Optimal Detection in a General Vector Channel Model#

The received signal is modeled as an \(N\)-dimensional vector, \(\vec{r}\) that depends statistically on the transmitted signal vector, \(\vec{s}_m\).

In this section, we consider a general vector channel model—not limited to the AWGN scenario—and develop the concepts underlying optimal detection.

Signal Transmission and the General Vector Channel#

Suppose the transmitter has a set of possible signal vectors

\[ \{\vec{s}_m, \, 1 \le m \le M\} \]

each corresponding to a message.

These vectors are transmitted according to certain a priori probabilities \(P_m\) (other notation \(p(\vec{s}_m)\) or \(p(m)\)), which capture the likelihood of each message being sent.

When a signal vector \(\vec{s}_m\) is transmitted, the received vector \(\vec{r}\) is a random quantity whose statistics are described by the conditional probability density function (pdf)

\[ \boxed{ \Pr(\vec{r} \mid \vec{s}_m) } \]

Thus, the overall channel model is characterized by the statistical relationship between \(\vec{s}_m\) and \(\vec{r}\). The receiver, upon observing \(\vec{r}\), must decide which message was most likely transmitted.

Decision Function and the Optimal Detector#

To formalize the detection process, we define a decision function (or decision rule) \(g(\vec{r})\), which is a mapping from the observation space \(\mathbb{R}^N\) to the set of messages. The decision rule is expressed as:

\[\begin{split} \begin{align} g: \mathbb{R}^N &\rightarrow \{1, 2, \ldots, M\} \\ \vec{r} &\mapsto \hat{m} \end{align} \end{split}\]

It indicates that the decision function \(g\) takes an \(N\)-dimensional vector (obtained from the projection of the received waveform \(r(t)\) onto an orthonormal basis) as its input and maps it to a message index \(\hat{m}\) in the set \(\{1, 2, \ldots, M\}\).

When the receiver observes a particular \(\vec{r}\), it declares the transmitted message as

\[ \hat{m} = g(\vec{r}) \]

The goal of optimal detection is to choose the decision rule \(g(\vec{r})\) so as to minimize the error probability or, equivalently, to maximize the probability of a correct decision.

Message Indices#

Note that in this context, \( m \) and \(\hat{m}\) are typically used as indices that label the possible messages.

Message Index:
The notation \( m \) (with \( 1 \le m \le M \)) identifies one of the \( M \) possible messages. Each message corresponds to a unique transmitted signal vector \( \vec{s}_m \). Although each message is originally derived from a sequence of bits (for example, a \( k \)-bit sequence when \( M = 2^k \)), in the detection and analysis framework we use the index \( m \) to refer to the message.

Detection Process:
The receiver observes the vector \(\vec{r}\) and makes a decision by choosing an index \(\hat{m}\) that maximizes the posterior probability (or likelihood) given the observation:

\[ \hat{m} = \arg\max_{1 \le m \le M} \Pr[m|\vec{r}] \]

Here, \(\hat{m}\) is the estimated message index that is declared as the transmitted message.

Mapping to Bit Sequences:
Although \( m \) (and thus \(\hat{m}\)) is an index, there is typically an established mapping between these indices and the actual sequences of bits. For example, if a \( k \)-bit message is mapped to an index \( m \), the receiver can convert the detected index \(\hat{m}\) back to its corresponding \( k \)-bit binary sequence using the inverse of the mapping used at the transmitter.

Probability of a Correct Decision#

If the receiver decides \(\hat{m}\) upon receiving \(\vec{r}\), the probability that this decision is correct is given by

\[ \boxed{ \Pr[\text{correct decision} \mid \vec{r}] = \Pr[\hat{m} \text{ was sent} \mid \vec{r}] } \]

That is, given the observation \(\vec{r}\), the probability of being correct is the conditional probability that the transmitted message is \(\hat{m}\).

To obtain the overall probability of a correct decision, we average this conditional probability over all possible received vectors \(\vec{r}\) weighted by their marginal probability \(p(\vec{r})\):

\[ \boxed{ \Pr[\text{correct decision}] = \int \Pr[\hat{m} \text{ was sent} \mid \vec{r}] \, p(\vec{r}) \, d\vec{r} } \]

Since \(p(\vec{r})\) is nonnegative for all \(\vec{r}\), the overall probability of a correct decision is maximized if, for every received \(\vec{r}\), the decision rule maximizes \(\Pr[\hat{m} \text{ was sent} \mid \vec{r}]\).

Design Criterion for the Optimal Detector#

We can see that since \(p(\vec{r})\) is nonnegative for all \(\vec{r}\), the overall probability of a correct decision is maximized if, for every received \(\vec{r}\), the decision rule maximizes \(\Pr[\hat{m} \text{ was sent} \mid \vec{r}]\).

This observation leads directly to the formulation of the optimal detection rule. The optimal detector selects the message \(m\) that maximizes the posterior probability \(\Pr[m|\vec{r}]\). Formally, the decision function \(g_{\text{opt}}(\vec{r})\) is defined as:

\[ \hat{m} = g_{\text{opt}}(\vec{r}) = \underset{1 \leq m \leq M}{\arg\max} \ \Pr[m|\vec{r}] \]

In practice, this means that upon receiving \(\vec{r}\), the detector computes \(\Pr[m|\vec{r}]\) for each \(m = 1, 2, \ldots, M\) and then declares the message corresponding to the largest value of this conditional probability.

