-// chain, pull in the full QR library so its template definition is available.
-namespace QR {
-template <uint8_t N>
-void EigenQR(Matrix<N, N> &matrixToDecompose, Matrix<N, N> &eigenVectors,
-             Matrix<N, 1> &eigenValues, uint32_t maxIterations,
-             float tolerance);
-}
-#ifndef QR_H_
-#include "QR.hpp"
-#endif
-
-// Forward-declare SVD::SVD so the Matrix::SVD implementation below can call
-// it even when Matrix.cpp is pulled in through SVD.hpp's own include chain
-// (SVD.hpp -> Matrix.hpp -> Matrix.cpp), where the SVD namespace has not
-// been declared yet at this point. If we are not already inside that chain,
-// pull in the full SVD library so its template definition is available.
-namespace SVD {
-template <uint8_t rows, uint8_t columns>
-void SVD(Matrix<rows, columns> &matrixToDecompose, Matrix<rows, columns> &U,
-         Matrix<columns, 1> &sigma, Matrix<columns, columns> &Vt);
-}
-#ifndef SVD_H_
-#include "SVD.hpp"
-#endif
-
 #ifdef MATRIX_H_ // since the .cpp file has to be included by the .hpp file this
                  // will evaluate to true
 #include "Matrix.hpp"
@@ -41,29 +5,29 @@ void SVD(Matrix<rows, columns> &matrixToDecompose, Matrix<rows, columns> &U,
 #include <algorithm>
 #include <cmath>
 #include <cstdlib>
+#include <type_traits>
 #include <cstring>
 
 template <uint8_t rows, uint8_t columns>
-Matrix<rows, columns>::Matrix(const std::array<float, rows * columns> &array) {
+Matrix<rows, columns>::Matrix(float value)
+{
+  this->Fill(value);
+}
+
+template <uint8_t rows, uint8_t columns>
+Matrix<rows, columns>::Matrix(const std::array<float, rows * columns> &array)
+{
   this->setMatrixToArray(array);
 }
 
 template <uint8_t rows, uint8_t columns>
-template <typename... Args,
-          std::enable_if_t<(std::is_arithmetic_v<Args> && ...), int>>
-Matrix<rows, columns>::Matrix(Args... args) {
+template <typename... Args>
+Matrix<rows, columns>::Matrix(Args... args)
+{
   constexpr uint16_t arraySize{static_cast<uint16_t>(rows) *
                                static_cast<uint16_t>(columns)};
 
   std::initializer_list<float> initList{static_cast<float>(args)...};
-  // if there is only one value, we actually want to do a fill
-  if (sizeof...(args) == 1) {
-    this->Fill(*initList.begin());
-  }
-  static_assert(sizeof...(args) == arraySize || sizeof...(args) == 1,
-                "You did not provide the right amount of initializers for this "
-                "matrix size");
-
   // choose whichever buffer size is smaller for the copy length
   uint32_t minSize =
       std::min(arraySize, static_cast<uint16_t>(initList.size()));
@@ -71,19 +35,22 @@ Matrix<rows, columns>::Matrix(Args... args) {
 }
 
 template <uint8_t rows, uint8_t columns>
-Matrix<rows, columns> Matrix<rows, columns>::Identity() {
-  Matrix<rows, columns> identityMatrix{0};
-  uint32_t minDimension = std::min(rows, columns);
-  for (uint8_t idx{0}; idx < minDimension; idx++) {
-    identityMatrix[idx][idx] = 1;
+void Matrix<rows, columns>::Identity()
+{
+  this->Fill(0);
+  for (uint8_t idx{0}; idx < rows; idx++)
+  {
+    this->matrix[idx * columns + idx] = 1;
   }
-  return identityMatrix;
 }
 
 template <uint8_t rows, uint8_t columns>
-Matrix<rows, columns>::Matrix(const Matrix<rows, columns> &other) {
-  for (uint8_t row_idx{0}; row_idx < rows; row_idx++) {
-    for (uint8_t column_idx{0}; column_idx < columns; column_idx++) {
+Matrix<rows, columns>::Matrix(const Matrix<rows, columns> &other)
+{
+  for (uint8_t row_idx{0}; row_idx < rows; row_idx++)
+  {
+    for (uint8_t column_idx{0}; column_idx < columns; column_idx++)
+    {
       this->matrix[row_idx * columns + column_idx] =
           other.Get(row_idx, column_idx);
     }
@@ -92,15 +59,21 @@ Matrix<rows, columns>::Matrix(const Matrix<rows, columns> &other) {
 
 template <uint8_t rows, uint8_t columns>
 void Matrix<rows, columns>::setMatrixToArray(
-    const std::array<float, rows * columns> &array) {
-  for (uint8_t row_idx{0}; row_idx < rows; row_idx++) {
-    for (uint8_t column_idx{0}; column_idx < columns; column_idx++) {
+    const std::array<float, rows * columns> &array)
+{
+  for (uint8_t row_idx{0}; row_idx < rows; row_idx++)
+  {
+    for (uint8_t column_idx{0}; column_idx < columns; column_idx++)
+    {
       uint16_t array_idx =
           static_cast<uint16_t>(row_idx) * static_cast<uint16_t>(columns) +
           static_cast<uint16_t>(column_idx);
-      if (array_idx < array.size()) {
+      if (array_idx < array.size())
+      {
         this->matrix[row_idx * columns + column_idx] = array[array_idx];
-      } else {
+      }
+      else
+      {
         this->matrix[row_idx * columns + column_idx] = 0;
       }
     }
@@ -110,9 +83,12 @@ void Matrix<rows, columns>::setMatrixToArray(
 template <uint8_t rows, uint8_t columns>
 Matrix<rows, columns> &
 Matrix<rows, columns>::Add(const Matrix<rows, columns> &other,
-                           Matrix<rows, columns> &result) const {
-  for (uint8_t row_idx{0}; row_idx < rows; row_idx++) {
-    for (uint8_t column_idx{0}; column_idx < columns; column_idx++) {
+                           Matrix<rows, columns> &result) const
+{
+  for (uint8_t row_idx{0}; row_idx < rows; row_idx++)
+  {
+    for (uint8_t column_idx{0}; column_idx < columns; column_idx++)
+    {
       result[row_idx][column_idx] =
           this->Get(row_idx, column_idx) + other.Get(row_idx, column_idx);
     }
@@ -123,9 +99,12 @@ Matrix<rows, columns>::Add(const Matrix<rows, columns> &other,
 template <uint8_t rows, uint8_t columns>
 Matrix<rows, columns> &
 Matrix<rows, columns>::Sub(const Matrix<rows, columns> &other,
-                           Matrix<rows, columns> &result) const {
-  for (uint8_t row_idx{0}; row_idx < rows; row_idx++) {
-    for (uint8_t column_idx{0}; column_idx < columns; column_idx++) {
+                           Matrix<rows, columns> &result) const
+{
+  for (uint8_t row_idx{0}; row_idx < rows; row_idx++)
+  {
+    for (uint8_t column_idx{0}; column_idx < columns; column_idx++)
+    {
       result[row_idx][column_idx] =
           this->Get(row_idx, column_idx) - other.Get(row_idx, column_idx);
     }
@@ -138,15 +117,18 @@ template <uint8_t rows, uint8_t columns>
 template <uint8_t other_columns>
 Matrix<rows, other_columns> &
 Matrix<rows, columns>::Mult(const Matrix<columns, other_columns> &other,
-                            Matrix<rows, other_columns> &result) const {
+                            Matrix<rows, other_columns> &result) const
+{
   // allocate some buffers for all of our dot products
   Matrix<1, columns> this_row;
   Matrix<columns, 1> other_column;
 
-  for (uint8_t row_idx{0}; row_idx < rows; row_idx++) {
+  for (uint8_t row_idx{0}; row_idx < rows; row_idx++)
+  {
     // get our row
     this->GetRow(row_idx, this_row);
-    for (uint8_t column_idx{0}; column_idx < other_columns; column_idx++) {
+    for (uint8_t column_idx{0}; column_idx < columns; column_idx++)
+    {
       // get the other matrix'ss column
       other.GetColumn(column_idx, other_column);
 
@@ -161,9 +143,12 @@ Matrix<rows, columns>::Mult(const Matrix<columns, other_columns> &other,
 
 template <uint8_t rows, uint8_t columns>
 Matrix<rows, columns> &
-Matrix<rows, columns>::Mult(float scalar, Matrix<rows, columns> &result) const {
-  for (uint8_t row_idx{0}; row_idx < rows; row_idx++) {
-    for (uint8_t column_idx{0}; column_idx < columns; column_idx++) {
+Matrix<rows, columns>::Mult(float scalar, Matrix<rows, columns> &result) const
+{
+  for (uint8_t row_idx{0}; row_idx < rows; row_idx++)
+  {
+    for (uint8_t column_idx{0}; column_idx < columns; column_idx++)
+    {
       result[row_idx][column_idx] = this->Get(row_idx, column_idx) * scalar;
     }
   }
@@ -172,7 +157,9 @@ Matrix<rows, columns>::Mult(float scalar, Matrix<rows, columns> &result) const {
 }
 
 template <uint8_t rows, uint8_t columns>
-Matrix<rows, columns> Matrix<rows, columns>::Invert() const {
+Matrix<rows, columns>
+Matrix<rows, columns>::Invert() const
+{
   // since all matrix sizes have to be statically specified at compile time we
   // can do this
   static_assert(rows == columns,
@@ -182,7 +169,8 @@ Matrix<rows, columns> Matrix<rows, columns>::Invert() const {
   // unfortunately we can't calculate this at compile time so we'll just reurn
   // zeros
   float determinant{this->Det()};
-  if (determinant == 0) {
+  if (determinant == 0)
+  {
     // you can't invert a matrix with a negative determinant
     result.Fill(0);
     return result;
@@ -207,10 +195,14 @@ Matrix<rows, columns> Matrix<rows, columns>::Invert() const {
 }
 
 template <uint8_t rows, uint8_t columns>
-Matrix<columns, rows> Matrix<rows, columns>::Transpose() const {
+Matrix<columns, rows>
+Matrix<rows, columns>::Transpose() const
+{
   Matrix<columns, rows> result{};
-  for (uint8_t column_idx{0}; column_idx < rows; column_idx++) {
-    for (uint8_t row_idx{0}; row_idx < columns; row_idx++) {
+  for (uint8_t column_idx{0}; column_idx < rows; column_idx++)
+  {
+    for (uint8_t row_idx{0}; row_idx < columns; row_idx++)
+    {
       result[row_idx][column_idx] = this->Get(column_idx, row_idx);
     }
   }
@@ -222,19 +214,24 @@ Matrix<columns, rows> Matrix<rows, columns>::Transpose() const {
 // the fastest way to calculate a 2x2 matrix determinant
 // template <>
 // inline float Matrix<0, 0>::Det() const { return 1e+6; }
-template <> inline float Matrix<1, 1>::Det() const { return this->matrix[0]; }
-template <> inline float Matrix<2, 2>::Det() const {
+template <>
+inline float Matrix<1, 1>::Det() const { return this->matrix[0]; }
+template <>
+inline float Matrix<2, 2>::Det() const
+{
   return this->matrix[0] * this->matrix[3] - this->matrix[1] * this->matrix[2];
 }
 
 template <uint8_t rows, uint8_t columns>
-float Matrix<rows, columns>::Det() const {
+float Matrix<rows, columns>::Det() const
+{
   static_assert(rows == columns,
                 "You can't take the determinant of a non-square matrix.");
 
   Matrix<rows - 1, columns - 1> MinorMatrix{};
   float determinant{0};
-  for (uint8_t column_idx{0}; column_idx < columns; column_idx++) {
+  for (uint8_t column_idx{0}; column_idx < columns; column_idx++)
+  {
     // for odd indices the sign is negative
     float sign = (column_idx % 2 == 0) ? 1 : -1;
     determinant += sign * this->matrix[column_idx] *
@@ -247,9 +244,12 @@ float Matrix<rows, columns>::Det() const {
 template <uint8_t rows, uint8_t columns>
 Matrix<rows, columns> &
 Matrix<rows, columns>::ElementMultiply(const Matrix<rows, columns> &other,
-                                       Matrix<rows, columns> &result) const {
-  for (uint8_t row_idx{0}; row_idx < rows; row_idx++) {
-    for (uint8_t column_idx{0}; column_idx < columns; column_idx++) {
+                                       Matrix<rows, columns> &result) const
+{
+  for (uint8_t row_idx{0}; row_idx < rows; row_idx++)
+  {
+    for (uint8_t column_idx{0}; column_idx < columns; column_idx++)
+    {
       result[row_idx][column_idx] =
           this->Get(row_idx, column_idx) * other.Get(row_idx, column_idx);
     }
@@ -261,9 +261,12 @@ Matrix<rows, columns>::ElementMultiply(const Matrix<rows, columns> &other,
 template <uint8_t rows, uint8_t columns>
 Matrix<rows, columns> &
 Matrix<rows, columns>::ElementDivide(const Matrix<rows, columns> &other,
-                                     Matrix<rows, columns> &result) const {
-  for (uint8_t row_idx{0}; row_idx < rows; row_idx++) {
-    for (uint8_t column_idx{0}; column_idx < columns; column_idx++) {
+                                     Matrix<rows, columns> &result) const
+{
+  for (uint8_t row_idx{0}; row_idx < rows; row_idx++)
+  {
+    for (uint8_t column_idx{0}; column_idx < columns; column_idx++)
+    {
       result[row_idx][column_idx] =
           this->Get(row_idx, column_idx) / other.Get(row_idx, column_idx);
     }
@@ -274,8 +277,10 @@ Matrix<rows, columns>::ElementDivide(const Matrix<rows, columns> &other,
 
 template <uint8_t rows, uint8_t columns>
 float Matrix<rows, columns>::Get(uint8_t row_index,
-                                 uint8_t column_index) const {
-  if (row_index > rows - 1 || column_index > columns - 1) {
+                                 uint8_t column_index) const
+{
+  if (row_index > rows - 1 || column_index > columns - 1)
+  {
     return 1e+10; // TODO: We should throw something here instead of failing
                   // quietly
   }
@@ -285,7 +290,8 @@ float Matrix<rows, columns>::Get(uint8_t row_index,
 template <uint8_t rows, uint8_t columns>
 Matrix<1, columns> &
 Matrix<rows, columns>::GetRow(uint8_t row_index,
-                              Matrix<1, columns> &row) const {
+                              Matrix<1, columns> &row) const
+{
   memcpy(&(row[0]), this->matrix.begin() + row_index * columns,
          columns * sizeof(float));
 
@@ -295,8 +301,10 @@ Matrix<rows, columns>::GetRow(uint8_t row_index,
 template <uint8_t rows, uint8_t columns>
 Matrix<rows, 1> &
 Matrix<rows, columns>::GetColumn(uint8_t column_index,
-                                 Matrix<rows, 1> &column) const {
-  for (uint8_t row_idx{0}; row_idx < rows; row_idx++) {
+                                 Matrix<rows, 1> &column) const
+{
+  for (uint8_t row_idx{0}; row_idx < rows; row_idx++)
+  {
     column[row_idx][0] = this->Get(row_idx, column_index);
   }
 
@@ -304,13 +312,17 @@ Matrix<rows, columns>::GetColumn(uint8_t column_index,
 }
 
 template <uint8_t rows, uint8_t columns>
-void Matrix<rows, columns>::ToString(std::string &stringBuffer) const {
-  for (uint8_t row_idx{0}; row_idx < rows; row_idx++) {
+void Matrix<rows, columns>::ToString(std::string &stringBuffer) const
+{
+  for (uint8_t row_idx{0}; row_idx < rows; row_idx++)
+  {
     stringBuffer += "|";
-    for (uint8_t column_idx{0}; column_idx < columns; column_idx++) {
+    for (uint8_t column_idx{0}; column_idx < columns; column_idx++)
+    {
       stringBuffer +=
           std::to_string(this->matrix[row_idx * columns + column_idx]);
-      if (column_idx != columns - 1) {
+      if (column_idx != columns - 1)
+      {
         stringBuffer += "\t";
       }
     }
@@ -319,14 +331,11 @@ void Matrix<rows, columns>::ToString(std::string &stringBuffer) const {
 }
 
 template <uint8_t rows, uint8_t columns>
-const float *Matrix<rows, columns>::ToArray() const {
-  return this->matrix.data();
-}
-
-template <uint8_t rows, uint8_t columns>
-std::array<float, columns> &
-Matrix<rows, columns>::operator[](uint8_t row_index) {
-  if (row_index > rows - 1) {
+std::array<float, columns> &Matrix<rows, columns>::
+operator[](uint8_t row_index)
+{
+  if (row_index > rows - 1)
+  {
     // TODO: We should throw something here instead of failing quietly.
     row_index = 0;
   }
@@ -337,8 +346,9 @@ Matrix<rows, columns>::operator[](uint8_t row_index) {
 }
 
 template <uint8_t rows, uint8_t columns>
-Matrix<rows, columns> &
-Matrix<rows, columns>::operator=(const Matrix<rows, columns> &other) {
+Matrix<rows, columns> &Matrix<rows, columns>::
+operator=(const Matrix<rows, columns> &other)
+{
   memcpy(this->matrix.begin(), other.matrix.begin(),
          rows * columns * sizeof(float));
 
@@ -347,16 +357,18 @@ Matrix<rows, columns>::operator=(const Matrix<rows, columns> &other) {
 }
 
 template <uint8_t rows, uint8_t columns>
-Matrix<rows, columns>
-Matrix<rows, columns>::operator+(const Matrix<rows, columns> &other) const {
+Matrix<rows, columns> Matrix<rows, columns>::
+operator+(const Matrix<rows, columns> &other) const
+{
   Matrix<rows, columns> buffer{};
   this->Add(other, buffer);
   return buffer;
 }
 
 template <uint8_t rows, uint8_t columns>
-Matrix<rows, columns>
-Matrix<rows, columns>::operator-(const Matrix<rows, columns> &other) const {
+Matrix<rows, columns> Matrix<rows, columns>::
+operator-(const Matrix<rows, columns> &other) const
+{
   Matrix<rows, columns> buffer{};
   this->Sub(other, buffer);
   return buffer;
@@ -364,42 +376,30 @@ Matrix<rows, columns>::operator-(const Matrix<rows, columns> &other) const {
 
 template <uint8_t rows, uint8_t columns>
 template <uint8_t other_columns>
-Matrix<rows, other_columns> Matrix<rows, columns>::operator*(
-    const Matrix<columns, other_columns> &other) const {
+Matrix<rows, other_columns> Matrix<rows, columns>::
+operator*(const Matrix<columns, other_columns> &other) const
+{
   Matrix<rows, other_columns> buffer{};
   this->Mult(other, buffer);
   return buffer;
 }
 
 template <uint8_t rows, uint8_t columns>
-Matrix<rows, columns> Matrix<rows, columns>::operator*(float scalar) const {
+Matrix<rows, columns> Matrix<rows, columns>::operator*(float scalar) const
+{
   Matrix<rows, columns> buffer{};
   this->Mult(scalar, buffer);
   return buffer;
 }
 
-template <uint8_t rows, uint8_t columns>
-Matrix<rows, columns> Matrix<rows, columns>::operator/(float scalar) const {
-  Matrix<rows, columns> buffer = *this;
-  if (scalar == 0) {
-    buffer.Fill(1e+10);
-    return buffer;
-  }
-
-  for (uint8_t row = 0; row < rows; row++) {
-    for (uint8_t column = 0; column < columns; column++) {
-      buffer[row][column] /= scalar;
-    }
-  }
-  return buffer;
-}
-
 template <uint8_t rows, uint8_t columns>
 template <uint8_t vector_size>
 float Matrix<rows, columns>::DotProduct(const Matrix<1, vector_size> &vec1,
-                                        const Matrix<1, vector_size> &vec2) {
+                                        const Matrix<1, vector_size> &vec2)
+{
   float sum{0};
-  for (uint8_t i{0}; i < vector_size; i++) {
+  for (uint8_t i{0}; i < vector_size; i++)
+  {
     sum += vec1.Get(0, i) * vec2.Get(0, i);
   }
 
@@ -409,9 +409,11 @@ float Matrix<rows, columns>::DotProduct(const Matrix<1, vector_size> &vec1,
 template <uint8_t rows, uint8_t columns>
 template <uint8_t vector_size>
 float Matrix<rows, columns>::DotProduct(const Matrix<vector_size, 1> &vec1,
-                                        const Matrix<vector_size, 1> &vec2) {
+                                        const Matrix<vector_size, 1> &vec2)
+{
   float sum{0};
-  for (uint8_t i{0}; i < vector_size; i++) {
+  for (uint8_t i{0}; i < vector_size; i++)
+  {
     sum += vec1.Get(i, 0) * vec2.Get(i, 0);
   }
 
@@ -419,9 +421,12 @@ float Matrix<rows, columns>::DotProduct(const Matrix<vector_size, 1> &vec1,
 }
 
 template <uint8_t rows, uint8_t columns>
-void Matrix<rows, columns>::Fill(float value) {
-  for (uint8_t row_idx{0}; row_idx < rows; row_idx++) {
-    for (uint8_t column_idx{0}; column_idx < columns; column_idx++) {
+void Matrix<rows, columns>::Fill(float value)
+{
+  for (uint8_t row_idx{0}; row_idx < rows; row_idx++)
+  {
+    for (uint8_t column_idx{0}; column_idx < columns; column_idx++)
+    {
       this->matrix[row_idx * columns + column_idx] = value;
     }
   }
@@ -429,11 +434,14 @@ void Matrix<rows, columns>::Fill(float value) {
 
 template <uint8_t rows, uint8_t columns>
 Matrix<rows, columns> &
-Matrix<rows, columns>::MatrixOfMinors(Matrix<rows, columns> &result) const {
+Matrix<rows, columns>::MatrixOfMinors(Matrix<rows, columns> &result) const
+{
   Matrix<rows - 1, columns - 1> MinorMatrix{};
 
-  for (uint8_t row_idx{0}; row_idx < rows; row_idx++) {
-    for (uint8_t column_idx{0}; column_idx < columns; column_idx++) {
+  for (uint8_t row_idx{0}; row_idx < rows; row_idx++)
+  {
+    for (uint8_t column_idx{0}; column_idx < columns; column_idx++)
+    {
       this->MinorMatrix(MinorMatrix, row_idx, column_idx);
       result[row_idx][column_idx] = MinorMatrix.Det();
     }
@@ -445,15 +453,20 @@ Matrix<rows, columns>::MatrixOfMinors(Matrix<rows, columns> &result) const {
 template <uint8_t rows, uint8_t columns>
 Matrix<rows - 1, columns - 1> &
 Matrix<rows, columns>::MinorMatrix(Matrix<rows - 1, columns - 1> &result,
-                                   uint8_t row_idx, uint8_t column_idx) const {
+                                   uint8_t row_idx, uint8_t column_idx) const
+{
   std::array<float, (rows - 1) * (columns - 1)> subArray{};
   uint16_t array_idx{0};
-  for (uint8_t row_iter{0}; row_iter < rows; row_iter++) {
-    if (row_iter == row_idx) {
+  for (uint8_t row_iter{0}; row_iter < rows; row_iter++)
+  {
+    if (row_iter == row_idx)
+    {
       continue;
     }
-    for (uint8_t column_iter{0}; column_iter < columns; column_iter++) {
-      if (column_iter == column_idx) {
+    for (uint8_t column_iter{0}; column_iter < columns; column_iter++)
+    {
+      if (column_iter == column_idx)
+      {
         continue;
       }
       subArray[array_idx] = this->Get(row_iter, column_iter);
@@ -467,9 +480,12 @@ Matrix<rows, columns>::MinorMatrix(Matrix<rows - 1, columns - 1> &result,
 
 template <uint8_t rows, uint8_t columns>
 Matrix<rows, columns> &
-Matrix<rows, columns>::adjugate(Matrix<rows, columns> &result) const {
-  for (uint8_t row_iter{0}; row_iter < rows; row_iter++) {
-    for (uint8_t column_iter{0}; column_iter < columns; column_iter++) {
+Matrix<rows, columns>::adjugate(Matrix<rows, columns> &result) const
+{
+  for (uint8_t row_iter{0}; row_iter < rows; row_iter++)
+  {
+    for (uint8_t column_iter{0}; column_iter < columns; column_iter++)
+    {
       float sign = ((row_iter + 1) % 2) == 0 ? -1 : 1;
       sign *= ((column_iter + 1) % 2) == 0 ? -1 : 1;
       result[column_iter][row_iter] = this->Get(row_iter, column_iter) * sign;
@@ -480,34 +496,55 @@ Matrix<rows, columns>::adjugate(Matrix<rows, columns> &result) const {
 }
 
 template <uint8_t rows, uint8_t columns>
-float Matrix<rows, columns>::EuclideanNorm() const {
-
+Matrix<rows, columns> &
+Matrix<rows, columns>::Normalize(Matrix<rows, columns> &result) const
+{
   float sum{0};
-  for (uint8_t row_idx{0}; row_idx < rows; row_idx++) {
-    for (uint8_t column_idx{0}; column_idx < columns; column_idx++) {
+  for (uint8_t row_idx{0}; row_idx < rows; row_idx++)
+  {
+    for (uint8_t column_idx{0}; column_idx < columns; column_idx++)
+    {
       float val{this->Get(row_idx, column_idx)};
       sum += val * val;
     }
   }
 
