Accounting is normally described as a measurement and reporting system.
But accounting can also be understood as an information system.
A firm contains an enormous amount of economically relevant information. It owns resources, enters contracts, makes investments, incurs obligations, develops knowledge, bears risks, receives cash flows and forms expectations about an uncertain future.
Financial statements compress this complex economic reality into a comparatively small collection of numbers:
\[ \text{Economic reality} \longrightarrow \text{Accounting system} \longrightarrow \text{Financial statements}. \]
From this perspective, accounting performs a problem remarkably similar to the problem studied by information theory.
There is an underlying state of the world.
Information about that state must be encoded.
Only a restricted representation is transmitted.
The receiver observes the representation and attempts to infer something about the underlying state.
The analogy can be written as
\[ E \xrightarrow{\mathcal{A}} A \xrightarrow{} I, \]
where \(E\) represents economic reality, \(\mathcal{A}\) the accounting transformation, \(A\) the resulting accounting representation, and \(I\) the information received by users of financial statements.
The Accounting Bias framework begins with
\[ AB=A-E. \]
Information theory allows this relationship to be interpreted more deeply.
Accounting Bias can be viewed not only as a measurement divergence but also as a consequence of the transformation, compression and loss of economic information produced when economic reality is converted into accounting representation.
Modern information theory begins with the problem of transmitting information through a communication system.
In its simplest form, a communication process contains a source, an encoder, a channel, a decoder and a receiver.
Abstractly,
\[ X \rightarrow \text{Encoder} \rightarrow \text{Channel} \rightarrow \text{Decoder} \rightarrow Y. \]
The source generates information.
The encoder transforms that information into a transmissible representation.
The channel carries the representation.
The receiver observes the resulting signal.
Accounting possesses a remarkably similar architecture.
We can write
\[ E \rightarrow \mathcal{A} \rightarrow FS \rightarrow U, \]
where
\[ E=\text{economic reality}, \]
\[ \mathcal{A}=\text{accounting system}, \]
\[ FS=\text{financial statements}, \]
and
\[ U=\text{user}. \]
The accounting system encodes economic phenomena into standardized categories and monetary amounts.
A building becomes property, plant and equipment.
A contractual obligation becomes a liability.
Economic deterioration becomes an impairment loss.
A sale becomes revenue.
An uncertain future payment may become a provision.
Economic phenomena are therefore not transmitted directly.
They are encoded.
Consider an underlying economic state
\[ E=(e_1,e_2,\ldots,e_n). \]
This state may contain information about productive resources, technological knowledge, contractual rights, obligations, risks, expectations, market conditions and future cash-generating capacity.
Accounting applies a transformation:
\[ A=\mathcal{A}(E). \]
The function \(\mathcal{A}\) consists of accounting rules concerning
\[ \mathcal{A} = \{ R,M,T,C,S \}, \]
where
\[ R=\text{recognition}, \]
\[ M=\text{measurement}, \]
\[ T=\text{timing and allocation}, \]
\[ C=\text{classification}, \]
and
\[ S=\text{estimation}. \]
The output is a much smaller representation:
\[ A=(a_1,a_2,\ldots,a_k). \]
Usually,
\[ k\ll n. \]
This is unavoidable.
No financial reporting system could reproduce every economically relevant characteristic of a firm.
Accounting therefore performs information reduction.
Claude Shannon formalized uncertainty using entropy.
Suppose a random variable \(X\) can take states
\[ x_1,x_2,\ldots,x_n \]
with probabilities
\[ p_1,p_2,\ldots,p_n. \]
The entropy of \(X\) is
\[ H(X) = -\sum_{i=1}^{n} p_i\log_2 p_i. \]
Entropy measures uncertainty concerning which state will occur.
If one outcome is certain,
\[ p_i=1, \]
then
\[ H(X)=0. \]
There is no uncertainty to resolve.
If several outcomes are possible and their probabilities are dispersed, entropy is higher.
Information can therefore be understood as a reduction in uncertainty.
This provides an immediate connection to accounting.
Before observing financial statements, an investor possesses uncertainty about the economic condition of a firm.
