Hash Fusing
data, hash, sha256, matrix, fusing, merkle-tree
This is an article about fusing hashes together while having the key properties of associativity and non-commutativity. I describe the basic approach of using matrix multiplication of Upper Triangle Matrices and then show the results of two experiments showing the quality and limits of this hash fusing approach.
The basic question this article tries to answer is: can fusing hashes together provide unique identifiers for arbitrary data? The answer is no. Specifically, the approach in this article can not provide unique identifiers for sufficiently large runs of repeated values. That is, sufficiently low entropy data is not uniquely represented by fusing hashes together.
The good news is that low entropy data may always be converted into high entropy data either via compression or injecting noise. A practical fusing operator can easily detect fuses of low entropy data and throw an exception when fusing produces those ‘bad’ hash values.
Introduction
I have begun working on implementing distributed immutable data structures as a way to efficiently share structured data between multiple clients. One of the key challenges in this area is being able to efficiently reference ordered collections of data. The usual approach is to use Merkle Trees which have the limitation that they are non-associative and the shape of the tree determines the final hash value at the root of the tree. But, since I plan on using Finger Trees to efficiently represent ordered collections of data, I need computed hashes that are insensitive to the tree shape. This means that I need to be able to fuse hashes together in a way that is associative and also non-commutative. As a starting point, I have been reading the HP paper describing hash fusing via matrix multiplication: https://www.labs.hpe.com/techreports/2017/HPE-2017-08.pdf
Hash Fusing via Upper Triangular Matrix Multiplication
The basic approach of fusing two hashes is to represent each hash as an upper triangular matrix and then to multiply the two matrices together. Since matrix multiplication is associative but not commutative. The following sections define hashes and build up the necessary functions to perform hash fusing via upper triangular matrix multiplication.
Hash Representation
A hash is represented as a string of 64 hexadecimal values (256 bits). To start with, here is a zero hash and a function to generate random hashes.
(def zero-hex (apply str (vec (repeat 64 0))))"0000000000000000000000000000000000000000000000000000000000000000"(defn random-hex
"Generate a random hex string representing length `n` bytes."
[n]
(let [hex-chars "0123456789abcdef"]
(apply str (repeatedly (* 2 n) #(rand-nth hex-chars)))))Generate some random hashes for testing:
(def a-hex (random-hex 32))"6c11426110fdae2d89adb5b40a3df66ef6ae68b4645a7447319b6501bd0314cb"(def b-hex (random-hex 32))"f800427d051c13b8eaa423a0ef6eeb90154fc6a13328d9de928ac328c8d681ca"(def c-hex (random-hex 32))"ab4c8a375464e6f2ea71b8b13afd8c5ebd4d03decf1311dcee3db2588e7c08bc"To fuse two hashes, we convert each hash to an upper triangular matrix and then multiply the two matrices together. The result is another upper triangular matrix which can be converted back to a hash by taking the elements above the main diagonal. The following several sections defines this mapping between hashes and upper triangular matrices. For the experiments below four different bit sizes of cells and four corresponding matrices are defined.
8 bit cells in a 9x9 matrix:
