LinAlgKit

Release Notes

This document contains detailed release notes for each version of LinAlgKit.

---

v0.2.1 - Performance Optimizations (2025-12-24)

🚀 Performance Improvements

This release introduces significant performance optimizations using Numba JIT compilation.

Benchmark Results (5M elements)

Function	Standard	Fast (JIT)	Speedup
mse_loss	33.75ms	2.82ms	12.0x
mae_loss	34.86ms	2.67ms	13.1x
leaky_relu	30.16ms	6.81ms	4.4x
gelu	200.78ms	76.44ms	2.6x
tanh	37.33ms	15.29ms	2.4x
swish	100.27ms	54.62ms	1.8x
elu	65.14ms	40.06ms	1.6x
sigmoid	80.98ms	54.71ms	1.5x
softplus	108.80ms	70.88ms	1.5x

New Features

fast Module

A new high-performance module with Numba JIT-compiled functions:

from LinAlgKit import fast

# 12x faster loss computation
loss = fast.fast_mse_loss(predictions, targets)

# 4.4x faster activation
output = fast.fast_leaky_relu(x, alpha=0.01)

Available fast functions:

• fast_sigmoid, fast_relu, fast_leaky_relu, fast_elu
• fast_gelu, fast_swish, fast_tanh, fast_softplus
• fast_mse_loss, fast_mae_loss, fast_normalize

In-Place Operations

New in-place methods that avoid memory allocation:

A.add_(B)      # A += B in-place
A.sub_(B)      # A -= B in-place
A.mul_(2.0)    # A *= 2.0 in-place
A.hadamard_(B) # Element-wise multiply in-place

Zero-Copy Access

• .T property - Transpose view (no copy)
• to_numpy_view() - Direct array access (no copy)
• transpose(copy=False) - Optional zero-copy transpose

Dependencies

• Optional: numba>=0.57.0 (for JIT acceleration)

---

v0.2.0 - Deep Learning Expansion (2025-12-24)

🧠 Deep Learning Functions

Major expansion with 50+ mathematical functions for deep learning.

Activation Functions

• sigmoid, relu, leaky_relu, elu, gelu, swish
• softplus, tanh, softmax, log_softmax
• Derivative functions for backpropagation

Loss Functions

• mse_loss - Mean Squared Error
• mae_loss - Mean Absolute Error
• huber_loss - Robust loss
• cross_entropy_loss - Multi-class classification
• binary_cross_entropy - Binary classification

Normalization

• batch_norm - Batch normalization
• layer_norm - Layer normalization (transformers)
• instance_norm - Instance normalization (style transfer)

Convolution Operations

• conv2d - 2D convolution
• max_pool2d, avg_pool2d - Pooling
• global_avg_pool2d - Global average pooling

Weight Initialization

• xavier_uniform, xavier_normal - For tanh/sigmoid
• he_uniform, he_normal - For ReLU networks

Utility Functions

• dropout, one_hot, clip, flatten, reshape
• normalize, cosine_similarity, euclidean_distance
• pairwise_distances, numerical_gradient

Matrix Decompositions

• lu() - LU decomposition with partial pivoting
• qr() - QR decomposition
• cholesky() - Cholesky decomposition
• svd() - Singular Value Decomposition

Eigenvalue Methods

• eig() - Eigenvalues and eigenvectors
• eigvals() - Eigenvalues only
• eigh() - Symmetric eigenvalue decomposition

Linear Solvers

• solve() - Solve Ax = b
• inv() - Matrix inverse
• pinv() - Moore-Penrose pseudoinverse
• lstsq() - Least-squares solution

Matrix Analysis

• norm() - Multiple norm types
• cond() - Condition number
• rank() - Matrix rank

---

v0.1.0 - Initial Release (2025-12-24)

Core Features

Matrix Classes

• Matrix - Double-precision (float64)
• MatrixF - Single-precision (float32)
• MatrixI - Integer matrix

Static Constructors

Matrix.identity(3)   # 3x3 identity
Matrix.zeros(2, 3)   # 2x3 zeros
Matrix.ones(4, 4)    # 4x4 ones

Matrix Operations

• transpose() - Transposition
• trace() - Sum of diagonal
• determinant() - LU-based O(n³)
• determinant_naive() - Recursive (small matrices)

Arithmetic Operators

• A + B - Element-wise addition
• A - B - Element-wise subtraction
• A * B - Matrix multiplication
• 2 * A - Scalar multiplication

NumPy Interoperability

# From NumPy
A = Matrix.from_numpy(np_array)

# To NumPy
np_array = A.to_numpy()

Element Access

val = A[i, j]    # Get element
A[i, j] = 5.0    # Set element

---

Upgrade Guide

From v0.1.0 to v0.2.0

No breaking changes. All v0.1.0 code works unchanged.

New imports available:

import LinAlgKit as lk

# New activation functions
output = lk.relu(x)
output = lk.softmax(logits)

# New matrix methods
L = A.cholesky()
eigenvalues, eigenvectors = A.eig()
x = A.solve(b)

From v0.2.0 to v0.2.1

No breaking changes. Install numba for automatic acceleration:

pip install numba

Use fast module for best performance:

from LinAlgKit import fast

# Check if Numba is available
import LinAlgKit as lk
print(lk.HAS_NUMBA)  # True if numba installed

---

Roadmap

See EXPANSION_PLAN.md for the complete roadmap.

Upcoming Features

v0.3.0 (Q1 2025)

• Sparse matrix support (CSR, COO)
• Iterative solvers (CG, GMRES)
• Matrix functions (expm, logm, sqrtm)

v0.4.0 (Q2 2025)

• C++/Rust native core
• BLAS/LAPACK integration
• GPU acceleration (CUDA)

v1.0.0 (2027)

• Production-ready
• Multi-backend support
• Distributed computing

This site is open source. Improve this page.