Torus Compression

Torus compression is a technique for reducing proof size by compressing elliptic curve points in Dory polynomial commitments. This optimization can achieve an estimated 3x reduction in proof size with minimal computational overhead.

Torus compression is mentioned in Jolt’s profiling benchmarks but may not be fully implemented in the current version. This page describes the theoretical technique and expected performance impact.

Problem: Large Curve Points

Dory commitments use elliptic curve points from BN254:

// G1 point (uncompressed): 64 bytes
// - x-coordinate: 32 bytes (field element)
// - y-coordinate: 32 bytes (field element)

// G2 point (uncompressed): 128 bytes
// - x-coordinate: 64 bytes (Fp2 element)
// - y-coordinate: 64 bytes (Fp2 element)

A typical Jolt proof contains:

Polynomial commitments: 50-100 G1 points
Opening proofs: 20-40 G1 points (Dory tier-1/tier-2)
BlindFold (ZK mode): 100-200 G1 points (Pedersen commitments, Hyrax row commitments)

Total: ~10-20 KB of curve points

Solution: Torus-Based Compression

Torus compression exploits the algebraic structure of elliptic curves to represent points more compactly.

Mathematical Background

For an elliptic curve point P = (x, y) on a curve y² = x³ + ax + b: Standard compressed format:

Store x (32 bytes) + 1 bit for y sign
Total: 33 bytes
Decompression: Solve for y = ±√(x³ + ax + b)

Torus compression:

Represent point in projective coordinates: [X : Y : Z]
Exploit torus structure: T² ≅ E[r] × E[r] (for r-torsion points)
Compress to ~21 bytes (estimated)
Decompression: More complex but still efficient

Compression Ratio

Format	G1 Size	G2 Size	Compression Ratio
Uncompressed	64 bytes	128 bytes	1x (baseline)
Standard compressed	33 bytes	65 bytes	~2x
Torus compressed	~21 bytes	~42 bytes	~3x

Implementation Strategy

Compression During Serialization

impl<C: JoltCurve> JoltProof<C> {
    pub fn serialize_compressed(&self) -> Vec<u8> {
        let mut bytes = Vec::new();
        
        // Compress all curve points
        for commitment in &self.commitments {
            bytes.extend(torus_compress(commitment));
        }
        
        // Scalar values (no compression)
        for scalar in &self.scalars {
            bytes.extend(scalar.to_bytes());
        }
        
        bytes
    }
    
    pub fn deserialize_compressed(bytes: &[u8]) -> Result<Self, DecompressionError> {
        let mut offset = 0;
        let mut commitments = Vec::new();
        
        // Decompress curve points
        while offset < commitments_end {
            let point = torus_decompress(&bytes[offset..offset + TORUS_G1_SIZE])?;
            commitments.push(point);
            offset += TORUS_G1_SIZE;
        }
        
        // ... deserialize scalars ...
        
        Ok(Self { commitments, scalars, ... })
    }
}

Compression Levels

Different proof components may use different compression:

pub enum CompressionLevel {
    None,          // Uncompressed (64 bytes G1)
    Standard,      // Standard compressed (33 bytes G1)
    Torus,         // Torus compressed (~21 bytes G1)
}

pub struct CompressionConfig {
    commitments: CompressionLevel,    // Polynomial commitments
    opening_proofs: CompressionLevel, // Dory opening proofs
    blindfold: CompressionLevel,      // BlindFold ZK proofs
}

Trade-off: Torus decompression is slower than standard decompression, so use it selectively for components that dominate proof size.

Performance Impact

Proof Size Reduction

For a typical Jolt proof:

Component	Count	Uncompressed	Torus Compressed	Savings
Polynomial commitments	80	5,120 bytes	1,680 bytes	3,440 bytes
Dory opening proof	30	1,920 bytes	630 bytes	1,290 bytes
BlindFold (ZK)	150	9,600 bytes	3,150 bytes	6,450 bytes
Sumcheck polynomials	—	4,000 bytes	4,000 bytes	0 (no compression)
Total	—	20,640 bytes	9,460 bytes	~2.2x reduction

Note: Actual compression ratio depends on the specific proof components. Estimated ~3x for curve points alone, ~2-2.5x for full proof (including scalars).

Computational Cost

Compression time (per G1 point):

Standard compression: ~5 μs
Torus compression: ~10 μs (estimated)

Decompression time (per G1 point):

Standard decompression: ~20 μs (requires square root)
Torus decompression: ~50 μs (estimated, more complex reconstruction)

Impact on prover:

Proof generation: Minimal overhead (~1-2% increase)
Proof serialization: Slight overhead from compression

Impact on verifier:

Proof deserialization: ~2-3x slower decompression
Verification: No change (works with decompressed points)

Net effect: Worthwhile trade-off for bandwidth-constrained applications (blockchain L1 submission, storage).