\[%% Example: 9x9 Upper Triangular Matrix for 8 bit cells %% representing a 256 bit hash hash \to [ h_0, h_1, h_2, h_3, \ldots, h_{31} ] \to {\begin{bmatrix} 1 & h_0 & h_8 & h_{15} & h_{21} & h_{26} & h_{29} & h_{31} & h_{25} \\ 0 & 1 & h_1 & h_9 & h_{16} & h_{22} & h_{27} & h_{30} & h_{20} \\ 0 & 0 & 1 & h_2 & h_{10} & h_{17} & h_{23} & h_{28} & h_{14} \\ 0 & 0 & 0 & 1 & h_3 & h_{11} & h_{18} & h_{24} & h_7 \\ 0 & 0 & 0 & 0 & 1 & h_4 & h_{12} & h_{19} & 0 \\ 0 & 0 & 0 & 0 & 0 & 1 & h_5 & h_{13} & 0 \\ 0 & 0 & 0 & 0 & 0 & 0 & 1 & h_6 & 0 \\ 0 & 0 & 0 & 0 & 0 & 0 & 0 & 1 & 0 \\ 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 1 \end{bmatrix}} \]
16 bit cells in a 7x7 matrix:
\[%% Example: 7x7 Upper Triangular Matrix for 16 bit cells %% representing a 256 bit hash hash \to [ h_0, h_1, h_2, h_3, \ldots, h_{15} ] \to {\begin{bmatrix} 1 & h_0 & h_6 & h_{10} & h_{13} & h_{15} & h_5 \\ 0 & 1 & h_1 & h_7 & h_{11} & h_{14} & 0 \\ 0 & 0 & 1 & h_2 & h_8 & h_{12} & 0 \\ 0 & 0 & 0 & 1 & h_3 & h_9 & 0 \\ 0 & 0 & 0 & 0 & 1 & h_4 & 0 \\ 0 & 0 & 0 & 0 & 0 & 1 & 0 \\ 0 & 0 & 0 & 0 & 0 & 0 & 1 \end{bmatrix}} \]
32 bit cells in a 5x5 matrix:
\[%% Example: 5x5 Upper Triangular Matrix for 32 bit cells %% representing a 256 bit hash hash \to [ h_0, h_1, h_2, h_3, \ldots, h_7 ] \to {\begin{bmatrix} 1 & h_0 & h_4 & h_7 & h_6 \\ 0 & 1 & h_1 & h_5 & h_3 \\ 0 & 0 & 1 & h_2 & 0 \\ 0 & 0 & 0 & 1 & 0 \end{bmatrix}} \]
64 bit cells in a 4x4 matrix:
\[%% Example: 4x4 Upper Triangular Matrix for 64 bit cells %% representing a 256 bit hash hash \to [ h_0, h_1, h_2, h_3 ] \to {\begin{bmatrix} 1 & h_0 & h_3 & h_2 \\ 0 & 1 & h_1 & 0 \\ 0 & 0 & 1 & 0 \\ 0 & 0 & 0 & 1 \end{bmatrix}} \]
Hex and Byte Vector Conversion Functions
The first type of conversion is between hex strings and byte vectors.
(defn hex->hash8
"Convert a hex string to a byte vector."
[s]
(with-meta
(->> s
(partition 2)
(map #(-> (apply str %)
(Integer/parseInt 16)))
(vec))
{:cell-size 8}))(defn hash8->hex
"Convert a byte vector to a hex string."
[bytes]
(->> bytes
(map #(format "%02x" %))
(apply str)))(= a-hex (-> a-hex hex->hash8 hash8->hex))trueHash Conversion Functions
The next type of conversion is from bytes and larger word sizes.
(defn hash8->hash16
"Convert a vector of 32 bytes into a vector of 16 2-byte words."
[hash8]
(with-meta
(vec (map (fn [i]
(+ (bit-shift-left (hash8 (* 2 i)) 8)
(hash8 (+ 1 (* 2 i)))))
(range 16)))
{:cell-size 16}))(defn hash16->hash8
"Convert a vector 2-byte words into a vector of bytes."
[hash16]
(with-meta
(vec (mapcat (fn [word]
[(bit-and (bit-shift-right word 8) 0xFF)
(bit-and word 0xFF)])
hash16))
{:cell-size 8}))(= a-hex (-> a-hex hex->hash8 hash8->hash16 hash16->hash8 hash8->hex))true(defn hash8->hash32
"Convert a vector of 32 bytes into a vector of 8 4-byte words."
[hash8]
(with-meta
(vec (map (fn [i]
(+ (bit-shift-left (hash8 (* 4 i)) 24)
(bit-shift-left (hash8 (+ 1 (* 4 i))) 16)
(bit-shift-left (hash8 (+ 2 (* 4 i))) 8)
(hash8 (+ 3 (* 4 i)))))
(range 8)))
{:cell-size 32}))(defn hash32->hash8
"Convert a vector of 4-byte words into a vector of bytes."
[hash32]
(with-meta
(vec (mapcat (fn [word]
[(bit-and (bit-shift-right word 24) 0xFF)
(bit-and (bit-shift-right word 16) 0xFF)
(bit-and (bit-shift-right word 8) 0xFF)
(bit-and word 0xFF)])
hash32))
{:cell-size 8}))(= a-hex (-> a-hex hex->hash8 hash8->hash32 hash32->hash8 hash8->hex))true(defn hash8->hash64
"Convert a vector of 32 bytes into a vector of 4 8-byte words."
[hash8]
(with-meta
(vec (map (fn [i]
(+ (bit-shift-left (hash8 (* 8 i)) 56)
(bit-shift-left (hash8 (+ 1 (* 8 i))) 48)
(bit-shift-left (hash8 (+ 2 (* 8 i))) 40)
(bit-shift-left (hash8 (+ 3 (* 8 i))) 32)
(bit-shift-left (hash8 (+ 4 (* 8 i))) 24)
(bit-shift-left (hash8 (+ 5 (* 8 i))) 16)
(bit-shift-left (hash8 (+ 6 (* 8 i))) 8)
(hash8 (+ 7 (* 8 i)))))
(range 4)))
{:cell-size 64}))(defn hash64->hash8
"Convert a vector of 8-byte words into a vector of bytes."
