Status: explanatory and normative for notation. This document does not add a security claim beyond the one in threat-model.md.
| Symbol | Meaning |
|---|---|
{0,1}^n |
Bit strings of length n. |
{0,1}^* |
Finite bit strings. |
[a,b) |
Half-open integer interval {a, ..., b-1}. |
x[i] |
The i-th byte of byte string x, zero-based. |
x[a:b] |
Bytes a through b-1, inclusive of a, exclusive of b. |
A || B |
Concatenation of byte strings. |
len(x) |
Length of x in bytes. |
LE_t(x) |
The t-bit unsigned integer x encoded as t/8 bytes, least-significant byte first. |
0x... |
Hexadecimal literal. |
byte(i) |
The byte whose unsigned value is i, 0 <= i < 256. |
repeat(b,n) |
n copies of byte b. |
Unless stated otherwise, multi-byte integers use little-endian byte order. This is the normative convention for the KEM packing, the DS address structure, the CA state, and the AEAD length suffixes. ChaCha20 uses the little-endian word loading of RFC 8439. BLAKE3 uses its specified little-endian compression encoding.
A byte string is not implicitly padded. All hashes absorb exactly the bytes listed in the relevant input expression. When a construction needs unambiguous concatenation of variable-length fields, the protocol either uses a fixed layout (all fields fixed length) or an explicit length suffix. The AEAD MAC is the principal example of the latter:
nonce || ad || ct || LE64(len(ad)) || LE64(len(ct)).
The KEM and CA-PRF inputs use fixed-length fields; their domain strings are distinct ASCII constants.
Hex strings in this documentation are lowercase, without a 0x prefix,
and encode bytes in order. Ellipses in a checksum mean that the middle
is omitted for readability; the complete value is in the named file.
ASCII domain strings are byte strings. For example,
"HYP-G/v07/ss" denotes the 12 bytes 48 59 50 2d 47 2f 76 30 37 2f 73 73.
Let
R_q = Z_q[X] / (X^256 + 1), q = 2^12 = 4096.
An element of R_q is represented by its coefficient vector
(a_0, ..., a_255), with a_i in [0,q). Addition and subtraction are
coefficient-wise modulo q. Multiplication is negacyclic:
(a*b)_k = sum_{i+j=k} a_i b_j - sum_{i+j=k+256} a_i b_j (mod q).
The implementation uses the right-hand expression directly, with an
i64 accumulator and a final low-12-bit reduction. Because q is a
power of two, reduction modulo q is a mask; no division or Barrett
reduction is used.
For module rank k, a vector is written a = (a_0, ..., a_{k-1}) and
<a,b> = sum_{i=0}^{k-1} a_i b_i in R_q.
The KEM uses k = 4. Matrices are written A = (A_{i,j}) with rows
indexed by i and columns by j.
CBD(eta) is the centered binomial distribution obtained from 4*eta
independent uniform bits. For eta = 2, the implementation uses four
bits per coefficient:
v = (b_0 + b_1) - (b_2 + b_3), b_i in {0,1} uniform.
The support is {-2,-1,0,1,2} with probabilities
Pr[-2] = 1/16, Pr[-1] = 4/16, Pr[0] = 6/16,
Pr[ 1] = 4/16, Pr[ 2] = 1/16,
mean 0, and variance 1. Coefficients are mapped to [0,q) by adding
q and masking. The 128-byte input to CBD(2) supplies exactly
256 * 4 bits.
U(R_q) denotes the uniform distribution on R_q, represented by 384
uniform bytes unpacked as 12-bit coefficients. U({0,1}^{256}) is the
uniform distribution on 32-byte strings.
The following functions are used throughout the specifications.
| Notation | Concrete instantiation |
|---|---|
H(x) |
Unkeyed BLAKE3-256 of x; 32-byte output. |
H_k(x) |
Keyed BLAKE3-256 with 32-byte key k; 32-byte output. |
XOF(x, L) |
First L bytes of the BLAKE3 XOF applied to x. |
XOF_D(domain, parts, L) |
First L bytes of XOF(domain || parts..., ...). |
DK(context, key) |
BLAKE3 derive_key(context, key) with a 32-byte key and 32-byte output. |
CA-PRF(seed, domain) is the deterministic expander defined in
spec.md. It is not an additional hardness assumption:
its final stage is BLAKE3 in XOF mode, so any security argument that
treats BLAKE3 as a random oracle or secure XOF already covers the
expander output. The cellular-automaton stage is a design hedge, not a
proven independent primitive.