Let this uncertainty be represented by
\[ H(E). \]
After observing accounting information \(A\), uncertainty becomes
\[ H(E\mid A). \]
The information supplied by accounting about economic reality can then be represented by the mutual information
\[ I(E;A) = H(E)-H(E\mid A). \]
The greater the reduction in uncertainty about \(E\), the greater the information contained in \(A\).
This gives us a formal way to think about accounting informativeness.
Suppose two accounting systems produce representations
\[ A_1 \]
and
\[ A_2. \]
If
\[ H(E\mid A_1) < H(E\mid A_2), \]
then
\[ I(E;A_1) > I(E;A_2). \]
System \(A_1\) conveys more information about economic reality than system \(A_2\).
This suggests a theoretical criterion for accounting quality:
\[ \boxed{ \text{A more informative accounting system reduces uncertainty about economically relevant states.} } \]
Notice that this criterion is different from asking whether accounting reports a single "true value".
As developed in the analysis of the non-measurability of economic value, economic value itself may be latent and model-dependent.
Information theory allows us to ask a weaker but potentially more operational question:
Does the accounting representation help users distinguish among economically different states of the firm?
That question does not require perfect observation of economic value.
Financial statements are necessarily compressed representations.
Imagine that the complete economic state of a corporation contains millions of relevant facts:
\[ E= (e_1,e_2,\ldots,e_n). \]
Accounting reduces them to a finite collection of recognized amounts:
\[ A= (a_1,a_2,\ldots,a_k). \]
Thus,
\[ \mathcal{A}:E\rightarrow A. \]
Many economically different states may therefore produce the same accounting representation.
Suppose
\[ E_1\neq E_2, \]
but
\[ \mathcal{A}(E_1) = \mathcal{A}(E_2). \]
Then the accounting system cannot distinguish between these two economic states.
From the perspective of the financial statements,
\[ E_1 \sim_{\mathcal{A}} E_2. \]
They belong to the same accounting equivalence class.
This is one of the deepest informational consequences of accounting rules.
Recognition provides the clearest example.
Suppose a firm develops valuable knowledge internally.
The economic state contains the resource:
\[ E_K>0. \]
But suppose the accounting recognition criteria prevent the resource from appearing as an asset:
\[ A_K=0. \]
Now consider two firms.
Firm 1 possesses relatively little internally generated knowledge:
\[ E_K^{(1)}=10. \]
Firm 2 possesses much more:
\[ E_K^{(2)}=100. \]
If neither amount is recognized,
\[ A_K^{(1)} = A_K^{(2)} = 0. \]
The accounting system maps economically different states onto the same accounting representation:
\[ 10\rightarrow0, \]
\[ 100\rightarrow0. \]
Information has been lost.
The recognition rule acts as a filter.
This does not mean that the rule is necessarily incorrect. Recognition may require reliability, identifiability, control or other institutional characteristics.
But informationally, the consequence is clear:
\[ \boxed{ \text{Non-recognition reduces the ability of accounting numbers to discriminate among economic states.} } \]
Recognition determines whether information enters the system.
Measurement determines how recognized information is represented.
Suppose an economic resource can take a continuous range of values:
\[ E_i\in\mathbb{R}. \]
Accounting may map these states into a more restricted set of measurements.
For example,
\[ A_i = \mathcal{M}(E_i). \]
This resembles quantization in information processing: a continuous or highly detailed signal is represented using a smaller set of possible values.
Historical-cost accounting provides an intuitive example.
Suppose two otherwise identical assets were acquired at different dates and prices.
Their current economic conditions may be similar:
\[ E_1\approx E_2, \]
while their carrying amounts differ:
\[ A_1\neq A_2. \]
Alternatively, two assets with identical historical costs may subsequently develop very different economic values:
\[ A_1=A_2 \]
while
\[ E_1\neq E_2. \]
Accounting measurement therefore determines which differences in economic reality remain visible in the accounting signal.
Financial reporting also aggregates information.
Suppose a firm possesses \(n\) assets:
\[ A_1,A_2,\ldots,A_n. \]
The balance sheet may report only their aggregate:
\[ A_{\text{total}} = \sum_{i=1}^{n}A_i. \]
Many different underlying portfolios can produce the same total.