Equivalence in Terms of Signal Vectors#

Since transmitting message (index) \(m\) is equivalent to transmitting the signal vector \(\vec{s}_m\), the optimal decision rule can equivalently be written in terms of these vectors:

\[ \hat{m} = g_{\text{opt}}(\vec{r}) = \underset{1 \leq m \leq M}{\arg\max} \ \Pr[\vec{s}_m|\vec{r}] \]

This formulation emphasizes that the receiver’s decision is based on the likelihood of having received \(\vec{r}\) given that the signal \(\vec{s}_m\) was transmitted.

Summary of Types of Probabilities#

Prior Probability \(P_m\)#

  • Definition:
    The prior probability, denoted by \(P_m\), is the probability that the transmitter sends the \(m\)-th message before any observation is made at the receiver. In other words,

    \[ P_m \equiv p(\vec{s}_m) \equiv p(m) = \Pr[m \text{ is selected to send}] = \Pr{\vec{s}_m \text{ is selected to send}} \]

    which reflects the statistical likelihood of the \(m\)-th message being transmitted.

  • Role in Detection:
    The prior probability encapsulates any inherent bias in the message selection process. In many systems, messages are chosen uniformly at random, in which case

    \[ P_m = \frac{1}{M} \quad \text{for all } m \]

    where \(M\) is the total number of messages. However, if some messages are more likely to occur than others, \(P_m\) will vary accordingly.

Likelihood Probability \(\Pr(\vec{r}|\vec{s}_m)\) or \(\Pr(\vec{r}|m)\)#

  • Definition:
    The likelihood probability represents the probability density (or probability, in the discrete case) of receiving the vector \(\vec{r}\) given that the signal corresponding to the \(m\)-th message (i.e., \(\vec{s}_m\)) was transmitted. It is denoted by:

    \[ \Pr(\vec{r}|\vec{s}_m) \quad \text{or} \quad \Pr(\vec{r}|m) \]
  • Role in Detection:
    This probability is determined by the channel’s statistical behavior. For example, in an AWGN channel, the likelihood is typically modeled as a multivariate Gaussian density centered at \(\vec{s}_m\) with a covariance matrix that depends on the noise variance. It tells us how likely it is to observe \(\vec{r}\) if \(\vec{s}_m\) were transmitted.

Posterior Probability \(\Pr(\vec{s}_m|\vec{r})\) or \(\Pr(m|\vec{r})\)#

  • Definition:
    The posterior probability is the probability that the \(m\)-th message (or equivalently, the signal \(\vec{s}_m\)) was transmitted given that the received vector is \(\vec{r}\). It is denoted by:

    \[ \Pr(m|\vec{r}) \quad \text{or} \quad \Pr(\vec{s}_m|\vec{r}) \]
  • Derivation Using Bayes’ Rule:
    The posterior probability is computed by applying Bayes’ theorem:

    \[ \Pr(m|\vec{r}) = \frac{\Pr(\vec{r}|m) \, P_m}{p(\vec{r})} \]

    where:

    • \(\Pr(\vec{r}|m)\) is the likelihood.

    • \(P_m\) is the prior probability.

    • \(p(\vec{r})\) is the marginal probability (or evidence), given by

      \[ p(\vec{r}) = \sum_{m=1}^{M} \Pr(\vec{r}|m) \, P_m \]
  • Role in Detection:
    The posterior probability represents the updated belief about which message was transmitted after taking into account the observation \(\vec{r}\). It is the key quantity in the Maximum a Posteriori (MAP) detection rule.

Below is an extended explanation of the marginal probability \(p(\vec{r})\) and an updated Markdown table that includes it.

Marginal Probability \(p(\vec{r})\)#

  • Definition: The marginal probability \(p(\vec{r})\) represents the unconditional probability density of receiving the vector \(\vec{r}\) regardless of which message was transmitted. It is calculated by averaging (or “marginalizing”) the likelihood \(\Pr(\vec{r}|m)\) over all possible messages, each weighted by its corresponding prior probability \(P_m\). Mathematically, it is given by:

\[ p(\vec{r}) = \sum_{m=1}^{M} \Pr(\vec{r}|m) \, P_m \]
  • Role in Detection: This term plays a crucial role in Bayes’ theorem, serving as a normalizing constant to ensure that the posterior probabilities sum to one. It provides a measure of how likely the observation \(\vec{r}\) is under the entire statistical model of the transmission process.

Conditional Probability (Correction Detection Probability) \(\Pr(\hat{m} \text{ was sent} \,|\, \vec{r})\)#

  • Definition:
    Once the receiver processes \(\vec{r}\), it makes a decision and declares \(\hat{m}\) as the transmitted message. The probability

    \[ \Pr(\hat{m} \text{ was sent} \,|\, \vec{r}) \]

    is the probability that the message corresponding to the decision \(\hat{m}\) is indeed the one that was transmitted, given the received vector \(\vec{r}\).

    This probability is often referred to as the conditional probability of correct detection (or simply the conditional correctness probability) given \(\vec{r}\).

  • Role in Performance Analysis:
    This quantity directly reflects the confidence of the detector in its decision for a given observation \(\vec{r}\). The overall performance (i.e., the overall probability of a correct decision) is obtained by averaging \(\Pr(\hat{m} \text{ was sent} \,|\, \vec{r})\) over all possible received vectors:

    \[ \boxed{ \Pr[\text{correct decision}] = \int \Pr(\hat{m} \text{ was sent} \,|\, \vec{r}) \, p(\vec{r}) \, d\vec{r} } \]