For a domain string d, write d_init = d || "-init" and
d_out = d || "-out". The exact CA-PRF construction is
init = XOF(seed || d_init, 256)
rule = XOF(seed || d, 64)
state = CA^32(init; rule)
out = XOF(d || state || d_out, L).
The CA^32 operation is specified in spec.md. All
three inputs use fixed-length components except the public domain string,
which is a compile-time constant at every call site.
The following standard definitions are used only to state the intended targets. They are not proofs, and none of them has been established for HypG by an external party.
A key-encapsulation mechanism is a triple
KeyGen() -> (pk, sk),
Encaps(pk) -> (ct, K),
Decaps(sk, ct) -> K',
with K, K' in {0,1}^32.
IND-CCA2. For every probabilistic polynomial-time adversary A with
a decapsulation oracle, the advantage
Adv^{IND-CCA2}_KEM(A)
= |Pr[b' = b] - 1/2|
is negligible, where the challenger runs (pk,sk) <- KeyGen(), gives
pk to A, computes (ct*,K_0) <- Encaps(pk), sets K_1 uniform,
gives (ct*,K_b) for a hidden bit b, and answers decapsulation queries
for ct != ct*.
Implicit rejection. On an invalid ciphertext, Decaps returns a
pseudorandom value rather than a distinguished error. This avoids a
public failure oracle but is not by itself a substitute for a CCA proof.
A signature scheme is a triple
KeyGen() -> (pk, sk),
Sign(sk, m) -> sigma,
Verify(pk, m, sigma) -> {0,1}.
UF-CMA. For every probabilistic polynomial-time forger F with a
signing oracle, the probability that F outputs (m*, sigma*) with
m* never queried and Verify(pk,m*,sigma*) = 1 is negligible.
Strong unforgeability (SUF-CMA) additionally forbids a new signature
on a previously queried message.
HypG-DS is deterministic. Determinism does not imply strong unforgeability; the reserved zero padding and exact-length check are intended to remove trivial encoding malleability, but no SUF-CMA proof exists for this composition.
An AEAD scheme is a pair
Seal(K, N, AD, M) -> (C, T),
Open(K, N, AD, C, T) -> M or error.
IND-CPA and INT-CTXT are the expected confidentiality and
ciphertext-integrity goals. The construction is Encrypt-then-MAC, so
integrity is checked before any plaintext is released by the one-shot
API. The streaming opener releases tentative plaintext into the caller's
buffer, but marks it valid only when finish returns Some(()).
A family F_k is a pseudorandom function if no efficient adversary can
distinguish F_k from a random function with a non-negligible advantage.
An extendable-output function is treated as a random oracle producing
arbitrarily many bytes. The instantiation is BLAKE3; the documentation
does not claim that BLAKE3 has been proven secure.
For a random variable X, E[X] is its expectation and Var[X] its
variance. sigma denotes a standard deviation. A working correctness
margin is written as
margin = threshold / sigma,
with the threshold stated explicitly. A margin is a heuristic summary, not a failure bound.
Q(x) is the Gaussian tail probability
Pr[Z > x] for a standard normal Z. Any use of Q in this
documentation is explicitly labelled a working estimate.
Ais the Module-LWE public matrix.s, eare the long-term secret and error vectors.r, e1, e2are per-encapsulation randomness.mis the 32-byte FO message. In the public API, theencapsargument namedcoinsis exactly thism; it is both the embedded message and the input to the FO randomness generator. It must be fresh and uniformly random for every encapsulation.H(pk)is the 32-byte unkeyed hash of the public key used in the KEM shared-secret derivation.fakeis the 32-byte implicit-rejection seed embedded insk.ADRSis the 32-byte DS address structure.pubseedis the 32-byte public DS hash key.eskis the 32-byte master DS secret;prfis a derived DS secret.Ris the 32-byte DS randomizer.mdis the 64-byte DS message digest.
In testplan.md and metrics.md:
| Term | Meaning |
|---|---|
n |
Number of timed samples or test trials, as stated. |
pm |
Per-mille ratio: 1000 means identical. |
|t| |
Absolute Welch t-statistic; a smoke signal, not a leakage proof. |
E0–E4 |
Evidence levels defined in the test plan. |
| "recorded" | Produced by the command shown, on the machine shown, and archived locally. |
| "not reproduced in this pass" | Prior local evidence retained for context, not re-run during the documentation change. |
If this document conflicts with a crate's Rust documentation, the code and frozen KAT files take precedence. If a future change alters bytes, the version map in README.md and the frozen files must be updated together; silently changing a wire format is a specification violation.