-  return sqrt(sum);
+  if (sum == 0)
+  {
+    // this wouldn't do anything anyways
+    result.Fill(1e+6);
+    return result;
+  }
+
+  sum = sqrt(sum);
+
+  for (uint8_t row_idx{0}; row_idx < rows; row_idx++)
+  {
+    for (uint8_t column_idx{0}; column_idx < columns; column_idx++)
+    {
+      result[row_idx][column_idx] = this->Get(row_idx, column_idx) / sum;
+    }
+  }
+
+  return result;
 }
 
 template <uint8_t rows, uint8_t columns>
-template <uint8_t sub_rows, uint8_t sub_columns, uint8_t row_offset,
-          uint8_t column_offset>
-Matrix<sub_rows, sub_columns> Matrix<rows, columns>::SubMatrix() const {
+template <uint8_t sub_rows, uint8_t sub_columns, uint8_t row_offset, uint8_t column_offset>
+Matrix<sub_rows, sub_columns> Matrix<rows, columns>::SubMatrix() const
+{
   // static assert that sub_rows + row_offset <= rows
   // static assert that sub_columns + column_offset <= columns
   static_assert(sub_rows + row_offset <= rows,
                 "The submatrix you're trying to get is out of bounds (rows)");
-  static_assert(
-      sub_columns + column_offset <= columns,
-      "The submatrix you're trying to get is out of bounds (columns)");
+  static_assert(sub_columns + column_offset <= columns,
+                "The submatrix you're trying to get is out of bounds (columns)");
 
   Matrix<sub_rows, sub_columns> buffer{};
-  for (uint8_t row_idx{0}; row_idx < sub_rows; row_idx++) {
-    for (uint8_t column_idx{0}; column_idx < sub_columns; column_idx++) {
+  for (uint8_t row_idx{0}; row_idx < sub_rows; row_idx++)
+  {
+    for (uint8_t column_idx{0}; column_idx < sub_columns; column_idx++)
+    {
       buffer[row_idx][column_idx] =
           this->Get(row_idx + row_offset, column_idx + column_offset);
     }
@@ -516,108 +553,21 @@ Matrix<sub_rows, sub_columns> Matrix<rows, columns>::SubMatrix() const {
 }
 
 template <uint8_t rows, uint8_t columns>
-template <uint8_t sub_rows, uint8_t sub_columns>
-void Matrix<rows, columns>::SetSubMatrix(
-    uint8_t rowOffset, uint8_t columnOffset,
-    const Matrix<sub_rows, sub_columns> &sub_matrix) {
-  int16_t adjustedSubRows = sub_rows;
-  int16_t adjustedSubColumns = sub_columns;
-  int16_t adjustedRowOffset = rowOffset;
-  int16_t adjustedColumnOffset = columnOffset;
+template <uint8_t sub_rows, uint8_t sub_columns, uint8_t row_offset, uint8_t column_offset>
+void Matrix<rows, columns>::SetSubMatrix(const Matrix<sub_rows, sub_columns> &sub_matrix)
+{
+  static_assert(sub_rows + row_offset <= rows,
+                "The submatrix you're trying to set is out of bounds (rows)");
+  static_assert(sub_columns + column_offset <= columns,
+                "The submatrix you're trying to set is out of bounds (columns)");
 
-  // a bunch of safety checks to make sure we don't overflow the matrix
-  if (sub_rows > rows) {
-    adjustedSubRows = rows;
-  }
-  if (sub_columns > columns) {
-    adjustedSubColumns = columns;
-  }
-
-  if (adjustedSubRows + adjustedRowOffset >= rows) {
-    adjustedRowOffset =
-        std::max(0, static_cast<int16_t>(rows) - adjustedSubRows);
-  }
-
-  if (adjustedSubColumns + adjustedColumnOffset >= columns) {
-    adjustedColumnOffset =
-        std::max(0, static_cast<int16_t>(columns) - adjustedSubColumns);
-  }
-
-  for (uint8_t row_idx{0}; row_idx < adjustedSubRows; row_idx++) {
-    for (uint8_t column_idx{0}; column_idx < adjustedSubColumns; column_idx++) {
-      this->matrix[(row_idx + adjustedRowOffset) * columns + column_idx +
-                   adjustedColumnOffset] = sub_matrix.Get(row_idx, column_idx);
+  for (uint8_t row_idx{0}; row_idx < sub_rows; row_idx++)
+  {
+    for (uint8_t column_idx{0}; column_idx < sub_columns; column_idx++)
+    {
+      this->matrix[(row_idx + row_offset) * columns + column_idx + column_offset] = sub_matrix.Get(row_idx, column_idx);
     }
   }
 }
 
-// QR decomposition: decomposes this matrix A into Q and R
-// Assumes square matrix
-template <uint8_t rows, uint8_t columns>
-void Matrix<rows, columns>::QRDecomposition(Matrix<rows, columns> &Q,
-                                            Matrix<columns, columns> &R) const {
-  static_assert(columns <= rows, "QR decomposition requires columns <= rows");
-
-  Q.Fill(0);
-  R.Fill(0);
-  Matrix<rows, 1> a_col, e, u, Q_column_k{};
-  Matrix<1, rows> e_T{};
-
-  for (uint8_t column = 0; column < columns; column++) {
-    this->GetColumn(column, a_col);
-    u = a_col;
-    // -----------------------
-    // ----- CALCULATE Q -----
-    // -----------------------
-    for (uint8_t k = 0; k <= column; k++) {
-      Q.GetColumn(k, Q_column_k);
-      Matrix<1, rows> Q_column_k_T = Q_column_k.Transpose();
-      u = u - Q_column_k * (Q_column_k_T * a_col);
-    }
-    float norm = u.EuclideanNorm();
-    if (norm > 1e-4) {
-      u = u / norm;
-    } else {
-      u.Fill(0);
-    }
-    Q.SetSubMatrix(0, column, u);
-
-    // -----------------------
-    // ----- CALCULATE R -----
-    // -----------------------
-    for (uint8_t k = 0; k <= column; k++) {
-      Q.GetColumn(k, e);
-      R[k][column] = (a_col.Transpose() * e).Get(0, 0);
-    }
-  }
-}
-
-template <uint8_t rows, uint8_t columns>
-void Matrix<rows, columns>::EigenQR(Matrix<rows, rows> &eigenVectors,
-                                    Matrix<rows, 1> &eigenValues,
-                                    uint32_t maxIterations,
-                                    float tolerance) const {
-  static_assert(rows > 1, "Matrix size must be > 1 for QR iteration");
-  static_assert(rows == columns, "Matrix size must be square for QR iteration");
-
-  // Delegate to the QR library: implicit shifted QR iteration with
-  // Wilkinson shift (see src/QR.hpp for the algorithm and conventions).
-  Matrix<rows, rows> A = *this; // QR::EigenQR does not modify its input
-  QR::EigenQR(A, eigenVectors, eigenValues, maxIterations, tolerance);
-}
-
-template <uint8_t rows, uint8_t columns>
-void Matrix<rows, columns>::SVD(Matrix<rows, columns> &U,
-                                Matrix<columns, 1> &sigma,
-                                Matrix<columns, columns> &Vt) const {
-  // Delegate to the SVD library (see src/SVD.hpp for the algorithm and
-  // conventions). NB: the fully-qualified ::SVD is required here — inside
-  // this member the unqualified name SVD refers to this method, which
-  // would shadow the namespace in a qualified lookup. SVD::SVD takes its
-  // input by non-const reference but does not modify it; pass a copy so
-  // the const-ness of *this is preserved.
-  Matrix<rows, columns> A = *this;
-  ::SVD::SVD<rows, columns>(A, U, sigma, Vt);
-}
-
 #endif // MATRIX_H_

src/Matrix.hpp

@@ -1,11 +1,13 @@
-#pragma once
+#ifndef MATRIX_H_
+#define MATRIX_H_
 
 #include <array>
 #include <cstdint>
 #include <string>
-#include <type_traits>
 
+// TODO: Add a function to calculate eigenvalues/vectors
 // TODO: Add a function to compute RREF
+// TODO: Add a function for SVD decomposition
 // TODO: Add a function for LQ decomposition
 
 template <uint8_t rows, uint8_t columns> class Matrix {
@@ -17,6 +19,11 @@ public:
    */
   Matrix() = default;
 
+  /**
+   * @brief Create a matrix but fill all of its entries with one value
+   */
+  Matrix(float value);
+
   /**
    * @brief Initialize a matrix with an array
    */
@@ -28,17 +35,14 @@ public:
   Matrix(const Matrix<rows, columns> &other);
 
   /**
-   * @brief Initialize a matrix directly with scalar values
-   * Uses SFINAE to only accept arithmetic types (int, float, double, etc.)
+   * @brief Initialize a matrix directly with any number of arguments
    */
-  template <typename... Args,
-            std::enable_if_t<(std::is_arithmetic_v<Args> && ...), int> = 0>
-  Matrix(Args... args);
+  template <typename... Args> Matrix(Args... args);
 
   /**
-   * @brief Create an identity matrix
+   * @brief set the matrix diagonals to 1 and all other values to 0
    */
-  static Matrix<rows, columns> Identity();
+  void Identity();
 
   /**
    * @brief Set all elements in this to value
@@ -125,11 +129,10 @@ public:
   Matrix<columns, rows> Transpose() const;
 
   /**
-   * @brief Returns the euclidean magnitude of the matrix. Also known as the L2
-   * norm
+   * @brief reduce the matrix so the sum of its elements equal 1
    * @param result a buffer to store the result into
    */
-  float EuclideanNorm() const;
+  Matrix<rows, columns> &Normalize(Matrix<rows, columns> &result) const;
 
   /**
    * @brief Get a row from the matrix
@@ -156,16 +159,8 @@ public:
    */
   constexpr uint8_t GetColumnSize() { return columns; }
 
-  /**
-   * @brief Write a string representation of the matrix into the buffer
-   */
   void ToString(std::string &stringBuffer) const;
 
-  /**
-   * @brief Returns the internal representation of the matrix as an array
-   */
-  const float *ToArray() const;
-
   /**
    * @brief Get an element from the matrix
    * @param row the row index of the element
@@ -198,15 +193,13 @@ public:
 
   Matrix<rows, columns> operator*(float scalar) const;
 
-  Matrix<rows, columns> operator/(float scalar) const;
-
   template <uint8_t sub_rows, uint8_t sub_columns, uint8_t row_offset,
             uint8_t column_offset>
   Matrix<sub_rows, sub_columns> SubMatrix() const;
 
-  template <uint8_t sub_rows, uint8_t sub_columns>
-  void SetSubMatrix(uint8_t rowOffset, uint8_t columnOffset,
-                    const Matrix<sub_rows, sub_columns> &sub_matrix);
+  template <uint8_t sub_rows, uint8_t sub_columns, uint8_t row_offset,
+            uint8_t column_offset>
+  void SetSubMatrix(const Matrix<sub_rows, sub_columns> &sub_matrix);
 
   /**
    * @brief take the dot product of the two vectors
@@ -223,53 +216,6 @@ public:
     return vec1.Get(0, 0) * vec2.Get(0, 0);
   }
 
-  /**
-   * @brief Performs QR decomposition on this matrix
-   * @param Q a buffer that will contain Q after the function completes
-   * @param R a buffer that will contain R after the function completes
-   */
-  void QRDecomposition(Matrix<rows, columns> &Q,
-                       Matrix<columns, columns> &R) const;
-
-  /**
-   * @brief Calculates the eigenvectors and values of this matrix using the
-   * implicit shifted QR iteration (Wilkinson shift, Givens bulge chasing);
-   * see src/QR.hpp in the QR library for the full algorithm.
-   * @note For a matrix larger than 2x2 the matrix MUST be symmetric.
-   *       A general (nonsymmetric) 2x2 is handled via the closed-form
-   *       solution.
-   * @note The eigenvalues come out sorted DESCENDING (largest first); the
-   *       eigenvector columns are swapped to match. Eigenvector signs are
-   *       arbitrary.
-   * @param eigenVectors a buffer that will contain the eigenvectors of this
-   * matrix in its columns (column i pairs with eigenValues[i])
-   * @param eigenValues a buffer that will contain the eigenvalues of this
-   * matrix, sorted descending
-   * @param maxIterations the number of iterations to perform before giving
-   * up on reaching the given tolerance
-   * @param tolerance the level of accuracy to obtain before stopping.
-   */
-  void EigenQR(Matrix<rows, rows> &eigenVectors, Matrix<rows, 1> &eigenValues,
-               uint32_t maxIterations = 1000, float tolerance = 1e-6f) const;
-
-  /**
-   * @brief Compute the Singular Value Decomposition (SVD) of this matrix.
-   *
-   * Wrapper around SVD::SVD (see SVD.hpp for the full algorithm
-   * description, output storage conventions, and stack-usage notes).
-   * Decomposes A = U · Σ · Vᵀ where U is rows×columns, Σ is the vector
-   * of singular values (columns×1, sorted descending), and Vᵀ is
-   * columns×columns. Works for any shape (wide matrices are handled
-   * internally by computing SVD(Aᵀ) and swapping the factors back).
-   * This matrix is not modified.
-   *
-   * @param U Output: left singular vectors (rows×columns)
-   * @param sigma Output: singular values in descending order (columns×1)
-   * @param Vt Output: right singular vectors, transposed (columns×columns)
-   */
-  void SVD(Matrix<rows, columns> &U, Matrix<columns, 1> &sigma,
-           Matrix<columns, columns> &Vt) const;
-
 protected:
   std::array<float, rows * columns> matrix;
 
@@ -279,6 +225,6 @@ private:
   void setMatrixToArray(const std::array<float, rows * columns> &array);
 };
 
-#ifndef MATRIX_H_
 #include "Matrix.cpp"
+
 #endif // MATRIX_H_

src/QR.cpp

@@ -1,422 +0,0 @@
-// This #ifndef section makes clangd happy so that it can properly do type hints
-// in this file
-#ifndef QR_H_
-#define QR_H_
-#include "QR.hpp"
-#endif
-
-#ifdef QR_H_ // since the .cpp file has to be included by the .hpp file this
-             // will evaluate to true
-#include "QR.hpp"
-#include <cmath>
-#include <cstdint>
-
-namespace QR {
-
-// ============================================================================
-// QR Building Block Implementations (fully templated, heap-free)
-// ============================================================================
-
-/**
- * GivensRotation: R * (a, b)^T = (r, 0)^T with R = [[c, s], [-s, c]],
- * r = +hypot(a, b), c = a/r, s = b/r.
- */
-// [[maybe_unused]]: this helper is only referenced from template
-// (EigenQR/Tridiagonalize), so in translation units that include this file
-// but never instantiate those templates, the definition is legitimately
-// unused. The attribute silences -Wunused-function there without hiding
-// real dead code in TUs that do use the algorithm.
-[[maybe_unused]] static void GivensRotation(float a, float b, float &c,
-                                            float &s) {
-  float r = sqrtf(a * a + b * b);
-  if (r == 0.0f) {
-    c = 1.0f;
-    s = 0.0f;
-    return;
-  }
-  c = a / r;
-  s = b / r;
-}
-
-/**
- * ApplyRotationBothSides: A <- G A G^T (similarity transform) with
- * G = [[c, s], [-s, c]] on the (i, i+1) block, i.e. G is the ZEROING
- * rotation G*(x, y)^T = (r, 0)^T (the orientation used by the implicit QR
- * chase: A = Q R with Q = G^T gives the next iterate R Q = G A G^T).
- * With (c, s) = GivensRotation(A[i][i], A[i+1][i]) this zeroes
- * A[i+1][i] after the LEFT multiplication; the right multiplication then
- * chases the bulge along the superdiagonal (tridiagonal chase).
- *
- * A must be symmetric on entry; the result stays symmetric, so both
- * triangles are written.
- *
- * Block updates (with a00 = A[i][i], a01 = A[i][i+1], a11 = A[i+1][i+1]):
- *   A[i][i]     = c^2 a00 + 2 c s a01 + s^2 a11
- *   A[i][i+1]   = (c^2 - s^2) a01 + c s (a11 - a00)
- *   A[i+1][i+1] = s^2 a00 - 2 c s a01 + c^2 a11
- * Off-block updates (uniform for both sides, since the left factor G and
- * the right factor G^T mix each side with the pattern (a, b) -> (c a + s b,
- * -s a + c b) after transposition):
- *   for j not in {i, i+1}:
- *     A[i][j]   = A[j][i]   = c A[i][j]   + s A[i+1][j]
- *     A[i+1][j] = A[j][i+1] = -s A[i][j]  + c A[i+1][j]
- */
-template <uint8_t N>
-static void ApplyRotationBothSides(Matrix<N, N> &A, uint8_t i, float c,
-                                   float s) {
-  float a00 = A.Get(i, i);
-  float a01 = A.Get(i, i + 1);
-  float a11 = A.Get(i + 1, i + 1);
-  float c2 = c * c;
-  float s2 = s * s;
-  float cs = c * s;
-
-  A[i][i] = c2 * a00 + 2.0f * cs * a01 + s2 * a11;
-  A[i][i + 1] = (c2 - s2) * a01 + cs * (a11 - a00);
-  A[i + 1][i + 1] = s2 * a00 - 2.0f * cs * a01 + c2 * a11;
-  A[i + 1][i] = A[i][i + 1]; // keep both triangles in sync
-
-  for (uint8_t j = 0; j < N; ++j) {
-    if (j == i || j == i + 1)
-      continue;
-    float x = A.Get(i, j);
-    float y = A.Get(i + 1, j);
-    A[i][j] = c * x + s * y;
-    A[j][i] = A[i][j];
-    A[i + 1][j] = -s * x + c * y;
-    A[j][i + 1] = A[i + 1][j];
-  }
-}
-
-/**
- * ApplyRotationToVectors: V <- V G^T with G = [[c, s], [-s, c]] on columns
- * (i, i+1), applied to every row. G^T = [[c, -s], [s, c]], so
- *   V[r][i]     <- c V[r][i]   + s V[r][i+1]
- *   V[r][i+1]   <- -s V[r][i]  + c V[r][i+1]
- *
- * Convention pairing: if A evolves as A <- G A G^T (ApplyRotationBothSides
- * with the SAME c, s), then V accumulates V <- V G^T. With V0 = I the
- * invariant A0 = V A V^T is preserved at every step, so at convergence
- * A0 = V D V^T and the columns of V are the eigenvectors. (Rationale:
- * each chase step is A <- R Q with R = G A the upper-triangular factor and
- * Q = G^T the orthogonal factor of A = Q R, so A = G^T A' G and the
- * orthogonal factors multiply as G1^T G2^T ... in application order.)
- */
-template <uint8_t N>
-static void ApplyRotationToVectors(Matrix<N, N> &V, uint8_t i, float c,
-                                   float s) {
-  for (uint8_t r = 0; r < N; ++r) {
-    float x = V.Get(r, i);
-    float y = V.Get(r, i + 1);
-    V[r][i] = c * x + s * y;
-    V[r][i + 1] = -s * x + c * y;
-  }
-}
-
-/**
- * WilkinsonShift: eigenvalue of [[a, b], [b, d]] closest to d.
- * mu = (a+d)/2 - sign(a-d) * sqrt(((a-d)/2)^2 + b^2), sign(0) = +1.
- */
-[[maybe_unused]] static float WilkinsonShift(float a, float b, float d) {
-  float delta = 0.5f * (a - d);
-  float spread = sqrtf(delta * delta + b * b);
-  return 0.5f * (a + d) - (delta >= 0.0f ? spread : -spread);
-}
-
-/**
- * Solve2x2Eigen: closed-form eigen-decomposition of the 2x2 block at
- * (lo, lo+1). Works for symmetric blocks and for general 2x2 blocks with
- * real eigenvalues (used by the N == 2 entry point).
- *
- * lambdaHi/lambdaLo come from the characteristic polynomial
- *   lambda^2 - trace*lambda + det = 0.
- * The eigenvector for lambdaHi is v = (b, lambdaHi - a) (from the first
- * row of (A - lambda*I)v = 0), normalized to unit length. If b == 0 the
- * block is triangular and the eigenvectors are coordinate vectors:
- * e1 for the larger of {a, d}, e2 for the other.
- */
-template <uint8_t N>
-static void Solve2x2Eigen(const Matrix<N, N> &A, uint8_t lo, float &lambdaHi,
-                          float &lambdaLo, float &c, float &s) {
-  float a = A.Get(lo, lo);
-  float b = A.Get(lo, lo + 1);
-  float e = A.Get(lo + 1, lo);
-  float d = A.Get(lo + 1, lo + 1);
-
-  float trace = a + d;
-  float det = a * d - b * e;
-  float disc = trace * trace - 4.0f * det;
-  if (disc < 0.0f)
-    disc = 0.0f; // round-off clamp: real 2x2 blocks have disc >= 0
-  float sqrtDisc = sqrtf(disc);
-  lambdaHi = 0.5f * (trace + sqrtDisc);
-  lambdaLo = 0.5f * (trace - sqrtDisc);
-
-  if (b != 0.0f) {
-    float v1 = lambdaHi - a;
-    float n = sqrtf(b * b + v1 * v1);
-    c = b / n;
-    s = v1 / n;
-  } else if (a >= d) {
-    c = 1.0f; // e1 is the eigenvector of a = lambdaHi
-    s = 0.0f;
-  } else {
-    c = 0.0f; // e2 is the eigenvector of d = lambdaHi
-    s = 1.0f;
-  }
-}
-
-/**
- * Deflate: zero subdiagonal entries i in [lo, hi) whose magnitude is at or
- * below tolerance * (|A[i][i]| + |A[i+1][i+1]|).
- */
-template <uint8_t N>
-static void Deflate(Matrix<N, N> &A, uint8_t lo, uint8_t hi, float tolerance) {
-  for (uint8_t i = lo; i < hi; ++i) {
-    float t = A.Get(i + 1, i);
-    float scale = fabsf(A.Get(i, i)) + fabsf(A.Get(i + 1, i + 1));
-    if (fabsf(t) <= tolerance * scale) {
-      A[i + 1][i] = 0.0f;
-      A[i][i + 1] = 0.0f;
-    }
-  }
-}
-
-// ============================================================================
-// QR::EigenQR driver (implicit Wilkinson-shifted QR, bulge chasing)
-// ============================================================================
-
-/**
- * Tridiagonalize: Givens tridiagonalization (Golub & Van Loan 8.3.1).
- *
- * For column k = 0..N-3 the entries A[k+2..N-1, k] are eliminated by
- * rotations on (i, i+1) applied BOTTOM-UP, i = N-2 down to k+1, each
- * formed from the CURRENT (already-updated) pair (A[i][k], A[i+1][k]).
- * Bottom-up is essential: a top-down pass zeros A[i+1][k] with a rotation
- * that would later be undone when the next rotation (i+1, i+2) is formed
- * from an entry below, reviving A[i][k]. Each bottom-up rotation zeros the
- * bottom of the remaining nonzero pair and the entries below stay zero
- * (they are not mixed again, only rows i-1/i are mixed next).
- *
- * Already-tridiagonalized leading columns j < k are untouched: the mixed
- * rows are both >= k+1 > j+1, so A[i][j] and A[i+1][j] are both zero there.
- * The rotation on (i, i+1) also keeps column k+1..k+2 structure intact and
- * does not destroy earlier columns, so after column k is done the leading
- * (k+1)x(k+1) block is tridiagonal forever.
- *
- * On return: A is symmetric tridiagonal and A_orig = U A U^T (U = product
- * of every rotation applied, in application order, as U <- U G^T).
- */
-template <uint8_t N>
-static void Tridiagonalize(Matrix<N, N> &A, Matrix<N, N> &U) {
-  U = Matrix<N, N>{0};
-  for (uint8_t i = 0; i < N; ++i) {
-    U[i][i] = 1.0f;
-  }
-  float c = 0.0f, s = 0.0f;
-  for (uint8_t k = 0; k + 2 < N; ++k) {
-    for (int i = (int)N - 2; i >= (int)k + 1; --i) {
-      GivensRotation(A.Get(i, k), A.Get(i + 1, k), c, s);
-      ApplyRotationBothSides(A, (uint8_t)i, c, s);
-      ApplyRotationToVectors(U, (uint8_t)i, c, s);
-    }
-  }
-}
-
-/**
- * See QR.hpp for the full contract. Implementation sketch:
- *
- *   Phase 0 (N >= 3): Tridiagonalize(A, U)  // A_orig = U A U^T
- *   V = I.
- *   while (hi > 0):
- *     Deflate(A, 0, hi, tol); peel exact-zero trailing subdiagonals (hi--)
- *     lo = top of the trailing unreduced block (scan down, stop at first
- *          exact zero subdiagonal)
- *     if lo == hi - 1:  closed-form 2x2 eigen-solve; fold Vblock into V
- *     else:             one implicit Wilkinson-shifted QR step:
- *                        mu = WilkinsonShift(A[hi-1][hi-1], A[hi][hi-1], A[hi][hi])
- *                        A[lo..hi diagonal] -= mu          // whole block!
- *                        G1 = Givens(A[lo][lo], A[lo+1][lo])
- *                        for i = lo..hi-1:
- *                          (i > lo: Gi = Givens(A[i][i], A[i+1][i]))
- *                          ApplyRotationBothSides(A, i, Gi)   // A <- Gi A Gi^T
- *                          ApplyRotationToVectors(V, i, Gi)   // V <- V Gi^T
- *                        A[lo..hi diagonal] += mu
- *   eigenvalues = diag(A), sorted descending with matching V column swaps.
- *   eigenvectors = U * V.
- *
- * Invariant maintained for N >= 3 (symmetric input): A is symmetric
- * tridiagonal (up to deflated zeros and ~1e-7 float roundoff in the
- * off-tridiagonal corners) at the top of every loop iteration, and
- * A_orig = U A U^T = (U V) A (U V)^T throughout (V = product of every
- * rotation applied so far, in application order, as V <- V Gi^T). At
- * convergence A = V D V^T and therefore A_orig = (U V) D (U V)^T.
- *
- * Orientation note: each chase rotation Gi is the ZEROING rotation
- * (Gi * (x, y)^T = (r, 0)^T). The step A <- Gi A Gi^T equals R Q with
- * R = Gi A upper-triangular (on the block) and Q = Gi^T -- i.e. it IS the
- * standard QR update Q(A - mu I)Q^T with Q the orthogonal QR factor. The
- * eigenvector accumulator therefore collects the Q factors: V <- V Gi^T.
- */
-template <uint8_t N>
-void EigenQR(Matrix<N, N> &matrixToDecompose, Matrix<N, N> &eigenVectors,
-             Matrix<N, 1> &eigenValues, uint32_t maxIterations, float tolerance) {
-  static_assert(N >= 2, "QR::EigenQR requires N >= 2 (N = 1 is trivial)");
-
-  Matrix<N, N> A = matrixToDecompose; // input is not modified
-  Matrix<N, N> V{0};
-  // NB: Matrix::Identity() is a static factory that returns by value; a
-  // bare call would be a no-op. Set the diagonal explicitly.
-  for (uint8_t i = 0; i < N; ++i) {
-    V[i][i] = 1.0f;
-  }
-
-  // ------------------------------------------------------------------
-  // N == 2: closed-form solution (works for nonsymmetric input too)
-  // ------------------------------------------------------------------
-  if (N == 2) {
-    float l1 = 0.0f, l2 = 0.0f, c = 0.0f, s = 0.0f;
-    Solve2x2Eigen(A, 0, l1, l2, c, s);
-    // V = I * Vblock = [[c, -s], [s, c]]
-    V[0][0] = c;
-    V[0][1] = -s;
-    V[1][0] = s;
-    V[1][1] = c;
-    eigenValues[0][0] = l1;
-    eigenValues[1][0] = l2;
-    for (uint8_t r = 0; r < N; ++r)
-      for (uint8_t col = 0; col < N; ++col)
-        eigenVectors[r][col] = V.Get(r, col);
-    return;
-  }
-
-  // ------------------------------------------------------------------
-  // N >= 3: implicit shifted QR iteration (symmetric input required)
-  // ------------------------------------------------------------------
-  // Phase 0: general symmetric -> symmetric tridiagonal. The implicit
-  // QR bulge chase only preserves a tridiagonal structure, so the input
-  // must be reduced first: A_orig = U A U^T with A tridiagonal.
-  Matrix<N, N> U{};
-  Tridiagonalize(A, U);
-
-  uint32_t iter = 0;
-  uint8_t hi = N - 1;
-  while (hi > 0) {
-    Deflate(A, 0, hi, tolerance);
-    // Peel trailing rows whose subdiagonal is exactly zero (deflated or
-    // already solved). Must be re-done every iteration: a peel is only
-    // meaningful once the subdiagonal beneath it has converged.
-    while (hi > 0 && A.Get(hi, hi - 1) == 0.0f) {
-      --hi;
-    }
-    if (hi == 0) {
-      break; // fully diagonal (within tolerance)
-    }
-
-    // Find the top of the trailing unreduced block: scan down from hi-1
-    // and stop at the first exact zero subdiagonal. A[hi][hi-1] != 0 here
-    // (just peeled), so lo < hi.
-    uint8_t lo = hi;
-    for (int i = (int)hi - 1; i >= 0; --i) {
-      if (A.Get(i + 1, i) == 0.0f) {
-        break;
-      }
-      lo = (uint8_t)i;
-    }
-
-    if (lo + 1 == hi) {
-      // Trailing unreduced block is 2x2: solve in closed form.
-      float l1 = 0.0f, l2 = 0.0f, c = 0.0f, s = 0.0f;
-      Solve2x2Eigen(A, lo, l1, l2, c, s);
-      A[lo][lo] = l1;
-      A[lo + 1][lo + 1] = l2;
-      A[lo][lo + 1] = 0.0f;
-      A[lo + 1][lo] = 0.0f;
-      // Fold Vblock = [[c, -s], [s, c]] into V: V <- V * Vblock on
-      // columns (lo, lo+1). NOTE the sign convention differs from
-      // ApplyRotationToVectors (which applies [[c, s], [-s, c]]):
-      // here column 0 of Vblock is (c, s)^T, column 1 is (-s, c)^T.
-      for (uint8_t r = 0; r < N; ++r) {
-        float x = V.Get(r, lo);
-        float y = V.Get(r, lo + 1);
-        V[r][lo] = c * x + s * y;
-        V[r][lo + 1] = -s * x + c * y;
-      }
-      if (lo == 0) {
-        break; // block reached the top: matrix is fully solved
-      }
-      hi = (uint8_t)(lo - 1);
-      continue;
-    }
-
-    // One implicit Wilkinson-shifted QR step on block [lo, hi].
-    float mu = WilkinsonShift(A.Get(hi - 1, hi - 1), A.Get(hi, hi - 1),
-                              A.Get(hi, hi));
-
-    // The shift applies to the ENTIRE active block: bulge chasing
-    // triangularizes (A - mu*I), and the first Givens rotation is formed
-    // from (A[lo][lo] - mu, A[lo+1][lo]).
-    for (uint8_t i = lo; i <= hi; ++i) {
-      A[i][i] -= mu;
-    }
-
-    float c = 0.0f, s = 0.0f;
-    for (uint8_t i = lo; i < hi; ++i) {
-      if (i == lo) {
-        GivensRotation(A.Get(lo, lo), A.Get(lo + 1, lo), c, s);
-      } else {
-        GivensRotation(A.Get(i, i), A.Get(i + 1, i), c, s);
-      }
-      ApplyRotationBothSides(A, i, c, s);
-      ApplyRotationToVectors(V, i, c, s);
-    }
-
-    for (uint8_t i = lo; i <= hi; ++i) {
-      A[i][i] += mu;
-    }
-
-    if (++iter >= maxIterations) {
-      // Best-effort: fall through with the partially diagonalized A.
-      break;
-    }
-  }
-
-  // ------------------------------------------------------------------
-  // Collect eigenvalues and sort DESCENDING (swap eigenvectors to match)
-  // ------------------------------------------------------------------
-  for (uint8_t i = 0; i < N; ++i) {
-    eigenValues[i][0] = A.Get(i, i);
-  }
-  for (uint8_t i = 0; i < N - 1; ++i) {
-    uint8_t k = i;
-    for (uint8_t j = i + 1; j < N; ++j) {
-      if (eigenValues.Get(j, 0) > eigenValues.Get(k, 0)) {
-        k = j;
-      }
-    }
-    if (k != i) {
-      float t = eigenValues[i][0];
-      eigenValues[i][0] = eigenValues[k][0];
-      eigenValues[k][0] = t;
-      for (uint8_t r = 0; r < N; ++r) {
-        float x = V.Get(r, i);
-        V[r][i] = V.Get(r, k);
-        V[r][k] = x;
-      }
-    }
-  }
-
-  // True eigenvectors of the original matrix: U * V. Reuse the A buffer
-  // (its diagonal has already been collected into eigenValues).
-  U.Mult(V, A);
-
-  for (uint8_t r = 0; r < N; ++r) {
-    for (uint8_t col = 0; col < N; ++col) {
-      eigenVectors[r][col] = A.Get(r, col);
-    }
-  }
-}
-
-} // namespace QR
-
-#endif // QR_H_