Use Cases

When to Use Torus Compression

✅ Good for:

On-chain proof submission (minimize calldata costs)
Proof archival (long-term storage)
Network transmission over slow connections
Applications where verification is infrequent

❌ Not ideal for:

High-throughput verification (latency-sensitive)
Local proof generation and verification
Applications that re-verify proofs frequently

Hybrid Approach

Use different compression for different stages:

// Stage 1: Generate proof (no compression)
let proof = jolt_prover.prove();

// Stage 2: Serialize for transmission (torus compression)
let compressed_bytes = proof.serialize_compressed(CompressionLevel::Torus);

// Stage 3: Transmit compressed proof
send_to_verifier(compressed_bytes);

// Stage 4: Verifier decompresses and verifies
let proof = JoltProof::deserialize_compressed(&compressed_bytes)?;
verifier.verify(&proof)?;

Implementation Status

Torus compression is not yet fully implemented in the current Jolt codebase. The compression estimates are based on theoretical analysis and benchmarking projections.

Current State

// From jolt-core/benches/e2e_profiling.rs

// Stage 8: Dory opening proof (curve points - benefits from compression)
// println!("  Dory opening: ~{} bytes", dory_bytes);

// Commitments (curve points - benefits from compression)
// println!("  Commitments: ~{} bytes", commitment_bytes);

// Estimate proof size with full Dory compression (assuming ~3x compression ratio)
// let compressed_size = (dory_bytes + commitment_bytes) / 3 + other_bytes;

These comments indicate planned compression support.

Roadmap

Phase 1: Standard Compression (likely implemented)

Use arkworks’ built-in point compression
Serialize G1 points as 33 bytes (x-coordinate + sign bit)
Serialize G2 points as 65 bytes

Phase 2: Torus Compression (planned)

Implement torus-based compression algorithm
Integrate with Dory commitment scheme
Benchmarking and optimization

Phase 3: Adaptive Compression (future)

Automatically choose compression level based on use case
Profile-guided compression (compress hot proof components less aggressively)

Other Proof Compression Methods

1. Recursive SNARKs:

Compress proof by proving its verification in another SNARK
Can achieve constant-size proofs (~1-2 KB)
Much higher prover cost (10-100x)

2. Batch Verification:

Amortize verification cost across multiple proofs
Doesn’t reduce individual proof size
Complementary to compression

3. Aggregation:

Combine multiple proofs into one
Useful for blockchain rollups
Can be combined with torus compression

Best Practices

Compression Configuration

// For on-chain submission (minimize size)
let config = CompressionConfig {
    commitments: CompressionLevel::Torus,
    opening_proofs: CompressionLevel::Torus,
    blindfold: CompressionLevel::Torus,
};

// For local/fast verification (minimize latency)
let config = CompressionConfig {
    commitments: CompressionLevel::Standard,
    opening_proofs: CompressionLevel::Standard,
    blindfold: CompressionLevel::None,
};

Error Handling

Torus decompression can fail if the compressed data is invalid:

match JoltProof::deserialize_compressed(&bytes) {
    Ok(proof) => verifier.verify(&proof),
    Err(DecompressionError(invalid_point)) => {
        // Handle invalid compressed point
        // This could indicate:
        // - Corrupted proof data
        // - Malicious proof
        // - Wrong compression format
    }
}

Batched Openings: Reduce number of curve points in proof
Batched Sumcheck: Reduce number of sumcheck rounds

References

BN254 Curve Specification — Pairing-friendly elliptic curve
Torus-Based Cryptography — Theoretical foundation (example paper)
Profiling benchmark: jolt-core/benches/e2e_profiling.rs

Overview

Core Components

Protocols

Optimizations

Theory

Torus Compression

Problem: Large Curve Points

Solution: Torus-Based Compression

Mathematical Background

Compression Ratio

Implementation Strategy

Compression During Serialization

Compression Levels

Performance Impact

Proof Size Reduction

Computational Cost

Use Cases

When to Use Torus Compression

Hybrid Approach

Implementation Status

Current State

Roadmap

Other Proof Compression Methods

Best Practices

Compression Configuration

Error Handling

References

Build docs developers (and LLMs) love

Overview

Core Components

Protocols

Optimizations

Theory

Documentation Index

​Problem: Large Curve Points

​Solution: Torus-Based Compression

​Mathematical Background

​Compression Ratio

​Implementation Strategy

​Compression During Serialization

​Compression Levels

​Performance Impact

​Proof Size Reduction

​Computational Cost

​Use Cases

​When to Use Torus Compression

​Hybrid Approach

​Implementation Status

​Current State

​Roadmap

​Related Techniques

​Other Proof Compression Methods

​Best Practices

​Compression Configuration

​Error Handling

​Related Optimizations

​References

Build docs developers (and LLMs) love

Problem: Large Curve Points

Solution: Torus-Based Compression

Mathematical Background

Compression Ratio

Implementation Strategy

Compression During Serialization

Compression Levels

Performance Impact

Proof Size Reduction

Computational Cost

Use Cases

When to Use Torus Compression

Hybrid Approach

Implementation Status

Current State

Roadmap

Related Techniques

Other Proof Compression Methods

Best Practices

Compression Configuration

Error Handling

Related Optimizations

References