[hash64]
(with-meta
(vec (mapcat (fn [word]
[(bit-and (bit-shift-right word 56) 0xFF)
(bit-and (bit-shift-right word 48) 0xFF)
(bit-and (bit-shift-right word 40) 0xFF)
(bit-and (bit-shift-right word 32) 0xFF)
(bit-and (bit-shift-right word 24) 0xFF)
(bit-and (bit-shift-right word 16) 0xFF)
(bit-and (bit-shift-right word 8) 0xFF)
(bit-and word 0xFF)])
hash64))
{:cell-size 8}))(= a-hex (-> a-hex hex->hash8 hash8->hash64 hash64->hash8 hash8->hex))trueMatrix Conversion Functions
The following functions convert between byte vectors representing hashes and four different upper triangular matrix sizes based on cell bit size.
(defn hash8->utm8
"Convert a vector of 32 bytes into a 9x9 upper triangular matrix."
[[h00 h01 h02 h03 h04 h05 h06 h07
h08 h09 h10 h11 h12 h13 h14 h15
h16 h17 h18 h19 h20 h21 h22 h23
h24 h25 h26 h27 h28 h29 h30 h31]]
(with-meta
[[1 h00 h08 h15 h21 h26 h29 h31 h25]
[0 1 h01 h09 h16 h22 h27 h30 h20]
[0 0 1 h02 h10 h17 h23 h28 h14]
[0 0 0 1 h03 h11 h18 h24 h07]
[0 0 0 0 1 h04 h12 h19 0]
[0 0 0 0 0 1 h05 h13 0]
[0 0 0 0 0 0 1 h06 0]
[0 0 0 0 0 0 0 1 0]
[0 0 0 0 0 0 0 0 1]]
{:cell-size 8}))(defn utm8->hash8
"Convert a 9x9 upper triangular matrix back to a vector of 32 bytes."
[[[_ h00 h08 h15 h21 h26 h29 h31 h25]
[_ _ h01 h09 h16 h22 h27 h30 h20]
[_ _ _ h02 h10 h17 h23 h28 h14]
[_ _ _ _ h03 h11 h18 h24 h07]
[_ _ _ _ _ h04 h12 h19 _]
[_ _ _ _ _ _ h05 h13 _]
[_ _ _ _ _ _ _ h06 _]
[_ _ _ _ _ _ _ _ _]
[_ _ _ _ _ _ _ _ _]]]
(with-meta
(mapv #(bit-and % 0xFF)
[h00 h01 h02 h03 h04 h05 h06 h07
h08 h09 h10 h11 h12 h13 h14 h15
h16 h17 h18 h19 h20 h21 h22 h23
h24 h25 h26 h27 h28 h29 h30 h31])
{:cell-size 8}))(defn hash16->utm16
"Convert a vector of 16 2-byte words into a 7x7 upper triangular matrix."
[[h00 h01 h02 h03 h04 h05 h06 h07
h08 h09 h10 h11 h12 h13 h14 h15]]
(with-meta
[[1 h00 h06 h10 h13 h15 h05]
[0 1 h01 h07 h11 h14 0]
[0 0 1 h02 h08 h12 0]
[0 0 0 1 h03 h09 0]
[0 0 0 0 1 h04 0]
[0 0 0 0 0 1 0]
[0 0 0 0 0 0 1]]
{:cell-size 16}))(defn utm16->hash16
"Convert a 7x7 upper triangular matrix back to a vector of 16 2-byte words."
[[[_ h00 h06 h10 h13 h15 h05]
[_ _ h01 h07 h11 h14 _]
[_ _ _ h02 h08 h12 _]
[_ _ _ _ h03 h09 _]
[_ _ _ _ _ h04 _]
[_ _ _ _ _ _ _]
[_ _ _ _ _ _ _]]]
(with-meta
(mapv #(bit-and % 0xFFFF)
[h00 h01 h02 h03 h04 h05 h06 h07
h08 h09 h10 h11 h12 h13 h14 h15])
{:cell-size 16}))(defn hash32->utm32
"Convert a vector of 8 4-byte words into a 5x5 upper triangular matrix."
[[h00 h01 h02 h03 h04 h05 h06 h07]]
(with-meta
[[1 h00 h04 h07 h06]
[0 1 h01 h05 h03]
[0 0 1 h02 0]
[0 0 0 1 0]
[0 0 0 0 1]]
{:cell-size 32}))(defn utm32->hash32
"Convert a 5x5 upper triangular matrix back to a vector of 8 4-byte words."