src/QR.hpp

@@ -1,196 +0,0 @@
-#pragma once
-#include "Matrix.hpp"
-
-/**
- * @brief Library that uses Matrix.hpp and computes the eigenvalues and
- * eigenvectors of a square matrix with the implicit shifted QR iteration
- * (Wilkinson shift, Givens bulge chasing).
- *
- * @note Fully templated: QR::EigenQR works for ANY Matrix<N,N> with N in
- *       2..255 (the uint8_t range of Matrix). There is no 5x5 limit.
- *
- * @note N >= 3: the input matrix MUST be symmetric (A[i][j] == A[j][i]).
- *       The implicit QR bulge chase maintains a symmetric tridiagonal
- *       structure, which only exists for symmetric input. N = 2 handles
- *       a general (nonsymmetric) 2x2 via the closed-form solution, so
- *       nonsymmetric 2x2 inputs also work.
- *
- * @note The input matrix is NOT modified (the iteration runs on a local
- *       copy), mirroring the SVD::SVD convention.
- *
- * @note EMBEDDED CONSTRAINT -- no heap. All working storage is stack
- *       allocated as templated Matrix<N,N> buffers. Peak stack usage per
- *       call is 3 * N^2 floats (A working copy + U and V accumulators) =
- *       12 * N^2 bytes:
- *         N = 5   -> ~0.3 KB
- *         N = 10  -> ~1.2 KB
- *         N = 20  -> ~4.8 KB
- *         N = 50  -> ~30 KB
- *         N = 100 -> ~120 KB
- *         N = 255 -> ~783 KB
- *       Instantiate only the sizes that fit your call-stack budget.
- *
- * @note Conventions:
- *       - Eigenvalues come out sorted DESCENDING (largest first); the
- *         eigenvector columns are swapped to match.
- *       - Eigenvector signs are arbitrary (v and -v are both valid);
- *         tests must be sign-invariant.
- *       - Wilkinson shift: the eigenvalue of the trailing 2x2 block
- *         closest to the bottom-right corner (Trefethen & Bau 13.4.1).
- *
- * @note Algorithm (Trefethen & Bau 13.4, Golub & Van Loan 8.4.3):
- *       Phase 0 (N >= 3): Givens tridiagonalization. A general symmetric
- *         matrix is NOT suitable for implicit QR (the bulge chase only
- *         preserves the tridiagonal structure), so first reduce A with
- *         adjacent Givens similarities A <- G A G^T (rotations applied
- *         BOTTOM-UP, i = N-2 down to k+1, per column k), accumulating
- *         U <- U G^T, until A is symmetric tridiagonal and
- *         A_orig = U A U^T. (N = 2 needs no reduction.)
- *       Phase 1: iterate until A is diagonal:
- *         1. Deflate: zero out subdiagonal entries at/under the tolerance
- *            (scaled by the adjacent diagonal magnitudes).
- *         2. Scan for the trailing unreduced block [lo, hi].
- *            - block of size 1: A[hi][hi] is a converged eigenvalue, done.
- *            - block of size 2: solve the 2x2 eigenproblem in closed form
- *              and fold its eigenvector matrix into V.
- *            - block larger: one implicit Wilkinson-shifted QR step
- *              (bulge chasing with Givens rotations; the shift is applied
- *              to the ENTIRE active block [lo, hi], not just the trailing
- *              2x2 -- the first Givens rotation must be formed from
- *              (A[lo][lo] - mu, A[lo+1][lo])). Every rotation is folded
- *              into V.
- *       Phase 2: eigenvalues = diag(A), sorted DESCENDING (eigenvector
- *         columns swapped to match), and the true eigenvectors of the
- *         ORIGINAL matrix are U * V.
- *
- * @note If maxIterations is exhausted before convergence the best-effort
- *       (partially diagonalized) values on the diagonal are returned.
- */
-namespace QR {
-
-/**
- * @brief Compute the eigenvalues and eigenvectors of a square matrix
- *
- * @param matrixToDecompose The matrix to take eigenvalues of (not
- *        modified). MUST be symmetric for N >= 3.
- * @param eigenVectors a buffer that will contain the eigenvectors in its
- *        COLUMNS, sorted by descending eigenvalue (column i is the
- *        eigenvector for eigenValues[i]).
- * @param eigenValues a buffer that will contain the eigenvalues sorted
- *        DESCENDING (largest first).
- * @param maxIterations the number of QR steps to perform before giving up
- *        on reaching the given tolerance
- * @param tolerance the level of accuracy to obtain before stopping; a
- *        subdiagonal entry is deflated when |A[i+1][i]| <= tolerance *
- *        (|A[i][i]| + |A[i+1][i+1]|). For float32 arithmetic, values
- *        around 1e-6 are a sensible choice (single-precision epsilon is
- *        ~1.2e-7).
- */
-template <uint8_t N>
-void EigenQR(Matrix<N, N> &matrixToDecompose, Matrix<N, N> &eigenVectors,
-             Matrix<N, 1> &eigenValues, uint32_t maxIterations, float tolerance);
-
-/**
- * @brief Apply the similarity transform A <- G A G^T on rows/cols (i, i+1)
- *
- * G = [ c  s ] on the (i, i+1) block, identity elsewhere, where G is the
- *     [ -s c ]
- * ZEROING rotation (G * (x, y)^T = (r, 0)^T) -- the orientation used by
- * the implicit QR chase: A = Q R with Q = G^T gives the next iterate
- * R Q = G A G^T. With (c, s) = GivensRotation(A[i][i], A[i+1][i]) the
- * (i+1, i) entry is zeroed by the left multiplication and the bulge is
- * chased along the superdiagonal by the right one. The matrix must be
- * symmetric on entry (guaranteed by construction in the QR iteration:
- * symmetric input stays symmetric under similarity by an orthogonal
- * matrix). Updates the full matrix, not just the tridiagonal structure.
- */
-template <uint8_t N>
-static void ApplyRotationBothSides(Matrix<N, N> &A, uint8_t i, float c,
-                                   float s);
-
-/**
- * @brief Accumulate eigenvectors: V <- V G^T on columns (i, i+1)
- *
- * G^T = [ c  -s ] on columns (i, i+1), identity elsewhere, where G =
- *       [ s   c ]
- * [ c, s ] / [ -s, c ] is the zeroing rotation paired with
- * ApplyRotationBothSides. Applied to all rows:
- *   V[r][i]     -> c V[r][i]   + s V[r][i+1]
- *   V[r][i+1]   -> -s V[r][i]  + c V[r][i+1]
- *
- * Every QR step's rotation is folded into V this way so that, together
- * with A <- G A G^T, the invariant A_orig = V A V^T is preserved at every
- * step (each step is A <- R Q with Q = G^T the orthogonal factor, and
- * the orthogonal factors multiply as G1^T G2^T ... in application order).
- * At convergence A_orig = V D V^T and the columns of V are the
- * eigenvectors.
- */
-template <uint8_t N>
-static void ApplyRotationToVectors(Matrix<N, N> &V, uint8_t i, float c,
-                                   float s);
-
-/**
- * @brief Solve the 2x2 eigenproblem of block rows/cols (lo, lo+1)
- *
- * Solves the (possibly nonsymmetric) 2x2 block
- *        [ A[lo][lo]     A[lo][lo+1]   ]
- *        [ A[lo+1][lo]   A[lo+1][lo+1] ]
- * in closed form (characteristic polynomial + eigenvector back-substitution).
- *
- * @param A the matrix containing the block (not modified)
- * @param lo the row/col index of the top-left corner of the block
- * @param lambdaHi (out) the LARGER eigenvalue
- * @param lambdaLo (out) the smaller eigenvalue
- * @param c (out), s (out) eigenvector pair as an orthogonal matrix
- *        Vblock = [ c  -s ] whose columns are the eigenvectors: column 0
- *                 [ s   c ]
- *        (c, s) is the unit eigenvector for lambdaHi, column 1 (-s, c) is
- *        the unit eigenvector for lambdaLo.
- *
- * Note: the caller applies Vblock to its eigenvector accumulator with
- * V <- V * Vblock (i.e. V[r][lo]   = c*x + s*y,
- *                              V[r][lo+1] = -s*x + c*y). Vblock has the
- * SAME [ c -s; s c ] form as the G^T factor used by
- * ApplyRotationToVectors, so both folding operations follow one uniform
- * convention.
- */
-template <uint8_t N>
-static void Solve2x2Eigen(const Matrix<N, N> &A, uint8_t lo, float &lambdaHi,
-                          float &lambdaLo, float &c, float &s);
-
-/**
- * @brief Deflate (zero out) subdiagonal entries that are at/under tolerance
- *
- * For each i in [lo, hi): if |A[i+1][i]| <= tolerance *
- * (|A[i][i]| + |A[i+1][i+1]|), sets A[i+1][i] = A[i][i+1] = 0, splitting
- * the matrix into smaller independent blocks.
- */
-template <uint8_t N>
-static void Deflate(Matrix<N, N> &A, uint8_t lo, uint8_t hi, float tolerance);
-
-/**
- * @brief Reduce a symmetric matrix to symmetric tridiagonal form
- *
- * Chases each column's entries below the subdiagonal to zero with
- * adjacent Givens similarities (Golub & Van Loan 8.3.1, Givens variant):
- * for column k = 0..N-3, rotations on (N-2, N-1), (N-3, N-2), ...
- * (k+1, k+2) -- BOTTOM-UP, each formed from the current (A[i][k],
- * A[i+1][k]) -- zero A[k+2..N-1, k] one by one. A top-down pass would not
- * work: the rotation that zeros A[i+1][k] would be undone by the later
- * rotation on (i+1, i+2) forming a new nonzero at A[i][k]. Each rotation
- * is applied to A as a similarity (A <- G A G^T) and accumulated into U
- * (U <- U G^T), so on return:
- *   - A is symmetric tridiagonal (off-tridiagonal entries EXACTLY zero),
- *   - A_orig = U A U^T (i.e. U^T A_orig U = A).
- *
- * U is initialized to the identity internally (its input contents are
- * ignored).
- */
-template <uint8_t N>
-static void Tridiagonalize(Matrix<N, N> &A, Matrix<N, N> &U);
-
-} // namespace QR
-
-#ifndef QR_H_
-#include "QR.cpp"
-#endif

src/Quaternion.cpp

@@ -6,115 +6,115 @@
  * @param angle The angle to rotate by
  * @param axis The axis to rotate around
  */
-Quaternion Quaternion::FromAngleAndAxis(float angle, const Matrix<1, 3> &axis) {
-  const float halfAngle = angle / 2;
-  const float sinHalfAngle = sin(halfAngle);
-  Matrix<1, 3> normalizedAxis = axis / axis.EuclideanNorm();
-  return Quaternion{static_cast<float>(cos(halfAngle)),
-                    normalizedAxis.Get(0, 0) * sinHalfAngle,
-                    normalizedAxis.Get(0, 1) * sinHalfAngle,
-                    normalizedAxis.Get(0, 2) * sinHalfAngle};
+Quaternion Quaternion::FromAngleAndAxis(float angle, const Matrix<1, 3> &axis)
+{
+    const float halfAngle = angle / 2;
+    const float sinHalfAngle = sin(halfAngle);
+    Matrix<1, 3> normalizedAxis{};
+    axis.Normalize(normalizedAxis);
+    return Quaternion{
+        static_cast<float>(cos(halfAngle)),
+        normalizedAxis.Get(0, 0) * sinHalfAngle,
+        normalizedAxis.Get(0, 1) * sinHalfAngle,
+        normalizedAxis.Get(0, 2) * sinHalfAngle};
 }
 
-float Quaternion::operator[](uint8_t index) const {
-  if (index < 4) {
-    return this->matrix[index];
-  }
+float Quaternion::operator[](uint8_t index) const
+{
+    if (index < 4)
+    {
+        return this->matrix[index];
+    }
 
-  // index out of bounds
-  return 1e+6;
+    // index out of bounds
+    return 1e+6;
 }
 
-void Quaternion::operator=(const Quaternion &other) {
-  memcpy(&(this->matrix), &(other.matrix), 4 * sizeof(float));
+void Quaternion::operator=(const Quaternion &other)
+{
+    memcpy(&(this->matrix), &(other.matrix), 4 * sizeof(float));
 }
 
-Quaternion Quaternion::operator*(const Quaternion &other) const {
-  Quaternion result{};
-  this->Q_Mult(other, result);
-  return result;
+Quaternion Quaternion::operator*(const Quaternion &other) const
+{
+    Quaternion result{};
+    this->Q_Mult(other, result);
+    return result;
 }
 
-Quaternion Quaternion::operator*(float scalar) const {
-  return Quaternion{this->w * scalar, this->v1 * scalar, this->v2 * scalar,
-                    this->v3 * scalar};
+Quaternion Quaternion::operator*(float scalar) const
+{
+    return Quaternion{this->w * scalar, this->v1 * scalar, this->v2 * scalar, this->v3 * scalar};
 }
 
-Quaternion Quaternion::operator+(const Quaternion &other) const {
-  return Quaternion{this->w + other.w, this->v1 + other.v1, this->v2 + other.v2,
-                    this->v3 + other.v3};
+Quaternion Quaternion::operator+(const Quaternion &other) const
+{
+    return Quaternion{this->w + other.w, this->v1 + other.v1, this->v2 + other.v2, this->v3 + other.v3};
 }
 
-Quaternion &Quaternion::Q_Mult(const Quaternion &other,
-                               Quaternion &buffer) const {
+Quaternion &
+Quaternion::Q_Mult(const Quaternion &other, Quaternion &buffer) const
+{
 
-  // eq. 6
-  buffer.w = (other.w * this->w - other.v1 * this->v1 - other.v2 * this->v2 -
-              other.v3 * this->v3);
-  buffer.v1 = (other.w * this->v1 + other.v1 * this->w - other.v2 * this->v3 +
-               other.v3 * this->v2);
-  buffer.v2 = (other.w * this->v2 + other.v1 * this->v3 + other.v2 * this->w -
-               other.v3 * this->v1);
-  buffer.v3 = (other.w * this->v3 - other.v1 * this->v2 + other.v2 * this->v1 +
-               other.v3 * this->w);
-  return buffer;
+    // eq. 6
+    buffer.w = (other.w * this->w - other.v1 * this->v1 - other.v2 * this->v2 - other.v3 * this->v3);
+    buffer.v1 = (other.w * this->v1 + other.v1 * this->w - other.v2 * this->v3 + other.v3 * this->v2);
+    buffer.v2 = (other.w * this->v2 + other.v1 * this->v3 + other.v2 * this->w - other.v3 * this->v1);
+    buffer.v3 = (other.w * this->v3 - other.v1 * this->v2 + other.v2 * this->v1 + other.v3 * this->w);
+    return buffer;
 }
 
-Quaternion &Quaternion::Rotate(Quaternion &other, Quaternion &buffer) const {
-  Quaternion prime{this->w, -this->v1, -this->v2, -this->v3};
-  buffer.v1 = other.v1;
-  buffer.v2 = other.v2;
-  buffer.v3 = other.v3;
-  buffer.w = 0;
+Quaternion &Quaternion::Rotate(Quaternion &other, Quaternion &buffer) const
+{
+    Quaternion prime{this->w, -this->v1, -this->v2, -this->v3};
+    buffer.v1 = other.v1;
+    buffer.v2 = other.v2;
+    buffer.v3 = other.v3;
+    buffer.w = 0;
 
-  Quaternion temp{};
-  this->Q_Mult(buffer, temp);
-  temp.Q_Mult(prime, buffer);
-  return buffer;
+    Quaternion temp{};
+    this->Q_Mult(buffer, temp);
+    temp.Q_Mult(prime, buffer);
+    return buffer;
 }
 