For example,
\[ (90,10) \]
and
\[ (50,50) \]
both produce
\[ A_{\text{total}}=100. \]
But the two portfolios may have completely different risk, liquidity and return characteristics.
Aggregation therefore reduces dimensionality:
\[ (A_1,A_2,\ldots,A_n) \rightarrow A_{\text{total}}. \]
Again, this is necessary.
Without aggregation, financial statements would become unusably complex.
But aggregation carries an informational cost.
Accounting therefore faces a fundamental trade-off.
At one extreme, the reporting system could attempt to preserve enormous amounts of economic detail.
Then
\[ I(E;A) \]
might increase.
But users would face excessive complexity.
At the other extreme, accounting could compress everything into a handful of numbers.
The representation would be easy to process, but substantial economic information would disappear.
Thus accounting confronts a trade-off between
\[ \text{information preservation} \]
and
\[ \text{compression}. \]
Conceptually,
\[ \boxed{ \text{Useful accounting} = \text{economic information} - \text{unnecessary complexity}. } \]
This resembles a general information-design problem.
The objective is not to transmit everything.
It is to transmit economically relevant information efficiently.
We can now reinterpret Accounting Bias.
The original framework defines
\[ AB=A-E. \]
This expresses the difference between accounting representation and an economic benchmark.
But suppose we focus not on the numerical difference alone, but on the information contained in the accounting representation.
Then accounting transformation produces
\[ E \xrightarrow{\mathcal{A}} A. \]
The information about \(E\) preserved in \(A\) is
\[ I(E;A). \]
The uncertainty remaining after observing accounting is
\[ H(E\mid A). \]
This residual uncertainty can be interpreted as an information gap:
\[ IG = H(E\mid A). \]
Accounting Bias and the information gap are not identical.
\[ AB\neq IG. \]
Accounting Bias concerns divergence in representation.
Information loss concerns uncertainty remaining about economic reality after observing that representation.
But the two concepts are related.
Accounting rules capable of generating Accounting Bias may also reduce the information accounting provides about the underlying economic state.
Suppose recognition rules exclude an economically relevant resource.
Then
\[ A_i=0 \]
despite
\[ E_i\neq0. \]
This produces recognition-related Accounting Bias:
\[ AB_i^{R}=A_i-E_i. \]
At the same time, users cannot infer much about variation in \(E_i\) from \(A_i\), because every excluded state maps to zero.
Thus,
\[ I(E_i;A_i) \]
may be low.
Recognition bias therefore possesses both a measurement interpretation and an information interpretation.
The measurement interpretation asks:
\[ A_i-E_i=? \]
The information interpretation asks:
\[ I(E_i;A_i)=? \]
These are different but complementary questions.
Timing creates another form of informational distortion.
Suppose an economic event occurs at time \(t\):
\[ E_t. \]
Accounting recognizes the event only at
\[ t+k. \]
Then during
\[ [t,t+k), \]
users observe accounting information that does not yet incorporate the event.
The problem is not permanent non-recognition.
It is information delay.
We can represent the accounting signal as
\[ A_t = \mathcal{A}(E_{t-k}). \]
The accounting system contains information about economic reality, but with a lag.
This suggests another informational dimension:
\[ \text{timeliness}. \]
A perfectly accurate signal delivered too late may have limited decision usefulness.
Thus accounting information depends not only on
\[ I(E;A), \]
but on how quickly that information becomes available.
Accounting frequently relies on estimates.
Examples include
\[ \text{expected credit losses}, \]
\[ \text{provisions}, \]
\[ \text{useful lives}, \]
\[ \text{impairment assumptions}, \]
\[ \text{fair values}, \]
and
\[ \text{discount rates}. \]
Suppose the economic quantity is
\[ E. \]
Accounting produces an estimate
\[ A=E+\varepsilon, \]
where
\[ \varepsilon \]
represents estimation error.
This resembles a noisy communication channel.
The underlying signal is \(E\).
The user observes
\[ A. \]
The noise \(\varepsilon\) prevents perfect recovery of the signal.
If
\[ Var(\varepsilon)\uparrow, \]
then the accounting signal becomes less informative about \(E\).