[[[_ h00 h04 h07 h06]
[_ _ h01 h05 h03]
[_ _ _ h02 _]
[_ _ _ _ _]
[_ _ _ _ _]]]
(with-meta
(mapv #(bit-and % 0xFFFFFFFF)
[h00 h01 h02 h03 h04 h05 h06 h07])
{:cell-size 32}))(defn hash64->utm64
"Convert a vector of 4 8-byte words into a 4x4 upper triangular matrix."
[[h00 h01 h02 h03]]
(with-meta
[[1 h00 h03 h02]
[0 1 h01 0]
[0 0 1 0]
[0 0 0 1]]
{:cell-size 64}))(defn utm64->hash64
"Convert a 4x4 upper triangular matrix back to a vector of 4 8-byte words."
[[[_ h00 h03 h02]
[_ _ h01 _]
[_ _ _ _]
[_ _ _ _]]]
(with-meta
[h00 h01 h02 h03]
{:cell-size 64}))Combined Conversion Functions
The following combined conversion functions convert between hex strings into upper triangular matrices and back for the four different cell bit sizes.
(def hex->utm8
"Convert a hex string to an upper triangular matrix with 8-bit cells."
(comp hash8->utm8 hex->hash8))(def utm8->hex
"Convert an upper triangular matrix with 8-bit cells to a hex string."
(comp hash8->hex utm8->hash8))(= a-hex (-> a-hex hex->utm8 utm8->hex))true(def hex->utm16
"Convert a hex string to an upper triangular matrix with 16-bit cells."
(comp hash16->utm16 hash8->hash16 hex->hash8))(def utm16->hex
"Convert an upper triangular matrix with 16-bit cells to a hex string."
(comp hash8->hex hash16->hash8 utm16->hash16))(= a-hex (-> a-hex hex->utm16 utm16->hex))true(def hex->utm32
"Convert a hex string to an upper triangular matrix with 32-bit cells."
(comp hash32->utm32 hash8->hash32 hex->hash8))(def utm32->hex
"Convert an upper triangular matrix with 32-bit cells to a hex string."
(comp hash8->hex hash32->hash8 utm32->hash32))(= a-hex (-> a-hex hex->utm32 utm32->hex))true(def hex->utm64
"Convert a hex string to an upper triangular matrix with 64-bit cells."
(comp hash64->utm64 hash8->hash64 hex->hash8))(def utm64->hex
"Convert an upper triangular matrix with 64-bit cells to a hex string."
(comp hash8->hex hash64->hash8 utm64->hash64))(= a-hex (-> a-hex hex->utm64 utm64->hex))true(defn hex->utm
"Convert a hex string to an upper triangular matrix with the given cell size
(8, 16, 32, or 64 bits)."
[hex cell-size]
(case cell-size
8 (hex->utm8 hex)
16 (hex->utm16 hex)
32 (hex->utm32 hex)
64 (hex->utm64 hex)
(throw (ex-info "Unsupported cell size for upper triangular matrix."
{:cell-size cell-size}))))(defn utm->hex
"Convert an upper triangular matrix with the given cell size (8, 16, 32,
or 64 bits) to a hex string."
[utm cell-size]
(case cell-size
8 (utm8->hex utm)
16 (utm16->hex utm)
32 (utm32->hex utm)
64 (utm64->hex utm)
(throw (ex-info "Unsupported cell size for upper triangular matrix."
{:cell-size cell-size}))))(= a-hex (-> a-hex (hex->utm 32) (utm->hex 32)))trueUpper Triangular Matrix Multiplication
This is the core function that performs the multiplication of two upper triangular matrices. The multiplication ignores the lower triangular part of of the matrices since they are always 0 (and always 1 on the main diagonal).
Note that unchecked math is enabled to ignore integer overflow since the cells are treated as fixed size bit fields.
(set! *unchecked-math* true)true(defn utm-multiply
"Multiply two upper triangular matrices `a` and `b`."
[a b]
(let [dim (count a)
cell-size (-> a meta :cell-size)
bit-mask (if (= cell-size 64)
-1
(dec (bit-shift-left 1 cell-size)))]
(with-meta
(vec (for [i (range dim)]
(vec (for [j (range dim)]
(cond
(< j i) 0
(= j i) 1
:else
(reduce (fn [sum k]
(-> sum
(+ (* (get-in a [i k])
(get-in b [k j])))
(bit-and bit-mask)))
0
(range i (inc j))))))))
{:cell-size cell-size})))Associativity & Non-Commutativity Properties
Show that upper triangular matrix multiplication is associative and non-commutative. Associativity is necessary for hash fusing to work with Finger Trees so that different tree shapes produce the same fused hash. Non-commutativity is necessary for sequences of data where the order of data affects the fused hash.