-void Quaternion::Normalize() {
-  float magnitude = sqrt(this->v1 * this->v1 + this->v2 * this->v2 +
-                         this->v3 * this->v3 + this->w * this->w);
-  if (magnitude == 0) {
-    return;
-  }
-  this->v1 /= magnitude;
-  this->v2 /= magnitude;
-  this->v3 /= magnitude;
-  this->w /= magnitude;
+void Quaternion::Normalize()
+{
+    float magnitude = sqrt(this->v1 * this->v1 + this->v2 * this->v2 + this->v3 * this->v3 + this->w * this->w);
+    if (magnitude == 0)
+    {
+        return;
+    }
+    this->v1 /= magnitude;
+    this->v2 /= magnitude;
+    this->v3 /= magnitude;
+    this->w /= magnitude;
 }
 
-Matrix<3, 3> Quaternion::ToRotationMatrix() const {
-  float xx = this->v1 * this->v1;
-  float yy = this->v2 * this->v2;
-  float zz = this->v3 * this->v3;
-  Matrix<3, 3> rotationMatrix{1 - 2 * (yy - zz),
-                              2 * (this->v1 * this->v2 - this->v3 * this->w),
-                              2 * (this->v1 * this->v3 + this->v2 * this->w),
-                              2 * (this->v1 * this->v2 + this->v3 * this->w),
-                              1 - 2 * (xx - zz),
-                              2 * (this->v2 * this->v3 - this->v1 * this->w),
-                              2 * (this->v1 * this->v3 - this->v2 * this->w),
-                              2 * (this->v2 * this->v3 + this->v1 * this->w),
-                              1 - 2 * (xx - yy)};
-  return rotationMatrix;
+Matrix<3, 3> Quaternion::ToRotationMatrix() const
+{
+    float xx = this->v1 * this->v1;
+    float yy = this->v2 * this->v2;
+    float zz = this->v3 * this->v3;
+    Matrix<3, 3> rotationMatrix{
+        1 - 2 * (yy - zz), 2 * (this->v1 * this->v2 - this->v3 * this->w), 2 * (this->v1 * this->v3 + this->v2 * this->w),
+        2 * (this->v1 * this->v2 + this->v3 * this->w), 1 - 2 * (xx - zz), 2 * (this->v2 * this->v3 - this->v1 * this->w),
+        2 * (this->v1 * this->v3 - this->v2 * this->w), 2 * (this->v2 * this->v3 + this->v1 * this->w), 1 - 2 * (xx - yy)};
+    return rotationMatrix;
 };
 
-Matrix<3, 1> Quaternion::ToEulerAngle() const {
-  float sqv1 = this->v1 * this->v1;
-  float sqv2 = this->v2 * this->v2;
-  float sqv3 = this->v3 * this->v3;
-  float sqw = this->w * this->w;
+Matrix<3, 1> Quaternion::ToEulerAngle() const
+{
+    float sqv1 = this->v1 * this->v1;
+    float sqv2 = this->v2 * this->v2;
+    float sqv3 = this->v3 * this->v3;
+    float sqw = this->w * this->w;
 
-  Matrix<3, 1> eulerAngle;
-  {
-    atan2(2.0 * (this->v1 * this->v2 + this->v3 * this->w),
-          (sqv1 - sqv2 - sqv3 + sqw));
-    asin(-2.0 * (this->v1 * this->v3 - this->v2 * this->w) /
-         (sqv1 + sqv2 + sqv3 + sqw));
-    atan2(2.0 * (this->v2 * this->v3 + this->v1 * this->w),
-          (-sqv1 - sqv2 + sqv3 + sqw));
-  };
-  return eulerAngle;
+    Matrix<3, 1> eulerAngle;
+    {
+        atan2(2.0 * (this->v1 * this->v2 + this->v3 * this->w), (sqv1 - sqv2 - sqv3 + sqw));
+        asin(-2.0 * (this->v1 * this->v3 - this->v2 * this->w) / (sqv1 + sqv2 + sqv3 + sqw));
+        atan2(2.0 * (this->v2 * this->v3 + this->v1 * this->w), (-sqv1 - sqv2 + sqv3 + sqw));
+    };
+    return eulerAngle;
 }

src/Quaternion.h

@@ -2,89 +2,90 @@
 #define QUATERNION_H_
 
 #include "Matrix.hpp"
-class Quaternion : public Matrix<1, 4> {
+class Quaternion : public Matrix<1, 4>
+{
 public:
-  Quaternion() : Matrix<1, 4>() {}
-  Quaternion(float w, float v1, float v2, float v3)
-      : Matrix<1, 4>(w, v1, v2, v3) {}
-  Quaternion(const Quaternion &q) : Matrix<1, 4>(q.w, q.v1, q.v2, q.v3) {}
-  Quaternion(const Matrix<1, 4> &matrix) : Matrix<1, 4>(matrix) {}
-  Quaternion(const std::array<float, 4> &array) : Matrix<1, 4>(array) {}
+    Quaternion() : Matrix<1, 4>() {}
+    Quaternion(float fillValue) : Matrix<1, 4>(fillValue) {}
+    Quaternion(float w, float v1, float v2, float v3) : Matrix<1, 4>(w, v1, v2, v3) {}
+    Quaternion(const Quaternion &q) : Matrix<1, 4>(q.w, q.v1, q.v2, q.v3) {}
+    Quaternion(const Matrix<1, 4> &matrix) : Matrix<1, 4>(matrix) {}
+    Quaternion(const std::array<float, 4> &array) : Matrix<1, 4>(array) {}
 
-  /**
-   * @brief Create a quaternion from an angle and axis
-   * @param angle The angle to rotate by
-   * @param axis The axis to rotate around
-   */
-  static Quaternion FromAngleAndAxis(float angle, const Matrix<1, 3> &axis);
+    /**
+     * @brief Create a quaternion from an angle and axis
+     * @param angle The angle to rotate by
+     * @param axis The axis to rotate around
+     */
+    static Quaternion FromAngleAndAxis(float angle, const Matrix<1, 3> &axis);
 
-  /**
-   * @brief Access the elements of the quaternion
-   * @param index The index of the element to access
-   * @return The value of the element at the index
-   */
-  float operator[](uint8_t index) const;
+    /**
+     * @brief Access the elements of the quaternion
+     * @param index The index of the element to access
+     * @return The value of the element at the index
+     */
+    float operator[](uint8_t index) const;
 
-  /**
-   * @brief Assign one quaternion to another
-   */
-  void operator=(const Quaternion &other);
+    /**
+     * @brief Assign one quaternion to another
+     */
+    void operator=(const Quaternion &other);
 
-  /**
-   * @brief Do quaternion multiplication
-   */
-  Quaternion operator*(const Quaternion &other) const;
+    /**
+     * @brief Do quaternion multiplication
+     */
+    Quaternion operator*(const Quaternion &other) const;
 
-  /**
-   * @brief Multiply the quaternion by a scalar
-   */
-  Quaternion operator*(float scalar) const;
+    /**
+     * @brief Multiply the quaternion by a scalar
+     */
+    Quaternion operator*(float scalar) const;
 
-  /**
-   * @brief Add two quaternions together
-   * @param other The quaternion to add to this one
-   * @return The net quaternion
-   */
-  Quaternion operator+(const Quaternion &other) const;
+    /**
+     * @brief Add two quaternions together
+     * @param other The quaternion to add to this one
+     * @return The net quaternion
+     */
+    Quaternion operator+(const Quaternion &other) const;
 
-  /**
-   * @brief Q_Mult a quaternion by another quaternion
-   * @param other The quaternion to rotate by
-   * @param buffer The buffer to store the result in
-   * @return A reference to the buffer
-   */
-  Quaternion &Q_Mult(const Quaternion &other, Quaternion &buffer) const;
+    /**
+     * @brief Q_Mult a quaternion by another quaternion
+     * @param other The quaternion to rotate by
+     * @param buffer The buffer to store the result in
+     * @return A reference to the buffer
+     */
+    Quaternion &Q_Mult(const Quaternion &other, Quaternion &buffer) const;
 
-  /**
-   * @brief Rotate a quaternion by this quaternion
-   * @param other The quaternion to rotate
-   * @param buffer The buffer to store the result in
-   *
-   */
-  Quaternion &Rotate(Quaternion &other, Quaternion &buffer) const;
+    /**
+     * @brief Rotate a quaternion by this quaternion
+     * @param other The quaternion to rotate
+     * @param buffer The buffer to store the result in
+     *
+     */
+    Quaternion &Rotate(Quaternion &other, Quaternion &buffer) const;
 
-  /**
-   * @brief Normalize the quaternion to a magnitude of 1
-   */
-  void Normalize();
+    /**
+     * @brief Normalize the quaternion to a magnitude of 1
+     */
+    void Normalize();
 
-  /**
-   * @brief Convert the quaternion to a rotation matrix
-   * @return The rotation matrix
-   */
-  Matrix<3, 3> ToRotationMatrix() const;
+    /**
+     * @brief Convert the quaternion to a rotation matrix
+     * @return The rotation matrix
+     */
+    Matrix<3, 3> ToRotationMatrix() const;
 
-  /**
-   * @brief Convert the quaternion to an Euler angle representation
-   * @return The Euler angle representation of the quaternion
-   */
-  Matrix<3, 1> ToEulerAngle() const;
+    /**
+     * @brief Convert the quaternion to an Euler angle representation
+     * @return The Euler angle representation of the quaternion
+     */
+    Matrix<3, 1> ToEulerAngle() const;
 
-  // Give people an easy way to access the elements
-  float &w{matrix[0]};
-  float &v1{matrix[1]};
-  float &v2{matrix[2]};
-  float &v3{matrix[3]};
+    // Give people an easy way to access the elements
+    float &w{matrix[0]};
+    float &v1{matrix[1]};
+    float &v2{matrix[2]};
+    float &v3{matrix[3]};
 };
 
 #endif // QUATERNION_H_

src/SVD.hpp

@@ -1,411 +0,0 @@
-#pragma once
-#include "Matrix.hpp"
-
-/**
- * @brief library that uses Matrix.hpp and performs SVD on a matrix
- *
- * @note Fully templated: SVD works for ANY Matrix<R, C> with R, C in
- *       1..255 (the uint8_t range of Matrix). There is no 5×5 limit.
- *
- * @note EMBEDDED CONSTRAINT — no heap. All working storage is stack
- *       allocated as templated Matrix<N,N> buffers where
- *       N = max(R, C). Peak stack usage per SVD call is
- *       ≈ 11·N² floats (≈ 44·N² bytes):
- *         N = 5   → ~1.1 KB
- *         N = 10  → ~4.4 KB
- *         N = 20  → ~18 KB
- *         N = 50  → ~110 KB
- *         N = 100 → ~440 KB
- *         N = 255 → ~2.9 MB
- *       Instantiate only the sizes that fit your call-stack budget.
- */
-namespace SVD {
-/**
- * @brief Compute the Singular Value Decomposition (SVD) of this matrix.
- *
- * Decomposes A into U × Σ × Vᵀ where:
- *   - U is an m×k orthogonal matrix (left singular vectors)
- *   - Σ is a k×k diagonal matrix with non-negative singular values
- *     (stored as a k×1 column vector)
- *   - Vᵀ is a k×n orthogonal matrix (right singular vectors, transposed)
- *   - k = min(m, n)
- *
- * The decomposition satisfies: A ≈ U × diag(Σ) × Vᵀ
- * Singular values are returned in descending order.
- *
- * Output storage conventions:
- *   - U: Matrix<rows, columns> — first k columns are meaningful
- *     (rows k..columns−1 are zero in the wide case)
- *   - sigma: Matrix<columns, 1> — first k entries are the singular
- *     values; entries beyond k (wide matrices only) are zero
- *   - Vt: Matrix<columns, columns> — first k rows are meaningful
- *     (zero-padded in the tall case)
- *
- * For wide matrices (rows < columns) the SVD is computed on Aᵀ and the
- * factors are swapped back.
- *
- * @tparam rows Number of rows in A (1..255)
- * @tparam columns Number of columns in A (1..255)
- * @param matrixToDecompose Input: the matrix A
- * @param U Output: left singular vectors (rows×columns matrix)
- * @param sigma Output: singular values (columns×1 vector, sorted descending)
- * @param Vt Output: right singular vectors transposed (columns×columns)
- *
- * @note This implementation uses Householder bidiagonalization followed
- *       by block reduction: 2×2 blocks via closed form, larger blocks
- *       via cyclic Jacobi eigen-decomposition of BᵀB with residual
- *       singular values σᵢ = ‖B·vᵢ‖ (see docs/svd-refactor.md).
- */
-template <uint8_t rows, uint8_t columns>
-void SVD(Matrix<rows, columns> &matrixToDecompose, Matrix<rows, columns> &U,
-         Matrix<columns, 1> &sigma, Matrix<columns, columns> &Vt);
-
-// ========================================================================
-// SVD Building Block Functions (for unit testing)
-//
-// Templated on the working-buffer size N. All block operations work on
-// N×N matrices with runtime bounds (m, n, p, blockSize, ...) — the
-// regions beyond the bounds are zero-padded working space.
-//
-// N is deduced from the Matrix arguments at the call site, e.g.
-//   Matrix<8, 8> W, QL, QR;
-//   SVD::Bidiagonalize(W, 6, 8, 6, QL, QR);   // N = 8 deduced
-// ========================================================================
-
-/**
- * @brief Compute a Householder reflector vector.
- *
- * Given input vector x, computes normalized v and scalar alpha such that:
- *   (I - 2·v·vᵀ) · x = [alpha, 0, 0, ...]ᵀ
- *
- * @param x Input vector (up to len elements)
- * @param len Number of valid elements in x
- * @param v Output: normalized Householder vector (length ≥ len)
- * @param alpha Output: the resulting first element after reflection
- * @return The norm of the input vector x
- */
-static float ComputeHouseholder(const float *x, uint8_t len, float *v,
-                                float &alpha);
-
-/**
- * @brief Apply a Householder reflection from the left.
- *
- * Transforms W = (I - 2·v·vᵀ) · W where v operates on rows [startRow..endRow]
- * and is applied across all N columns (zero-padded columns are a no-op).
- *
- * @tparam N Working buffer size
- * @param W Input/output: matrix to transform
- * @param v Householder vector (length = endRow - startRow + 1)
- * @param startRow First row index
- * @param endRow Last row index
- */
-template <uint8_t N>
-static void ApplyHouseholderLeft(Matrix<N, N> &W, const float *v,
-                                 uint8_t startRow, uint8_t endRow);
-
-/**
- * @brief Apply a Householder reflection from the right.
- *
- * Transforms W = W · (I - 2·v·vᵀ) where v operates on columns
- * [startCol..endCol] and is applied across all N rows (zero-padded rows
- * are a no-op).
- *
- * @tparam N Working buffer size
- * @param W Input/output: matrix to transform
- * @param v Householder vector (length = endCol - startCol + 1)
- * @param startCol First column index
- * @param endCol Last column index
- */
-template <uint8_t N>
-static void ApplyHouseholderRight(Matrix<N, N> &W, const float *v,
-                                  uint8_t startCol, uint8_t endCol);
-
-/**
- * @brief Reduce a matrix to upper bidiagonal form using Householder reflections.
- *
- * Applies a sequence of Householder reflections to reduce the input matrix
- * W (m×q, where q ≥ p, stored in N×N working space) to upper bidiagonal
- * form B (p×q), accumulating the left and right transformation matrices
- * in QL and QR respectively.
- *
- * Algorithm (Golub-Kahan bidiagonalization):
- *   For k = 0 to p-1:
- *     1. Left HH on column k, rows k..m-1: zero out subdiagonal below B[k+1][k]
- *     2. Right HH on row k, cols k+2..q-1: zero out superdiagonal above B[k][k+1]
- *
- * The accumulated transformations satisfy:
- *   QLᵀ · W_original · QR = B (upper bidiagonal)
- *
- * @tparam N Working buffer size (≥ m and ≥ q)
- * @param W Input/output: matrix to bidiagonalize (first m×q used)
- * @param m Number of rows in the working matrix
- * @param q Number of columns in the working matrix (q ≥ p)
- * @param p Rank = min(m, original_columns) — number of bidiagonalization steps
- * @param QL Input/output: left Householder accumulation (initialized to identity)
- * @param QR Input/output: right Householder accumulation (initialized to identity)
- */
-template <uint8_t N>
-static void Bidiagonalize(Matrix<N, N> &W, uint8_t m, uint8_t q, uint8_t p,
-                          Matrix<N, N> &QL, Matrix<N, N> &QR);
-
-/**
- * @brief Deflate a bidiagonal matrix by zeroing negligible superdiagonals.
- *
- * Scans the p×p upper-bidiagonal matrix stored in W and zeros out any
- * superdiagonal element W[i][i+1] whose magnitude is negligible relative
- * to the local diagonal scale (|W[i][i]| + |W[i+1][i+1]|). Deflating
- * splits the matrix into independent unreduced blocks that can each be
- * solved separately.
- *
- * @tparam N Working buffer size
- * @param W Input/output: bidiagonal matrix (first p×p used)
- * @param p Size of the bidiagonal matrix (min(rows, columns))
- * @param tol Relative deflation tolerance (e.g. 1e-8f)
- */
-template <uint8_t N>
-static void DeflateBidiagonal(Matrix<N, N> &W, uint8_t p, float tol);
-
-/**
- * @brief Check whether a bidiagonal matrix has fully reduced to diagonal.
- *
- * Returns true when every superdiagonal element of the p×p bidiagonal
- * matrix in W is (numerically) zero, i.e. the diagonal entries are the
- * (unsorted) singular values and no unreduced blocks remain.
- *
- * @tparam N Working buffer size
- * @param W Input: bidiagonal matrix (first p×p used)
- * @param p Size of the bidiagonal matrix (min(rows, columns))
- * @param tol Numerical zero threshold multiplier
- * @return true when all superdiagonal elements are ~0
- */
-template <uint8_t N>
-static bool BidiagonalIsDiagonal(const Matrix<N, N> &W, uint8_t p, float tol);
-
-/**
- * @brief Compute the full SVD of a 2×2 upper-bidiagonal block (pure).
- *
- * Decomposes B = [[a, b], [0, d]] as:
- *   B = Ublock · diag(sigma[0], sigma[1]) · Vblockᵀ
- *
- * Guarantees:
- *   - sigma[0] ≥ sigma[1] ≥ 0 (singular values, from eigenvalues of BᵀB)
- *   - Ublock and Vblock are orthogonal (columns are the left/right
- *     singular vectors respectively; Vblock = scipy's Vᵀᵀ)
- *   - Ublock · diag(sigma) · Vblockᵀ == B (within float tolerance)
- *
- * Math: eigenvectors of BᵀB = [[a², ab], [ab, b²+d²]] give the right
- * singular vectors (v1 = normalize(ab, σ1²−a²) with a safe fallback when
- * that vector is ~0; v2 = (−v1y, v1x)); left singular vectors are
- * uᵢ = B·vᵢ/σᵢ with a rank-deficiency guard: when σᵢ ≈ 0 (i.e. ~1e-30),
- * that U column is filled with the signed orthogonal complement of the
- * other U column instead of dividing by ~0.
- *
- * @param a B[0][0] (first diagonal element)
- * @param b B[0][1] (superdiagonal element)
- * @param d B[1][1] (second diagonal element)
- * @param Ublock Output: 2×2 left singular vectors (columns)
- * @param Vblock Output: 2×2 right singular vectors (columns)
- * @param sigma Output: singular values, sigma[0] ≥ sigma[1] ≥ 0
- */
-static void SolveBidiagonalBlock2x2(float a, float b, float d,
-                                    float Ublock[2][2], float Vblock[2][2],
-                                    float sigma[2]);
-
-/**
- * @brief Cyclic Jacobi eigenvalue algorithm for a symmetric matrix (pure).
- *
- * Reduces symmetric n×n matrix T to (near-)diagonal form IN PLACE using
- * cyclic Jacobi rotations, accumulating the eigenvectors in V.
- *
- * On return:
- *   - T's diagonal entries are the eigenvalues (off-diagonals ~0)
- *   - evals[i] = T[i][i], UNSORTED, SIGNED (this is a general symmetric
- *     eigen solver, not just for PSD matrices like T = BᵀB)
- *   - columns of V are the corresponding eigenvectors (T·V = V·Λ)
- *
- * Convergence: relative off-diagonal tolerance 1e-10, hard-capped at
- * 100 sweeps.
- *
- * @tparam N Working buffer size (≥ n)
- * @param T Input/output: symmetric matrix (first n×n used, destroyed in place)
- * @param n Matrix size
- * @param evals Output: eigenvalues, unsorted, length ≥ n
- * @param V Output: eigenvector matrix (first n×n used), columns are eigenvectors
- */
-template <uint8_t N>
-static void JacobiEigenSymmetric(Matrix<N, N> &T, uint8_t n, float *evals,
-                                 Matrix<N, N> &V);
-
-/**
- * @brief Fold a block SVD's factors into the QL/QR accumulators.
- *
- * Given the block SVD of a bidiagonal block, B = Ublock·Σ·Vblockᵀ, the
- * accumulated Householder matrices must absorb the block factors:
- *   QL[:, blockStart..blockStart+blockSize−1] ← QL[:, ...] · Ublock
- *     (rows 0..rowsQL−1)
- *   QR[:, blockStart..blockStart+blockSize−1] ← QR[:, ...] · Vblock
- *     (rows 0..rowsQR−1)
- *
- * rowsQL / rowsQR are the meaningful row extents of the accumulators
- * (e.g. for a wide matrix W = Aᵀ, QL carries n = rows(W) meaningful
- * rows while QR is read back over its first m rows).
- *
- * In-place update is done through temporary buffers (updating QL's block
- * columns while still reading them corrupts the result).
- *
- * @tparam N Working buffer size
- * @param blockStart First column/row index of the block in W
- * @param blockSize Size of the block (2, or > 2 for the Jacobi path)
- * @param Ublock Left singular-vector factor of the block (first blockSize×blockSize used)
- * @param Vblock Right singular-vector factor of the block (first blockSize×blockSize used)
- * @param rowsQL Number of meaningful rows of QL
- * @param rowsQR Number of meaningful rows of QR
- * @param QL Input/output: left transformation accumulator
- * @param QR Input/output: right transformation accumulator
- */
-template <uint8_t N>
-static void ApplyBlockFactorsToAccumulators(uint8_t blockStart,
-                                            uint8_t blockSize,
-                                            const Matrix<N, N> &Ublock,
-                                            const Matrix<N, N> &Vblock,
-                                            uint8_t rowsQL, uint8_t rowsQR,
-                                            Matrix<N, N> &QL,
-                                            Matrix<N, N> &QR);
-
-/**
- * @brief Solve a bidiagonal block larger than 2×2 via Jacobi eigen of BᵀB.
- *
- * Computes the full SVD of the unreduced upper-bidiagonal block
- * W[blockStart..blockStart+blockSize−1] via eigendecomposition of the
- * tridiagonal T = BᵀB:
- *   1. Snapshot the ORIGINAL block diagonal/superdiagonal from W
- *   2. Form T = BᵀB (tridiagonal symmetric)
- *   3. JacobiEigenSymmetric on T → eigenvalues (unsorted) + V
- *   4. Sort eigenvalues descending, reordering V columns
- *   5. Compute RESIDUAL singular values: σᵢ = ‖B_orig · vᵢ‖
- *      (NOT sqrt(eigenvalue) — forming BᵀB squares the condition number,
- *       causing float noise to swamp true tiny eigenvalues for
- *       rank-deficient blocks)
- *   6. Re-sort σ descending, keeping V and B·v consistent
- *   7. Build Ublock: uᵢ = B_orig · vᵢ / σᵢ (unit norm); for σᵢ ≈ 0,
- *      use Gram-Schmidt orthogonal completion against prior U columns
- *   8. Fold Ublock/Vblock into QL/QR via ApplyBlockFactorsToAccumulators
- *   9. Write residual norms into W's diagonal and zero the block's
- *      superdiagonals
- *
- * @tparam N Working buffer size (≥ blockSize)
- * @param W Input/output: bidiagonal matrix; the block's diagonal holds
- *        the singular values and its superdiagonals are zeroed on return
- * @param blockStart First column/row index of the block
- * @param blockSize Size of the block (> 2)
- * @param rowsQL Number of meaningful rows of QL
- * @param rowsQR Number of meaningful rows of QR
- * @param QL Input/output: left transformation accumulator
- * @param QR Input/output: right transformation accumulator
- */
-template <uint8_t N>
-static void SolveBidiagonalBlockJacobi(Matrix<N, N> &W, uint8_t blockStart,
-                                       uint8_t blockSize, uint8_t rowsQL,
-                                       uint8_t rowsQR, Matrix<N, N> &QL,
-                                       Matrix<N, N> &QR);
-
-/**
- * @brief Extract singular values from bidiagonal matrix diagonal and sort.
- *
- * Extracts absolute values of diagonal elements of W as singular values,
- * then sorts them in descending order while reordering columns of QL
- * and QR to maintain consistency. A negative diagonal element flips the
- * sign of the corresponding QL column to keep A = U·Σ·Vᵀ.
- *
- * @tparam N Working buffer size
- * @param W Input: bidiagonal matrix (first p×p used)
- * @param sigma Output: sorted singular values (N×1 column vector, only first p used)
- * @param p Number of singular values (min(rows, columns))
- * @param QL Input/output: left transformation matrix (modified during sort)
- * @param QR Input/output: right transformation matrix (modified during sort)
- */
-template <uint8_t N>
-static void ExtractAndSortSingularValues(Matrix<N, N> &W, Matrix<N, 1> &sigma,
-                                         uint8_t p, Matrix<N, N> &QL,
-                                         Matrix<N, N> &QR);
-
-/**
- * @brief Assemble final U and Vt matrices from QL/QR.
- *
- * Computes the final left singular vectors (U) and right singular vectors
- * transposed (Vt) from the accumulated Householder transformations.
- *
- * For non-transpose case: U = QL[:,0:p], Vt = QR[:,0:p]ᵀ
- * For transpose case:    U = QR[:,0:p], Vt = full QLᵀ (all n rows)
- *
- * @tparam N Working buffer size (≥ m and ≥ n)
- * @param m Number of rows in original matrix
- * @param n Number of columns in original matrix
- * @param p Rank = min(m, n)
- * @param transposeNeeded True if we computed SVD(Aᵀ) instead of SVD(A)
- * @param QL Left Householder accumulation (N×N)
- * @param QR Right Householder accumulation (N×N)
- * @param U Output: left singular vectors (N×N, first m×p used)
- * @param Vt Output: right singular vectors transposed (N×N, first p×n used)
- */
-template <uint8_t N>
-static void AssembleUAndVt(uint8_t m, uint8_t n, uint8_t p,
-                           bool transposeNeeded, const Matrix<N, N> &QL,
-                           const Matrix<N, N> &QR, Matrix<N, N> &U,
-                           Matrix<N, N> &Vt);
-
-/**
- * @brief Compute a Givens rotation that zeros out y.
- *
- * Computes c, s such that:
- *   [c  s] [x] = [r]
- *   [-s c] [y]   [0]
- * where r = sqrt(x² + y²).
- *
- * @param x First element
- * @param y Second element (to be zeroed)
- * @param c Output: cosine of rotation angle
- * @param s Output: sine of rotation angle
- */
-static void ComputeGivens(float x, float y, float &c, float &s);
-
-/**
- * @brief Apply a Givens rotation from the left to rows i and j.
- *
- * Applies [c s; -s c] to rows i, j of W (columns startCol..endCol).
- *
- * @tparam N Working buffer size
- * @param W Input/output: matrix to transform
- * @param i First row index
- * @param j Second row index
- * @param c Cosine of rotation angle
- * @param s Sine of rotation angle
- * @param startCol First column to transform
- * @param endCol Last column to transform
- */
-template <uint8_t N>
-static void ApplyGivensLeft(Matrix<N, N> &W, uint8_t i, uint8_t j, float c,
-                            float s, uint8_t startCol, uint8_t endCol);
-
-/**
- * @brief Apply a Givens rotation from the right to columns i and j.
- *
- * Applies [c -s; s c]ᵀ to columns i, j of W (rows startRow..endRow).
- *
- * @tparam N Working buffer size
- * @param W Input/output: matrix to transform
- * @param i First column index
- * @param j Second column index
- * @param c Cosine of rotation angle
- * @param s Sine of rotation angle
- * @param startRow First row to transform
- * @param endRow Last row to transform
- */
-template <uint8_t N>
-static void ApplyGivensRight(Matrix<N, N> &W, uint8_t i, uint8_t j, float c,
-                             float s, uint8_t startRow, uint8_t endRow);
-} // namespace SVD
-
-#ifndef SVD_H_
-#include "SVD.cpp"
-#endif