Thus estimation uncertainty can be interpreted as accounting-channel noise.
The analogy can be developed further.
Suppose accounting information contains
\[ A=S+N, \]
where
\[ S \]
is economically informative signal and
\[ N \]
is noise introduced by measurement uncertainty, estimation, aggregation or classification.
The quality of the accounting representation depends partly on the relationship between the two.
Conceptually,
\[ \text{Accounting information quality} \uparrow \]
when
\[ \frac{S}{N}\uparrow. \]
This does not mean accounting should eliminate estimation.
Often estimation is necessary precisely because direct measurement is impossible.
The relevant question is whether the estimation process increases or decreases the information users possess about economic reality.
An important result follows.
Increasing disclosure does not necessarily increase useful information proportionally.
Suppose a financial report expands from
\[ 100 \]
reported variables to
\[ 1{,}000. \]
The amount of data has increased enormously.
But useful information increases only if the additional variables reduce uncertainty about economically relevant states.
Thus,
\[ \text{Data}\neq\text{Information}. \]
A disclosure that is perfectly predictable from existing disclosures may add almost no new information.
In information-theoretic language, what matters is not simply the number of messages transmitted but their incremental information.
This provides a theoretical foundation for distinguishing disclosure quantity from disclosure informativeness.
Redundancy is not necessarily useless.
Suppose two disclosures contain similar information.
Then
\[ I(A_1;A_2) \]
may be high.
At first sight, this appears inefficient.
But redundancy can serve important purposes.
It may improve verification.
It may make errors easier to detect.
It may allow users to reconstruct relationships between statements.
Double-entry bookkeeping itself contains a powerful form of structural redundancy.
The accounting identity
\[ Assets=Liabilities+Equity \]
imposes consistency on the system.
Every transaction affects multiple accounts according to a constrained structure.
From a pure compression perspective, some of this information may appear redundant.
From a control perspective, redundancy is valuable.
This reveals an important distinction between informational efficiency and system reliability.
Accounting is designed for both.
This insight can be pushed further.
In communication systems, redundancy can help detect transmission errors.
Accounting possesses a comparable property.
Because
\[ A=L+E, \]
an inconsistent set of entries violates the accounting identity.
Double-entry bookkeeping therefore imposes structural constraints on admissible messages.
A transaction cannot arbitrarily change one side of the accounting system without a corresponding entry elsewhere.
Conceptually,
\[ \text{economic event} \rightarrow \text{multiple linked entries}. \]
These linked entries create internal consistency checks.
Thus one of the oldest accounting technologies can be interpreted using a modern informational concept:
\[ \boxed{ \text{Double entry introduces structured redundancy that increases the reliability of the accounting information system.} } \]
Accounting standards can now be interpreted as a coding protocol.
Without common accounting rules, two firms could encode identical economic events differently.
Suppose
\[ E_1=E_2, \]
but firm 1 uses coding rule \(\mathcal{A}_1\) while firm 2 uses \(\mathcal{A}_2\).
Then
\[ \mathcal{A}_1(E) \neq \mathcal{A}_2(E). \]
Users would have difficulty determining whether observed differences arose from economic differences or coding differences.
Standardization attempts to reduce this problem by imposing a common mapping:
\[ \mathcal{A}_1 \approx \mathcal{A}_2 \approx \mathcal{A}. \]
Then differences in reported numbers are more likely to correspond to differences in underlying economic states.
This provides an information-theoretic interpretation of comparability.
Comparability reduces uncertainty about whether differences between accounting signals arise from economic reality or from the encoding system itself.
Accounting rules are not always symmetric.
Suppose bad economic news is recognized according to one threshold
\[ \tau_{-}, \]
while good economic news requires another threshold
\[ \tau_{+}. \]
If
\[ \tau_{-}<\tau_{+}, \]
negative information enters accounting more readily than positive information.
The accounting encoder is asymmetric.
This can be represented as
\[ \mathcal{A}(E^{+}) \neq -\mathcal{A}(-E^{+}). \]
Conservatism can therefore be interpreted not merely as a valuation tendency but as an asymmetric information filter.
The accounting system assigns different transmission rules to different types of economic information.