(-> (for [cell-size [8 16 32 64]]
(let [a (hex->utm a-hex cell-size)
b (hex->utm b-hex cell-size)
c (hex->utm c-hex cell-size)
ab (utm-multiply a b)
ba (utm-multiply b a)
bc (utm-multiply b c)
ab*c (utm-multiply ab c)
a*bc (utm-multiply a bc)]
{:cell-size cell-size
:associative? (= ab*c a*bc)
:commutative? (= ab ba)}))
(kind/table))| cell-size | associative? | commutative? |
|---|---|---|
| 8 | true | false |
| 16 | true | false |
| 32 | true | false |
| 64 | true | false |
Experiment 1: Random Fuses
This experiment is run with two different hashes which are fused onto an accumulator iteratively. For each iteration, one of the two hashes is randomly selected and is fused to the accumulator. This random fusing is repeated many times and the quality of the accumulator is measured both by keeping track of global uniqueness after each fuse and by the uniform distribution of bit values.
(defn random-fuses
"Perform random fuses of two hashes onto an accumulator keeping track of all
unique hashes produced and all duplicate hashes produced. Repeat this for all
four utm sizes based on cell bit size."
[]
(let [hashes {:a (random-hex 32)
:b (random-hex 32)}
;; convert hashes to utms for each bit size and store in a map
utms {8 (update-vals hashes hex->utm8)
16 (update-vals hashes hex->utm16)
32 (update-vals hashes hex->utm32)
64 (update-vals hashes hex->utm64)}
results {8 {:acc (hex->utm8 zero-hex)
:uniques #{}
:duplicates #{}}
16 {:acc (hex->utm16 zero-hex)
:uniques #{}
:duplicates #{}}
32 {:acc (hex->utm32 zero-hex)
:uniques #{}
:duplicates #{}}
64 {:acc (hex->utm64 zero-hex)
:uniques #{}
:duplicates #{}}}
results (reduce
(fn [results _]
(reduce
(fn [results cell-size]
(let [curr-acc (get-in results [cell-size :acc])
;; randomly select one of the two hashes
selected-hash (if (< (rand) 0.5)
(get-in utms [cell-size :a])
(get-in utms [cell-size :b]))
;; fuse the selected hash onto the accumulator
new-acc (utm-multiply curr-acc selected-hash)]
;; update results with new accumulator and uniqueness info
(if (contains? (get-in results [cell-size :uniques]) new-acc)
(update-in results [cell-size :duplicates] conj new-acc)
(-> results
(assoc-in [cell-size :acc] new-acc)
(update-in [cell-size :uniques] conj new-acc)))))
results
[8 16 32 64]))
results
(range 10000))]
;; convert final results to totals of unique and duplicate hashes
(->> results
(map (fn [[cell-size {:keys [uniques duplicates]}]]
{:cell-size cell-size
:uniques (count uniques)
:duplicates (count duplicates)}))
(kind/table))))Random Fuses Results
(random-fuses)| cell-size | uniques | duplicates |
|---|---|---|
| 8 | 10000 | 0 |
| 16 | 10000 | 0 |
| 32 | 10000 | 0 |
| 64 | 10000 | 0 |
These results show that high entropy data can be well represented by fusing hashes together. All four bit sizes for the cells of the upper triangular matrices performed well with high entroy data. I have rerun this experiment with millions of fuses and have never observed a duplicate hash being produced. Since 64 bit cells fit in the smallest matrix, this is the fasted to compute.
Experiment 2: Folded Fuses
This experiment is run with a hash fused together with itself (i.e. folding) and then the result in turn fused with itself and so on. This folding is repeated many times and the quality of the accumulator is measured both by keeping track of global uniqueness after each fuse and by the uniform distribution of bit values. Once the accumulator becomes zero the folding stops and the number of folds to reach this state is recorded. Also, the number of lower bits that are zero across all cells is recorded after each fold.
(defn calc-zero-lower-bits
"Calculate the number of lower bits that are zero across all upper cells in
the upper triangular matrix."
[utm]
(let [cell-size (-> utm meta :cell-size)]
(loop [mask 1 zero-lower-bits 0]
(if (and (< zero-lower-bits cell-size)
(->> (for [i (range (count utm))
j (range (inc i) (count utm))]
(get-in utm [i j]))
(map #(bit-and mask %))
(every? zero?)))