unit-tests/CMakeLists.txt

@@ -13,7 +13,6 @@ add_executable(matrix-tests matrix-tests.cpp)
 target_link_libraries(matrix-tests 
     PRIVATE 
     matrix
-    qr
     Catch2::Catch2WithMain
 )
 
@@ -33,34 +32,4 @@ target_link_libraries(vector-3d-tests
     PRIVATE
     vector-3d
     Catch2::Catch2WithMain
-)
-
-# SVD building block tests
-add_executable(svd-build-blocks-tests svd-build-blocks-tests.cpp)
-
-target_link_libraries(svd-build-blocks-tests
-    PRIVATE
-    matrix
-    svd
-    Catch2::Catch2WithMain
-)
-
-# SVD integration tests
-add_executable(svd-integration-test svd-integration-test.cpp)
-
-target_link_libraries(svd-integration-test
-    PRIVATE
-    matrix
-    svd
-    Catch2::Catch2WithMain
-)
-
-# QR building block tests
-add_executable(qr-build-blocks-tests qr-build-blocks-tests.cpp)
-
-target_link_libraries(qr-build-blocks-tests
-    PRIVATE
-    matrix
-    qr
-    Catch2::Catch2WithMain
 )

unit-tests/matrix-timing-tests.cpp

@@ -8,7 +8,6 @@
 // any other libraries
 #include <array>
 #include <cmath>
-#include <cstdint>
 
 // basically re-run all of the matrix tests with huge matrices and time the
 // results.
@@ -30,13 +29,13 @@ TEST_CASE("Timing Tests", "Matrix") {
   Matrix<4, 4> mat5{};
 
   SECTION("Addition") {
-    for (uint32_t i{0}; i < 100000; i++) {
+    for (uint32_t i{0}; i < 10000; i++) {
       mat3 = mat1 + mat2;
     }
   }
 
   SECTION("Subtraction") {
-    for (uint32_t i{0}; i < 100000; i++) {
+    for (uint32_t i{0}; i < 10000; i++) {
       mat3 = mat1 - mat2;
     }
   }
@@ -48,19 +47,19 @@ TEST_CASE("Timing Tests", "Matrix") {
   }
 
   SECTION("Scalar Multiplication") {
-    for (uint32_t i{0}; i < 100000; i++) {
+    for (uint32_t i{0}; i < 10000; i++) {
       mat3 = mat1 * 3;
     }
   }
 
   SECTION("Element Multiply") {
-    for (uint32_t i{0}; i < 100000; i++) {
+    for (uint32_t i{0}; i < 10000; i++) {
       mat1.ElementMultiply(mat2, mat3);
     }
   }
 
   SECTION("Element Divide") {
-    for (uint32_t i{0}; i < 100000; i++) {
+    for (uint32_t i{0}; i < 10000; i++) {
       mat1.ElementDivide(mat2, mat3);
     }
   }
@@ -69,60 +68,52 @@ TEST_CASE("Timing Tests", "Matrix") {
     // what about matrices of 0,0 or 1,1?
     // minor matrix for 2x2 matrix
     Matrix<49, 49> minorMat1{};
-    for (uint32_t i{0}; i < 100000; i++) {
+    for (uint32_t i{0}; i < 10000; i++) {
       mat1.MinorMatrix(minorMat1, 0, 0);
     }
   }
 
   SECTION("Determinant") {
-    for (uint32_t i{0}; i < 1000000; i++) {
-      float det = mat4.Det();
-      (void)det;
+    for (uint32_t i{0}; i < 100000; i++) {
+      float det1 = mat4.Det();
     }
   }
 
   SECTION("Matrix of Minors") {
-    for (uint32_t i{0}; i < 1000000; i++) {
+    for (uint32_t i{0}; i < 100000; i++) {
       mat4.MatrixOfMinors(mat5);
     }
   }
 
   SECTION("Invert") {
-    for (uint32_t i{0}; i < 1000000; i++) {
+    for (uint32_t i{0}; i < 100000; i++) {
       mat5 = mat4.Invert();
     }
   };
 
   SECTION("Transpose") {
-    for (uint32_t i{0}; i < 100000; i++) {
+    for (uint32_t i{0}; i < 10000; i++) {
       mat3 = mat1.Transpose();
     }
   }
 
   SECTION("Normalize") {
-    for (uint32_t i{0}; i < 100000; i++) {
-      mat3 = mat1 / mat1.EuclideanNorm();
+    for (uint32_t i{0}; i < 10000; i++) {
+      mat1.Normalize(mat3);
     }
   }
 
   SECTION("GET ROW") {
     Matrix<1, 50> mat1Rows{};
-    for (uint32_t i{0}; i < 100000000; i++) {
+    for (uint32_t i{0}; i < 1000000; i++) {
       mat1.GetRow(0, mat1Rows);
     }
   }
 
   SECTION("GET COLUMN") {
     Matrix<50, 1> mat1Columns{};
-    for (uint32_t i{0}; i < 100000000; i++) {
+    for (uint32_t i{0}; i < 1000000; i++) {
       mat1.GetColumn(0, mat1Columns);
     }
   }
-
-  SECTION("QR Decomposition") {
-    Matrix<50, 50> Q, R{};
-    for (uint32_t i{0}; i < 500; i++) {
-      mat1.QRDecomposition(Q, R);
-    }
-  }
 }