This may generate systematic Accounting Bias.
But it may simultaneously serve institutional purposes such as creditor protection or verification.
Once again, bias does not necessarily mean error.
Accounting communication does not end when financial statements are published.
The user must interpret them.
Thus the full process is
\[ E \rightarrow \mathcal{A} \rightarrow A \rightarrow \mathcal{D}_u \rightarrow \hat{E}_u, \]
where
\[ \mathcal{D}_u \]
is the user's decoding process and
\[ \hat{E}_u \]
is the user's inferred economic state.
Even if accounting transmission were perfect, users could interpret the same information differently.
Thus
\[ \hat{E}_1 \neq \hat{E}_2 \]
may occur despite
\[ A_1=A_2. \]
Financial literacy, analytical models, expectations and incentives affect decoding.
The final information problem therefore contains two transformations:
\[ \boxed{ E \xrightarrow{\mathcal{A}} A \xrightarrow{\mathcal{D}_u} \hat{E}. } \]
Accounting Bias concerns primarily the first transformation.
Investor interpretation introduces a second potential source of divergence.
Information theory contains an important general principle.
Processing information cannot create information about the original state that was not present in the input.
If
\[ E\rightarrow A\rightarrow B \]
forms a processing chain, then, under the relevant conditions,
\[ I(E;B)\leq I(E;A). \]
This is the intuition behind the data-processing inequality.
The accounting implication is significant.
Once economically relevant information has been removed during recognition or aggregation, subsequent manipulation of the published accounting numbers cannot fully reconstruct it without additional information.
Suppose internally generated knowledge is completely excluded from \(A\).
An analyst can calculate
\[ ROA, \]
\[ ROE, \]
\[ EBITDA, \]
and hundreds of other ratios from \(A\).
But mathematical transformation of \(A\) alone cannot recover information about \(E\) that accounting never transmitted.
This places a fundamental limit on financial-statement analysis.
We can now formulate the central proposition.
Economic reality contains a high-dimensional state:
\[ E. \]
Accounting maps this state into a lower-dimensional representation:
\[ A=\mathcal{A}(E). \]
Financial statements therefore operate as an information bottleneck:
\[ \boxed{ E \longrightarrow \underbrace{\mathcal{A}}_{\text{recognition, measurement, timing, classification}} \longrightarrow A. } \]
Some economic information passes through.
Some is aggregated.
Some is delayed.
Some is estimated.
Some is never recognized.
Accounting Bias can therefore be interpreted partly as the systematic consequence of this bottleneck.
The central question becomes not simply
\[ A-E=? \]
but also
\[ \boxed{ \text{What information about }E \text{ survives the transformation into }A? } \]
The Accounting Bias framework can consequently be extended along two dimensions.
The first is the original measurement dimension:
\[ AB=A-E. \]
The second is an informational dimension:
\[ IG=H(E\mid A). \]
We can then describe an accounting system through
\[ \boxed{ \mathcal{S} = (AB,IG). } \]
Here,
\[ AB \]
describes divergence between accounting representation and the economic benchmark, while
\[ IG \]
describes uncertainty about economic reality remaining after observing accounting information.
A system could theoretically exhibit relatively small numerical bias while still transmitting little information.
Conversely, a systematically conservative system could exhibit directional bias while nevertheless transmitting substantial information about changes in economic conditions.
This distinction is important.
Bias and informativeness are not opposites.
Suppose two firms possess economic assets
\[ E_1=100 \]
and
\[ E_2=200. \]
Accounting system \(A\) reports
\[ A_1=80, \]
\[ A_2=160. \]
The system is systematically conservative:
\[ AB_1=-20, \]
\[ AB_2=-40. \]
But the accounting signal perfectly preserves the ranking and proportional relationship between the two economic states.
Now consider another system reporting
\[ A_1=100, \]
\[ A_2=100. \]
For the first firm,
\[ AB_1=0. \]
Yet the system completely fails to distinguish the two firms.
Thus an accounting system can sometimes produce a numerically unbiased observation while being informationally poor.
This shows why accounting quality cannot be reduced to
\[ AB=0. \]
A broader criterion must consider both measurement and information.
We can therefore define an abstract accounting information-efficiency objective.