(recur (bit-shift-left mask 1)
(inc zero-lower-bits))
zero-lower-bits))))(defn folded-fuses
"Perform folded fuses starting from the fixed hash `a-hex`, converted to an
upper triangular matrix (UTM) for each cell size. After each fold, track how
many lower bits of each cell are zero across all upper cells using
`calc-zero-lower-bits`. Stop when all those lower bits are zero (i.e.,
`zero-lower-bits` equals the cell size) or when 1000 folds have been
performed. Report, for each cell size and fold, the fold count, the number
of zero lower bits, and the resulting hash."
[]
(let [;; convert hashes to utms for each bit size and store in a map
results (reduce
(fn [results cell-size]
(let [fold (hex->utm a-hex cell-size)
zero-lower-bits (calc-zero-lower-bits fold)
fold-result
(loop [folds [{:fold fold
:zero-lower-bits zero-lower-bits}]]
(let [{:keys [fold zero-lower-bits]} (last folds)]
(if (or (= zero-lower-bits cell-size)
(>= (count folds) 1000))
folds
(let [fold (utm-multiply fold fold)
zero-lower-bits (calc-zero-lower-bits fold)]
(recur (conj folds
{:fold fold
:zero-lower-bits zero-lower-bits}))))))]
(assoc results cell-size fold-result)))
{}
[8 16 32 64])]
;; format results into a table
(->> results
(mapcat (fn [[cell-size folds]]
(map-indexed (fn [index {:keys [fold zero-lower-bits]}]
{:cell-size cell-size
:fold-count index
:zero-lower-bits zero-lower-bits
:fold (-> [:pre (utm->hex fold cell-size)]
(kind/hiccup))})
folds)))
(kind/table))))Folded Fuses Results
(folded-fuses)| cell-size | fold-count | zero-lower-bits | fold |
|---|---|---|---|
| 8 | 0 | 0 | |
| 8 | 1 | 0 | |
| 8 | 2 | 1 | |
| 8 | 3 | 2 | |
| 8 | 4 | 3 | |
| 8 | 5 | 4 | |
| 8 | 6 | 5 | |
| 8 | 7 | 6 | |
| 8 | 8 | 7 | |
| 8 | 9 | 8 | |
| 16 | 0 | 0 | |
| 16 | 1 | 0 | |
| 16 | 2 | 0 | |
| 16 | 3 | 1 | |
| 16 | 4 | 2 | |
| 16 | 5 | 3 | |
| 16 | 6 | 4 | |
| 16 | 7 | 5 | |
| 16 | 8 | 6 | |
| 16 | 9 | 7 | |
| 16 | 10 | 8 | |
| 16 | 11 | 9 | |
| 16 | 12 | 10 | |
| 16 | 13 | 11 | |
| 16 | 14 | 12 | |
| 16 | 15 | 13 | |
| 16 | 16 | 14 | |
| 16 | 17 | 15 | |
| 16 | 18 | 16 | |
| 32 | 0 | 0 | |
| 32 | 1 | 0 | |
| 32 | 2 | 1 | |
| 32 | 3 | 2 | |
| 32 | 4 | 3 | |
| 32 | 5 | 4 | |
| 32 | 6 | 5 | |
| 32 | 7 | 6 | |
| 32 | 8 | 7 | |
| 32 | 9 | 8 | |
| 32 | 10 | 9 | |
| 32 | 11 | 10 | |
| 32 | 12 | 11 | |
| 32 | 13 | 12 | |
| 32 | 14 | 13 | |
| 32 | 15 | 14 | |
| 32 | 16 | 15 | |
| 32 | 17 | 16 | |
| 32 | 18 | 17 | |
| 32 | 19 | 18 | |
| 32 | 20 | 19 | |
| 32 | 21 | 20 | |
| 32 | 22 | 21 | |
| 32 | 23 | 22 | |
| 32 | 24 | 23 | |
| 32 | 25 | 24 | |
| 32 | 26 | 25 | |
| 32 | 27 | 26 | |
| 32 | 28 | 27 | |
| 32 | 29 | 28 | |
| 32 | 30 | 29 | |
| 32 | 31 | 30 | |
| 32 | 32 | 31 | |
| 32 | 33 | 32 | |
| 64 | 0 | 0 | |
| 64 | 1 | 1 | |
| 64 | 2 | 2 | |
| 64 | 3 | 3 | |
| 64 | 4 | 4 | |
| 64 | 5 | 5 | |
| 64 | 6 | 6 | |
| 64 | 7 | 7 | |
| 64 | 8 | 8 | |
| 64 | 9 | 9 | |
| 64 | 10 | 10 | |
| 64 | 11 | 11 | |
| 64 | 12 | 12 | |
| 64 | 13 | 13 | |
| 64 | 14 | 14 | |
| 64 | 15 | 15 | |
| 64 | 16 | 16 | |
| 64 | 17 | 17 | |
| 64 | 18 | 18 | |
| 64 | 19 | 19 | |
| 64 | 20 | 20 | |
| 64 | 21 | 21 | |
| 64 | 22 | 22 | |
| 64 | 23 | 23 | |
| 64 | 24 | 24 | |
| 64 | 25 | 25 | |
| 64 | 26 | 26 | |
| 64 | 27 | 27 | |
| 64 | 28 | 28 | |
| 64 | 29 | 29 | |
| 64 | 30 | 30 | |
| 64 | 31 | 31 | |
| 64 | 32 | 32 | |
| 64 | 33 | 33 | |
| 64 | 34 | 34 | |
| 64 | 35 | 35 | |
| 64 | 36 | 36 | |
| 64 | 37 | 37 | |
| 64 | 38 | 38 | |