unit-tests/qr-build-blocks-tests.cpp

@@ -1,581 +0,0 @@
-// include the unit test framework first
-#include <catch2/catch_test_macros.hpp>
-#include <catch2/matchers/catch_matchers_floating_point.hpp>
-
-// include the module you're going to test next
-#include "Matrix.hpp"
-#include "QR.hpp"
-
-// any other libraries
-#include <array>
-#include <cmath>
-#include <iostream>
-
-// ============================================================================
-// Helpers
-// ============================================================================
-
-/**
- * @brief Frobenius norm of an N x N matrix.
- */
-template <uint8_t N>
-static float frob(const Matrix<N, N> &M) {
-  float sum = 0.0f;
-  for (uint8_t i = 0; i < N; i++)
-    for (uint8_t j = 0; j < N; j++) {
-      float v = M.Get(i, j);
-      sum += v * v;
-    }
-  return sqrtf(sum);
-}
-
-/**
- * @brief Check M is orthogonal (M^T M ~ I).
- */
-template <uint8_t N>
-static bool isOrthogonal(const Matrix<N, N> &M, float tol = 1e-5f) {
-  Matrix<N, N> Mt = M.Transpose();
-  Matrix<N, N> MtM{};
-  Mt.Mult(M, MtM);
-  for (uint8_t i = 0; i < N; i++)
-    for (uint8_t j = 0; j < N; j++) {
-      float expected = (i == j) ? 1.0f : 0.0f;
-      if (fabsf(MtM.Get(i, j) - expected) > tol)
-        return false;
-    }
-  return true;
-}
-
-/**
- * @brief 3x3 trace.
- */
-static float trace3(const Matrix<3, 3> &A) {
-  return A.Get(0, 0) + A.Get(1, 1) + A.Get(2, 2);
-}
-
-/**
- * @brief 3x3 sum of principal 2x2 minors (2nd elementary invariant).
- */
-static float e2_3x3(const Matrix<3, 3> &A) {
-  return A.Get(0, 0) * A.Get(1, 1) - A.Get(0, 1) * A.Get(0, 1) +
-         A.Get(0, 0) * A.Get(2, 2) - A.Get(0, 2) * A.Get(0, 2) +
-         A.Get(1, 1) * A.Get(2, 2) - A.Get(1, 2) * A.Get(1, 2);
-}
-
-/**
- * @brief 3x3 determinant.
- */
-static float det3(const Matrix<3, 3> &A) {
-  return A.Get(0, 0) *
-             (A.Get(1, 1) * A.Get(2, 2) - A.Get(1, 2) * A.Get(2, 1)) -
-         A.Get(0, 1) *
-             (A.Get(1, 0) * A.Get(2, 2) - A.Get(1, 2) * A.Get(2, 0)) +
-         A.Get(0, 2) *
-             (A.Get(1, 0) * A.Get(2, 1) - A.Get(1, 1) * A.Get(2, 0));
-}
-
-/**
- * @brief Sign-invariant comparison of |actual| against refAbs.
- */
-static bool matchesAbs(float actual, float refAbs, float relTol = 1e-5f,
-                       float absTol = 1e-6f) {
-  float a = fabsf(actual);
-  if (refAbs < 1e-3f)
-    return a < absTol + relTol;
-  return fabsf(a - refAbs) <= relTol * refAbs;
-}
-
-// ============================================================================
-// TEST 1: GivensRotation
-// ============================================================================
-TEST_CASE("QR Building Block: GivensRotation", "[Matrix][QR]") {
-  // R = [[c, s], [-s, c]] must satisfy R * (a, b)^T = (r, 0)^T.
-
-  {
-    // Reference: hypot(2, 1) = sqrt(5) = 2.236067977
-    float c = 0, s = 0;
-    QR::GivensRotation(2.0f, 1.0f, c, s);
-    REQUIRE_THAT(c, Catch::Matchers::WithinRel(0.894427191f, 1e-6f));
-    REQUIRE_THAT(s, Catch::Matchers::WithinRel(0.447213595f, 1e-6f));
-    REQUIRE_THAT(c * 2.0f + s * 1.0f,
-                 Catch::Matchers::WithinRel(2.236067977f, 1e-6f));
-    REQUIRE_THAT(-s * 2.0f + c * 1.0f, Catch::Matchers::WithinAbs(0.0f, 1e-6f));
-  }
-
-  {
-    // Reference: hypot(3, 4) = 5 exactly
-    float c = 0, s = 0;
-    QR::GivensRotation(3.0f, 4.0f, c, s);
-    REQUIRE_THAT(c, Catch::Matchers::WithinRel(0.6f, 1e-6f));
-    REQUIRE_THAT(s, Catch::Matchers::WithinRel(0.8f, 1e-6f));
-    REQUIRE_THAT(c * 3.0f + s * 4.0f, Catch::Matchers::WithinRel(5.0f, 1e-6f));
-    REQUIRE_THAT(-s * 3.0f + c * 4.0f, Catch::Matchers::WithinAbs(0.0f, 1e-6f));
-  }
-
-  {
-    // Pure second component: c = 0, s = 1
-    float c = 1, s = 1;
-    QR::GivensRotation(0.0f, 5.0f, c, s);
-    REQUIRE_THAT(c, Catch::Matchers::WithinAbs(0.0f, 1e-7f));
-    REQUIRE_THAT(s, Catch::Matchers::WithinRel(1.0f, 1e-6f));
-  }
-
-  {
-    // Zero vector: identity rotation
-    float c = 0, s = 0;
-    QR::GivensRotation(0.0f, 0.0f, c, s);
-    REQUIRE_THAT(c, Catch::Matchers::WithinRel(1.0f, 1e-7f));
-    REQUIRE_THAT(s, Catch::Matchers::WithinAbs(0.0f, 1e-7f));
-  }
-
-  {
-    // Negative first component preserves the sign of c
-    float c = 0, s = 0;
-    QR::GivensRotation(-2.0f, 1.0f, c, s);
-    REQUIRE_THAT(c, Catch::Matchers::WithinRel(-0.894427191f, 1e-6f));
-    REQUIRE_THAT(s, Catch::Matchers::WithinRel(0.447213595f, 1e-6f));
-    REQUIRE_THAT(-s * -2.0f + c * 1.0f, Catch::Matchers::WithinAbs(0.0f, 1e-6f));
-  }
-}
-
-// ============================================================================
-// TEST 2: ApplyRotationBothSides (similarity A <- G A G^T)
-// ============================================================================
-TEST_CASE("QR Building Block: ApplyRotationBothSides", "[Matrix][QR]") {
-  // Reference (numpy, float64): A = [[2,1,0],[1,3,1],[0,1,4]], i = 0,
-  // Givens(2,1) -> G A G^T =
-  // [[ 3.0,         1.0,           0.447213595],
-  //  [ 1.0,         2.0,           0.894427191],
-  //  [ 0.447213595, 0.894427191,   4.0]]
-  // (Note: G A G^T with G zeroing (2,1) sends the A[0][1] coupling into the
-  // (0,2) corner, NOT into the subdiagonal -- the subdiagonal-zeroing happens
-  // in the QR chase context where the bulge column has the right shape.)
-  {
-    Matrix<3, 3> A{2, 1, 0, 1, 3, 1, 0, 1, 4};
-    float c = 0.894427191f, s = 0.447213595f;
-
-    QR::ApplyRotationBothSides(A, 0, c, s);
-
-    REQUIRE_THAT(A.Get(0, 0), Catch::Matchers::WithinRel(3.0f, 1e-5f));
-    REQUIRE_THAT(A.Get(0, 1), Catch::Matchers::WithinRel(1.0f, 1e-5f));
-    REQUIRE_THAT(A.Get(0, 2),
-                 Catch::Matchers::WithinRel(0.447213595f, 1e-5f));
-    REQUIRE_THAT(A.Get(1, 1), Catch::Matchers::WithinRel(2.0f, 1e-5f));
-    REQUIRE_THAT(A.Get(1, 2),
-                 Catch::Matchers::WithinRel(0.894427191f, 1e-5f));
-    REQUIRE_THAT(A.Get(2, 2), Catch::Matchers::WithinRel(4.0f, 1e-5f));
-
-    // Symmetry must be preserved exactly in both triangles
-    for (uint8_t i = 0; i < 3; i++)
-      for (uint8_t j = 0; j < 3; j++)
-        REQUIRE(A.Get(i, j) == A.Get(j, i));
-  }
-
-  // Same check at i = 1.
-  // Reference (numpy, float64): B = [[5,0,1],[0,6,2],[1,2,7]], i = 1,
-  // Givens(6,2) -> G B G^T =
-  // [[ 5.0,          0.316227766,  0.948683298],
-  //  [ 0.316227766,   7.3,          1.9],
-  //  [ 0.948683298,   1.9,          5.7]]
-  {
-    Matrix<3, 3> B{5, 0, 1, 0, 6, 2, 1, 2, 7};
-    float c = 0.948683298f, s = 0.316227766f;
-
-    QR::ApplyRotationBothSides(B, 1, c, s);
-
-    REQUIRE_THAT(B.Get(0, 0), Catch::Matchers::WithinRel(5.0f, 1e-5f));
-    REQUIRE_THAT(B.Get(0, 1),
-                 Catch::Matchers::WithinRel(0.316227766f, 1e-5f));
-    REQUIRE_THAT(B.Get(0, 2),
-                 Catch::Matchers::WithinRel(0.948683298f, 1e-5f));
-    REQUIRE_THAT(B.Get(1, 1), Catch::Matchers::WithinRel(7.3f, 1e-5f));
-    REQUIRE_THAT(B.Get(1, 2), Catch::Matchers::WithinRel(1.9f, 1e-5f));
-    REQUIRE_THAT(B.Get(2, 2), Catch::Matchers::WithinRel(5.7f, 1e-5f));
-
-    for (uint8_t i = 0; i < 3; i++)
-      for (uint8_t j = 0; j < 3; j++)
-        REQUIRE(B.Get(i, j) == B.Get(j, i));
-  }
-
-  // Identity rotation leaves the matrix unchanged
-  {
-    Matrix<3, 3> C{1, 2, 3, 2, 4, 5, 3, 5, 6};
-    QR::ApplyRotationBothSides(C, 1, 1.0f, 0.0f);
-    REQUIRE(C.Get(0, 0) == 1.0f);
-    REQUIRE(C.Get(0, 1) == 2.0f);
-    REQUIRE(C.Get(0, 2) == 3.0f);
-    REQUIRE(C.Get(1, 1) == 4.0f);
-    REQUIRE(C.Get(1, 2) == 5.0f);
-    REQUIRE(C.Get(2, 2) == 6.0f);
-  }
-
-  // Spectrum invariants (trace, Frobenius norm) are preserved. (c, s)
-  // must be a unit vector for G A G^T to be a similarity transform.
-  {
-    Matrix<3, 3> D{1, 2, 3, 2, 5, 8, 3, 8, 9};
-    float tr = trace3(D);
-    float fn = frob(D);
-    float c = 0.6f, s = 0.8f;
-    QR::ApplyRotationBothSides(D, 0, c, s);
-    REQUIRE_THAT(trace3(D), Catch::Matchers::WithinRel(tr, 1e-5f));
-    REQUIRE_THAT(frob(D), Catch::Matchers::WithinRel(fn, 1e-5f));
-  }
-}
-// ============================================================================
-// TEST 3: ApplyRotationToVectors (V <- V G^T)
-// ============================================================================
-TEST_CASE("QR Building Block: ApplyRotationToVectors", "[Matrix][QR]") {
-  // V = I, i = 0, Givens(2,1): V <- I * G^T with G^T = [[c, -s], [s, c]] =
-  // [[ c,  -s,  0],
-  //  [ s,   c,  0],
-  //  [ 0,   0,  1]]
-  {
-    Matrix<3, 3> V{0};
-    V[0][0] = 1;
-    V[1][1] = 1;
-    V[2][2] = 1;
-    float c = 0.894427191f, s = 0.447213595f;
-
-    QR::ApplyRotationToVectors(V, 0, c, s);
-
-    REQUIRE_THAT(V.Get(0, 0), Catch::Matchers::WithinRel(0.894427191f, 1e-6f));
-    REQUIRE_THAT(V.Get(0, 1), Catch::Matchers::WithinRel(-0.447213595f, 1e-6f));
-    REQUIRE_THAT(V.Get(0, 2), Catch::Matchers::WithinAbs(0.0f, 1e-7f));
-    REQUIRE_THAT(V.Get(1, 0), Catch::Matchers::WithinRel(0.447213595f, 1e-6f));
-    REQUIRE_THAT(V.Get(1, 1), Catch::Matchers::WithinRel(0.894427191f, 1e-6f));
-    REQUIRE_THAT(V.Get(1, 2), Catch::Matchers::WithinAbs(0.0f, 1e-7f));
-    REQUIRE_THAT(V.Get(2, 0), Catch::Matchers::WithinAbs(0.0f, 1e-7f));
-    REQUIRE_THAT(V.Get(2, 1), Catch::Matchers::WithinAbs(0.0f, 1e-7f));
-    REQUIRE_THAT(V.Get(2, 2), Catch::Matchers::WithinRel(1.0f, 1e-7f));
-
-    // Product of rotations must stay orthogonal
-    REQUIRE(isOrthogonal(V));
-  }
-
-  // Two successive rotations accumulate (V <- V G1^T G2^T)
-  // Reference (numpy, float64):
-  // [[ 0.894427191, -0.424264069,  0.141421356],
-  //  [ 0.447213595,  0.848528137, -0.282842712],
-  //  [ 0.0,          0.316227766,  0.948683298]]
-  {
-    Matrix<3, 3> V{0};
-    V[0][0] = 1;
-    V[1][1] = 1;
-    V[2][2] = 1;
-    QR::ApplyRotationToVectors(V, 0, 0.894427191f, 0.447213595f);
-    QR::ApplyRotationToVectors(V, 1, 0.948683298f, 0.316227766f);
-    REQUIRE(isOrthogonal(V));
-    // Column 0 was only touched by the first rotation
-    REQUIRE_THAT(V.Get(0, 0), Catch::Matchers::WithinRel(0.894427191f, 1e-5f));
-    REQUIRE_THAT(V.Get(1, 0), Catch::Matchers::WithinRel(0.447213595f, 1e-5f));
-    REQUIRE_THAT(V.Get(2, 0), Catch::Matchers::WithinAbs(0.0f, 1e-7f));
-    REQUIRE_THAT(V.Get(0, 1), Catch::Matchers::WithinRel(-0.424264069f, 1e-5f));
-    REQUIRE_THAT(V.Get(0, 2), Catch::Matchers::WithinRel(0.141421356f, 1e-5f));
-    REQUIRE_THAT(V.Get(1, 2), Catch::Matchers::WithinRel(-0.282842712f, 1e-5f));
-    REQUIRE_THAT(V.Get(2, 1), Catch::Matchers::WithinRel(0.316227766f, 1e-5f));
-    REQUIRE_THAT(V.Get(2, 2), Catch::Matchers::WithinRel(0.948683298f, 1e-5f));
-  }
-}
-
-// ============================================================================
-// TEST 4: WilkinsonShift
-// ============================================================================
-TEST_CASE("QR Building Block: WilkinsonShift", "[Matrix][QR]") {
-  // mu = (a+d)/2 - sign(a-d) * sqrt(((a-d)/2)^2 + b^2)
-  // Reference: eigenvalues of [[2,1],[1,4]] are 1.5858, 4.4142; closest
-  // to d = 4 is 4.414213562.
-  REQUIRE_THAT(QR::WilkinsonShift(2.0f, 1.0f, 4.0f),
-               Catch::Matchers::WithinRel(4.414213562f, 1e-6f));
-
-  // [[5,2],[2,1]]: eigenvalues 0.1716, 5.8284; closest to d = 1 is 0.171572875
-  REQUIRE_THAT(QR::WilkinsonShift(5.0f, 2.0f, 1.0f),
-               Catch::Matchers::WithinRel(0.171572875f, 1e-5f));
-
-  // Zero off-diagonal: returns d itself (sign(0) = +1 picks d, not a)
-  REQUIRE_THAT(QR::WilkinsonShift(3.0f, 0.0f, 7.0f),
-               Catch::Matchers::WithinRel(7.0f, 1e-7f));
-  REQUIRE_THAT(QR::WilkinsonShift(7.0f, 0.0f, 3.0f),
-               Catch::Matchers::WithinRel(3.0f, 1e-7f));
-
-  // a == d: shift is the larger-magnitude off-diagonal combination
-  // [[1,3],[3,1]]: eigenvalues -2, 4; closest to d = 1 is -2
-  REQUIRE_THAT(QR::WilkinsonShift(1.0f, 3.0f, 1.0f),
-               Catch::Matchers::WithinRel(-2.0f, 1e-6f));
-}
-
-// ============================================================================
-// TEST 5: Solve2x2Eigen
-// ============================================================================
-TEST_CASE("QR Building Block: Solve2x2Eigen", "[Matrix][QR]") {
-  // Symmetric block [[2,1],[1,3]]:
-  // eigenvalues 1.381966011, 3.618033989;
-  // eigenvector of 3.618033989 is +/- (0.525731112, 0.850650808)
-  {
-    Matrix<2, 2> A{2, 1, 1, 3};
-    float lHi = 0, lLo = 0, c = 0, s = 0;
-    QR::Solve2x2Eigen(A, 0, lHi, lLo, c, s);
-
-    REQUIRE_THAT(lHi, Catch::Matchers::WithinRel(3.618033989f, 1e-6f));
-    REQUIRE_THAT(lLo, Catch::Matchers::WithinRel(1.381966011f, 1e-6f));
-    REQUIRE(matchesAbs(c, 0.525731112f));
-    REQUIRE(matchesAbs(s, 0.850650808f));
-
-    // Residual: A * vHi = lHi * vHi with vHi = (c, s)
-    REQUIRE_THAT(c * 2.0f + s * 1.0f,
-                 Catch::Matchers::WithinRel(lHi * c, 1e-5f));
-    REQUIRE_THAT(c * 1.0f + s * 3.0f,
-                 Catch::Matchers::WithinRel(lHi * s, 1e-5f));
-    // Second eigenvector vLo = (-s, c)
-    REQUIRE_THAT(-s * 2.0f + c * 1.0f,
-                 Catch::Matchers::WithinRel(lLo * -s, 1e-5f));
-    REQUIRE_THAT(-s * 1.0f + c * 3.0f,
-                 Catch::Matchers::WithinRel(lLo * c, 1e-5f));
-  }
-
-  // Nonsymmetric block [[1,2],[3,4]] (used by the N == 2 entry point):
-  // eigenvalues 5.372281323, -0.372281323;
-  // eigenvector of 5.372281323 is +/- (0.415973558, 0.909376709)
-  {
-    Matrix<2, 2> A{1, 2, 3, 4};
-    float lHi = 0, lLo = 0, c = 0, s = 0;
-    QR::Solve2x2Eigen(A, 0, lHi, lLo, c, s);
-
-    REQUIRE_THAT(lHi, Catch::Matchers::WithinRel(5.372281323f, 1e-6f));
-    REQUIRE_THAT(lLo, Catch::Matchers::WithinRel(-0.372281323f, 1e-6f));
-    REQUIRE(matchesAbs(c, 0.415973558f));
-    REQUIRE(matchesAbs(s, 0.909376709f));
-
-    // Both-row residual with vHi = (c, s): A v = l v
-    REQUIRE_THAT(c * 1.0f + s * 2.0f,
-                 Catch::Matchers::WithinRel(lHi * c, 1e-5f));
-    REQUIRE_THAT(c * 3.0f + s * 4.0f,
-                 Catch::Matchers::WithinRel(lHi * s, 1e-5f));
-  }
-
-  // Diagonal blocks: eigenvectors are coordinate vectors
-  {
-    Matrix<2, 2> A{5, 0, 0, 2};
-    float lHi = 0, lLo = 0, c = 0, s = 0;
-    QR::Solve2x2Eigen(A, 0, lHi, lLo, c, s);
-    REQUIRE_THAT(lHi, Catch::Matchers::WithinRel(5.0f, 1e-7f));
-    REQUIRE_THAT(lLo, Catch::Matchers::WithinRel(2.0f, 1e-7f));
-    REQUIRE_THAT(c, Catch::Matchers::WithinRel(1.0f, 1e-7f));
-    REQUIRE_THAT(s, Catch::Matchers::WithinAbs(0.0f, 1e-7f));
-
-    A = Matrix<2, 2>{2, 0, 0, 5};
-    QR::Solve2x2Eigen(A, 0, lHi, lLo, c, s);
-    REQUIRE_THAT(lHi, Catch::Matchers::WithinRel(5.0f, 1e-7f));
-    REQUIRE_THAT(lLo, Catch::Matchers::WithinRel(2.0f, 1e-7f));
-    REQUIRE_THAT(c, Catch::Matchers::WithinAbs(0.0f, 1e-7f));
-    REQUIRE_THAT(s, Catch::Matchers::WithinRel(1.0f, 1e-7f));
-  }
-}
-// ============================================================================
-// TEST 6: Deflate
-// ============================================================================
-TEST_CASE("QR Building Block: Deflate", "[Matrix][QR]") {
-  // subdiag[0] = 1e-9 <= 1e-6 * (|2| + |3|) = 5e-6  -> deflated
-  // subdiag[1] = 0.5  > 1e-6 * (|3| + |4|) = 7e-6  -> kept
-  {
-    Matrix<3, 3> A{2, 1e-9f, 0, 1e-9f, 3, 0.5f, 0, 0.5f, 4};
-    QR::Deflate(A, 0, 2, 1e-6f);
-
-    REQUIRE(A.Get(1, 0) == 0.0f);
-    REQUIRE(A.Get(0, 1) == 0.0f);
-    REQUIRE_THAT(A.Get(2, 1), Catch::Matchers::WithinRel(0.5f, 1e-7f));
-    REQUIRE_THAT(A.Get(1, 2), Catch::Matchers::WithinRel(0.5f, 1e-7f));
-    // Diagonals untouched
-    REQUIRE_THAT(A.Get(0, 0), Catch::Matchers::WithinRel(2.0f, 1e-7f));
-    REQUIRE_THAT(A.Get(1, 1), Catch::Matchers::WithinRel(3.0f, 1e-7f));
-    REQUIRE_THAT(A.Get(2, 2), Catch::Matchers::WithinRel(4.0f, 1e-7f));
-  }
-
-  // Nothing deflated when all subdiagonals are well above tolerance
-  {
-    Matrix<3, 3> A{2, 0.1f, 0, 0.1f, 3, 0.2f, 0, 0.2f, 4};
-    QR::Deflate(A, 0, 2, 1e-6f);
-    REQUIRE_THAT(A.Get(1, 0), Catch::Matchers::WithinRel(0.1f, 1e-7f));
-    REQUIRE_THAT(A.Get(2, 1), Catch::Matchers::WithinRel(0.2f, 1e-7f));
-  }
-}
-
-// ============================================================================
-// TEST 7: Tridiagonalize
-// ============================================================================
-TEST_CASE("QR Building Block: Tridiagonalize", "[Matrix][QR]") {
-  // 4x4 symmetric with a full (0,3) corner coupling
-  {
-    Matrix<4, 4> A{2, 1, 0, 1, 1, 3, 1, 0, 0, 1, 4, 1, 1, 0, 1, 5};
-    Matrix<4, 4> Aorig = A;
-    Matrix<4, 4> U{0};
-
-    QR::Tridiagonalize(A, U);
-
-    // Off-tridiagonal entries must be zero up to float32 roundoff (the
-    // Givens zeroing cancels only in exact arithmetic; residuals are
-    // ~1e-7 for O(1) entries).
-    REQUIRE_THAT(A.Get(0, 2), Catch::Matchers::WithinAbs(0.0f, 1e-5f));
-    REQUIRE_THAT(A.Get(2, 0), Catch::Matchers::WithinAbs(0.0f, 1e-5f));
-    REQUIRE_THAT(A.Get(0, 3), Catch::Matchers::WithinAbs(0.0f, 1e-5f));
-    REQUIRE_THAT(A.Get(3, 0), Catch::Matchers::WithinAbs(0.0f, 1e-5f));
-    REQUIRE_THAT(A.Get(1, 3), Catch::Matchers::WithinAbs(0.0f, 1e-5f));
-    REQUIRE_THAT(A.Get(3, 1), Catch::Matchers::WithinAbs(0.0f, 1e-5f));
-
-    // Symmetry preserved exactly
-    for (uint8_t i = 0; i < 4; i++)
-      for (uint8_t j = 0; j < 4; j++)
-        REQUIRE(A.Get(i, j) == A.Get(j, i));
-
-    // U must be orthogonal
-    REQUIRE(isOrthogonal(U));
-
-    // Reconstruction: U * A_tri * U^T == Aorig (absolute check for
-    // originally-zero entries: WithinRel has no absolute fallback there)
-    Matrix<4, 4> UAt{};
-    U.Mult(A, UAt);
-    Matrix<4, 4> UAtU{};
-    UAt.Mult(U.Transpose(), UAtU);
-    for (uint8_t i = 0; i < 4; i++)
-      for (uint8_t j = 0; j < 4; j++) {
-        float actual = UAtU.Get(i, j);
-        float expected = Aorig.Get(i, j);
-        if (fabsf(expected) < 1e-3f)
-          REQUIRE_THAT(actual, Catch::Matchers::WithinAbs(0.0f, 1e-5f));
-        else
-          REQUIRE_THAT(actual,
-                       Catch::Matchers::WithinRel(expected, 1e-5f));
-      }
-
-    // Spectrum invariants match the original
-    {
-      float tr0 = Aorig.Get(0, 0) + Aorig.Get(1, 1) + Aorig.Get(2, 2) +
-                  Aorig.Get(3, 3);
-      float tr1 = A.Get(0, 0) + A.Get(1, 1) + A.Get(2, 2) + A.Get(3, 3);
-      REQUIRE_THAT(tr1, Catch::Matchers::WithinRel(tr0, 1e-6f));
-      REQUIRE_THAT(frob(A), Catch::Matchers::WithinRel(frob(Aorig), 1e-6f));
-    }
-
-    // Eigenvalues of the tridiagonal match the original (scipy reference):
-    // 6.0, 4.0, 3.0, 1.0
-    {
-      Matrix<4, 1> vals{};
-      Matrix<4, 4> vecs{};
-      QR::EigenQR(A, vecs, vals, 10000, 1e-6f);
-      REQUIRE_THAT(vals[0][0], Catch::Matchers::WithinRel(6.0f, 1e-4f));
-      REQUIRE_THAT(vals[1][0], Catch::Matchers::WithinRel(4.0f, 1e-4f));
-      REQUIRE_THAT(vals[2][0], Catch::Matchers::WithinRel(3.0f, 1e-4f));
-      REQUIRE_THAT(vals[3][0], Catch::Matchers::WithinRel(1.0f, 1e-4f));
-    }
-  }
-
-  // 5x5 symmetric
-  {
-    Matrix<5, 5> A{3, 1, 0, 0, 1, 1, 4, 1, 0, 0, 0, 1, 5, 1, 0, 0, 0, 1, 6, 1,
-                   1, 0, 0, 1, 7};
-    Matrix<5, 5> Aorig = A;
-    Matrix<5, 5> U{0};
-
-    QR::Tridiagonalize(A, U);
-
-    // All |i - j| >= 2 entries zero up to float32 roundoff
-    for (uint8_t i = 0; i < 5; i++)
-      for (uint8_t j = 0; j < 5; j++)
-        if (i > j + 1 || j > i + 1)
-          REQUIRE_THAT(A.Get(i, j), Catch::Matchers::WithinAbs(0.0f, 1e-5f));
-
-    REQUIRE(isOrthogonal(U));
-
-    Matrix<5, 5> UAt{};
-    U.Mult(A, UAt);
-    Matrix<5, 5> UAtU{};
-    UAt.Mult(U.Transpose(), UAtU);
-    for (uint8_t i = 0; i < 5; i++)
-      for (uint8_t j = 0; j < 5; j++) {
-        float actual = UAtU.Get(i, j);
-        float expected = Aorig.Get(i, j);
-        if (fabsf(expected) < 1e-3f)
-          REQUIRE_THAT(actual, Catch::Matchers::WithinAbs(0.0f, 1e-5f));
-        else
-          REQUIRE_THAT(actual,
-                       Catch::Matchers::WithinRel(expected, 1e-5f));
-      }
-  }
-
-  // Already tridiagonal: U must come out as the identity
-  {
-    Matrix<3, 3> A{1, 2, 0, 2, 5, 2, 0, 2, 9};
-    Matrix<3, 3> U{0};
-    QR::Tridiagonalize(A, U);
-    for (uint8_t i = 0; i < 3; i++)
-      for (uint8_t j = 0; j < 3; j++) {
-        float expected = (i == j) ? 1.0f : 0.0f;
-        REQUIRE_THAT(U.Get(i, j), Catch::Matchers::WithinAbs(expected, 1e-7f));
-      }
-  }
-}
-// ============================================================================
-// TEST 8: One full shifted QR step (integration of the blocks)
-// ============================================================================
-TEST_CASE("QR Building Block: Full Shifted QR Step", "[Matrix][QR]") {
-  // One Wilkinson-shifted QR step on the whole 3x3 block is a similarity
-  // transform, so all spectrum invariants (trace, sum of principal 2x2
-  // minors, determinant) must be preserved.
-  //
-  // A = [[1,2,3],[2,5,8],[3,8,9]]:  tr = 15, e2 = -18, det = -4
-  {
-    Matrix<3, 3> A{1, 2, 3, 2, 5, 8, 3, 8, 9};
-    float tr0 = trace3(A);   // 15
-    float e20 = e2_3x3(A);   // -18
-    float det0 = det3(A);    // -4
-
-    // mu from the trailing 2x2 [[5,8],[8,9]]: eigenvalues
-    // -1.246211251, 15.246211251; closest to d = 9 is 15.246211251 (Wilkinson)
-    float mu = QR::WilkinsonShift(A.Get(1, 1), A.Get(2, 1), A.Get(2, 2));
-    REQUIRE_THAT(mu, Catch::Matchers::WithinRel(15.246211251f, 1e-5f));
-
-    for (uint8_t i = 0; i < 3; i++)
-      A[i][i] -= mu;
-
-    // Bulge chase: rotations on (0,1) then (1,2)
-    float c = 0, s = 0;
-    QR::GivensRotation(A.Get(0, 0), A.Get(1, 0), c, s);
-    QR::ApplyRotationBothSides(A, 0, c, s);
-    QR::GivensRotation(A.Get(1, 1), A.Get(2, 1), c, s);
-    QR::ApplyRotationBothSides(A, 1, c, s);
-
-    for (uint8_t i = 0; i < 3; i++)
-      A[i][i] += mu;
-
-    // Symmetry preserved
-    for (uint8_t i = 0; i < 3; i++)
-      for (uint8_t j = 0; j < 3; j++)
-        REQUIRE(A.Get(i, j) == A.Get(j, i));
-
-    // Spectrum invariants preserved
-    REQUIRE_THAT(trace3(A), Catch::Matchers::WithinRel(tr0, 1e-5f));
-    REQUIRE_THAT(e2_3x3(A), Catch::Matchers::WithinRel(e20, 1e-5f));
-    REQUIRE_THAT(det3(A), Catch::Matchers::WithinRel(det0, 1e-5f));
-  }
-
-  // For TRIDIAGONAL input a single step keeps the tridiagonal structure
-  {
-    Matrix<3, 3> T{1, 2, 0, 2, 5, 2, 0, 2, 9};
-    float mu = QR::WilkinsonShift(T.Get(1, 1), T.Get(2, 1), T.Get(2, 2));
-    for (uint8_t i = 0; i < 3; i++)
-      T[i][i] -= mu;
-    float c = 0, s = 0;
-    QR::GivensRotation(T.Get(0, 0), T.Get(1, 0), c, s);
-    QR::ApplyRotationBothSides(T, 0, c, s);
-    QR::GivensRotation(T.Get(1, 1), T.Get(2, 1), c, s);
-    QR::ApplyRotationBothSides(T, 1, c, s);
-    for (uint8_t i = 0; i < 3; i++)
-      T[i][i] += mu;
-
-    // Corners must vanish up to float32 roundoff: tridiagonal form
-    // maintained. The cancellation is exact in exact arithmetic (the
-    // corner is s1*a - c1*b times a factor, and Givens gives s1*a = c1*b),
-    // so the residual is pure rounding, ~1e-6 for O(1) entries.
-    REQUIRE_THAT(T.Get(0, 2), Catch::Matchers::WithinAbs(0.0f, 1e-5f));
-    REQUIRE_THAT(T.Get(2, 0), Catch::Matchers::WithinAbs(0.0f, 1e-5f));
-  }
-}

unit-tests/qr-reference-values.py

@@ -1,246 +0,0 @@
-#!/usr/bin/env python3
-"""
-Reference values for the QR eigen-decomposition building block tests
-(unit-tests/qr-build-blocks-tests.cpp). Run this to verify/implement the
-C++ implementation in src/QR.hpp / src/QR.cpp against numpy/scipy.
-
-Conventions (match the C++ exactly):
-  * Givens zeroing rotation: G = [[c, s], [-s, c]], c = x/r, s = y/r,
-    r = hypot(x, y).  G * (x, y)^T = (r, 0)^T.
-  * Similarity transform: A <- G A G^T (ApplyRotationBothSides).
-  * Eigenvector accumulation: V <- V G^T (ApplyRotationToVectors).
-    Vblock in the 2x2 closed form is [[c, -s], [s, c]] (same shape as G^T).
-  * Tridiagonalization: bottom-up Givens (i = N-2 down to k+1 per column k).
-  * Shifted QR loop: Wilkinson shift mu from the trailing 2x2, chase on the
-    trailing unreduced block [lo, hi], deflate by relative tolerance, peel
-    exact-zero subdiagonals, 2x2 closed-form termination.
-  * Pipeline: M0 = U * Mtri * U^T and Mtri = V * D * V^T  =>
-    eigenvectors of M0 = U * V (columns), eigenvalues = diag(D).
-
-Usage: python3 qr-reference-values.py
-"""
-
-import numpy as np
-import scipy.linalg as sla
-
-np.set_printoptions(precision=9, linewidth=120)
-
-
-def givens(x, y):
-    """c = x/r, s = y/r with r = hypot(x, y)."""
-    r = np.hypot(x, y)
-    if r == 0.0:
-        return 1.0, 0.0
-    return x / r, y / r
-
-
-def rot(n, i, c, s):
-    """G = I with [[c, s], [-s, c]] embedded at (i, i+1)."""
-    G = np.eye(n)
-    G[i:i + 2, i:i + 2] = np.array([[c, s], [-s, c]])
-    return G
-
-
-def tridiagonalize(M0):
-    """Bottom-up Givens tridiagonalization. Returns (Mtri, U) with
-    M0 = U Mtri U^T."""
-    n = len(M0)
-    M = M0.copy()
-    U = np.eye(n)
-    for k in range(n - 2):
-        for i in range(n - 2, k, -1):
-            c, s = givens(M[i, k], M[i + 1, k])
-            G = rot(n, i, c, s)
-            M = G @ M @ G.T
-            U = U @ G.T
-    return M, U
-
-
-def wilkinson(a, b, d):
-    """Eigenvalue of [[a, b], [b, d]] closest to d."""
-    delta = 0.5 * (a - d)
-    spread = np.sqrt(delta * delta + b * b)
-    return 0.5 * (a + d) - (spread if delta >= 0 else -spread)
-
-
-def solve2x2(A, lo):
-    """Closed form for the block at (lo, lo+1): (lHi, lLo, c, s) with
-    vHi = (c, s), vLo = (-s, c)."""
-    a = A[lo, lo]
-    b = A[lo, lo + 1]
-    e = A[lo + 1, lo]
-    d = A[lo + 1, lo + 1]
-    tr = a + d
-    det = a * d - b * e
-    disc = max(0.0, tr * tr - 4 * det)
-    lhi = 0.5 * (tr + np.sqrt(disc))
-    llo = 0.5 * (tr - np.sqrt(disc))
-    if b != 0.0:
-        v1 = lhi - a
-        nn = np.hypot(b, v1)
-        c, s = b / nn, v1 / nn
-    elif a >= d:
-        c, s = 1.0, 0.0
-    else:
-        c, s = 0.0, 1.0
-    return lhi, llo, c, s
-
-
-def eigenqr(M0, tol=1e-12, max_iter=100000):
-    """Full pipeline mirroring QR::EigenQR. Returns (eigs, W) where W has
-    the eigenvectors of M0 as columns."""
-    n = len(M0)
-    if n == 2:
-        l1, l2, c, s = solve2x2(M0, 0)
-        return np.array([l1, l2]), np.array([[c, -s], [s, c]])
-    M, U = tridiagonalize(M0)
-    V = np.eye(n)
-    hi = n - 1
-    for _ in range(max_iter):
-        # deflate: zero tiny subdiagonals (relative test)
-        for i in range(hi):
-            t = M[i + 1, i]
-            scale = abs(M[i, i]) + abs(M[i + 1, i + 1])
-            if abs(t) <= tol * scale:
-                M[i + 1, i] = M[i, i + 1] = 0.0
-        # peel exact-zero trailing subdiagonals
-        while hi > 0 and M[hi, hi - 1] == 0.0:
-            hi -= 1
-        if hi == 0:
-            break
-        # find start of trailing unreduced block
-        lo = hi
-        for i in range(hi - 1, -1, -1):
-            if M[i + 1, i] == 0.0:
-                break
-            lo = i
-        if lo + 1 == hi:
-            # closed-form 2x2 termination: set diagonal, fold Vblock in
-            l1, l2, c, s = solve2x2(M, lo)
-            Vb = np.eye(n)
-            Vb[lo:lo + 2, lo:lo + 2] = np.array([[c, -s], [s, c]])
-            V = V @ Vb
-            M[lo, lo] = l1
-            M[lo + 1, lo + 1] = l2
-            M[lo + 1, lo] = M[lo, lo + 1] = 0.0
-            if lo == 0:
-                break
-            hi = lo - 1
-            continue
-        # full shifted step on [lo, hi] (shift applies to the active block)
-        mu = wilkinson(M[hi - 1, hi - 1], M[hi, hi - 1], M[hi, hi])
-        diag = M.diagonal().copy()
-        diag[lo:hi + 1] -= mu
-        np.fill_diagonal(M, diag)
-        c, s = givens(M[lo, lo], M[lo + 1, lo])
-        G = rot(n, lo, c, s)
-        M = G @ M @ G.T
-        V = V @ G.T
-        for i in range(lo + 1, hi):
-            c, s = givens(M[i, i], M[i + 1, i])
-            G = rot(n, i, c, s)
-            M = G @ M @ G.T
-            V = V @ G.T
-        diag = M.diagonal().copy()
-        diag[lo:hi + 1] += mu
-        np.fill_diagonal(M, diag)
-
-    eigs = np.diag(M).astype(float)
-    order = np.argsort(eigs)[::-1]  # descending, like the C++ test harness
-    eigs = eigs[order]
-    W = U @ V
-    W = W[:, order]
-    return eigs, W
-
-
-def report(name, val, ref=None, tol=1e-6):
-    ok = "OK " if ref is None or np.allclose(val, ref, rtol=tol, atol=tol) else "FAIL"
-    print(f"[{ok}] {name} = {val}")
-    if ref is not None:
-        print(f"        scipy/numpy ref = {ref}")
-
-
-def main():
-    print("=== TEST 1: GivensRotation ===")
-    c, s = givens(2.0, 1.0)
-    print(f"  c = {c}  s = {s}")
-    # G * (x, y)^T = (r, 0)^T:  G = [[c, s], [-s, c]]
-    assert abs(c * 2 + s * 1 - np.sqrt(5)) < 1e-15
-    assert abs(-s * 2 + c * 1) < 1e-15
-
-    print("\n=== TEST 2: ApplyRotationBothSides A <- G A G^T ===")
-    A = np.array([[3.0, 4.0, 5.0], [6.0, 7.0, 8.0], [9.0, 10.0, 11.0]])
-    G = rot(3, 0, 0.6, 0.8)
-    B = G @ A @ G.T
-    print(B)
-
-    A = np.array([[5.0, 0.0, 1.0], [0.0, 6.0, 2.0], [1.0, 2.0, 7.0]])
-    c, s = givens(6.0, 2.0)
-    G = rot(3, 1, c, s)
-    B = G @ A @ G.T
-    print(B)
-
-    print("\n=== TEST 3: V accumulation V <- V G^T ===")
-    V = np.eye(3)
-    G = rot(3, 0, 0.894427191, 0.447213595)
-    V = V @ G.T
-    print(V)
-    V2 = V @ rot(3, 1, 0.848874681, 0.528748047).T
-    print(V2)
-
-    print("\n=== TEST 4: Solve2x2Eigen ===")
-    for A in (np.array([[5.0, 8.0], [8.0, 9.0]]), np.array([[1.0, 2.0], [3.0, 4.0]])):
-        l1, l2, c, s = solve2x2(A, 0)
-        ref = np.linalg.eigvalsh(A) if np.allclose(A, A.T) else np.linalg.eigvals(A)
-        print(f"  A={A.ravel()}  lHi={l1} lLo={l2}  c={c} s={s}  ref={np.sort(ref)[::-1]}")
-
-    print("\n=== TEST 8: WilkinsonShift ===")
-    print(f"  W(5, 8, 9) = {wilkinson(5, 8, 9)}")
-    print(f"  W(4, 2, 7) = {wilkinson(4, 2, 7)}")
-    print(f"  W(9, 2, 5) = {wilkinson(9, 2, 5)}")
-
-    print("\n=== TEST 8b: one full shifted chase step on tridiagonal 3x3 ===")
-    T = np.array([[1.0, 2.0, 0.0], [2.0, 5.0, 2.0], [0.0, 2.0, 9.0]])
-    mu = wilkinson(5, 2, 9)
-    M = T - mu * np.eye(3)
-    c, s = givens(M[0, 0], M[1, 0])
-    M = rot(3, 0, c, s) @ M @ rot(3, 0, c, s).T
-    c, s = givens(M[1, 1], M[2, 1])
-    M = rot(3, 1, c, s) @ M @ rot(3, 1, c, s).T
-    M = M + mu * np.eye(3)
-    print(f"  mu = {mu}")
-    print(M)
-    print(f"  corners: {M[0, 2]}, {M[2, 0]} (exact-arithmetic zeros)")
-    print(f"  trace {M.trace():.15f} (was {T.trace()})")
-
-    print("\n=== TEST 7: Tridiagonalize ===")
-    M4 = np.array([[2.0, 1, 0, 1], [1, 3, 1, 0], [0, 1, 4, 1], [1, 0, 1, 5]])
-    M, U = tridiagonalize(M4)
-    print("  M4 tridiagonalized:\n", M)
-    print(f"  reconstruction U M U^T == M4: {np.allclose(U @ M @ U.T, M4, atol=1e-9)}")
-    M5 = np.array([[3.0, 1, 0, 0, 1], [1, 4, 1, 0, 0], [0, 1, 5, 1, 0],
-                   [0, 0, 1, 6, 1], [1, 0, 0, 1, 7]])
-    M, U = tridiagonalize(M5)
-    print("  M5 tridiagonalized:\n", M)
-    print(f"  reconstruction: {np.allclose(U @ M @ U.T, M5, atol=1e-9)}")
-
-    print("\n=== End-to-end: random symmetric vs scipy.linalg.eigh ===")
-    rng = np.random.default_rng(12345)
-    worst = 0.0
-    for n in range(3, 9):
-        M0 = rng.normal(size=(n, n))
-        M0 = (M0 + M0.T) / 2
-        eigs, W = eigenqr(M0.astype(float))
-        ref = sla.eigh(M0)
-        e_err = np.max(np.abs(np.sort(eigs) - ref[0]))
-        resid = np.linalg.norm(W @ np.diag(eigs) @ W.T - M0)
-        ortho = np.linalg.norm(W.T @ W - np.eye(n))
-        print(f"  n={n}: eigs_err={e_err:.2e}  resid={resid:.2e}  ortho={ortho:.2e}")
-        worst = max(worst, e_err, resid, ortho)
-    print(f"\nworst over all n: {worst:.2e}")
-    assert worst < 1e-10, "end-to-end reference FAILED"
-    print("ALL REFERENCES OK")
-
-
-if __name__ == "__main__":
-    main()