Let
\[ I(E;A) \]
measure information preserved about economic reality, while
\[ C(A) \]
represents the cost or complexity of producing, auditing and processing the accounting representation.
Then an accounting system might be interpreted as solving
\[ \max_{\mathcal{A}} \left[ I(E;A)-\lambda C(A) \right], \]
where
\[ \lambda>0 \]
represents the importance assigned to reporting and processing costs.
This captures an essential institutional problem.
Perfect information is neither attainable nor costless.
Accounting must select which information to preserve.
Recognition, measurement and disclosure rules therefore represent choices about the allocation of limited informational capacity.
This leads to perhaps the simplest information-theoretic description of accounting.
Accounting is a form of lossy compression.
It transforms
\[ \text{high-dimensional economic reality} \]
into
\[ \text{low-dimensional standardized financial representation}. \]
The compression is lossy because
\[ A \]
generally does not contain enough information to reconstruct
\[ E \]
perfectly.
Formally,
\[ \mathcal{A}^{-1}(A) \]
is generally not unique.
Many possible economic states may correspond to the same accounting representation.
Yet this loss is not necessarily a defect.
Lossy compression is useful precisely because transmitting the complete underlying object would be impractical.
The relevant question is therefore:
Which information should accounting preserve, and which information can it afford to discard?
That is simultaneously an accounting question and an information-theoretic question.
The entire problem can now be summarized.
Economic reality is
\[ E. \]
Accounting produces
\[ A=\mathcal{A}(E). \]
Users infer
\[ \hat{E} = \mathcal{D}(A). \]
Therefore,
\[ \boxed{ E \xrightarrow{\text{accounting encoding}} A \xrightarrow{\text{user decoding}} \hat{E}. } \]
Three distinct divergences can arise.
First, measurement divergence:
\[ A-E. \]
Second, information loss:
\[ H(E\mid A). \]
Third, interpretation divergence:
\[ \hat{E}-E. \]
Accounting Bias primarily studies the first.
Information theory illuminates the second.
Behavioral and financial economics can study the third.
Together they provide a more complete theory of financial reporting.
Accounting is not merely a collection of valuation rules.
It is an information-processing architecture.
Economic reality contains vastly more information than can appear in financial statements.
Accounting therefore selects, classifies, measures, aggregates and compresses economic phenomena into a standardized representation:
\[ E \xrightarrow{\mathcal{A}} A. \]
Recognition determines which information enters the system.
Measurement determines how that information is encoded.
Aggregation compresses it.
Timing determines when it is transmitted.
Estimation introduces noise.
Standardization creates a common coding protocol.
Double-entry bookkeeping introduces structured redundancy and internal consistency.
Financial statements are the resulting signal.
From this perspective, Accounting Bias acquires an informational interpretation.
The conventional relation
\[ AB=A-E \]
describes divergence between accounting representation and an economic benchmark.
Information theory adds another question:
\[ I(E;A)=? \]
How much information about economic reality survives accounting transformation?
Equivalently, we can ask how much uncertainty remains:
\[ H(E\mid A). \]
This distinction produces an important theoretical result:
\[ \boxed{ \text{Accounting Bias and accounting informativeness are related, but they are not the same thing.} } \]
A biased accounting system may still be highly informative.
An apparently unbiased accounting number may contain very little information.
Accounting quality therefore cannot be evaluated solely by asking whether reported numbers approximate economic values.
It must also be evaluated by asking whether those numbers allow users to distinguish economically different states.
This leads to a broader interpretation of accounting.
Accounting is necessarily selective because economic reality is too complex to transmit in its entirety. It is necessarily reductive because financial reporting requires standardization and aggregation. And it is necessarily imperfect because some information is lost whenever a high-dimensional economic system is compressed into a finite set of accounting numbers.
The central problem of accounting can therefore be restated:
\[ \boxed{ \textbf{Accounting is the problem of compressing economic reality while preserving the information that matters.} } \]
Under this interpretation, Accounting Bias is not merely a deviation between two numbers.
It is one manifestation of a more fundamental problem:
\[ \boxed{ \textbf{What economic information survives when reality is transformed into accounts?} } \]