| 64 | 39 | 39 | |
| 64 | 40 | 40 | |
| 64 | 41 | 41 | |
| 64 | 42 | 42 | |
| 64 | 43 | 43 | |
| 64 | 44 | 44 | |
| 64 | 45 | 45 | |
| 64 | 46 | 46 | |
| 64 | 47 | 47 | |
| 64 | 48 | 48 | |
| 64 | 49 | 49 | |
| 64 | 50 | 50 | |
| 64 | 51 | 51 | |
| 64 | 52 | 52 | |
| 64 | 53 | 53 | |
| 64 | 54 | 54 | |
| 64 | 55 | 55 | |
| 64 | 56 | 56 | |
| 64 | 57 | 57 | |
| 64 | 58 | 58 | |
| 64 | 59 | 59 | |
| 64 | 60 | 60 | |
| 64 | 61 | 61 | |
| 64 | 62 | 62 | |
| 64 | 63 | 63 | |
| 64 | 64 | 64 | |
These results show that repeated folding devolves into a zero hash after only a few folds. The number of bits in a cell nearly exactly defines the number of folds needed before reaching a zero hash. Since each fold represents a doubling of the number of fuses of the same value, folding shows that fusing can not reliably represent long runs of repeated values.
Practical Usage
Based on the experiments above, here are recommendations for using hash fusing in practice.
Recommended Configuration
Use 64-bit cells (4x4 matrix). This configuration offers:
- The smallest matrix size, making it the fastest to compute
- The highest tolerance for repeated values (64 folds before zero)
- No observed collisions with high-entropy data
Detecting Low-Entropy Failures
The key insight from Experiment 2 is that low-entropy data causes the lower bits to become zero progressively. We can detect this condition and fail fast rather than silently producing degenerate hashes.
The following function fuses two hashes and raises an error if the result shows signs of low-entropy degeneration (lower 32 bits all zero):
(defn high-entropy-fuse
"Fuse two 256-bit hashes together via upper triangular matrix multiplication
with 64-bit cells. Raise an error when the lower 32 bits of the result are
all zero."
[a-hex b-hex]
(let [a (hex->utm64 a-hex)
b (hex->utm64 b-hex)
ab (utm-multiply a b)]
(if (->> (for [i (range (count a))
j (range (inc i) (count a))]
(get-in ab [i j]))
(map #(bit-and 0xFFFFFFFF %))
(every? zero?))
(throw (ex-info "Low entropy data detected during hash fusing."
{:a a-hex :b b-hex :fused (utm64->hex ab)}))
(utm64->hex ab))))Example of high entropy data fusing successfully:
(high-entropy-fuse a-hex b-hex)"641184de1619c1e57451d954f9ace1fe0bfe2f5597834e258e8dbf19cdc4dee5"Example of low entropy data causing an error:
(try
(high-entropy-fuse zero-hex zero-hex)
(catch Exception e
(.getMessage e)))"Low entropy data detected during hash fusing."Handling Low-Entropy Data
When working with data that may contain long runs of repeated values, consider these strategies:
Compression: Run-length encode or compress the data before hashing. This converts low-entropy patterns into high-entropy representations.
Salting: Inject position-dependent noise into the hash sequence. For example, XOR each element’s hash with a hash of its index.
Chunking: Break long sequences into smaller chunks and hash each chunk separately before fusing. This limits the damage from repetition.
Conclusion
This article described a method for fusing hashes together using upper triangular matrix multiplication. The method is associative and non-commutative.
Experiment 1 showed that high entropy data can be well represented by fusing hashes together and that the size of cell fields, from 8 to 64 bits, all performed well with high entropy data. Since 64 bit cells fit in the smallest matrix, this is the fastest to compute.