unit-tests/svd-integration-test.cpp

@@ -1,363 +0,0 @@
-#include "Matrix.hpp"
-#include "SVD.hpp"
-#include <catch2/catch_test_macros.hpp>
-#include <catch2/matchers/catch_matchers_floating_point.hpp>
-#include <iostream>
-
-// Generic helper functions for any matrix size
-template <uint8_t rows, uint8_t columns>
-static float frobeniusNorm(const Matrix<rows, columns> &M) {
-  float sum = 0.0f;
-  for (int i = 0; i < rows; i++)
-    for (int j = 0; j < columns; j++) {
-      float v = M.Get(i, j);
-      sum += v * v;
-    }
-  return sqrtf(sum);
-}
-
-template <uint8_t n>
-static bool isOrthogonal(const Matrix<n, n> &M, float tol = 1e-4f) {
-  Matrix<n, n> Mt = M.Transpose();
-  Matrix<n, n> MtM{0};
-  Mt.Mult(M, MtM);
-  for (int i = 0; i < n; i++)
-    for (int j = 0; j < n; j++) {
-      float expected = (i == j) ? 1.0f : 0.0f;
-      if (fabsf(MtM.Get(i, j) - expected) > tol)
-        return false;
-    }
-  return true;
-}
-
-TEST_CASE("SVD Integration: 2x2 [[1,2],[3,4]]", "[Matrix][SVD][Integration]") {
-  Matrix<2, 2> A{1, 2, 3, 4};
-  Matrix<2, 2> U{0};
-  Matrix<2, 1> sigma{0};
-  Matrix<2, 2> Vt{0};
-
-  SVD::SVD(A, U, sigma, Vt);
-
-  // Reference singular values from scipy: [5.464985704219, 0.365966190626]
-  REQUIRE_THAT(sigma.Get(0, 0), Catch::Matchers::WithinRel(5.4649857f, 1e-3f));
-  REQUIRE_THAT(sigma.Get(1, 0), Catch::Matchers::WithinRel(0.3659662f, 1e-3f));
-
-  // Check orthogonality of U and Vt (first 2x2 blocks)
-  REQUIRE(isOrthogonal<2>(U));
-  REQUIRE(isOrthogonal<2>(Vt));
-
-  // Check reconstruction: A ≈ U · diag(sigma) · Vt
-  Matrix<2, 2> recon{0};
-  Matrix<2, 2> Usig{0};
-  for (int i = 0; i < 2; i++)
-    for (int j = 0; j < 2; j++)
-      Usig[i][j] = U.Get(i, j) * sigma.Get(j, 0);
-
-  Usig.Mult(Vt, recon);
-
-  float err = 0.0f;
-  for (int i = 0; i < 2; i++)
-    for (int j = 0; j < 2; j++) {
-      float diff = recon.Get(i, j) - A.Get(i, j);
-      err += diff * diff;
-    }
-  err = sqrtf(err);
-  REQUIRE_THAT(err, Catch::Matchers::WithinAbs(0.0f, 1e-3f));
-
-  std::cout << "SVD 2x2 [[1,2],[3,4]]:\n";
-  std::cout << "Sigma: [" << sigma.Get(0, 0) << ", " << sigma.Get(1, 0)
-            << "]\n";
-}
-
-TEST_CASE("SVD Integration: 3x3 diagonal [10,5,2]",
-          "[Matrix][SVD][Integration]") {
-  Matrix<3, 3> A{10, 0, 0, 0, 5, 0, 0, 0, 2};
-  Matrix<3, 3> U{0};
-  Matrix<3, 1> sigma{0};
-  Matrix<3, 3> Vt{0};
-
-  SVD::SVD(A, U, sigma, Vt);
-
-  // Singular values should be [10, 5, 2] (already diagonal)
-  REQUIRE_THAT(sigma.Get(0, 0), Catch::Matchers::WithinRel(10.0f, 1e-3f));
-  REQUIRE_THAT(sigma.Get(1, 0), Catch::Matchers::WithinRel(5.0f, 1e-3f));
-  REQUIRE_THAT(sigma.Get(2, 0), Catch::Matchers::WithinRel(2.0f, 1e-3f));
-
-  // U and Vt should be identity (or close) for diagonal matrix
-  float uErr = frobeniusNorm(U - Matrix<3, 3>{1, 0, 0, 0, 1, 0, 0, 0, 1});
-  float vtErr = frobeniusNorm(Vt - Matrix<3, 3>{1, 0, 0, 0, 1, 0, 0, 0, 1});
-  REQUIRE_THAT(uErr, Catch::Matchers::WithinAbs(0.0f, 1e-2f));
-  REQUIRE_THAT(vtErr, Catch::Matchers::WithinAbs(0.0f, 1e-2f));
-}
-
-TEST_CASE("SVD Integration: 3x3 rank-deficient [[1,2,3],[4,5,6],[7,8,9]]",
-          "[Matrix][SVD][Integration]") {
-  Matrix<3, 3> A{1, 2, 3, 4, 5, 6, 7, 8, 9};
-  Matrix<3, 3> U{0};
-  Matrix<3, 1> sigma{0};
-  Matrix<3, 3> Vt{0};
-
-  SVD::SVD(A, U, sigma, Vt);
-
-  // Reference: [16.848103352614, 1.068369514555, 0.0]
-  REQUIRE_THAT(sigma.Get(0, 0), Catch::Matchers::WithinRel(16.8481f, 1e-2f));
-  REQUIRE_THAT(sigma.Get(1, 0), Catch::Matchers::WithinRel(1.06837f, 1e-2f));
-  // Third singular value should be ~0 (rank-deficient)
-  REQUIRE_THAT(sigma.Get(2, 0), Catch::Matchers::WithinAbs(0.0f, 1e-2f));
-
-  // Check reconstruction
-  Matrix<3, 3> recon{0};
-  Matrix<3, 3> Usig{0};
-  for (int i = 0; i < 3; i++)
-    for (int j = 0; j < 3; j++)
-      Usig[i][j] = U.Get(i, j) * sigma.Get(j, 0);
-  Usig.Mult(Vt, recon);
-
-  float err = 0.0f;
-  for (int i = 0; i < 3; i++)
-    for (int j = 0; j < 3; j++) {
-      float diff = recon.Get(i, j) - A.Get(i, j);
-      err += diff * diff;
-    }
-  err = sqrtf(err);
-  REQUIRE_THAT(err, Catch::Matchers::WithinAbs(0.0f, 1e-2f));
-
-  std::cout << "SVD 3x3 rank-deficient:\n";
-  std::cout << "Sigma: [" << sigma.Get(0, 0) << ", " << sigma.Get(1, 0) << ", "
-            << sigma.Get(2, 0) << "]\n";
-}
-
-TEST_CASE("SVD Integration: tall 4x3 matrix", "[Matrix][SVD][Integration]") {
-  Matrix<4, 3> A{1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12};
-  Matrix<4, 3> U{0};
-  Matrix<3, 1> sigma{0};
-  Matrix<3, 3> Vt{0};
-
-  SVD::SVD(A, U, sigma, Vt);
-
-  // Reference: [25.462407436036, 1.290661675761, 0.0]
-  REQUIRE_THAT(sigma.Get(0, 0), Catch::Matchers::WithinRel(25.4624f, 1e-2f));
-  REQUIRE_THAT(sigma.Get(1, 0), Catch::Matchers::WithinRel(1.29066f, 1e-2f));
-  REQUIRE_THAT(sigma.Get(2, 0), Catch::Matchers::WithinAbs(0.0f, 1e-2f));
-
-  // Check reconstruction
-  Matrix<4, 3> recon{0};
-  Matrix<4, 3> Usig{0};
-  for (int i = 0; i < 4; i++)
-    for (int j = 0; j < 3; j++)
-      Usig[i][j] = U.Get(i, j) * sigma.Get(j, 0);
-  Usig.Mult(Vt, recon);
-
-  float err = 0.0f;
-  for (int i = 0; i < 4; i++)
-    for (int j = 0; j < 3; j++) {
-      float diff = recon.Get(i, j) - A.Get(i, j);
-      err += diff * diff;
-    }
-  err = sqrtf(err);
-  REQUIRE_THAT(err, Catch::Matchers::WithinAbs(0.0f, 1e-2f));
-
-  std::cout << "SVD tall 4x3:\n";
-  std::cout << "Sigma: [" << sigma.Get(0, 0) << ", " << sigma.Get(1, 0) << ", "
-            << sigma.Get(2, 0) << "]\n";
-}
-
-TEST_CASE("SVD Integration: wide 3x5 matrix", "[Matrix][SVD][Integration]") {
-  Matrix<3, 5> A{1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15};
-  Matrix<3, 5> U{0};
-  Matrix<5, 1> sigma{0}; // sigma is columns x 1 = 5x1 for wide matrix
-  Matrix<5, 5> Vt{0};    // Vt is columns x columns = 5x5
-
-  SVD::SVD(A, U, sigma, Vt);
-
-  // Reference: [35.127223333575, 2.465396696917, 0.0]
-  REQUIRE_THAT(sigma.Get(0, 0), Catch::Matchers::WithinRel(35.1272f, 1e-2f));
-  REQUIRE_THAT(sigma.Get(1, 0), Catch::Matchers::WithinRel(2.46540f, 1e-2f));
-  REQUIRE_THAT(sigma.Get(2, 0), Catch::Matchers::WithinAbs(0.0f, 1e-2f));
-
-  // Check reconstruction: A (3x5) = U * Sigma * Vt, where U (3x5) has
-  // its meaningful part in the first 3 columns, sigma (5x1) in the
-  // first 3 entries, and Vt (5x5) in its first 3 rows (right
-  // singular vectors as rows). So:
-  //   A[i][j] = sum_k U[i][k] * sigma[k] * Vt[k][j]
-
-  float err2 = 0.0f;
-  for (int i = 0; i < 3; i++) {
-    for (int j = 0; j < 5; j++) {
-      float recon_val = 0.0f;
-      for (int k = 0; k < 3; k++) {
-        recon_val += U.Get(i, k) * sigma.Get(k, 0) * Vt.Get(k, j);
-      }
-      float diff = recon_val - A.Get(i, j);
-      err2 += diff * diff;
-    }
-  }
-  err2 = sqrtf(err2);
-  REQUIRE_THAT(err2, Catch::Matchers::WithinAbs(0.0f, 1e-2f));
-
-  std::cout << "SVD wide 3x5:\n";
-  std::cout << "Sigma: [" << sigma.Get(0, 0) << ", " << sigma.Get(1, 0) << ", "
-            << sigma.Get(2, 0) << "]\n";
-}
-
-TEST_CASE("SVD Integration: identity 3x3", "[Matrix][SVD][Integration]") {
-  Matrix<3, 3> A{1, 0, 0, 0, 1, 0, 0, 0, 1};
-  Matrix<3, 3> U{0};
-  Matrix<3, 1> sigma{0};
-  Matrix<3, 3> Vt{0};
-
-  SVD::SVD(A, U, sigma, Vt);
-
-  REQUIRE_THAT(sigma.Get(0, 0), Catch::Matchers::WithinRel(1.0f, 1e-3f));
-  REQUIRE_THAT(sigma.Get(1, 0), Catch::Matchers::WithinRel(1.0f, 1e-3f));
-  REQUIRE_THAT(sigma.Get(2, 0), Catch::Matchers::WithinRel(1.0f, 1e-3f));
-
-  float err = frobeniusNorm(U - Matrix<3, 3>{1, 0, 0, 0, 1, 0, 0, 0, 1});
-  REQUIRE_THAT(err, Catch::Matchers::WithinAbs(0.0f, 1e-2f));
-}
-
-TEST_CASE("SVD Integration: symmetric positive definite 2x2 [[5,3],[3,5]]",
-          "[Matrix][SVD][Integration]") {
-  Matrix<2, 2> A{5, 3, 3, 5};
-  Matrix<2, 2> U{0};
-  Matrix<2, 1> sigma{0};
-  Matrix<2, 2> Vt{0};
-
-  SVD::SVD(A, U, sigma, Vt);
-
-  // For SPD matrix, singular values = eigenvalues: [8, 2]
-  REQUIRE_THAT(sigma.Get(0, 0), Catch::Matchers::WithinRel(8.0f, 1e-3f));
-  REQUIRE_THAT(sigma.Get(1, 0), Catch::Matchers::WithinRel(2.0f, 1e-3f));
-
-  // Check reconstruction
-  Matrix<2, 2> recon{0};
-  Matrix<2, 2> Usig{0};
-  for (int i = 0; i < 2; i++)
-    for (int j = 0; j < 2; j++)
-      Usig[i][j] = U.Get(i, j) * sigma.Get(j, 0);
-  Usig.Mult(Vt, recon);
-
-  float err = 0.0f;
-  for (int i = 0; i < 2; i++)
-    for (int j = 0; j < 2; j++) {
-      float diff = recon.Get(i, j) - A.Get(i, j);
-      err += diff * diff;
-    }
-  err = sqrtf(err);
-  REQUIRE_THAT(err, Catch::Matchers::WithinAbs(0.0f, 1e-3f));
-
-  std::cout << "SVD SPD 2x2 [[5,3],[3,5]]:\n";
-  std::cout << "Sigma: [" << sigma.Get(0, 0) << ", " << sigma.Get(1, 0)
-            << "]\n";
-}
-
-// ----------------------------------------------------------------------------
-// Matrix::SVD member wrapper (delegates to SVD::SVD)
-// ----------------------------------------------------------------------------
-
-/**
- * Reconstruction error ‖U·diag(sigma)·Vᵀ − A‖_F. Zero-padded entries of
- * U/sigma/Vt (wide/tall cases) are zero by the output conventions, so the
- * full product equals U[:, :k]·diag(sigma[:k])·Vt[:k, :].
- */
-template <uint8_t rows, uint8_t columns>
-static float svdReconstructionError(const Matrix<rows, columns> &A,
-                                    const Matrix<rows, columns> &U,
-                                    const Matrix<columns, 1> &sigma,
-                                    const Matrix<columns, columns> &Vt) {
-  Matrix<rows, columns> recon{0};
-  Matrix<rows, columns> Usig{0};
-  for (int i = 0; i < rows; i++)
-    for (int j = 0; j < columns; j++)
-      Usig[i][j] = U.Get(i, j) * sigma.Get(j, 0);
-  Usig.Mult(Vt, recon);
-  float err = 0.0f;
-  for (int i = 0; i < rows; i++)
-    for (int j = 0; j < columns; j++) {
-      float diff = recon.Get(i, j) - A.Get(i, j);
-      err += diff * diff;
-    }
-  return sqrtf(err);
-}
-
-/**
- * Orthonormality of the first k columns of M: the k×k leading block of
- * MᵀM must equal I_k. (For a tall SVD, U has k = min(rows, cols)
- * meaningful columns and this is the full UᵀU.)
- */
-template <uint8_t r, uint8_t c>
-static bool leadingColumnsOrthonormal(const Matrix<r, c> &M, uint8_t k,
-                                      float tol = 1e-4f) {
-  Matrix<c, r> Mt = M.Transpose();
-  Matrix<c, c> MtM{0};
-  Mt.Mult(M, MtM);
-  for (int i = 0; i < k; i++)
-    for (int j = 0; j < k; j++) {
-      float expected = (i == j) ? 1.0f : 0.0f;
-      if (fabsf(MtM.Get(i, j) - expected) > tol)
-        return false;
-    }
-  return true;
-}
-
-/**
- * Orthonormality of the first k rows of M: the k×k leading block of
- * M·Mᵀ must equal I_k. (Vᵀ may have zero-padded trailing rows in the
- * wide case, so check only the meaningful leading block.)
- */
-template <uint8_t r, uint8_t c>
-static bool leadingRowsOrthonormal(const Matrix<r, c> &M, uint8_t k,
-                                   float tol = 1e-4f) {
-  Matrix<c, r> Mt = M.Transpose();
-  Matrix<r, r> MMt{0};
-  M.Mult(Mt, MMt);
-  for (int i = 0; i < k; i++)
-    for (int j = 0; j < k; j++) {
-      float expected = (i == j) ? 1.0f : 0.0f;
-      if (fabsf(MMt.Get(i, j) - expected) > tol)
-        return false;
-    }
-  return true;
-}
-
-TEST_CASE("Matrix::SVD wrapper: 3x2 tall [[1,2],[3,4],[5,6]]",
-          "[Matrix][SVD][Wrapper]") {
-  Matrix<3, 2> A{1, 2, 3, 4, 5, 6};
-  Matrix<3, 2> U{0};
-  Matrix<2, 1> sigma{0};
-  Matrix<2, 2> Vt{0};
-
-  A.SVD(U, sigma, Vt);
-
-  // Reference singular values from numpy: [9.52552, 0.514301]
-  REQUIRE_THAT(sigma.Get(0, 0), Catch::Matchers::WithinRel(9.52552f, 1e-3f));
-  REQUIRE_THAT(sigma.Get(1, 0), Catch::Matchers::WithinRel(0.514301f, 1e-3f));
-
-  REQUIRE(leadingColumnsOrthonormal(U, 2));
-  REQUIRE(leadingRowsOrthonormal(Vt, 2));
-
-  float err = svdReconstructionError(A, U, sigma, Vt);
-  REQUIRE_THAT(err, Catch::Matchers::WithinAbs(0.0f, 1e-3f));
-}
-
-TEST_CASE("Matrix::SVD wrapper: 2x3 wide [[1,2,3],[4,5,6]]",
-          "[Matrix][SVD][Wrapper]") {
-  Matrix<2, 3> A{1, 2, 3, 4, 5, 6};
-  Matrix<2, 3> U{0};
-  Matrix<3, 1> sigma{0};
-  Matrix<3, 3> Vt{0};
-
-  A.SVD(U, sigma, Vt);
-
-  // Reference singular values from numpy: [9.50803, 0.77287]; the third
-  // entry (wide-matrix padding) must be zero.
-  REQUIRE_THAT(sigma.Get(0, 0), Catch::Matchers::WithinRel(9.50803f, 1e-3f));
-  REQUIRE_THAT(sigma.Get(1, 0), Catch::Matchers::WithinRel(0.77287f, 1e-3f));
-  REQUIRE_THAT(sigma.Get(2, 0), Catch::Matchers::WithinAbs(0.0f, 1e-6f));
-
-  REQUIRE(leadingColumnsOrthonormal(U, 2));
-  REQUIRE(leadingRowsOrthonormal(Vt, 2));
-
-  float err = svdReconstructionError(A, U, sigma, Vt);
-  REQUIRE_THAT(err, Catch::Matchers::WithinAbs(0.0f, 1e-3f));
-}

