Shannon's Source Coding Theorem: The Fundamental Limit of Compression

Terminology

What & Why

In 1948, Claude Shannon published "A Mathematical Theory of Communication," founding information theory. His source coding theorem establishes the fundamental limit of lossless data compression: you cannot compress data below its entropy, and you can get arbitrarily close to it.

Entropy $H(X)$ measures the average "surprise" per symbol. A source that always outputs the same symbol has zero entropy (no surprise, infinite compressibility). A source that outputs each of $n$ symbols with equal probability has entropy $\log_2 n$ bits per symbol (maximum surprise, no compressibility beyond $\log_2 n$ bits per symbol).

Shannon's theorem says: for a source with entropy $H$ bits per symbol, any lossless compression scheme must use at least $H$ bits per symbol on average. Conversely, there exist codes that use at most $H + \epsilon$ bits per symbol for any $\epsilon > 0$ (by coding long blocks). Huffman coding achieves within 1 bit of entropy per symbol, and arithmetic coding gets arbitrarily close.

This theorem is the reason gzip, zstd, and every other compressor has a limit. It is the reason truly random data cannot be compressed. It is the mathematical foundation of every compression algorithm ever built.

How It Works

Entropy: The Measure of Information

For a discrete random variable $X$ with possible values $x_1, \ldots, x_n$ and probabilities $p_1, \ldots, p_n$:

Properties of entropy:

• $H(X) \geq 0$, with equality iff $X$ is deterministic.
• $H(X) \leq \log_2 n$, with equality iff all outcomes are equally likely.
• $H(X)$ is measured in bits (when using $\log_2$).

The Source Coding Theorem

Let $X_1, X_2, \ldots$ be i.i.d. random variables from a source with entropy $H$. For any lossless code:

And there exists a code achieving:

when encoding blocks of $n$ symbols at a time. As $n \to \infty$, the rate approaches $H(X)$.

Huffman Coding: The Optimal Symbol Code

Huffman coding assigns shorter codewords to more frequent symbols. It builds a binary tree bottom-up:

1. Create a leaf node for each symbol with its probability.
2. Repeatedly merge the two nodes with the smallest probabilities.
3. Assign 0 to the left branch and 1 to the right branch.

The resulting code is optimal among all prefix-free codes (minimum expected length). The expected length $L$ satisfies:

Why You Cannot Beat Entropy

The proof uses Kraft's inequality. For any prefix-free code with lengths $l_1, \ldots, l_n$:

The expected code length is $L = \sum p_i l_i$. Using the method of Lagrange multipliers (or the log-sum inequality), the minimum of $L$ subject to Kraft's inequality is achieved when $l_i = -\log_2 p_i$, giving $L = H(X)$. Since code lengths must be integers, the actual minimum is at most $H(X) + 1$.

Complexity Analysis

Example: a source with $P(A) = 0.5, P(B) = 0.25, P(C) = 0.125, P(D) = 0.125$:

Huffman code: A = 0, B = 10, C = 110, D = 111. Expected length: $0.5(1) + 0.25(2) + 0.125(3) + 0.125(3) = 1.75$ bits. This matches entropy exactly because the probabilities are all powers of 2.

Implementation

ALGORITHM ShannonEntropy(probabilities)
INPUT: probabilities: list of symbol probabilities (must sum to 1)
OUTPUT: entropy in bits
BEGIN
  H <- 0
  FOR EACH p IN probabilities DO
    IF p > 0 THEN
      H <- H - p * LOG2(p)
    END IF
  END FOR
  RETURN H
END

ALGORITHM HuffmanBuildTree(symbols, probabilities)
INPUT: symbols: list of symbols, probabilities: corresponding probabilities
OUTPUT: root of Huffman tree
BEGIN
  // Create leaf nodes
  heap <- MIN_PRIORITY_QUEUE
  FOR i FROM 0 TO LENGTH(symbols) - 1 DO
    node <- CREATE_LEAF(symbols[i], probabilities[i])
    INSERT node INTO heap WITH PRIORITY probabilities[i]
  END FOR