Experiment 2 showed that low entropy data, specifically long runs of repeated values, can not be well represented by fusing hashes together since they all approach a zero hash. The rate of approach is determined by the bit count in the cells. 64 bit cells were the most able to handle repeated values.
The Practical Usage section provides a working implementation with low-entropy detection, along with strategies for handling problematic data.
Appendix: Why Low Entropy Data Fails — Nilpotent Group Structure
The following explanation, based on interactions with a Claude-based AI assistant, is very well written and accurate.
The low-entropy failure observed in Experiment 2 is not an accident of implementation — it is a fundamental consequence of the algebraic structure we chose. Understanding this helps clarify both the limitations and the design space for alternatives.
Upper Triangular Matrices Form a Nilpotent Group
An n×n upper triangular matrix with 1s on the diagonal can be written as I + N, where I is the identity matrix and N is strictly upper triangular (all diagonal entries are 0). The key property of strictly upper triangular matrices is that they are nilpotent: repeated multiplication eventually yields zero.
Specifically, for an n×n strictly upper triangular matrix N:
N^n = 0 (the zero matrix)
This happens because each multiplication “shifts” the non-zero entries one diagonal further from the main diagonal. After n-1 multiplications, all entries have been shifted off the matrix.
How This Affects Hash Fusing
When we fuse a hash with itself (A × A), we are computing (I + N) × (I + N):
(I + N)² = I + 2N + N²
With our bit masking over w-bit cells, arithmetic is effectively done modulo 2^w. Multiplying by 2 shifts all bits left and, after masking back to w bits, forces the least significant bit of each cell to 0. After k squarings:
(I + N)(2k) = I + 2^k · N + (higher powers of N)
The coefficient 2^k means the lowest k bits of each cell are forced to zero under mod 2^w / bitmasking. This is exactly what Experiment 2 observed: each fold (squaring) zeros out one more bit, and after cell-size folds, all bits are zero.
Why This Is Fundamental
The nilpotent structure is intrinsic to upper triangular matrices. Any associative, non-commutative composition based on this structure will exhibit the same behavior. This is not a bug — it’s a mathematical property of the group we chose.
The Trade-off
We chose upper triangular matrices because they provide:
- Associativity: Essential for Finger Trees
- Non-commutativity: Essential for ordered sequences
- Efficient representation: The matrix structure maps cleanly to hash bits
The cost is nilpotency, which manifests as the low-entropy failure mode. This trade-off is acceptable for high-entropy data (which is the common case for content hashes) and can be mitigated with detection and preprocessing for edge cases.
Alternative Structures
For applications where low-entropy data is common, one could explore non-nilpotent groups such as GL(n, F_p) — the general linear group over a finite field. However, this would sacrifice the sparse matrix efficiency and require more complex encoding. The upper triangular approach remains practical for most content-addressed data structure applications.
Appendix B: Alternatives Explored
Before settling on upper triangular matrices, several alternative approaches were investigated. This section documents those explorations and explains why they were not suitable.
Quaternion Multiplication
Quaternions are a natural candidate: they form a non-commutative algebra and have efficient 4-component representations. However, experiments with quaternion-based fusing revealed the same nilpotent-like behavior observed with upper triangular matrices.
When quaternions are encoded similarly to the UTM approach — with values close to the identity element (1 + εi + εj + εk where ε represents small perturbations) — repeated self-multiplication causes the perturbation terms to degenerate. While pure unit quaternions form a proper group (isomorphic to SU(2)), the encoding required for hash values reintroduces the same fundamental problem.
Modified Fusing Operations
Various modifications to the basic fusing operation were attempted, including:
- XOR combined with rotations
- Nonlinear mixing functions
- Polynomial operations over finite fields
The consistent finding was that associativity is surprisingly fragile. Most intuitive “mixing” operations that might improve entropy preservation break associativity, which is essential for the Finger Tree use case.
Matrix multiplication is special precisely because it is one of the few operations that provides both associativity and non-commutativity naturally.
Why Upper Triangular Matrices Were Chosen
Despite the low-entropy limitation, upper triangular matrices remain the best practical choice because:
Guaranteed associativity: Matrix multiplication is inherently associative — this cannot be broken by implementation choices.
Guaranteed non-commutativity: For n > 1, matrix multiplication is non-commutative, preserving sequence order.
Efficient sparse encoding: The upper triangular structure means only n(n-1)/2 cells carry information, mapping cleanly to hash bits.
Predictable failure mode: The low-entropy degeneration is detectable and follows a known pattern (one bit per fold), making it possible to implement safeguards.
The alternatives explored either broke associativity (disqualifying them for Finger Trees) or exhibited the same degeneration behavior without the other benefits of the matrix approach.
source: src/math/hashing/hashfusing.clj