unit-tests/svd-reference-values.py

@@ -1,513 +0,0 @@
-#!/usr/bin/env python3
-"""
-Generate reference values for SVD building block unit tests.
-Run this to verify/implement the C++ SVD implementation against scipy/numpy.
-
-Usage: python3 svd-reference-values.py
-"""
-
-import numpy as np
-from scipy.linalg import svd, qr as scipy_qr
-import json
-
-def compute_householder(x):
-    """Compute Householder reflector: H*x = [alpha, 0, 0, ...]^T.
-    
-    Returns (v_normalized, alpha) where v is the normalized Householder vector.
-    H = I - 2*v*v^T / (v^T*v)
-    """
-    x = np.array(x, dtype=np.float64)
-    norm_x = np.linalg.norm(x)
-    
-    if norm_x < 1e-30:
-        return x.copy(), 0.0
-    
-    alpha = -np.sign(x[0]) * norm_x if x[0] != 0 else -norm_x
-    v = x.copy()
-    v[0] -= alpha
-    v_norm = np.linalg.norm(v)
-    
-    if v_norm < 1e-30:
-        return np.zeros_like(x), alpha
-    
-    v /= v_norm
-    return v, alpha
-
-def apply_householder_left(A, v, start_row):
-    """Apply Householder reflection from the left: A = (I - 2vv^T) @ A.
-    
-    v is the normalized Householder vector operating on rows [start_row:].
-    The length of v must match the number of rows affected.
-    """
-    A = A.copy()
-    k = len(v)
-    
-    for col in range(A.shape[1]):
-        dot = np.dot(v, A[start_row:start_row+k, col])
-        A[start_row:start_row+k, col] -= 2.0 * dot * v
-    
-    return A
-
-def apply_householder_right(A, v, start_col):
-    """Apply Householder reflection from the right: A = A @ (I - 2vv^T).
-    
-    v is the normalized Householder vector operating on columns [start_col:].
-    The length of v must match the number of columns affected.
-    """
-    A = A.copy()
-    k = len(v)
-    
-    for row in range(A.shape[0]):
-        dot = np.dot(A[row, start_col:start_col+k], v)
-        A[row, start_col:start_col+k] -= 2.0 * dot * v
-    
-    return A
-
-def compute_givens(x, y):
-    """Compute Givens rotation that zeros out y.
-    
-    Returns (c, s) such that [c s; -s c] @ [x; y] = [r; 0].
-    """
-    r = np.sqrt(x*x + y*y)
-    if r < 1e-30:
-        return 1.0, 0.0
-    c = x / r
-    s = y / r
-    return c, s
-
-def apply_givens_left(A, i, j, c, s):
-    """Apply Givens rotation from the left to rows i and j of A.
-    
-    [c  s]   [row_i]
-    [-s  c] @ [row_j] = [new_row_i]
-                          [new_row_j]
-    """
-    A = A.copy()
-    new_i = c * A[i] + s * A[j]
-    new_j = -s * A[i] + c * A[j]
-    A[i] = new_i
-    A[j] = new_j
-    return A
-
-def apply_givens_right(A, i, j, c, s):
-    """Apply Givens rotation from the right to columns i and j of A.
-    
-    [col_i  col_j] @ [c -s] = [new_col_i  new_col_j]
-                       [s  c]
-    """
-    A = A.copy()
-    new_i = c * A[:, i] + s * A[:, j]
-    new_j = -s * A[:, i] + c * A[:, j]
-    A[:, i] = new_i
-    A[:, j] = new_j
-    return A
-
-def householder_bidiagonalization(A):
-    """Full Householder bidiagonalization: A = Q_L @ B @ Q_R^T.
-    
-    Returns (B, Q_L, Q_R) where B is upper bidiagonal.
-    """
-    m, n = A.shape
-    p = min(m, n)
-    
-    QL = np.eye(m, dtype=np.float64)
-    QR = np.eye(n, dtype=np.float64)
-    W = A.copy()
-    
-    for k in range(p):
-        # Left HH: zero out W[k+1:, k]
-        if k < m - 1:
-            x = W[k+1:, k].copy()
-            v, alpha = compute_householder(x)
-            if np.linalg.norm(v) > 1e-30:
-                W = apply_householder_left(W, v, k + 1)
-                QL = apply_householder_right(QL, v, k + 1)
-        
-        # Right HH: zero out W[k, k+2:] (superdiagonal)
-        if k < p - 1 and k + 2 <= n:
-            x = W[k, k+2:].copy()
-            v, alpha = compute_householder(x)
-            if np.linalg.norm(v) > 1e-30:
-                W = apply_householder_right(W, v, k + 2)
-                QR = apply_householder_right(QR, v, k + 2)
-    
-    return W, QL, QR
-
-def implicit_qr_iteration(B, QR_acc):
-    """Implicit QR iteration on a bidiagonal matrix.
-    
-    Returns (Sigma, QR_acc) where Sigma is diagonal with singular values
-    and QR_acc contains the accumulated right transformations.
-    """
-    m, n = B.shape
-    p = min(m, n)
-    W = B.copy()
-    
-    max_iter = 1000
-    tol = 1e-10
-    
-    for iteration in range(max_iter):
-        # Deflate negligible subdiagonal elements
-        for i in range(p - 1, 0, -1):
-            if abs(W[i, i-1]) < tol * (abs(W[i-1, i-1]) + abs(W[i, i])):
-                W[i, i-1] = 0.0
-        
-        # Find smallest unreduced block [start, end]
-        start = 0
-        for i in range(p - 1):
-            if abs(W[i+1, i]) >= tol * (abs(W[i, i]) + abs(W[i+1, i+1])):
-                start = i + 1
-        
-        end = p - 1
-        for i in range(p - 2, -1, -1):
-            if abs(W[i+1, i]) >= tol * (abs(W[i, i]) + abs(W[i+1, i+1])):
-                end = i
-                break
-        
-        if start >= end:
-            continue
-        
-        # Wilkinson shift from bottom 2x2 corner
-        a, b = W[end-1, end-1], W[end-1, end]
-        c_val, d = W[end, end-1], W[end, end]
-        trace = a + d
-        det = a * d - b * c_val
-        disc = trace**2 - 4 * det
-        
-        if disc >= 0:
-            sqrt_disc = np.sqrt(disc)
-            e1, e2 = (trace + sqrt_disc) / 2, (trace - sqrt_disc) / 2
-            shift = e1 if abs(e1 - d) < abs(e2 - d) else e2
-        else:
-            shift = d
-        
-        # Implicit QR step using Givens rotations
-        # Process from top to bottom within the block
-        x = W[start, start] - shift
-        y = W[start + 1, start]
-        
-        for i in range(start, end):
-            r = np.sqrt(x*x + y*y)
-            if r < 1e-30:
-                x = W[i + 1, i]
-                y = W[i + 1, i + 1] if i + 2 <= end else 0.0
-                continue
-            
-            c_rot = x / r
-            s_rot = y / r
-            
-            # Apply from left to rows i, i+1 (columns i..n-1)
-            for j in range(i, n):
-                t1, t2 = W[i, j], W[i + 1, j]
-                W[i, j] = c_rot * t1 + s_rot * t2
-                W[i + 1, j] = -s_rot * t1 + c_rot * t2
-            
-            # Apply from right to columns i, i+1 (rows 0..i)
-            if i > start:
-                for j in range(i + 1):
-                    t1, t2 = W[j, i], W[j, i + 1]
-                    W[j, i] = c_rot * t1 + s_rot * t2
-                    W[j, i + 1] = -s_rot * t1 + c_rot * t2
-            
-            # Accumulate into QR_acc
-            for j in range(QR_acc.shape[0]):
-                t1, t2 = QR_acc[j, i], QR_acc[j, i + 1]
-                QR_acc[j, i] = c_rot * t1 + s_rot * t2
-                QR_acc[j, i + 1] = -s_rot * t1 + c_rot * t2
-            
-            # Prepare for next rotation
-            x = W[i + 1, i]
-            y = W[i + 1, i + 1] if i + 2 <= end else 0.0
-    
-    return W, QR_acc
-
-
-def main():
-    print("=" * 70)
-    print("SVB BUILDING BLOCK REFERENCE VALUES")
-    print("Generated with scipy/numpy for C++ unit test verification")
-    print("=" * 70)
-    
-    # ------------------------------------------------------------------
-    # Test 1: Householder Vector Computation
-    # ------------------------------------------------------------------
-    print("\n" + "=" * 70)
-    print("TEST 1: computeHouseholderVector")
-    print("=" * 70)
-    
-    test_vectors = [
-        ("2D [1,3]", [1.0, 3.0]),
-        ("2D [3,4] (norm=5)", [3.0, 4.0]),
-        ("3D [1,2,3]", [1.0, 2.0, 3.0]),
-        ("3D [0,0,1]", [0.0, 0.0, 1.0]),
-        ("4D [5,-3,2,1]", [5.0, -3.0, 2.0, 1.0]),
-    ]
-    
-    for name, vec in test_vectors:
-        v, alpha = compute_householder(vec)
-        x = np.array(vec)
-        Hx = x - 2 * np.dot(v, x) * v
-        
-        print(f"\n{name}:")
-        print(f"  Input:            {list(x)}")
-        print(f"  ||x||:            {np.linalg.norm(x):.15f}")
-        print(f"  alpha:            {alpha:.15f}")
-        print(f"  v (normalized):   {[round(float(vi), 12) for vi in v]}")
-        print(f"  H*x = [alpha,0..]: {[round(float(xi), 12) for xi in Hx]}")
-        print(f"  Off-diagonal ~0:  {np.allclose(Hx[1:], 0, atol=1e-12)}")
-    
-    # ------------------------------------------------------------------
-    # Test 2: Householder Apply Left
-    # ------------------------------------------------------------------
-    print("\n" + "=" * 70)
-    print("TEST 2: applyHouseholderLeft")
-    print("=" * 70)
-    
-    A_test = np.array([[1.0, 2.0, 3.0], [4.0, 5.0, 6.0], [7.0, 8.0, 9.0]], dtype=np.float64)
-    x_col = A_test[1:, 0].copy()
-    v_left, _ = compute_householder(x_col)
-    
-    print(f"\nInput matrix:\n{A_test}")
-    print(f"Householder vector (rows 1:3): {[round(float(vi), 12) for vi in v_left]}")
-    
-    A_result = apply_householder_left(A_test, v_left, 1)
-    print(f"\nAfter applyHouseholderLeft:\n{A_result}")
-    print(f"  A[1,0] = {A_result[1,0]:.2e}, A[2,0] = {A_result[2,0]:.2e} (should be ~0)")
-    
-    # ------------------------------------------------------------------
-    # Test 3: Householder Apply Right
-    # ------------------------------------------------------------------
-    print("\n" + "=" * 70)
-    print("TEST 3: applyHouseholderRight")
-    print("=" * 70)
-    
-    A_test = np.array([[1.0, 2.0, 3.0], [4.0, 5.0, 6.0], [7.0, 8.0, 9.0]], dtype=np.float64)
-    x_row = A_test[0, 1:].copy()
-    v_right, _ = compute_householder(x_row)
-    
-    print(f"\nInput matrix:\n{A_test}")
-    print(f"Householder vector (cols 1:3): {[round(float(vi), 12) for vi in v_right]}")
-    
-    A_result = apply_householder_right(A_test, v_right, 1)
-    print(f"\nAfter applyHouseholderRight:\n{A_result}")
-    print(f"  A[0,1] = {A_result[0,1]:.2e}, A[0,2] = {A_result[0,2]:.2e} (should be ~0)")
-    
-    # ------------------------------------------------------------------
-    # Test 4: Givens Rotation Computation
-    # ------------------------------------------------------------------
-    print("\n" + "=" * 70)
-    print("TEST 4: computeGivens")
-    print("=" * 70)
-    
-    givens_tests = [
-        ("3-4-5 triangle", 3.0, 4.0),
-        ("y already zero", 1.0, 0.0),
-        ("x is zero", 0.0, 5.0),
-        ("Both negative", -3.0, -4.0),
-        ("45 degree case", 1.0, -1.0),
-    ]
-    
-    for name, x, y in givens_tests:
-        c, s = compute_givens(x, y)
-        result_x = c * x + s * y
-        result_y = -s * x + c * y
-        
-        print(f"\n{name}: x={x}, y={y}")
-        print(f"  r = {np.sqrt(x*x+y*y):.12f}")
-        print(f"  c = {c:.12f}, s = {s:.12f}")
-        print(f"  [c s; -s c] @ [x;y] = [{result_x:.2e}, {result_y:.2e}]")
-    
-    # ------------------------------------------------------------------
-    # Test 5: Apply Givens Left/Right
-    # ------------------------------------------------------------------
-    print("\n" + "=" * 70)
-    print("TEST 5: applyGivensLeft / applyGivensRight")
-    print("=" * 70)
-    
-    A_test = np.array([[3.0, 4.0], [1.0, 2.0]], dtype=np.float64)
-    c, s = compute_givens(3.0, 1.0)
-    
-    print(f"\nInput matrix:\n{A_test}")
-    print(f"Givens rotation (rows 0,1): c={c:.12f}, s={s:.12f}")
-    
-    A_left = apply_givens_left(A_test, 0, 1, c, s)
-    print(f"\nAfter applyGivensLeft:\n{A_left}")
-    print(f"  A[1,0] = {A_left[1,0]:.2e} (should be ~0)")
-    
-    A_test = np.array([[3.0, 1.0], [4.0, 2.0]], dtype=np.float64)
-    c, s = compute_givens(3.0, 4.0)
-    
-    print(f"\nInput matrix:\n{A_test}")
-    print(f"Givens rotation (cols 0,1): c={c:.12f}, s={s:.12f}")
-    
-    A_right = apply_givens_right(A_test, 0, 1, c, s)
-    print(f"\nAfter applyGivensRight:\n{A_right}")
-    print(f"  A[0,1] = {A_right[0,1]:.2e} (should be ~0)")
-    
-    # ------------------------------------------------------------------
-    # Test 6: Full Bidiagonalization
-    # ------------------------------------------------------------------
-    print("\n" + "=" * 70)
-    print("TEST 6: householderBidiagonalization")
-    print("=" * 70)
-    
-    bidiag_tests = [
-        ("2x2 [[1,2],[3,4]]", np.array([[1.0, 2.0], [3.0, 4.0]])),
-        ("3x3 SPD [[5,3],[3,5]]", np.array([[5.0, 3.0], [3.0, 5.0]])),
-        ("3x3 diag [[10,0,0],[0,5,0],[0,0,2]]", 
-         np.array([[10.0, 0, 0], [0, 5.0, 0], [0, 0, 2.0]])),
-        ("3x3 full [[1,2,3],[4,5,6],[7,8,10]]", 
-         np.array([[1.0, 2.0, 3.0], [4.0, 5.0, 6.0], [7.0, 8.0, 10.0]])),
-        ("Tall 4x3", np.array([[1,2,3],[4,5,6],[7,8,9],[10,11,12]], dtype=np.float64)),
-    ]
-    
-    for name, A in bidiag_tests:
-        B, QL, QR = householder_bidiagonalization(A)
-        m, n = A.shape
-        p = min(m, n)
-        
-        print(f"\n{name}:")
-        print(f"  Original:\n{A}")
-        print(f"\n  Bidiagonal B:\n{B}")
-        print(f"  Diagonal: {[round(float(B[i,i]), 10) for i in range(p)]}")
-        print(f"  Superdiag: {[round(float(B[i,i+1]), 10) for i in range(min(p-1, n-1))]}")
-        
-        recon = QL @ B @ QR.T
-        err = np.linalg.norm(recon - A, 'fro')
-        print(f"  ||QL @ B @ QR^T - A||_F = {err:.2e}")
-    
-    # ------------------------------------------------------------------
-    # Test 7: Full SVD Reference Values
-    # ------------------------------------------------------------------
-    print("\n" + "=" * 70)
-    print("TEST 7: Full SVD Reference Values (scipy.linalg.svd)")
-    print("=" * 70)
-    
-    test_matrices = [
-        ("Simple 2x2", np.array([[1,2],[3,4]], dtype=np.float64)),
-        ("SPD 2x2", np.array([[5,3],[3,5]], dtype=np.float64)),
-        ("Full-rank 3x3", np.array([[1,2,3],[4,5,6],[7,8,10]], dtype=np.float64)),
-        ("Rank-deficient 3x3", np.array([[1,2,3],[4,5,6],[7,8,9]], dtype=np.float64)),
-        ("Diagonal 3x3", np.array([[10,0,0],[0,5,0],[0,0,2]], dtype=np.float64)),
-        ("Tall 4x3", np.array([[1,2,3],[4,5,6],[7,8,9],[10,11,12]], dtype=np.float64)),
-        ("Wide 3x5", np.array([[1,2,3,4,5],[6,7,8,9,10],[11,12,13,14,15]], dtype=np.float64)),
-        ("Symmetric tri 5x5", np.array([[2,-1,0,0,0],[-1,2,-1,0,0],[0,-1,2,-1,0],[0,0,-1,2,-1],[0,0,0,-1,2]], dtype=np.float64)),
-        ("Neg values 2x3", np.array([[0.5,-0.3,0.8],[-0.2,0.7,0.1]], dtype=np.float64)),
-        ("Near-singular 2x2", np.array([[1,0],[0,1e-6]], dtype=np.float64)),
-        ("Orthogonal 3x3", np.array([[np.cos(np.pi/4), -np.sin(np.pi/4), 0],
-                                      [np.sin(np.pi/4), np.cos(np.pi/4), 0],
-                                      [0, 0, 1]], dtype=np.float64)),
-        ("Identity 3x3", np.eye(3)),
-        ("Zero 3x3", np.zeros((3,3))),
-        ("Col vector 2x1", np.array([[3],[4]], dtype=np.float64)),
-        ("Row vector 1x2", np.array([[3,4]], dtype=np.float64)),
-        # Large-size instantiation cases (N > 5). Literals MUST match the
-        # C++ test matrices in unit-tests/matrix-tests.cpp exactly, and the
-        # C++ references use float32 inputs: cast to float32 before svd().
-        ("Tall 7x5", np.array([
-            [-0.7528, 2.7043, 1.392, 0.592, -2.0639],
-            [-2.064, -2.6515, 2.1971, 0.6067, 1.2484],
-            [-2.8765, 2.8195, 1.9947, -1.726, -1.9091],
-            [-1.8996, -1.1745, 0.1485, -0.4083, -1.2526],
-            [0.6711, -2.163, -1.2471, -0.8018, -0.2636],
-            [1.7111, -1.802, 0.0854, 0.5545, -2.7213],
-            [0.6453, -1.9769, -2.6097, 2.6933, 2.7938]], dtype=np.float32)),
-        ("Square 6x6", np.array([
-            [1.2336, -0.7815, -1.6093, 0.7369, -0.2394, -1.5118],
-            [-0.0193, -1.8624, 1.6373, -0.9649, 0.6501, -0.7532],
-            [0.0803, 0.1868, -1.2606, 1.8783, 1.1005, 1.758],
-            [1.5793, 0.3916, 1.6875, -1.646, -1.2161, -1.8191],
-            [-0.6987, -0.4453, -0.9146, 1.315, -0.573, -0.8763],
-            [0.1708, -1.4363, 1.2088, -1.7018, 1.089, 1.9475]], dtype=np.float32)),
-        ("Wide 5x8", np.array([
-            [-1.5064, -2.4724, 1.5773, 1.0343, 1.145, 1.3564, -2.1298, -0.7077],
-            [-1.9207, 1.8155, 0.6165, -0.8455, -2.1822, -0.9451, -0.8741, 1.148],
-            [0.6878, 1.9361, -0.1389, -1.902, 1.0662, 1.3039, 0.3064, 1.3548],
-            [-0.031, 0.1137, -0.3623, -2.3729, -1.9605, -2.3429, 0.6821, -0.9282],
-            [0.0429, 2.0378, -1.2535, -0.4481, 1.2778, -1.356, -2.1151, -1.0512]], dtype=np.float32)),
-        ("Tall 6x4 rank-def", np.array([
-            [-0.086904, 1.410225, 1.308323, 2.234762],
-            [0.022123, 0.896751, 0.324176, 0.773607],
-            [-0.473015, 1.555111, 0.290059, 1.157726],
-            [-0.78371, 1.398884, -1.930606, -1.548717],
-            [0.201518, -0.626835, 0.976596, 0.875294],
-            [-1.24206, 1.60595, -3.078089, -2.73695]], dtype=np.float32)),
-    ]
-    
-    for name, A in test_matrices:
-        U, s, Vt = svd(A, full_matrices=False)
-        
-        print(f"\n{name}: shape={A.shape}")
-        print(f"  Singular values: {[round(float(x), 12) for x in s]}")
-        print(f"  U:\n{np.array2string(U, precision=6, floatmode='maxprec_equal')}")
-        print(f"  Vt:\n{np.array2string(Vt, precision=6, floatmode='maxprec_equal')}")
-        recon_err = np.linalg.norm(A - U @ np.diag(s) @ Vt, 'fro')
-        print(f"  Reconstruction error: {recon_err:.2e}")
-    
-    # ------------------------------------------------------------------
-    # Test 8: Implicit QR Iteration on Bidiagonal
-    # ------------------------------------------------------------------
-    print("\n" + "=" * 70)
-    print("TEST 8: implicitQRIteration")
-    print("=" * 70)
-    
-    qr_tests = [
-        ("2x2 [[1,2],[3,4]]", np.array([[1.0, 2.0], [3.0, 4.0]])),
-        ("3x3 diag", np.array([[10.0, 0, 0], [0, 5.0, 0], [0, 0, 2.0]])),
-    ]
-    
-    for name, A in qr_tests:
-        B, QL, QR = householder_bidiagonalization(A)
-        Sigma, QR_final = implicit_qr_iteration(B.copy(), QR.copy())
-        
-        print(f"\n{name}:")
-        print(f"  Bidiagonal B:\n{B}")
-        print(f"  After QR iteration (Sigma):\n{Sigma}")
-        print(f"  Diagonal entries: {[round(float(Sigma[i,i]), 10) for i in range(min(Sigma.shape))]}")
-        
-        # Verify: QL @ Sigma @ QR_final^T ≈ A
-        recon = QL @ Sigma @ QR_final.T
-        err = np.linalg.norm(recon - A, 'fro')
-        print(f"  ||QL @ Sigma @ QR^T - A||_F = {err:.2e}")
-    
-    # ------------------------------------------------------------------
-    # JSON output for easy import into C++ tests
-    # ------------------------------------------------------------------
-    print("\n" + "=" * 70)
-    print("JSON OUTPUT (for easy C++ integration)")
-    print("=" * 70)
-    
-    json_data = {}
-    
-    # Householder test vectors
-    hh_tests = {}
-    for name, vec in test_vectors:
-        v, alpha = compute_householder(vec)
-        x = np.array(vec)
-        Hx = x - 2 * np.dot(v, x) * v
-        hh_tests[name] = {
-            "input": [float(xi) for xi in x],
-            "norm": float(np.linalg.norm(x)),
-            "alpha": float(alpha),
-            "v_normalized": [round(float(vi), 12) for vi in v],
-            "Hx": [round(float(xi), 12) for xi in Hx],
-        }
-    json_data["householder_vectors"] = hh_tests
-    
-    # Full SVD reference values
-    svd_tests = {}
-    for name, A in test_matrices:
-        U, s, Vt = svd(A, full_matrices=False)
-        svd_tests[name] = {
-            "shape": list(A.shape),
-            "singular_values": [round(float(x), 12) for x in s],
-            "U": [[round(float(U[i,j]), 8) for j in range(U.shape[1])] for i in range(U.shape[0])],
-            "Vt": [[round(float(Vt[i,j]), 8) for j in range(Vt.shape[1])] for i in range(Vt.shape[0])],
-        }
-    json_data["svd_reference"] = svd_tests
-    
-    print(json.dumps(json_data, indent=2))
-
-
-if __name__ == "__main__":
-    main()

unit-tests/timing-results/matrix-timing-tests.txt

@@ -1,36 +1,56 @@
-Running matrix-timing-tests with timing
-Randomness seeded to: 3567651885
-1.857 s: Addition
-1.857 s: Timing Tests
-1.788 s: Subtraction
-1.788 s: Timing Tests
-1.929 s: Multiplication
-1.929 s: Timing Tests
-1.268 s: Scalar Multiplication
-1.268 s: Timing Tests
-1.798 s: Element Multiply
-1.798 s: Timing Tests
-1.802 s: Element Divide
-1.803 s: Timing Tests
-1.553 s: Minor Matrix
-1.554 s: Timing Tests
-1.009 s: Determinant
-1.009 s: Timing Tests
-4.076 s: Matrix of Minors
-4.076 s: Timing Tests
-1.066 s: Invert
-1.066 s: Timing Tests
-1.246 s: Transpose
-1.246 s: Timing Tests
-2.284 s: Normalize
-2.284 s: Timing Tests
-0.606 s: GET ROW
-0.606 s: Timing Tests
-24.629 s: GET COLUMN
-24.630 s: Timing Tests
-3.064 s: QR Decomposition
-3.064 s: Timing Tests
+Randomness seeded to: 3183250018
+0.179 s: Addition
+0.179 s: Timing Tests
+0.177 s: Subtraction
+0.177 s: Timing Tests
+1.931 s: Multiplication
+1.931 s: Timing Tests
+0.129 s: Scalar Multiplication
+0.129 s: Timing Tests
+0.175 s: Element Multiply
+0.175 s: Timing Tests
+0.174 s: Element Divide
+0.174 s: Timing Tests
+0.155 s: Minor Matrix
+0.155 s: Timing Tests
+0.104 s: Determinant
+0.104 s: Timing Tests
+0.465 s: Matrix of Minors
+0.465 s: Timing Tests
+0.111 s: Invert
+0.111 s: Timing Tests
+0.128 s: Transpose
+0.128 s: Timing Tests
+0.203 s: Normalize
+0.203 s: Timing Tests
+0.006 s: GET ROW
+0.006 s: Timing Tests
+0.238 s: GET COLUMN
+0.238 s: Timing Tests
 ===============================================================================
 test cases: 1 | 1 passed
 assertions: - none -
 
+	Command being timed: "build/unit-tests/matrix-timing-tests -d yes"
+	User time (seconds): 4.16
+	System time (seconds): 0.00
+	Percent of CPU this job got: 99%
+	Elapsed (wall clock) time (h:mm:ss or m:ss): 0:04.17
+	Average shared text size (kbytes): 0
+	Average unshared data size (kbytes): 0
+	Average stack size (kbytes): 0
+	Average total size (kbytes): 0
+	Maximum resident set size (kbytes): 3200
+	Average resident set size (kbytes): 0
+	Major (requiring I/O) page faults: 184
+	Minor (reclaiming a frame) page faults: 172
+	Voluntary context switches: 1
+	Involuntary context switches: 84
+	Swaps: 0
+	File system inputs: 12
+	File system outputs: 1
+	Socket messages sent: 0
+	Socket messages received: 0
+	Signals delivered: 0
+	Page size (bytes): 4096
+	Exit status: 0