  // Build tree bottom-up
  WHILE SIZE(heap) > 1 DO
    left <- EXTRACT_MIN(heap)
    right <- EXTRACT_MIN(heap)
    parent <- CREATE_INTERNAL_NODE(left, right)
    parent.probability <- left.probability + right.probability
    INSERT parent INTO heap WITH PRIORITY parent.probability
  END WHILE

  RETURN EXTRACT_MIN(heap)  // root of the tree
END

ALGORITHM HuffmanEncode(message, codeTable)
INPUT: message: sequence of symbols
       codeTable: map from symbol to binary codeword
OUTPUT: compressed binary string
BEGIN
  encoded <- empty bit string
  FOR EACH symbol IN message DO
    APPEND codeTable[symbol] TO encoded
  END FOR
  RETURN encoded
END

ALGORITHM HuffmanDecode(bitString, tree)
INPUT: bitString: compressed binary string
       tree: Huffman tree root
OUTPUT: decoded message
BEGIN
  decoded <- empty list
  node <- tree

  FOR EACH bit IN bitString DO
    IF bit = 0 THEN
      node <- node.left
    ELSE
      node <- node.right
    END IF

    IF node IS LEAF THEN
      APPEND node.symbol TO decoded
      node <- tree  // restart from root
    END IF
  END FOR

  RETURN decoded
END

ALGORITHM VerifySourceCodingTheorem(source, codeTable)
INPUT: source: probability distribution over symbols
       codeTable: map from symbol to codeword
OUTPUT: comparison of expected length vs entropy
BEGIN
  H <- ShannonEntropy(source.probabilities)

  expectedLength <- 0
  FOR EACH (symbol, prob) IN source DO
    codeLength <- LENGTH(codeTable[symbol])
    expectedLength <- expectedLength + prob * codeLength
  END FOR

  ASSERT expectedLength >= H  // source coding theorem lower bound
  ASSERT expectedLength < H + 1  // Huffman guarantee

  RETURN (entropy: H, expectedLength: expectedLength, gap: expectedLength - H)
END

Real-World Applications

• Every lossless compression algorithm (gzip, zstd, brotli, lz4) is bounded by Shannon's theorem: they can approach entropy but never beat it, and their effectiveness depends on how far the source's entropy is below the raw bit rate
• Huffman coding is used in JPEG image compression (for encoding quantized DCT coefficients), DEFLATE (the algorithm behind gzip and PNG), and MP3 audio compression
• Arithmetic coding, which gets closer to entropy than Huffman, is used in H.264/H.265 video codecs (CABAC), JPEG 2000, and modern compression libraries like zstd and ANS
• Shannon's theorem explains why encrypted data and truly random data cannot be compressed: their entropy is maximal (1 bit per bit), so no compression is possible
• In communication systems, the source coding theorem determines the minimum bandwidth needed to transmit data from a source, informing the design of modems, satellite links, and streaming protocols
• The theorem connects to Kolmogorov complexity: Shannon entropy is the expected Kolmogorov complexity for strings from a probabilistic source, bridging the probabilistic and individual views of information

Key Takeaways

• Shannon's source coding theorem establishes that entropy $H(X)$ is the fundamental lower bound on lossless compression: no code can achieve fewer than $H$ bits per symbol on average
• Entropy $H(X) = -\sum p_i \log_2 p_i$ measures the average information content per symbol; it is zero for deterministic sources and maximal ($\log_2 n$) for uniform distributions
• Huffman coding constructs an optimal prefix-free code with expected length between $H$ and $H + 1$ bits per symbol; it matches entropy exactly when probabilities are powers of 2
• Arithmetic coding and block coding can get arbitrarily close to entropy by encoding longer sequences, approaching the theoretical limit as block length increases
• The theorem explains why random and encrypted data are incompressible (entropy equals the raw bit rate) and why structured data compresses well (entropy is much lower)
• Shannon's 1948 paper founded information theory and established the mathematical framework that underlies all modern data compression, communication, and coding theory